Authors: Lei Zhang, Markus Stricker
MatNexus is a specialized software for the automated collection, processing, and analysis of text from scientific articles. Through an integrated suite of modules, the MatNexus facilitates the retrieval of scientific articles, processes textual data for insights, generates vector representations suitable for machine learning, and offers visualization capabilities for word embeddings. With the vast volume of scientific publications, MatNexus stands out as an end-to-end tool for researchers aiming to gain insights from scientific literature in material science, making the exploration of materials, such as the electrocatalyst examples we show here, efficient and insightful.
Authors: Jinheon Baek, Nirupama Chandrasekaran, Silviu Cucerzan, Allen herring, Sujay Kumar Jauhar
Large Language Models (LLMs) excel at tackling various natural language tasks. However, due to the significant costs involved in re-training or fine-tuning them, they remain largely static and difficult to personalize. Nevertheless, a variety of applications could benefit from generations that are tailored to users' preferences, goals, and knowledge. Among them is web search, where knowing what a user is trying to accomplish, what they care about, and what they know can lead to improved search experiences. In this work, we propose a novel and general approach that augments an LLM with relevant context from users' interaction histories with a search engine in order to personalize its outputs. Specifically, we construct an entity-centric knowledge store for each user based on their search and browsing activities on the web, which is then leveraged to provide contextually relevant LLM prompt augmentations. This knowledge store is light-weight, since it only produces user-specific aggregate projections of interests and knowledge onto public knowledge graphs, and leverages existing search log infrastructure, thereby mitigating the privacy, compliance, and scalability concerns associated with building deep user profiles for personalization. We then validate our approach on the task of contextual query suggestion, which requires understanding not only the user's current search context but also what they historically know and care about. Through a number of experiments based on human evaluation, we show that our approach is significantly better than several other LLM-powered baselines, generating query suggestions that are contextually more relevant, personalized, and useful.
Authors: Aditi Singh
Text-to-Image and Text-to-Video AI generation models are revolutionary technologies that use deep learning and natural language processing (NLP) techniques to create images and videos from textual descriptions. This paper investigates cutting-edge approaches in the discipline of Text-to-Image and Text-to-Video AI generations. The survey provides an overview of the existing literature as well as an analysis of the approaches used in various studies. It covers data preprocessing techniques, neural network types, and evaluation metrics used in the field. In addition, the paper discusses the challenges and limitations of Text-to-Image and Text-to-Video AI generations, as well as future research directions. Overall, these models have promising potential for a wide range of applications such as video production, content creation, and digital marketing.
Authors: Ruolin Su, Ting-Wei Wu, Biing-Hwang Juang
Tracking dialogue states is an essential topic in task-oriented dialogue systems, which involve filling in the necessary information in pre-defined slots corresponding to a schema. While general pre-trained language models have been shown effective in slot-filling, their performance is limited when applied to specific domains. We propose a graph-based framework that learns domain-specific prompts by incorporating the dialogue schema. Specifically, we embed domain-specific schema encoded by a graph neural network into the pre-trained language model, which allows for relations in the schema to guide the model for better adaptation to the specific domain. Our experiments demonstrate that the proposed graph-based method outperforms other multi-domain DST approaches while using similar or fewer trainable parameters. We also conduct a comprehensive study of schema graph architectures, parameter usage, and module ablation that demonstrate the effectiveness of our model on multi-domain dialogue state tracking.
Authors: Yunting Yin, Steven Skiena
Dictionary definitions are historically the arbitrator of what words mean, but this primacy has come under threat by recent progress in NLP, including word embeddings and generative models like ChatGPT. We present an exploratory study of the degree of alignment between word definitions from classical dictionaries and these newer computational artifacts. Specifically, we compare definitions from three published dictionaries to those generated from variants of ChatGPT. We show that (i) definitions from different traditional dictionaries exhibit more surface form similarity than do model-generated definitions, (ii) that the ChatGPT definitions are highly accurate, comparable to traditional dictionaries, and (iii) ChatGPT-based embedding definitions retain their accuracy even on low frequency words, much better than GloVE and FastText word embeddings.
Authors: Maxime Delmas, Magdalena Wysocka, André Freitas
The sparsity of labelled data is an obstacle to the development of Relation Extraction models and the completion of databases in various biomedical areas. While being of high interest in drug-discovery, the natural-products literature, reporting the identification of potential bioactive compounds from organisms, is a concrete example of such an overlooked topic. To mark the start of this new task, we created the first curated evaluation dataset and extracted literature items from the LOTUS database to build training sets. To this end, we developed a new sampler inspired by diversity metrics in ecology, named Greedy Maximum Entropy sampler, or GME-sampler (https://github.com/idiap/gme-sampler). The strategic optimization of both balance and diversity of the selected items in the evaluation set is important given the resource-intensive nature of manual curation. After quantifying the noise in the training set, in the form of discrepancies between the input abstracts text and the expected output labels, we explored different strategies accordingly. Framing the task as an end-to-end Relation Extraction, we evaluated the performance of standard fine-tuning as a generative task and few-shot learning with open Large Language Models (LLaMA 7B-65B). In addition to their evaluation in few-shot settings, we explore the potential of open Large Language Models (Vicuna-13B) as synthetic data generator and propose a new workflow for this purpose. All evaluated models exhibited substantial improvements when fine-tuned on synthetic abstracts rather than the original noisy data. We provide our best performing (f1-score=59.0) BioGPT-Large model for end-to-end RE of natural-products relationships along with all the generated synthetic data and the evaluation dataset. See more details at https://github.com/idiap/abroad-re.
Authors: Uyen Lai, Gurjit S. Randhawa, Paul Sheridan
Heaps' law is an empirical relation in text analysis that predicts vocabulary growth as a function of corpus size. While this law has been validated in diverse human-authored text corpora, its applicability to large language model generated text remains unexplored. This study addresses this gap, focusing on the emulation of corpora using the suite of GPT-Neo large language models. To conduct our investigation, we emulated corpora of PubMed abstracts using three different parameter sizes of the GPT-Neo model. Our emulation strategy involved using the initial five words of each PubMed abstract as a prompt and instructing the model to expand the content up to the original abstract's length. Our findings indicate that the generated corpora adhere to Heaps' law. Interestingly, as the GPT-Neo model size grows, its generated vocabulary increasingly adheres to Heaps' law as as observed in human-authored text. To further improve the richness and authenticity of GPT-Neo outputs, future iterations could emphasize enhancing model size or refining the model architecture to curtail vocabulary repetition.
Authors: Simran Khanuja, Srinivas Gowriraj, Lucio Dery, Graham Neubig
We consider the task of optimally fine-tuning pre-trained multilingual models, given small amounts of unlabelled target data and an annotation budget. In this paper, we introduce DEMUX, a framework that prescribes the exact data-points to label from vast amounts of unlabelled multilingual data, having unknown degrees of overlap with the target set. Unlike most prior works, our end-to-end framework is language-agnostic, accounts for model representations, and supports multilingual target configurations. Our active learning strategies rely upon distance and uncertainty measures to select task-specific neighbors that are most informative to label, given a model. DeMuX outperforms strong baselines in 84% of the test cases, in the zero-shot setting of disjoint source and target language sets (including multilingual target pools), across three models and four tasks. Notably, in low-budget settings (5-100 examples), we observe gains of up to 8-11 F1 points for token-level tasks, and 2-5 F1 for complex tasks. Our code is released here: https://github.com/simran-khanuja/demux.
Authors: Lucio Dery, David Grangier, Awni Hannun
Large, pre-trained models are problematic to use in resource constrained applications. Fortunately, task-aware structured pruning methods offer a solution. These approaches reduce model size by dropping structural units like layers and attention heads in a manner that takes into account the end-task. However, these pruning algorithms require more task-specific data than is typically available. We propose a framework which combines structured pruning with transfer learning to reduce the need for task-specific data. Our empirical results answer questions such as: How should the two tasks be coupled? What parameters should be transferred? And, when during training should transfer learning be introduced? Leveraging these insights, we demonstrate that our framework results in pruned models with improved generalization over strong baselines.
Authors: Pouya Pezeshkpour, Hayate Iso, Thom Lake, Nikita Bhutani, Estevam Hruschka
Numerous HR applications are centered around resumes and job descriptions. While they can benefit from advancements in NLP, particularly large language models, their real-world adoption faces challenges due to absence of comprehensive benchmarks for various HR tasks, and lack of smaller models with competitive capabilities. In this paper, we aim to bridge this gap by introducing the Resume-Job Description Benchmark (RJDB). We meticulously craft this benchmark to cater to a wide array of HR tasks, including matching and explaining resumes to job descriptions, extracting skills and experiences from resumes, and editing resumes. To create this benchmark, we propose to distill domain-specific knowledge from a large language model (LLM). We rely on a curated skill-occupation graph to ensure diversity and provide context for LLMs generation. Our benchmark includes over 50 thousand triples of job descriptions, matched resumes and unmatched resumes. Using RJDB, we train multiple smaller student models. Our experiments reveal that the student models achieve near/better performance than the teacher model (GPT-4), affirming the effectiveness of the benchmark. Additionally, we explore the utility of RJDB on out-of-distribution data for skill extraction and resume-job description matching, in zero-shot and weak supervision manner. We release our datasets and code to foster further research and industry applications.
Authors: Benjamin C. Warner, Thomas Kannampallil, Seunghwan Kim
EHR audit logs are a highly granular stream of events that capture clinician activities, and is a significant area of interest for research in characterizing clinician workflow on the electronic health record (EHR). Existing techniques to measure the complexity of workflow through EHR audit logs (audit logs) involve time- or frequency-based cross-sectional aggregations that are unable to capture the full complexity of a EHR session. We briefly evaluate the usage of transformer-based tabular language model (tabular LM) in measuring the entropy or disorderedness of action sequences within workflow and release the evaluated models publicly.
Authors: Apoorv Khandelwal, Ellie Pavlick, Chen Sun
Modular neural networks without additional training have recently been shown to surpass end-to-end neural networks on challenging vision-language tasks. The latest such methods simultaneously introduce LLM-based code generation to build programs and a number of skill-specific, task-oriented modules to execute them. In this paper, we focus on ViperGPT and ask where its additional performance comes from and how much is due to the (state-of-art, end-to-end) BLIP-2 model it subsumes vs. additional symbolic components. To do so, we conduct a controlled study (comparing end-to-end, modular, and prompting-based methods across several VQA benchmarks). We find that ViperGPT's reported gains over BLIP-2 can be attributed to its selection of task-specific modules, and when we run ViperGPT using a more task-agnostic selection of modules, these gains go away. Additionally, ViperGPT retains much of its performance if we make prominent alterations to its selection of modules: e.g. removing or retaining only BLIP-2. Finally, we compare ViperGPT against a prompting-based decomposition strategy and find that, on some benchmarks, modular approaches significantly benefit by representing subtasks with natural language, instead of code.
Authors: Nedelina Teneva, Estevam Hruschka
Despite the recent popularity of knowledge graph (KG) related tasks and benchmarks such as KG embeddings, link prediction, entity alignment and evaluation of the reasoning abilities of pretrained language models as KGs, the structure and properties of real KGs are not well studied. In this paper, we perform a large scale comparative study of 29 real KG datasets from diverse domains such as the natural sciences, medicine, and NLP to analyze their properties and structural patterns. Based on our findings, we make several recommendations regarding KG-based model development and evaluation. We believe that the rich structural information contained in KGs can benefit the development of better KG models across fields and we hope this study will contribute to breaking the existing data silos between different areas of research (e.g., ML, NLP, AI for sciences).
Authors: Martino Pelucchi, Matias Valdenegro-Toro
ChatGPT took the world by storm for its impressive abilities. Due to its release without documentation, scientists immediately attempted to identify its limits, mainly through its performance in natural language processing (NLP) tasks. This paper aims to join the growing literature regarding ChatGPT's abilities by focusing on its performance in high-resource languages and on its capacity to predict its answers' accuracy by giving a confidence level. The analysis of high-resource languages is of interest as studies have shown that low-resource languages perform worse than English in NLP tasks, but no study so far has analysed whether high-resource languages perform as well as English. The analysis of ChatGPT's confidence calibration has not been carried out before either and is critical to learn about ChatGPT's trustworthiness. In order to study these two aspects, five high-resource languages and two NLP tasks were chosen. ChatGPT was asked to perform both tasks in the five languages and to give a numerical confidence value for each answer. The results show that all the selected high-resource languages perform similarly and that ChatGPT does not have a good confidence calibration, often being overconfident and never giving low confidence values.
Authors: Isaac Caswell, Lisa Wang, Isabel Papadimitriou
Data quality is a problem that perpetually resurfaces throughout the field of NLP, regardless of task, domain, or architecture, and remains especially severe for lower-resource languages. A typical and insidious issue, affecting both training data and model output, is data that is repetitive and dominated by linguistically uninteresting boilerplate, such as price catalogs or computer-generated log files. Though this problem permeates many web-scraped corpora, there has yet to be a benchmark to test against, or a systematic study to find simple metrics that generalize across languages and agree with human judgements of data quality. In the present work, we create and release BREAD, a human-labeled benchmark on repetitive boilerplate vs. plausible linguistic content, spanning 360 languages. We release several baseline CRED (Character REDundancy) scores along with it, and evaluate their effectiveness on BREAD. We hope that the community will use this resource to develop better filtering methods, and that our reference implementations of CRED scores can become standard corpus evaluation tools, driving the development of cleaner language modeling corpora, especially in low-resource languages.
Authors: Saad Almohaimeed, Saleh Almohaimeed, Ashfaq Ali Shafin, Bogdan Carbunar, Ladislau Bölöni
Detecting harmful content on social media, such as Twitter, is made difficult by the fact that the seemingly simple yes/no classification conceals a significant amount of complexity. Unfortunately, while several datasets have been collected for training classifiers in hate and offensive speech, there is a scarcity of datasets labeled with a finer granularity of target classes and specific targets. In this paper, we introduce THOS, a dataset of 8.3k tweets manually labeled with fine-grained annotations about the target of the message. We demonstrate that this dataset makes it feasible to train classifiers, based on Large Language Models, to perform classification at this level of granularity.
Authors: Vatsal Venkatkrishna, Durga Shree Nagabushanam, Emmanuel Iko-Ojo Simon, Fatemeh H. Fard, Melina Vidoni, Zadia Codabux
Documentation debt hinders the effective utilization of open-source software. Although code summarization tools have been helpful for developers, most would prefer a detailed account of each parameter in a function rather than a high-level summary. However, generating such a summary is too intricate for a single generative model to produce reliably due to the lack of high-quality training data. Thus, we propose a multi-step approach that combines multiple task-specific models, each adept at producing a specific section of a docstring. The combination of these models ensures the inclusion of each section in the final docstring. We compared the results from our approach with existing generative models using both automatic metrics and a human-centred evaluation with 17 participating developers, which proves the superiority of our approach over existing methods.
Authors: Vasudha Varadarajan, Sverker Sikström, Oscar N.E. Kjell, H. Andrew Schwartz
Mental health issues widely vary across individuals - the manifestations of signs and symptoms can be fairly heterogeneous. Recently, language-based depression and anxiety assessments have shown promise for capturing this heterogeneous nature by evaluating a patient's own language, but such approaches require a large sample of words per person to be accurate. In this work, we introduce adaptive language-based assessment - the task of iteratively estimating an individual's psychological score based on limited language responses to questions that the model also decides to ask. To this end, we explore two statistical learning-based approaches for measurement/scoring: classical test theory (CTT) and item response theory (IRT). We find that using adaptive testing in general can significantly reduce the number of questions required to achieve high validity (r ~ 0.7) with standardized tests, bringing down from 11 total questions down to 3 for depression and 5 for anxiety. Given the combinatorial nature of the problem, we empirically evaluate multiple strategies for both the ordering and scoring objectives, introducing two new methods: a semi-supervised item response theory based method (ALIRT), and a supervised actor-critic based model. While both of the models achieve significant improvements over random and fixed orderings, we find ALIRT to be a scalable model that achieves the highest accuracy with lower numbers of questions (e.g. achieves Pearson r ~ 0.93 after only 3 questions versus asking all 11 questions). Overall, ALIRT allows prompting a reduced number of questions without compromising accuracy or overhead computational costs.
Authors: Aidin Shiri, Kaushik Roy, Amit Sheth, Manas Gaur
Fine-tuning pre-trained foundational language models (FLM) for specific tasks is often impractical, especially for resource-constrained devices. This necessitates the development of a Lifelong Learning (L3) framework that continuously adapts to a stream of Natural Language Processing (NLP) tasks efficiently. We propose an approach that focuses on extracting meaningful representations from unseen data, constructing a structured knowledge base, and improving task performance incrementally. We conducted experiments on various NLP tasks to validate its effectiveness, including benchmarks like GLUE and SuperGLUE. We measured good performance across the accuracy, training efficiency, and knowledge transfer metrics. Initial experimental results show that the proposed L3 ensemble method increases the model accuracy by 4% ~ 36% compared to the fine-tuned FLM. Furthermore, L3 model outperforms naive fine-tuning approaches while maintaining competitive or superior performance (up to 15.4% increase in accuracy) compared to the state-of-the-art language model (T5) for the given task, STS benchmark.
Authors: Yichi Zhang, Zhuo Chen, Yin Fang, Lei Cheng, Yanxi Lu, Fangming Li, Wen Zhang, Huajun Chen
Recently, the development of large language models (LLMs) has attracted wide attention in academia and industry. Deploying LLMs to real scenarios is one of the key directions in the current Internet industry. In this paper, we present a novel pipeline to apply LLMs for domain-specific question answering (QA) that incorporates domain knowledge graphs (KGs), addressing an important direction of LLM application. As a real-world application, the content generated by LLMs should be user-friendly to serve the customers. Additionally, the model needs to utilize domain knowledge properly to generate reliable answers. These two issues are the two major difficulties in the LLM application as vanilla fine-tuning can not adequately address them. We think both requirements can be unified as the model preference problem that needs to align with humans to achieve practical application. Thus, we introduce Knowledgeable Preference AlignmenT (KnowPAT), which constructs two kinds of preference set called style preference set and knowledge preference set respectively to tackle the two issues. Besides, we design a new alignment objective to align the LLM preference with human preference, aiming to train a better LLM for real-scenario domain-specific QA to generate reliable and user-friendly answers. Adequate experiments and comprehensive with 15 baseline methods demonstrate that our KnowPAT is an outperforming pipeline for real-scenario domain-specific QA with LLMs. Our code is open-source at https://github.com/zjukg/KnowPAT.
Authors: Hsuan Su, Rebecca Qian, Chinnadhurai Sankar, Shahin Shayandeh, Shang-Tse Chen, Hung-yi Lee, Daniel M. Bikel
Recent works have shown considerable improvements in task-oriented dialogue (TOD) systems by utilizing pretrained large language models (LLMs) in an end-to-end manner. However, the biased behavior of each component in a TOD system and the error propagation issue in the end-to-end framework can lead to seriously biased TOD responses. Existing works of fairness only focus on the total bias of a system. In this paper, we propose a diagnosis method to attribute bias to each component of a TOD system. With the proposed attribution method, we can gain a deeper understanding of the sources of bias. Additionally, researchers can mitigate biased model behavior at a more granular level. We conduct experiments to attribute the TOD system's bias toward three demographic axes: gender, age, and race. Experimental results show that the bias of a TOD system usually comes from the response generation model.
Authors: Matan Abudy, Nur Lan, Emmanuel Chemla, Roni Katzir
Associative memory architectures are designed for memorization but also offer, through their retrieval method, a form of generalization to unseen inputs: stored memories can be seen as prototypes from this point of view. Focusing on Modern Hopfield Networks (MHN), we show that a large memorization capacity undermines the generalization opportunity. We offer a solution to better optimize this tradeoff. It relies on Minimum Description Length (MDL) to determine during training which memories to store, as well as how many of them.
Authors: Peiyu Liu, Junming Liu, Lirong Fu, Kangjie Lu, Yifan Xia, Xuhong Zhang, Wenzhi Chen, Haiqin Weng, Shouling Ji, Wenhai Wang
Recently, ChatGPT has attracted great attention from the code analysis domain. Prior works show that ChatGPT has the capabilities of processing foundational code analysis tasks, such as abstract syntax tree generation, which indicates the potential of using ChatGPT to comprehend code syntax and static behaviors. However, it is unclear whether ChatGPT can complete more complicated real-world vulnerability management tasks, such as the prediction of security relevance and patch correctness, which require an all-encompassing understanding of various aspects, including code syntax, program semantics, and related manual comments.
In this paper, we explore ChatGPT's capabilities on 6 tasks involving the complete vulnerability management process with a large-scale dataset containing 78,445 samples. For each task, we compare ChatGPT against SOTA approaches, investigate the impact of different prompts, and explore the difficulties. The results suggest promising potential in leveraging ChatGPT to assist vulnerability management. One notable example is ChatGPT's proficiency in tasks like generating titles for software bug reports. Furthermore, our findings reveal the difficulties encountered by ChatGPT and shed light on promising future directions. For instance, directly providing random demonstration examples in the prompt cannot consistently guarantee good performance in vulnerability management. By contrast, leveraging ChatGPT in a self-heuristic way -- extracting expertise from demonstration examples itself and integrating the extracted expertise in the prompt is a promising research direction. Besides, ChatGPT may misunderstand and misuse the information in the prompt. Consequently, effectively guiding ChatGPT to focus on helpful information rather than the irrelevant content is still an open problem.
Authors: Marta R. Costa-jussà, David Dale, Maha Elbayad, Bokai Yu
Added toxicity in the context of translation refers to the fact of producing a translation output with more toxicity than there exists in the input. In this paper, we present MinTox which is a novel pipeline to identify added toxicity and mitigate this issue which works at inference time. MinTox uses a toxicity detection classifier which is multimodal (speech and text) and works in languages at scale. The mitigation method is applied to languages at scale and directly in text outputs. MinTox is applied to SEAMLESSM4T, which is the latest multimodal and massively multilingual machine translation system. For this system, MinTox achieves significant added toxicity mitigation across domains, modalities and language directions. MinTox manages to approximately filter out from 25% to 95% of added toxicity (depending on the modality and domain) while keeping translation quality.
Authors: Elliott Ash, Aniket Kesari, Suresh Naidu, Lena Song, Dominik Stammbach
Written judicial opinions are an important tool for building public trust in court decisions, yet they can be difficult for non-experts to understand. We present a pipeline for using an AI assistant to generate simplified summaries of judicial opinions. These are more accessible to the public and more easily understood by non-experts, We show in a survey experiment that the simplified summaries help respondents understand the key features of a ruling. We discuss how to integrate legal domain knowledge into studies using large language models. Our results suggest a role both for AI assistants to inform the public, and for lawyers to guide the process of generating accessible summaries.
Authors: Maarten De Raedt, Semere Kiros Bitew, Fréderic Godin, Thomas Demeester, Chris Develder
The brittleness of finetuned language model performance on out-of-distribution (OOD) test samples in unseen domains has been well-studied for English, yet is unexplored for multi-lingual models. Therefore, we study generalization to OOD test data specifically in zero-shot cross-lingual transfer settings, analyzing performance impacts of both language and domain shifts between train and test data. We further assess the effectiveness of counterfactually augmented data (CAD) in improving OOD generalization for the cross-lingual setting, since CAD has been shown to benefit in a monolingual English setting. Finally, we propose two new approaches for OOD generalization that avoid the costly annotation process associated with CAD, by exploiting the power of recent large language models (LLMs). We experiment with 3 multilingual models, LaBSE, mBERT, and XLM-R trained on English IMDb movie reviews, and evaluate on OOD test sets in 13 languages: Amazon product reviews, Tweets, and Restaurant reviews. Results echo the OOD performance decline observed in the monolingual English setting. Further, (i) counterfactuals from the original high-resource language do improve OOD generalization in the low-resource language, and (ii) our newly proposed cost-effective approaches reach similar or up to +3.1% better accuracy than CAD for Amazon and Restaurant reviews.
Authors: Hanzhang Zhou, Junlang Qian, Zijian Feng, Hui Lu, Zixiao Zhu, Kezhi Mao
In this study, we investigate in-context learning (ICL) in document-level event argument extraction (EAE). The paper identifies key challenges in this problem, including example selection, context length limitation, abundance of event types, and the limitation of Chain-of-Thought (CoT) prompting in non-reasoning tasks. To address these challenges, we introduce the Heuristic-Driven Link-of-Analogy (HD-LoA) prompting method. Specifically, we hypothesize and validate that LLMs learn task-specific heuristics from demonstrations via ICL. Building upon this hypothesis, we introduce an explicit heuristic-driven demonstration construction approach, which transforms the haphazard example selection process into a methodical method that emphasizes task heuristics. Additionally, inspired by the analogical reasoning of human, we propose the link-of-analogy prompting, which enables LLMs to process new situations by drawing analogies to known situations, enhancing their adaptability. Extensive experiments show that our method outperforms the existing prompting methods and few-shot supervised learning methods, exhibiting F1 score improvements of 4.53% and 9.38% on the document-level EAE dataset. Furthermore, when applied to sentiment analysis and natural language inference tasks, the HD-LoA prompting achieves accuracy gains of 2.87% and 2.63%, indicating its effectiveness across different tasks.
Authors: Xiaoqian Li, Ercong Nie, Sheng Liang
The remarkable ability of Large Language Models (LLMs) to understand and follow instructions has sometimes been limited by their in-context learning (ICL) performance in low-resource languages. To address this, we introduce a novel approach that leverages cross-lingual retrieval-augmented in-context learning (CREA-ICL). By extracting semantically similar prompts from high-resource languages, we aim to improve the zero-shot performance of multilingual pre-trained language models (MPLMs) across diverse tasks. Though our approach yields steady improvements in classification tasks, it faces challenges in generation tasks. Our evaluation offers insights into the performance dynamics of retrieval-augmented in-context learning across both classification and generation domains.
Authors: Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Lovering, Chris Tanner
As large language models (LLMs) impact a growing number of complex domains, it is becoming increasingly important to have fair, accurate, and rigorous evaluation benchmarks. Evaluating the reasoning skills required for business and financial NLP stands out as a particularly difficult challenge. We introduce BizBench, a new benchmark for evaluating models' ability to reason about realistic financial problems. BizBench comprises 8 quantitative reasoning tasks. Notably, BizBench targets the complex task of question-answering (QA) for structured and unstructured financial data via program synthesis (i.e., code generation). We introduce three diverse financially-themed code-generation tasks from newly collected and augmented QA data. Additionally, we isolate distinct financial reasoning capabilities required to solve these QA tasks: reading comprehension of financial text and tables, which is required to extract correct intermediate values; and understanding domain knowledge (e.g., financial formulas) needed to calculate complex solutions. Collectively, these tasks evaluate a model's financial background knowledge, ability to extract numeric entities from financial documents, and capacity to solve problems with code. We conduct an in-depth evaluation of open-source and commercial LLMs, illustrating that BizBench is a challenging benchmark for quantitative reasoning in the finance and business domain.
Authors: Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, Xiang Bai
Large Multimodal Models have demonstrated impressive capabilities in understanding general vision-language tasks. However, due to the limitation of supported input resolution (e.g., 448 x 448) as well as the inexhaustive description of the training image-text pair, these models often encounter challenges when dealing with intricate scene understandings and narratives. Here we address the problem by proposing the Monkey. Our contributions are two-fold: 1) without pretraining from the start, our method can be built upon an existing vision encoder (e.g., vit-BigHuge) to effectively improve the input resolution capacity up to 896 x 1344 pixels; 2) we propose a multi-level description generation method, which automatically provides rich information that can guide model to learn contextual association between scenes and objects. Our extensive testing across more than 16 distinct datasets reveals that Monkey achieves consistently competitive performance over the existing LMMs on fundamental tasks, such as Image Captioning, General Visual Question Answering (VQA), and Document-oriented VQA. Models, interactive demo, and the source code are provided at the following https://github.com/Yuliang-Liu/Monkey.
Authors: Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, Tong Zhang
The integration of visual inputs with large language models (LLMs) has led to remarkable advancements in multi-modal capabilities, giving rise to visual large language models (VLLMs). However, effectively harnessing VLLMs for intricate visual perception tasks remains a challenge. In this paper, we present a novel end-to-end framework named PerceptionGPT, which efficiently and effectively equips the VLLMs with visual perception abilities by leveraging the representation power of LLMs' token embedding. Our proposed method treats the token embedding of the LLM as the carrier of spatial information, then leverage lightweight visual task encoders and decoders to perform visual perception tasks (e.g., detection, segmentation). Our approach significantly alleviates the training difficulty suffered by previous approaches that formulate the visual outputs as discrete tokens, and enables achieving superior performance with fewer trainable parameters, less training data and shorted training time. Moreover, as only one token embedding is required to decode the visual outputs, the resulting sequence length during inference is significantly reduced. Consequently, our approach enables accurate and flexible representations, seamless integration of visual perception tasks, and efficient handling of a multiple of visual outputs. We validate the effectiveness and efficiency of our approach through extensive experiments. The results demonstrate significant improvements over previous methods with much fewer trainable parameters and GPU hours, which facilitates future research in enabling LLMs with visual perception abilities.
Authors: Haoyuan Li, Hao Jiang, Tianke Zhang, Zhelun Yu, Aoxiong Yin, Hao Cheng, Siming Fu, Yuhao Zhang, Wanggui He
Training AI models has always been challenging, especially when there is a need for custom models to provide personalized services. Algorithm engineers often face a lengthy process to iteratively develop models tailored to specific business requirements, making it even more difficult for non-experts. The quest for high-quality and efficient model development, along with the emergence of Large Language Model (LLM) Agents, has become a key focus in the industry. Leveraging the powerful analytical, planning, and decision-making capabilities of LLM, we propose a TrainerAgent system comprising a multi-agent framework including Task, Data, Model and Server agents. These agents analyze user-defined tasks, input data, and requirements (e.g., accuracy, speed), optimizing them comprehensively from both data and model perspectives to obtain satisfactory models, and finally deploy these models as online service. Experimental evaluations on classical discriminative and generative tasks in computer vision and natural language processing domains demonstrate that our system consistently produces models that meet the desired criteria. Furthermore, the system exhibits the ability to critically identify and reject unattainable tasks, such as fantastical scenarios or unethical requests, ensuring robustness and safety. This research presents a significant advancement in achieving desired models with increased efficiency and quality as compared to traditional model development, facilitated by the integration of LLM-powered analysis, decision-making, and execution capabilities, as well as the collaboration among four agents. We anticipate that our work will contribute to the advancement of research on TrainerAgent in both academic and industry communities, potentially establishing it as a new paradigm for model development in the field of AI.
Authors: Florian Mai, Arnaud Pannatier, Fabio Fehr, Haolin Chen, Francois Marelli, Francois Fleuret, James Henderson
Transformer-based architectures are the model of choice for natural language understanding, but they come at a significant cost, as they have quadratic complexity in the input length, require a lot of training data, and can be difficult to tune. In the pursuit of lower costs, we investigate simple MLP-based architectures. We find that existing architectures such as MLPMixer, which achieves token mixing through a static MLP applied to each feature independently, are too detached from the inductive biases required for natural language understanding. In this paper, we propose a simple variant, HyperMixer, which forms the token mixing MLP dynamically using hypernetworks. Empirically, we demonstrate that our model performs better than alternative MLP-based models, and on par with Transformers. In contrast to Transformers, HyperMixer achieves these results at substantially lower costs in terms of processing time, training data, and hyperparameter tuning.
Authors: Amanpreet Singh, Mike D'Arcy, Arman Cohan, Doug Downey, Sergey Feldman
Learned representations of scientific documents can serve as valuable input features for downstream tasks without further fine-tuning. However, existing benchmarks for evaluating these representations fail to capture the diversity of relevant tasks. In response, we introduce SciRepEval, the first comprehensive benchmark for training and evaluating scientific document representations. It includes 24 challenging and realistic tasks, 8 of which are new, across four formats: classification, regression, ranking and search. We then use this benchmark to study and improve the generalization ability of scientific document representation models. We show how state-of-the-art models like SPECTER and SciNCL struggle to generalize across the task formats, and that simple multi-task training fails to improve them. However, a new approach that learns multiple embeddings per document, each tailored to a different format, can improve performance. We experiment with task-format-specific control codes and adapters and find they outperform the existing single-embedding state-of-the-art by over 2 points absolute. We release the resulting family of multi-format models, called SPECTER2, for the community to use and build on.
Authors: Albert Webson, Alyssa Marie Loo, Qinan Yu, Ellie Pavlick
Prompts have been the center of progress in advancing language models' zero-shot and few-shot performance. However, recent work finds that models can perform surprisingly well when given intentionally irrelevant or misleading prompts. Such results may be interpreted as evidence that model behavior is not "human like". In this study, we challenge a central assumption in such work: that humans would perform badly when given pathological instructions. We find that humans are able to reliably ignore irrelevant instructions and thus, like models, perform well on the underlying task despite an apparent lack of signal regarding the task they are being asked to do. However, when given deliberately misleading instructions, humans follow the instructions faithfully, whereas models do not. Our findings caution that future research should not idealize human behaviors as a monolith and should not train or evaluate models to mimic assumptions about these behaviors without first validating humans' behaviors empirically.
Authors: Michal Kosinski
We explore the intriguing possibility that theory of mind (ToM), or the uniquely human ability to impute unobservable mental states to others, might have spontaneously emerged in large language models (LLMs). We designed 40 false-belief tasks, considered a gold standard in testing ToM in humans, and administered them to several LLMs. Each task included a false-belief scenario, three closely matched true-belief controls, and the reversed versions of all four. Smaller and older models solved no tasks; GPT-3-davinci-003 (from November 2022) and ChatGPT-3.5-turbo (from March 2023) solved 20% of the tasks; ChatGPT-4 (from June 2023) solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. These findings suggest the intriguing possibility that ToM, previously considered exclusive to humans, may have spontaneously emerged as a byproduct of LLMs' improving language skills.
Authors: Kevin Scaria, Himanshu Gupta, Siddharth Goyal, Saurabh Arjun Sawant, Swaroop Mishra, Chitta Baral
We introduce InstructABSA, an instruction learning paradigm for Aspect-Based Sentiment Analysis (ABSA) subtasks. Our method introduces positive, negative, and neutral examples to each training sample, and instruction tune the model (Tk-Instruct) for ABSA subtasks, yielding significant performance improvements. Experimental results on the Sem Eval 2014, 15, and 16 datasets demonstrate that InstructABSA outperforms the previous state-of-the-art (SOTA) approaches on Term Extraction (ATE), Sentiment Classification(ATSC) and Sentiment Pair Extraction (ASPE) subtasks. In particular, InstructABSA outperforms the previous state-of-the-art (SOTA) on the Rest14 ATE subtask by 5.69% points, the Rest15 ATSC subtask by 9.59% points, and the Lapt14 AOPE subtask by 3.37% points, surpassing 7x larger models. We also get competitive results on AOOE, AOPE, and AOSTE subtasks indicating strong generalization ability to all subtasks. Exploring sample efficiency reveals that just 50% train data is required to get competitive results with other instruction tuning approaches. Lastly, we assess the quality of instructions and observe that InstructABSA's performance experiences a decline of ~10% when adding misleading examples.
Authors: Kaitlyn Zhou, Dan Jurafsky, Tatsunori Hashimoto
The increased deployment of LMs for real-world tasks involving knowledge and facts makes it important to understand model epistemology: what LMs think they know, and how their attitudes toward that knowledge are affected by language use in their inputs. Here, we study an aspect of model epistemology: how epistemic markers of certainty, uncertainty, or evidentiality like "I'm sure it's", "I think it's", or "Wikipedia says it's" affect models, and whether they contribute to model failures. We develop a typology of epistemic markers and inject 50 markers into prompts for question answering. We find that LMs are highly sensitive to epistemic markers in prompts, with accuracies varying more than 80%. Surprisingly, we find that expressions of high certainty result in a 7% decrease in accuracy as compared to low certainty expressions; similarly, factive verbs hurt performance, while evidentials benefit performance. Our analysis of a popular pretraining dataset shows that these markers of uncertainty are associated with answers on question-answering websites, while markers of certainty are associated with questions. These associations may suggest that the behavior of LMs is based on mimicking observed language use, rather than truly reflecting epistemic uncertainty.
Authors: Felix Faltings, Michel Galley, Baolin Peng, Kianté Brantley, Weixin Cai, Yizhe Zhang, Jianfeng Gao, Bill Dolan
Users interact with text, image, code, or other editors on a daily basis. However, machine learning models are rarely trained in the settings that reflect the interactivity between users and their editor. This is understandable as training AI models with real users is not only slow and costly, but what these models learn may be specific to user interface design choices. Unfortunately, this means most of the research on text, code, and image generation has focused on non-interactive settings, whereby the model is expected to get everything right without accounting for any input from a user who may be willing to help.
We introduce a new Interactive Text Generation task that allows training generation models interactively without the costs of involving real users, by using user simulators that provide edits that guide the model towards a given target text. We train our interactive models using Imitation Learning, and our experiments against competitive non-interactive generation models show that models trained interactively are superior to their non-interactive counterparts, even when all models are given the same budget of user inputs or edits.
Authors: Clemens Grabmayer
Axiomatization and expressibility problems for Milner's process semantics (1984) of regular expressions modulo bisimilarity have turned out to be difficult for the full class of expressions with deadlock 0 and empty step~1. We report on a phenomenon that arises from the added presence of 1 when 0 is available, and that brings a crucial reason for this difficulty into focus. To wit, while interpretations of 1-free regular expressions are closed under bisimulation collapse, this is not the case for the interpretations of arbitrary regular expressions.
Process graph interpretations of 1-free regular expressions satisfy the loop existence and elimination property LEE, which is preserved under bisimulation collapse. These features of LEE were applied for showing that an equational proof system for 1-free regular expressions modulo bisimilarity is complete, and that it is decidable in polynomial time whether a process graph is bisimilar to the interpretation of a 1-free regular expression.
While interpretations of regular expressions do not satisfy the property LEE in general, we show that LEE can be recovered by refined interpretations as graphs with 1-transitions refined interpretations with 1-transitions (which are similar to silent steps for automata). This suggests that LEE can be expedient also for the general axiomatization and expressibility problems. But a new phenomenon emerges that needs to be addressed: the property of a process graph `to can be refined into a process graph with 1-transitions and with LEE' is not preserved under bisimulation collapse. We provide a 10-vertex graph with two 1-transitions that satisfies LEE, and in which a pair of bisimilar vertices cannot be collapsed on to each other while preserving the refinement property. This implies that the image of the process interpretation of regular expressions is not closed under bisimulation collapse.
Authors: Borui Cai, Yong Xiang, Longxiang Gao, Di Wu, He Zhang, Jiong Jin, Tom Luan
Knowledge graph embedding (KGE) that maps entities and relations into vector representations is essential for downstream applications. Conventional KGE methods require high-dimensional representations to learn the complex structure of knowledge graph, but lead to oversized model parameters. Recent advances reduce parameters by low-dimensional entity representations, while developing techniques (e.g., knowledge distillation or reinvented representation forms) to compensate for reduced dimension. However, such operations introduce complicated computations and model designs that may not benefit large knowledge graphs. To seek a simple strategy to improve the parameter efficiency of conventional KGE models, we take inspiration from that deeper neural networks require exponentially fewer parameters to achieve expressiveness comparable to wider networks for compositional structures. We view all entity representations as a single-layer embedding network, and conventional KGE methods that adopt high-dimensional entity representations equal widening the embedding network to gain expressiveness. To achieve parameter efficiency, we instead propose a deeper embedding network for entity representations, i.e., a narrow entity embedding layer plus a multi-layer dimension lifting network (LiftNet). Experiments on three public datasets show that by integrating LiftNet, four conventional KGE methods with 16-dimensional representations achieve comparable link prediction accuracy as original models that adopt 512-dimensional representations, saving 68.4% to 96.9% parameters.
Authors: Namrata Shivagunde, Vladislav Lialin, Anna Rumshisky
Language model probing is often used to test specific capabilities of models. However, conclusions from such studies may be limited when the probing benchmarks are small and lack statistical power. In this work, we introduce new, larger datasets for negation (NEG-1500-SIMP) and role reversal (ROLE-1500) inspired by psycholinguistic studies. We dramatically extend existing NEG-136 and ROLE-88 benchmarks using GPT3, increasing their size from 18 and 44 sentence pairs to 750 each. We also create another version of extended negation dataset (NEG-1500-SIMP-TEMP), created using template-based generation. It consists of 770 sentence pairs. We evaluate 22 models on the extended datasets, seeing model performance dip 20-57% compared to the original smaller benchmarks. We observe high levels of negation sensitivity in models like BERT and ALBERT demonstrating that previous findings might have been skewed due to smaller test sets. Finally, we observe that while GPT3 has generated all the examples in ROLE-1500 is only able to solve 24.6% of them during probing. The datasets and code are available on $\href{https://github.com/text-machine-lab/extending_psycholinguistic_dataset}{Github}$.
Authors: Emilio Ferrara
As the capabilities of generative language models continue to advance, the implications of biases ingrained within these models have garnered increasing attention from researchers, practitioners, and the broader public. This article investigates the challenges and risks associated with biases in large-scale language models like ChatGPT. We discuss the origins of biases, stemming from, among others, the nature of training data, model specifications, algorithmic constraints, product design, and policy decisions. We explore the ethical concerns arising from the unintended consequences of biased model outputs. We further analyze the potential opportunities to mitigate biases, the inevitability of some biases, and the implications of deploying these models in various applications, such as virtual assistants, content generation, and chatbots. Finally, we review the current approaches to identify, quantify, and mitigate biases in language models, emphasizing the need for a multi-disciplinary, collaborative effort to develop more equitable, transparent, and responsible AI systems. This article aims to stimulate a thoughtful dialogue within the artificial intelligence community, encouraging researchers and developers to reflect on the role of biases in generative language models and the ongoing pursuit of ethical AI.
Authors: Bohan Li, Longxu Dou, Yutai Hou, Yunlong Feng, Honglin Mu, Qingfu Zhu, Qinghua Sun, Wanxiang Che
Prompt-based learning has shown considerable promise in reformulating various downstream tasks as cloze problems by combining original input with a predetermined template. This approach demonstrates its effectiveness, especially in few-shot learning scenarios, where the model is trained on a scarce amount of data. Despite its successes, the limited templates and text in few-shot prompt-based learning scenarios leave significant room for performance improvement. Moreover, existing methods sometimes resort to model ensembles, which, while effective, could potentially hamper model efficiency due to increased computational demands. To address these issues, we introduce MixPro, an augmentation method designed to augment both the vanilla input text and the templates. We implement this through the token-level, the sentence-level, and the template-level Mixup strategies. The experimental results on five few-shot datasets show that MixPro outperforms other augmentation baselines, improving model performance by an average of 5.08% compared to before augmentation.
Authors: Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, Daphne Ippolito
Pretraining is the preliminary and fundamental step in developing capable language models (LM). Despite this, pretraining data design is critically under-documented and often guided by empirically unsupported intuitions. To address this, we pretrain 28 1.5B parameter decoder-only models, training on data curated (1) at different times, (2) with varying toxicity and quality filters, and (3) with different domain compositions. First, we quantify the effect of pretraining data age. A temporal shift between evaluation data and pretraining data leads to performance degradation, which is not overcome by finetuning. Second, we explore the effect of quality and toxicity filters, showing a trade-off between performance on standard benchmarks and risk of toxic generations. Our findings indicate there does not exist a one-size-fits-all solution to filtering training data. We also find that the effects of different types of filtering are not predictable from text domain characteristics. Lastly, we empirically validate that the inclusion of heterogeneous data sources, like books and web, is broadly beneficial and warrants greater prioritization. These findings constitute the largest set of experiments to validate, quantify, and expose many undocumented intuitions about text pretraining, which we hope will help support more informed data-centric decisions in LM development.
Authors: Evgeniia Razumovskaia, Ivan Vulić, Anna Korhonen
Slot labeling (SL) is a core component of task-oriented dialogue (ToD) systems, where slots and corresponding values are usually language-, task- and domain-specific. Therefore, extending the system to any new language-domain-task configuration requires (re)running an expensive and resource-intensive data annotation process. To mitigate the inherent data scarcity issue, current research on multilingual ToD assumes that sufficient English-language annotated data are always available for particular tasks and domains, and thus operates in a standard cross-lingual transfer setup. In this work, we depart from this often unrealistic assumption. We examine challenging scenarios where such transfer-enabling English annotated data cannot be guaranteed, and focus on bootstrapping multilingual data-efficient slot labelers in transfer-free scenarios directly in the target languages without any English-ready data. We propose a two-stage slot labeling approach (termed TWOSL) which transforms standard multilingual sentence encoders into effective slot labelers. In Stage 1, relying on SL-adapted contrastive learning with only a handful of SL-annotated examples, we turn sentence encoders into task-specific span encoders. In Stage 2, we recast SL from a token classification into a simpler, less data-intensive span classification task. Our results on two standard multilingual TOD datasets and across diverse languages confirm the effectiveness and robustness of TWOSL. It is especially effective for the most challenging transfer-free few-shot setups, paving the way for quick and data-efficient bootstrapping of multilingual slot labelers for ToD.
Authors: Yaoting Wang, Yuanchao Li, Paul Pu Liang, Louis-Philippe Morency, Peter Bell, Catherine Lai
Fusing multiple modalities has proven effective for multimodal information processing. However, the incongruity between modalities poses a challenge for multimodal fusion, especially in affect recognition. In this study, we first analyze how the salient affective information in one modality can be affected by the other, and demonstrate that inter-modal incongruity exists latently in crossmodal attention. Based on this finding, we propose the Hierarchical Crossmodal Transformer with Dynamic Modality Gating (HCT-DMG), a lightweight incongruity-aware model, which dynamically chooses the primary modality in each training batch and reduces fusion times by leveraging the learned hierarchy in the latent space to alleviate incongruity. The experimental evaluation on five benchmark datasets: CMU-MOSI, CMU-MOSEI, and IEMOCAP (sentiment and emotion), where incongruity implicitly lies in hard samples, as well as UR-FUNNY (humour) and MUStaRD (sarcasm), where incongruity is common, verifies the efficacy of our approach, showing that HCT-DMG: 1) outperforms previous multimodal models with a reduced size of approximately 0.8M parameters; 2) recognizes hard samples where incongruity makes affect recognition difficult; 3) mitigates the incongruity at the latent level in crossmodal attention.
Authors: Zihan Wang, Jingbo Shang, Ruiqi Zhong
Unsupervised clustering is widely used to explore large corpora, but existing formulations neither consider the users' goals nor explain clusters' meanings. We propose a new task formulation, "Goal-Driven Clustering with Explanations" (GoalEx), which represents both the goal and the explanations as free-form language descriptions. For example, to categorize the errors made by a summarization system, the input to GoalEx is a corpus of annotator-written comments for system-generated summaries and a goal description "cluster the comments based on why the annotators think the summary is imperfect.''; the outputs are text clusters each with an explanation ("this cluster mentions that the summary misses important context information."), which relates to the goal and precisely explain which comments should (not) belong to a cluster. To tackle GoalEx, we prompt a language model with "[corpus subset] + [goal] + Brainstorm a list of explanations each representing a cluster."; then we classify whether each sample belongs to a cluster based on its explanation; finally, we use integer linear programming to select a subset of candidate clusters to cover most samples while minimizing overlaps. Under both automatic and human evaluation on corpora with or without labels, our method produces more accurate and goal-related explanations than prior methods. We release our data and implementation at https://github.com/ZihanWangKi/GoalEx.
Authors: Nicholas Deas, Jessi Grieser, Shana Kleiner, Desmond Patton, Elsbeth Turcan, Kathleen McKeown
We evaluate how well LLMs understand African American Language (AAL) in comparison to their performance on White Mainstream English (WME), the encouraged "standard" form of English taught in American classrooms. We measure LLM performance using automatic metrics and human judgments for two tasks: a counterpart generation task, where a model generates AAL (or WME) given WME (or AAL), and a masked span prediction (MSP) task, where models predict a phrase that was removed from their input. Our contributions include: (1) evaluation of six pre-trained, large language models on the two language generation tasks; (2) a novel dataset of AAL text from multiple contexts (social media, hip-hop lyrics, focus groups, and linguistic interviews) with human-annotated counterparts in WME; and (3) documentation of model performance gaps that suggest bias and identification of trends in lack of understanding of AAL features.
Authors: Vivek Verma, Eve Fleisig, Nicholas Tomlin, Dan Klein
We introduce Ghostbuster, a state-of-the-art system for detecting AI-generated text. Our method works by passing documents through a series of weaker language models, running a structured search over possible combinations of their features, and then training a classifier on the selected features to predict whether documents are AI-generated. Crucially, Ghostbuster does not require access to token probabilities from the target model, making it useful for detecting text generated by black-box models or unknown model versions. In conjunction with our model, we release three new datasets of human- and AI-generated text as detection benchmarks in the domains of student essays, creative writing, and news articles. We compare Ghostbuster to a variety of existing detectors, including DetectGPT and GPTZero, as well as a new RoBERTa baseline. Ghostbuster achieves 99.0 F1 when evaluated across domains, which is 5.9 F1 higher than the best preexisting model. It also outperforms all previous approaches in generalization across writing domains (+7.5 F1), prompting strategies (+2.1 F1), and language models (+4.4 F1). We also analyze the robustness of our system to a variety of perturbations and paraphrasing attacks and evaluate its performance on documents written by non-native English speakers.
Authors: Terry Yue Zhuo, Zhou Yang, Zhensu Sun, Yufei Wang, Li Li, Xiaoning Du, Zhenchang Xing, David Lo
The increasingly popular adoption of deep learning models in many critical source code tasks motivates the development of data augmentation (DA) techniques to enhance training data and improve various capabilities (e.g., robustness and generalizability) of these models. Although a series of DA methods have been proposed and tailored for source code models, there lacks a comprehensive survey and examination to understand their effectiveness and implications. This paper fills this gap by conducting a comprehensive and integrative survey of data augmentation for source code, wherein we systematically compile and encapsulate existing literature to provide a comprehensive overview of the field. We start with an introduction of data augmentation in source code and then provide a discussion on major representative approaches. Next, we highlight the general strategies and techniques to optimize the DA quality. Subsequently, we underscore techniques useful in real-world source code scenarios and downstream tasks. Finally, we outline the prevailing challenges and potential opportunities for future research. In essence, we aim to demystify the corpus of existing literature on source code DA for deep learning, and foster further exploration in this sphere. Complementing this, we present a continually updated GitHub repository that hosts a list of update-to-date papers on DA for source code modeling, accessible at \url{https://github.com/terryyz/DataAug4Code}.
Authors: Jonathan D. Chang, Kiante Brantley, Rajkumar Ramamurthy, Dipendra Misra, Wen Sun
Reinforcement learning (RL) has emerged as a powerful paradigm for fine-tuning Large Language Models (LLMs) for text generation. In particular, recent LLMs such as ChatGPT and GPT-4 can engage in fluent conversations with users after finetuning with RL. Capitalizing on key properties of text generation, we seek to investigate RL algorithms beyond general purpose algorithms like Proximal Policy Optimization (PPO). In particular, we extend RL algorithms to allow them to interact with a dynamic black-box guide LLM and propose RL with guided feedback (RLGF), a suite of RL algorithms for LLM fine-tuning. We provide two ways for the guide LLM to interact with the LLM to be optimized for maximizing rewards. The guide LLM can generate text which serves as additional starting states for the RL optimization procedure. The guide LLM can also be used to complete the partial sentences generated by the LLM that is being optimized, treating the guide LLM as an expert to imitate and surpass eventually. We experiment on the IMDB positive sentiment, CommonGen, and TL;DR summarization tasks. We show that our RL algorithms achieve higher performance than supervised learning (SL) and the RL baseline PPO, demonstrating the benefit of interaction with the guide LLM. On both CommonGen and TL;DR, we not only outperform our SL baselines but also improve upon PPO across a variety of metrics beyond the one we optimized for. Our code can be found at https://github.com/Cornell-RL/tril.
Authors: Minghao Wu, Alham Fikri Aji
As large language models (LLMs) continue to advance, accurately and comprehensively evaluating their performance becomes increasingly challenging. Ranking the relative performance of LLMs based on Elo ratings, according to human judgment, is gaining more popularity. However, the extent to which humans and LLMs are capable evaluators remains uncertain. This study investigates the behavior of crowd-sourced and expert annotators, as well as LLMs, when comparing outputs from different models. To achieve this, we curate a dataset of intentionally flawed machine-generated answers. Our findings reveal a concerning bias in the evaluation process, as answers with factual errors are rated more favorably than answers that are too short or contained grammatical errors. To address this issue, we propose independently evaluating machine-generated text across multiple dimensions, rather than merging all the evaluation aspects into a single score. We instantiate this idea with the Elo rating system, resulting in the Multi-Elo Rating System (MERS). Empirical results from our study reveal that this proposed approach significantly enhances the quality of LLM-based evaluations, particularly in terms of factual accuracy. However, there is no significant improvement in crowd-sourced-based evaluations, indicating the need for further investigation.
Authors: Jiate Liu, Yiqin Zhu, Kaiwen Xiao, Qiang Fu, Xiao Han, Wei Yang, Deheng Ye
The goal of program synthesis, or code generation, is to generate executable code based on given descriptions. Recently, there has been an increasing number of studies employing reinforcement learning (RL) to improve the performance of large language models (LLMs) for code. However, current representative works either rely solely on offline frameworks, limiting the exploration of new sample spaces, or fall short in the utilization of unit test signals, not accounting for specific error locations within the code. To address these issues, we propose RLTF, i.e., Reinforcement Learning from Unit Test Feedback, a novel online RL framework with unit test feedback of multi-granularity for refining code LLMs. Our approach generates data in real-time during training and simultaneously utilizes fine-grained feedback signals to guide the model towards producing higher-quality code. Extensive experiments show that RLTF achieves state-of-the-art performance on the APPS and the MBPP benchmarks. Our code is available at: https://github.com/Zyq-scut/RLTF.
Authors: Jean Kaddour, Oscar Key, Piotr Nawrot, Pasquale Minervini, Matt J. Kusner
The computation necessary for training Transformer-based language models has skyrocketed in recent years. This trend has motivated research on efficient training algorithms designed to improve training, validation, and downstream performance faster than standard training. In this work, we revisit three categories of such algorithms: dynamic architectures (layer stacking, layer dropping), batch selection (selective backprop, RHO loss), and efficient optimizers (Lion, Sophia). When pre-training BERT and T5 with a fixed computation budget using such methods, we find that their training, validation, and downstream gains vanish compared to a baseline with a fully-decayed learning rate. We define an evaluation protocol that enables computation to be done on arbitrary machines by mapping all computation time to a reference machine which we call reference system time. We discuss the limitations of our proposed protocol and release our code to encourage rigorous research in efficient training procedures: https://github.com/JeanKaddour/NoTrainNoGain.
Authors: Le Peng, Gaoxiang Luo, sicheng zhou, jiandong chen, Rui Zhang, Ziyue Xu, Ju Sun
Language models (LMs) such as BERT and GPT have revolutionized natural language processing (NLP). However, the medical field faces challenges in training LMs due to limited data access and privacy constraints imposed by regulations like the Health Insurance Portability and Accountability Act (HIPPA) and the General Data Protection Regulation (GDPR). Federated learning (FL) offers a decentralized solution that enables collaborative learning while ensuring data privacy. In this study, we evaluated FL on 2 biomedical NLP tasks encompassing 8 corpora using 6 LMs. Our results show that: 1) FL models consistently outperformed models trained on individual clients' data and sometimes performed comparably with models trained with polled data; 2) with the fixed number of total data, FL models training with more clients produced inferior performance but pre-trained transformer-based models exhibited great resilience. 3) FL models significantly outperformed large language models using zero-/one-shot learning and offered lightning inference speed.
Authors: Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, Xing Xie
Emotional intelligence significantly impacts our daily behaviors and interactions. Although Large Language Models (LLMs) are increasingly viewed as a stride toward artificial general intelligence, exhibiting impressive performance in numerous tasks, it is still uncertain if LLMs can genuinely grasp psychological emotional stimuli. Understanding and responding to emotional cues gives humans a distinct advantage in problem-solving. In this paper, we take the first step towards exploring the ability of LLMs to understand emotional stimuli. To this end, we first conduct automatic experiments on 45 tasks using various LLMs, including Flan-T5-Large, Vicuna, Llama 2, BLOOM, ChatGPT, and GPT-4. Our tasks span deterministic and generative applications that represent comprehensive evaluation scenarios. Our automatic experiments show that LLMs have a grasp of emotional intelligence, and their performance can be improved with emotional prompts (which we call "EmotionPrompt" that combines the original prompt with emotional stimuli), e.g., 8.00% relative performance improvement in Instruction Induction and 115% in BIG-Bench. In addition to those deterministic tasks that can be automatically evaluated using existing metrics, we conducted a human study with 106 participants to assess the quality of generative tasks using both vanilla and emotional prompts. Our human study results demonstrate that EmotionPrompt significantly boosts the performance of generative tasks (10.9% average improvement in terms of performance, truthfulness, and responsibility metrics). We provide an in-depth discussion regarding why EmotionPrompt works for LLMs and the factors that may influence its performance. We posit that EmotionPrompt heralds a novel avenue for exploring interdisciplinary knowledge for human-LLMs interaction.
Authors: Rui Zhang, Yixin Su, Bayu Distiawan Trisedya, Xiaoyan Zhao, Min Yang, Hong Cheng, Jianzhong Qi
The task of entity alignment between knowledge graphs (KGs) aims to identify every pair of entities from two different KGs that represent the same entity. Many machine learning-based methods have been proposed for this task. However, to our best knowledge, existing methods all require manually crafted seed alignments, which are expensive to obtain. In this paper, we propose the first fully automatic alignment method named AutoAlign, which does not require any manually crafted seed alignments. Specifically, for predicate embeddings, AutoAlign constructs a predicate-proximity-graph with the help of large language models to automatically capture the similarity between predicates across two KGs. For entity embeddings, AutoAlign first computes the entity embeddings of each KG independently using TransE, and then shifts the two KGs' entity embeddings into the same vector space by computing the similarity between entities based on their attributes. Thus, both predicate alignment and entity alignment can be done without manually crafted seed alignments. AutoAlign is not only fully automatic, but also highly effective. Experiments using real-world KGs show that AutoAlign improves the performance of entity alignment significantly compared to state-of-the-art methods.
Authors: Kadhim Hayawi, Sakib Shahriar, Sujith Samuel Mathew
The potential of artificial intelligence (AI)-based large language models (LLMs) holds considerable promise in revolutionizing education, research, and practice. However, distinguishing between human-written and AI-generated text has become a significant task. This paper presents a comparative study, introducing a novel dataset of human-written and LLM-generated texts in different genres: essays, stories, poetry, and Python code. We employ several machine learning models to classify the texts. Results demonstrate the efficacy of these models in discerning between human and AI-generated text, despite the dataset's limited sample size. However, the task becomes more challenging when classifying GPT-generated text, particularly in story writing. The results indicate that the models exhibit superior performance in binary classification tasks, such as distinguishing human-generated text from a specific LLM, compared to the more complex multiclass tasks that involve discerning among human-generated and multiple LLMs. Our findings provide insightful implications for AI text detection while our dataset paves the way for future research in this evolving area.
Authors: Seokjin Oh, Su Ah Lee, Woohwan Jung
Despite the rapid growth in model architecture, the scarcity of large parallel corpora remains the main bottleneck in Neural Machine Translation. Data augmentation is a technique that enhances the performance of data-hungry models by generating synthetic data instead of collecting new ones. We explore prompt-based data augmentation approaches that leverage large-scale language models such as ChatGPT. To create a synthetic parallel corpus, we compare 3 methods using different prompts. We employ two assessment metrics to measure the diversity of the generated synthetic data. This approach requires no further model training cost, which is mandatory in other augmentation methods like back-translation. The proposed method improves the unaugmented baseline by 0.68 BLEU score.
Authors: Minsoo Kim, Sihwa Lee, Janghwan Lee, Sukjin Hong, Du-Seong Chang, Wonyong Sung, Jungwook Choi
Generative Language Models (GLMs) have shown impressive performance in tasks such as text generation, understanding, and reasoning. However, the large model size poses challenges for practical deployment. To solve this problem, Quantization-Aware Training (QAT) has become increasingly popular. However, current QAT methods for generative models have resulted in a noticeable loss of accuracy. To counteract this issue, we propose a novel knowledge distillation method specifically designed for GLMs. Our method, called token-scaled logit distillation, prevents overfitting and provides superior learning from the teacher model and ground truth. This research marks the first evaluation of ternary weight quantization-aware training of large-scale GLMs with less than 1.0 degradation in perplexity and achieves enhanced accuracy in tasks like common-sense QA and arithmetic reasoning as well as natural language understanding. Our code is available at https://github.com/aiha-lab/TSLD.
Authors: Terufumi Morishita, Gaku Morio, Atsuki Yamaguchi, Yasuhiro Sogawa
We study a synthetic corpus based approach for language models (LMs) to acquire logical deductive reasoning ability. The previous studies generated deduction examples using specific sets of deduction rules. However, these rules were limited or otherwise arbitrary, limiting the generalizability of acquired reasoning ability. We rethink this and adopt a well-grounded set of deduction rules based on formal logic theory, which can derive any other deduction rules when combined in a multistep way. Then, using the proposed corpora, which we name FLD (Formal Logic Deduction), we first evaluate and analyze the logical reasoning ability of the latest LLMs. Even GPT-4 can solve only half of the problems, suggesting that pure logical reasoning isolated from knowledge is still challenging for the LLMs, and additional training specialized in logical reasoning is indeed essential. We next empirically verify that LMs trained on FLD corpora acquire more generalizable reasoning ability. Furthermore, we identify the aspects of reasoning ability on which deduction corpora can enhance LMs and those on which they cannot, and discuss future directions on each aspect. The released corpora serve both as learning resources and as challenging benchmarks.
Authors: Alexander Matt Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, Monte MacDiarmid
Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering and guided decoding. We instead investigate activation engineering: modifying activations at inference-time to predictably alter model behavior. We bias the forward pass with a 'steering vector' implicitly specified through natural language. Past work learned these steering vectors; our Activation Addition (ActAdd) method instead computes them by taking the activation differences which result from pairs of prompts.
We demonstrate ActAdd on GPT-2 on OpenWebText and ConceptNet, and replicate the effect on Llama-13B and GPT-J-6B. Our approach yields inference-time control over high-level properties of output & preserves performance on off-target topics. The method requires far less compute and implementation effort than finetuning and RLHF, allows for natural language specification by users, and its overhead scales naturally with model size.
Authors: Lucas Cassiel Jacaruso
Topic modeling and text mining are subsets of Natural Language Processing (NLP) with relevance for conducting meta-analysis (MA) and systematic review (SR). For evidence synthesis, the above NLP methods are conventionally used for topic-specific literature searches or extracting values from reports to automate essential phases of SR and MA. Instead, this work proposes a comparative topic modeling approach to analyze reports of contradictory results on the same general research question. Specifically, the objective is to identify topics exhibiting distinct associations with significant results for an outcome of interest by ranking them according to their proportional occurrence in (and consistency of distribution across) reports of significant effects. The proposed method was tested on broad-scope studies addressing whether supplemental nutritional compounds significantly benefit macular degeneration (MD). Six compounds were identified as having a particular association with reports of significant results for benefiting MD. Four of these were further supported in terms of effectiveness upon conducting a follow-up literature search for validation (omega-3 fatty acids, copper, zeaxanthin, and nitrates). The two not supported by the follow-up literature search (niacin and molybdenum) also had scores in the lowest range under the proposed scoring system, suggesting that the proposed methods score for a given topic may be a viable proxy for its degree of association with the outcome of interest and can be helpful in the search for potentially causal relationships. These results underpin the proposed methods potential to add specificity in understanding effects from broad-scope reports, elucidate topics of interest for future research, and guide evidence synthesis in a systematic and scalable way. All of this is accomplished while yielding valuable insights into the prevention of MD.
Authors: Zachary Levonian, Chenglu Li, Wangda Zhu, Anoushka Gade, Owen Henkel, Millie-Ellen Postle, Wanli Xing
For middle-school math students, interactive question-answering (QA) with tutors is an effective way to learn. The flexibility and emergent capabilities of generative large language models (LLMs) has led to a surge of interest in automating portions of the tutoring process - including interactive QA to support conceptual discussion of mathematical concepts. However, LLM responses to math questions can be incorrect or mismatched to the educational context - such as being misaligned with a school's curriculum. One potential solution is retrieval-augmented generation (RAG), which involves incorporating a vetted external knowledge source in the LLM prompt to increase response quality. In this paper, we designed prompts that retrieve and use content from a high-quality open-source math textbook to generate responses to real student questions. We evaluate the efficacy of this RAG system for middle-school algebra and geometry QA by administering a multi-condition survey, finding that humans prefer responses generated using RAG, but not when responses are too grounded in the textbook content. We argue that while RAG is able to improve response quality, designers of math QA systems must consider trade-offs between generating responses preferred by students and responses closely matched to specific educational resources.
Authors: Dasol Choi, Jooyoung Song, Eunsun Lee, Jinwoo Seo, Heejune Park, Dongbin Na
With the growth of online services, the need for advanced text classification algorithms, such as sentiment analysis and biased text detection, has become increasingly evident. The anonymous nature of online services often leads to the presence of biased and harmful language, posing challenges to maintaining the health of online communities. This phenomenon is especially relevant in South Korea, where large-scale hate speech detection algorithms have not yet been broadly explored. In this paper, we introduce "KoMultiText", a new comprehensive, large-scale dataset collected from a well-known South Korean SNS platform. Our proposed dataset provides annotations including (1) Preferences, (2) Profanities, and (3) Nine types of Bias for the text samples, enabling multi-task learning for simultaneous classification of user-generated texts. Leveraging state-of-the-art BERT-based language models, our approach surpasses human-level accuracy across diverse classification tasks, as measured by various metrics. Beyond academic contributions, our work can provide practical solutions for real-world hate speech and bias mitigation, contributing directly to the improvement of online community health. Our work provides a robust foundation for future research aiming to improve the quality of online discourse and foster societal well-being. All source codes and datasets are publicly accessible at https://github.com/Dasol-Choi/KoMultiText.
Authors: Neng Wang, Hongyang Yang, Christina Dan Wang
In the swiftly expanding domain of Natural Language Processing (NLP), the potential of GPT-based models for the financial sector is increasingly evident. However, the integration of these models with financial datasets presents challenges, notably in determining their adeptness and relevance. This paper introduces a distinctive approach anchored in the Instruction Tuning paradigm for open-source large language models, specifically adapted for financial contexts. Through this methodology, we capitalize on the interoperability of open-source models, ensuring a seamless and transparent integration. We begin by explaining the Instruction Tuning paradigm, highlighting its effectiveness for immediate integration. The paper presents a benchmarking scheme designed for end-to-end training and testing, employing a cost-effective progression. Firstly, we assess basic competencies and fundamental tasks, such as Named Entity Recognition (NER) and sentiment analysis to enhance specialization. Next, we delve into a comprehensive model, executing multi-task operations by amalgamating all instructional tunings to examine versatility. Finally, we explore the zero-shot capabilities by earmarking unseen tasks and incorporating novel datasets to understand adaptability in uncharted terrains. Such a paradigm fortifies the principles of openness and reproducibility, laying a robust foundation for future investigations in open-source financial large language models (FinLLMs).
Authors: Chengwen Qi, Bowen Li, Binyuan Hui, Bailin Wang, Jinyang Li, Jinwang Wu, Yuanjun Laili
Large Language Models (LLMs) have achieved remarkable success in many formal language oriented tasks, such as structural data-to-text and semantic parsing. However current benchmarks mostly follow the data distribution of the pre-training data of LLMs. Therefore, a natural question rises that do LLMs really understand the structured semantics of formal languages. In this paper, we investigate this problem on a special case, converse binary relation. We introduce a new benchmark ConvRe focusing on converse relations, which contains 17 relations and 1240 triples extracted from popular knowledge graph completion datasets. Our ConvRE features two tasks, Re2Text and Text2Re, which are formulated as multi-choice question answering to evaluate LLMs' ability to determine the matching between relations and associated text. For the evaluation protocol, apart from different prompting methods, we further introduce variants to the test text and few-shot example text. We conduct experiments on three popular LLM families and have observed various scaling trends. The results suggest that LLMs often resort to shortcut learning and still face challenges on our proposed benchmark.
Authors: Siyang Liu, Naihao Deng, Sahand Sabour, Yilin Jia, Minlie Huang, Rada Mihalcea
We propose task-adaptive tokenization as a way to adapt the generation pipeline to the specifics of a downstream task and enhance long-form generation in mental health. Inspired by insights from cognitive science, our task-adaptive tokenizer samples variable segmentations from multiple outcomes, with sampling probabilities optimized based on task-specific data. We introduce a strategy for building a specialized vocabulary and introduce a vocabulary merging protocol that allows for the integration of task-specific tokens into the pre-trained model's tokenization step. Through extensive experiments on psychological question-answering tasks in both Chinese and English, we find that our task-adaptive tokenization approach brings a significant improvement in generation performance while using up to 60% fewer tokens. Preliminary experiments point to promising results when using our tokenization approach with very large language models.
Authors: Joseph S. Boyle, Antanas Kascenas, Pat Lok, Maria Liakata, Alison Q. O'Neil
The task of assigning diagnostic ICD codes to patient hospital admissions is typically performed by expert human coders. Efforts towards automated ICD coding are dominated by supervised deep learning models. However, difficulties in learning to predict the large number of rare codes remain a barrier to adoption in clinical practice. In this work, we leverage off-the-shelf pre-trained generative large language models (LLMs) to develop a practical solution that is suitable for zero-shot and few-shot code assignment, with no need for further task-specific training. Unsupervised pre-training alone does not guarantee precise knowledge of the ICD ontology and specialist clinical coding task, therefore we frame the task as information extraction, providing a description of each coded concept and asking the model to retrieve related mentions. For efficiency, rather than iterating over all codes, we leverage the hierarchical nature of the ICD ontology to sparsely search for relevant codes.
Authors: Aviv Slobodkin, Avi Caciularu, Eran Hirsch, Ido Dagan
The recently introduced Controlled Text Reduction (CTR) task isolates the text generation step within typical summarization-style tasks. It does so by challenging models to generate coherent text conforming to pre-selected content within the input text (``highlights''). This framing enables increased modularity in summarization-like tasks, allowing to couple a single CTR model with various content-selection setups and modules. However, there are currently no reliable CTR models, while the performance of the existing baseline for the task is mediocre, falling short of practical utility. Here, we address this gap by introducing a high-quality, open-source CTR model that tackles two prior key limitations: inadequate enforcement of the content-preservation constraint, and suboptimal silver training data. Addressing these, we amplify the content-preservation constraint in both training, via RL, and inference, via a controlled decoding strategy. Further, we substantially improve the silver training data quality via GPT-4 distillation. Overall, pairing the distilled dataset with the highlight-adherence strategies yields marked gains over the current baseline, of up to 30 ROUGE-L points, providing a reliable CTR model for downstream use.
Authors: Alex Mei, Sharon Levy, William Yang Wang
As large language models are integrated into society, robustness toward a suite of prompts is increasingly important to maintain reliability in a high-variance environment.Robustness evaluations must comprehensively encapsulate the various settings in which a user may invoke an intelligent system. This paper proposes ASSERT, Automated Safety Scenario Red Teaming, consisting of three methods -- semantically aligned augmentation, target bootstrapping, and adversarial knowledge injection. For robust safety evaluation, we apply these methods in the critical domain of AI safety to algorithmically generate a test suite of prompts covering diverse robustness settings -- semantic equivalence, related scenarios, and adversarial. We partition our prompts into four safety domains for a fine-grained analysis of how the domain affects model performance. Despite dedicated safeguards in existing state-of-the-art models, we find statistically significant performance differences of up to 11% in absolute classification accuracy among semantically related scenarios and error rates of up to 19% absolute error in zero-shot adversarial settings, raising concerns for users' physical safety.
Authors: Ehsan Latif, Xiaoming Zhai
This study highlights the potential of fine-tuned ChatGPT (GPT-3.5) for automatically scoring student written constructed responses using example assessment tasks in science education. Recent studies on OpenAI's generative model GPT-3.5 proved its superiority in predicting the natural language with high accuracy and human-like responses. GPT-3.5 has been trained over enormous online language materials such as journals and Wikipedia; therefore, more than direct usage of pre-trained GPT-3.5 is required for automatic scoring as students utilize a different language than trained material. These imply that a domain-specific model, fine-tuned over data for specific tasks, can enhance model performance. In this study, we fine-tuned GPT-3.5 on six assessment tasks with a diverse dataset of middle-school and high-school student responses and expert scoring. The six tasks comprise two multi-label and four multi-class assessment tasks. We compare the performance of fine-tuned GPT-3.5 with the fine-tuned state-of-the-art Google's generated language model, BERT. The results show that in-domain training corpora constructed from science questions and responses for BERT achieved average accuracy = 0.838, SD = 0.069. GPT-3.5 shows a remarkable average increase (9.1%) in automatic scoring accuracy (mean = 9.15, SD = 0.042) for the six tasks, p =0.001 < 0.05. Specifically, for multi-label tasks (item 1 with 5 labels; item 2 with 10 labels), GPT-3.5 achieved significantly higher scoring accuracy than BERT across all the labels, with the second item achieving a 7.1% increase. The average scoring increase for the four multi-class items for GPT-3.5 was 10.6% compared to BERT. Our study confirmed the effectiveness of fine-tuned GPT-3.5 for automatic scoring of student responses on domain-specific data in education with high accuracy. We have released fine-tuned models for public use and community engagement.
Authors: Hao Zhao, Jie Fu, Zhaofeng He
Parameter-efficient fine-tuning (PEFT) has shown its effectiveness in adapting the pre-trained language models to downstream tasks while only updating a small number of parameters. Despite the success, most existing methods independently adapt to each task without considering knowledge transfer between tasks and are limited to low-data regimes. To overcome this issue, we propose Prototype-based HyperAdapter (PHA), a novel framework built on the adapter-tuning and hypernetwork. It introduces an instance-dense retriever and a prototypical hypernetwork to generate the conditional modules in a sample-efficient manner. This leads to comparable performance improvements against existing PEFT methods on multi-task learning and few-shot transfer learning. More importantly, when the available data size gets smaller, our method outperforms other strong baselines by a large margin. Based on our extensive empirical experiments across various datasets, we demonstrate that PHA strikes a better trade-off between trainable parameters, accuracy on stream tasks, and sample efficiency.
Authors: Jiefeng Chen, Jinsung Yoon, Sayna Ebrahimi, Sercan O Arik, Tomas Pfister, Somesh Jha
Large language models (LLMs) have recently shown great advances in a variety of tasks, including natural language understanding and generation. However, their use in high-stakes decision-making scenarios is still limited due to the potential for errors. Selective prediction is a technique that can be used to improve the reliability of the LLMs by allowing them to abstain from making predictions when they are unsure of the answer. In this work, we propose a novel framework for adaptation with self-evaluation to improve the selective prediction performance of LLMs. Our framework is based on the idea of using parameter-efficient tuning to adapt the LLM to the specific task at hand while improving its ability to perform self-evaluation. We evaluate our method on a variety of question-answering (QA) datasets and show that it outperforms state-of-the-art selective prediction methods. For example, on the CoQA benchmark, our method improves the AUACC from 91.23% to 92.63% and improves the AUROC from 74.61% to 80.25%.
Authors: Su Ah Lee, Seokjin Oh, Woohwan Jung
Named Entity Recognition (NER) frequently suffers from the problem of insufficient labeled data, particularly in fine-grained NER scenarios. Although $K$-shot learning techniques can be applied, their performance tends to saturate when the number of annotations exceeds several tens of labels. To overcome this problem, we utilize existing coarse-grained datasets that offer a large number of annotations. A straightforward approach to address this problem is pre-finetuning, which employs coarse-grained data for representation learning. However, it cannot directly utilize the relationships between fine-grained and coarse-grained entities, although a fine-grained entity type is likely to be a subcategory of a coarse-grained entity type. We propose a fine-grained NER model with a Fine-to-Coarse(F2C) mapping matrix to leverage the hierarchical structure explicitly. In addition, we present an inconsistency filtering method to eliminate coarse-grained entities that are inconsistent with fine-grained entity types to avoid performance degradation. Our experimental results show that our method outperforms both $K$-shot learning and supervised learning methods when dealing with a small number of fine-grained annotations.
Authors: Aviv Slobodkin, Omer Goldman, Avi Caciularu, Ido Dagan, Shauli Ravfogel
Large language models (LLMs) have been shown to possess impressive capabilities, while also raising crucial concerns about the faithfulness of their responses. A primary issue arising in this context is the management of (un)answerable queries by LLMs, which often results in hallucinatory behavior due to overconfidence. In this paper, we explore the behavior of LLMs when presented with (un)answerable queries. We ask: do models represent the fact that the question is (un)answerable when generating a hallucinatory answer? Our results show strong indications that such models encode the answerability of an input query, with the representation of the first decoded token often being a strong indicator. These findings shed new light on the spatial organization within the latent representations of LLMs, unveiling previously unexplored facets of these models. Moreover, they pave the way for the development of improved decoding techniques with better adherence to factual generation, particularly in scenarios where query (un)answerability is a concern.
Authors: Etsuko Ishii, Yan Xu, Bryan Wilie, Ziwei Ji, Holy Lovenia, Willy Chung, Pascale Fung
Inference, especially those derived from inductive processes, is a crucial component in our conversation to complement the information implicitly or explicitly conveyed by a speaker. While recent large language models show remarkable advances in inference tasks, their performance in inductive reasoning, where not all information is present in the context, is far behind deductive reasoning. In this paper, we analyze the behavior of the models based on the task difficulty defined by the semantic information gap -- which distinguishes inductive and deductive reasoning (Johnson-Laird, 1988, 1993). Our analysis reveals that the disparity in information between dialogue contexts and desired inferences poses a significant challenge to the inductive inference process. To mitigate this information gap, we investigate a contrastive learning approach by feeding negative samples. Our experiments suggest negative samples help models understand what is wrong and improve their inference generations.
Authors: Rishabh Bhardwaj, Soujanya Poria
Red-teaming has been a widely adopted way to evaluate the harmfulness of Large Language Models (LLMs). It aims to jailbreak a model's safety behavior to make it act as a helpful agent disregarding the harmfulness of the query. Existing methods are primarily based on input text-based red-teaming such as adversarial prompts, low-resource prompts, or contextualized prompts to condition the model in a way to bypass its safe behavior. Bypassing the guardrails uncovers hidden harmful information and biases in the model that are left untreated or newly introduced by its safety training. However, prompt-based attacks fail to provide such a diagnosis owing to their low attack success rate, and applicability to specific models. In this paper, we present a new perspective on LLM safety research i.e., parametric red-teaming through Unalignment. It simply (instruction) tunes the model parameters to break model guardrails that are not deeply rooted in the model's behavior. Unalignment using as few as 100 examples can significantly bypass commonly referred to as CHATGPT, to the point where it responds with an 88% success rate to harmful queries on two safety benchmark datasets. On open-source models such as VICUNA-7B and LLAMA-2-CHAT 7B AND 13B, it shows an attack success rate of more than 91%. On bias evaluations, Unalignment exposes inherent biases in safety-aligned models such as CHATGPT and LLAMA- 2-CHAT where the model's responses are strongly biased and opinionated 64% of the time.
Authors: Megha Sundriyal, Tanmoy Chakraborty, Preslav Nakov
With the rise of social media, users are exposed to many misleading claims. However, the pervasive noise inherent in these posts presents a challenge in identifying precise and prominent claims that require verification. Extracting the important claims from such posts is arduous and time-consuming, yet it is an underexplored problem. Here, we aim to bridge this gap. We introduce a novel task, Claim Normalization (aka ClaimNorm), which aims to decompose complex and noisy social media posts into more straightforward and understandable forms, termed normalized claims. We propose CACN, a pioneering approach that leverages chain-of-thought and claim check-worthiness estimation, mimicking human reasoning processes, to comprehend intricate claims. Moreover, we capitalize on the in-context learning capabilities of large language models to provide guidance and to improve claim normalization. To evaluate the effectiveness of our proposed model, we meticulously compile a comprehensive real-world dataset, CLAN, comprising more than 6k instances of social media posts alongside their respective normalized claims. Our experiments demonstrate that CACN outperforms several baselines across various evaluation measures. Finally, our rigorous error analysis validates CACN's capabilities and pitfalls.
Authors: Hans W. A. Hanley, Zakir Durumeric
Stance detection is important for understanding different attitudes and beliefs on the Internet. However, given that a passage's stance toward a given topic is often highly dependent on that topic, building a stance detection model that generalizes to unseen topics is difficult. In this work, we propose using contrastive learning as well as an unlabeled dataset of news articles that cover a variety of different topics to train topic-agnostic/TAG and topic-aware/TAW embeddings for use in downstream stance detection. Combining these embeddings in our full TATA model, we achieve state-of-the-art performance across several public stance detection datasets (0.771 $F_1$-score on the Zero-shot VAST dataset). We release our code and data at https://github.com/hanshanley/tata.
Authors: Stephanie Brandl, Emanuele Bugliarello, Ilias Chalkidis
In order to build reliable and trustworthy NLP applications, models need to be both fair across different demographics and explainable. Usually these two objectives, fairness and explainability, are optimized and/or examined independently of each other. Instead, we argue that forthcoming, trustworthy NLP systems should consider both. In this work, we perform a first study to understand how they influence each other: do fair(er) models rely on more plausible rationales? and vice versa. To this end, we conduct experiments on two English multi-class text classification datasets, BIOS and ECtHR, that provide information on gender and nationality, respectively, as well as human-annotated rationales. We fine-tune pre-trained language models with several methods for (i) bias mitigation, which aims to improve fairness; (ii) rationale extraction, which aims to produce plausible explanations. We find that bias mitigation algorithms do not always lead to fairer models. Moreover, we discover that empirical fairness and explainability are orthogonal.
Authors: Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, Gillian Hadfield, Jeff Clune, Tegan Maharaj, Frank Hutter, Atılım Güneş Baydin, Sheila McIlraith, Qiqi Gao, Ashwin Acharya, David Krueger, Anca Dragan, Philip Torr, Stuart Russell, Daniel Kahneman, Jan Brauner, Sören Mindermann
In this short consensus paper, we outline risks from upcoming, advanced AI systems. We examine large-scale social harms and malicious uses, as well as an irreversible loss of human control over autonomous AI systems. In light of rapid and continuing AI progress, we propose urgent priorities for AI R&D and governance.
Authors: Hanqing Wang, Yajing Luo, Boya Xiong, Guanhua Chen, Yun Chen
Stylistic headline generation is the task to generate a headline that not only summarizes the content of an article, but also reflects a desired style that attracts users. As style-specific article-headline pairs are scarce, previous researches focus on unsupervised approaches with a standard headline generation dataset and mono-style corpora. In this work, we follow this line and propose StyleBART, an unsupervised approach for stylistic headline generation. Our method decorates the pretrained BART model with adapters that are responsible for different styles and allows the generation of headlines with diverse styles by simply switching the adapters. Different from previous works, StyleBART separates the task of style learning and headline generation, making it possible to freely combine the base model and the style adapters during inference. We further propose an inverse paraphrasing task to enhance the style adapters. Extensive automatic and human evaluations show that StyleBART achieves new state-of-the-art performance in the unsupervised stylistic headline generation task, producing high-quality headlines with the desired style.
Authors: Yuchen Shen, Xiaojun Wan
Opinion summarization sets itself apart from other types of summarization tasks due to its distinctive focus on aspects and sentiments. Although certain automated evaluation methods like ROUGE have gained popularity, we have found them to be unreliable measures for assessing the quality of opinion summaries. In this paper, we present OpinSummEval, a dataset comprising human judgments and outputs from 14 opinion summarization models. We further explore the correlation between 24 automatic metrics and human ratings across four dimensions. Our findings indicate that metrics based on neural networks generally outperform non-neural ones. However, even metrics built on powerful backbones, such as BART and GPT-3/3.5, do not consistently correlate well across all dimensions, highlighting the need for advancements in automated evaluation methods for opinion summarization. The code and data are publicly available at https://github.com/A-Chicharito-S/OpinSummEval/tree/main.
Authors: Chloe Qinyu Zhu, Rickard Stureborg, Brandon Fain
Language Representation Models (LRMs) trained with real-world data may capture and exacerbate undesired bias and cause unfair treatment of people in various demographic groups. Several techniques have been investigated for applying interventions to LRMs to remove bias in benchmark evaluations on, for example, word embeddings. However, the negative side effects of debiasing interventions are usually not revealed in the downstream tasks. We propose xGAP-DEBIAS, a set of evaluations on assessing the fairness of debiasing. In this work, We examine four debiasing techniques on a real-world text classification task and show that reducing biasing is at the cost of degrading performance for all demographic groups, including those the debiasing techniques aim to protect. We advocate that a debiasing technique should have good downstream performance with the constraint of ensuring no harm to the protected group.
Authors: Pengyue Jia, Yiding Liu, Xiangyu Zhao, Xiaopeng Li, Changying Hao, Shuaiqiang Wang, Dawei Yin
Query expansion is a commonly-used technique in many search systems to better represent users' information needs with additional query terms. Existing studies for this task usually propose to expand a query with retrieved or generated contextual documents. However, both types of methods have clear limitations. For retrieval-based methods, the documents retrieved with the original query might not be accurate enough to reveal the search intent, especially when the query is brief or ambiguous. For generation-based methods, existing models can hardly be trained or aligned on a particular corpus, due to the lack of corpus-specific labeled data. In this paper, we propose a novel Large Language Model (LLM) based mutual verification framework for query expansion, which alleviates the aforementioned limitations. Specifically, we first design a query-query-document generation pipeline, which can effectively leverage the contextual knowledge encoded in LLMs to generate sub-queries and corresponding documents from multiple perspectives. Next, we employ a mutual verification method for both generated and retrieved contextual documents, where 1) retrieved documents are filtered with the external contextual knowledge in generated documents, and 2) generated documents are filtered with the corpus-specific knowledge in retrieved documents. Overall, the proposed method allows retrieved and generated documents to complement each other to finalize a better query expansion. We conduct extensive experiments on three information retrieval datasets, i.e., TREC-DL-2020, TREC-COVID, and MSMARCO. The results demonstrate that our method outperforms other baselines significantly.
Authors: Soon-Jae Hwang, Chang-Sung Jeong
Neural Machine Translation (NMT) has become a significant technology in natural language processing through extensive research and development. However, the deficiency of high-quality bilingual language pair data still poses a major challenge to improving NMT performance. Recent studies are exploring the use of contextual information from pre-trained language model (PLM) to address this problem. Yet, the issue of incompatibility between PLM and NMT model remains unresolved. This study proposes a PLM-integrated NMT (PiNMT) model to overcome the identified problems. The PiNMT model consists of three critical components, PLM Multi Layer Converter, Embedding Fusion, and Cosine Alignment, each playing a vital role in providing effective PLM information to NMT. Furthermore, two training strategies, Separate Learning Rates and Dual Step Training, are also introduced in this paper. By implementing the proposed PiNMT model and training strategy, we achieved state-of-the-art performance on the IWSLT'14 En$\leftrightarrow$De dataset. This study's outcomes are noteworthy as they demonstrate a novel approach for efficiently integrating PLM with NMT to overcome incompatibility and enhance performance.
Authors: Haryo Akbarianto Wibowo, Erland Hilman Fuadi, Made Nindyatama Nityasya, Radityo Eko Prasojo, Alham Fikri Aji
We present publicly available COPAL-ID, a novel Indonesian language common sense reasoning dataset. Unlike the previous Indonesian COPA dataset (XCOPA-ID), COPAL-ID incorporates Indonesian local and cultural nuances, and therefore, provides a more natural portrayal of day-to-day causal reasoning within the Indonesian cultural sphere. Professionally written by natives from scratch, COPAL-ID is more fluent and free from awkward phrases, unlike the translated XCOPA-ID. In addition, we present COPAL-ID in both standard Indonesian and in Jakartan Indonesian--a dialect commonly used in daily conversation. COPAL-ID poses a greater challenge for existing open-sourced and closed state-of-the-art multilingual language models, yet is trivially easy for humans. Our findings suggest that even the current best open-source, multilingual model struggles to perform well, achieving 65.47% accuracy on COPAL-ID, significantly lower than on the culturally-devoid XCOPA-ID (79.40%). Despite GPT-4's impressive score, it suffers the same performance degradation compared to its XCOPA-ID score, and it still falls short of human performance. This shows that these language models are still way behind in comprehending the local nuances of Indonesian.
Authors: Baisong Li, Xingwang Wang, Haixiao Xu
Large language models(LLMs) exhibit excellent performance across a variety of tasks, but they come with significant computational and storage costs. Quantizing these models is an effective way to alleviate this issue. However, existing methods struggle to strike a balance between model accuracy and hardware efficiency. This is where we introduce AWEQ, a post-training method that requires no additional training overhead. AWEQ excels in both ultra-low-bit quantization and 8-bit weight and activation (W8A8) quantization. There is an observation that weight quantization is less challenging than activation quantization. AWEQ transfers the difficulty of activation quantization to weights using channel equalization, achieving a balance between the quantization difficulties of both, and thereby maximizing performance. We have further refined the equalization method to mitigate quantization bias error, ensuring the robustness of the model. Extensive experiments on popular models such as LLaMA and OPT demonstrate that AWEQ outperforms all existing post-training quantization methods for large models.
Authors: Björn Deiseroth, Max Meuer, Nikolas Gritsch, Constantin Eichenberg, Patrick Schramowski, Matthias Aßenmacher, Kristian Kersting
Large Language Models (LLMs) have reshaped natural language processing with their impressive capabilities. Their ever-increasing size, however, raised concerns about their effective deployment and the need for LLM compressions. This study introduces the Divergent Token metrics (DTMs), a novel approach for assessing compressed LLMs, addressing the limitations of traditional perplexity or accuracy measures that fail to accurately reflect text generation quality. DTMs focus on token divergence, that allow deeper insights into the subtleties of model compression, i.p. when evaluating component's impacts individually. Utilizing the First Divergent Token metric (FDTM) in model sparsification reveals that a quarter of all attention components can be pruned beyond 90% on the Llama-2 model family, still keeping SOTA performance. For quantization FDTM suggests that over 80% of parameters can naively be transformed to int8 without special outlier management. These evaluations indicate the necessity of choosing appropriate compressions for parameters individually-and that FDTM can identify those-while standard metrics result in deteriorated outcomes.
Authors: Sean Xie, Soroush Vosoughi, Saeed Hassanpour
Large Language Models (LLMs) have significantly advanced the field of Natural Language Processing (NLP), but their lack of interpretability has been a major concern. Current methods for interpreting LLMs are post hoc, applied after inference time, and have limitations such as their focus on low-level features and lack of explainability at higher level text units. In this work, we introduce proto-lm, a prototypical network-based white-box framework that allows LLMs to learn immediately interpretable embeddings during the fine-tuning stage while maintaining competitive performance. Our method's applicability and interpretability are demonstrated through experiments on a wide range of NLP tasks, and our results indicate a new possibility of creating interpretable models without sacrificing performance. This novel approach to interpretability in LLMs can pave the way for more interpretable models without the need to sacrifice performance.
Authors: Shahbaz Syed, Ahmad Dawar Hakimi, Khalid Al-Khatib, Martin Potthast
Current approaches to automatic summarization of scientific papers generate informative summaries in the form of abstracts. However, abstracts are not intended to show the relationship between a paper and the references cited in it. We propose a new contextualized summarization approach that can generate an informative summary conditioned on a given sentence containing the citation of a reference (a so-called "citance"). This summary outlines the content of the cited paper relevant to the citation location. Thus, our approach extracts and models the citances of a paper, retrieves relevant passages from cited papers, and generates abstractive summaries tailored to each citance. We evaluate our approach using $\textbf{Webis-Context-SciSumm-2023}$, a new dataset containing 540K~computer science papers and 4.6M~citances therein.
Authors: Yann Hicke, Anmol Agarwal, Qianou Ma, Paul Denny
Responding to the thousands of student questions on online QA platforms each semester has a considerable human cost, particularly in computing courses with rapidly growing enrollments. To address the challenges of scalable and intelligent question-answering (QA), we introduce an innovative solution that leverages open-source Large Language Models (LLMs) from the LLaMA-2 family to ensure data privacy. Our approach combines augmentation techniques such as retrieval augmented generation (RAG), supervised fine-tuning (SFT), and learning from human preferences data using Direct Preference Optimization (DPO). Through extensive experimentation on a Piazza dataset from an introductory CS course, comprising 10,000 QA pairs and 1,500 pairs of preference data, we demonstrate a significant 30% improvement in the quality of answers, with RAG being a particularly impactful addition. Our contributions include the development of a novel architecture for educational QA, extensive evaluations of LLM performance utilizing both human assessments and LLM-based metrics, and insights into the challenges and future directions of educational data processing. This work paves the way for the development of CHATA, an intelligent QA assistant customizable for courses with an online QA platform
Authors: Ao Zhang, Wei Ji, Tat-Seng Chua
The development of large language models (LLMs) has greatly advanced the field of multimodal understanding, leading to the emergence of large multimodal models (LMMs). In order to enhance the level of visual comprehension, recent studies have equipped LMMs with region-level understanding capabilities by representing object bounding box coordinates as a series of text sequences (pixel2seq). In this paper, we introduce a novel paradigm for object location modeling called pixel2emb method, where we ask the LMM to output the location embeddings and then decoded by different decoders. This paradigm allows for different location formats (such as bounding boxes and masks) to be used in multimodal conversations Furthermore, this kind of embedding based location modeling enables the utilization of existing practices in localization tasks, such as detection and segmentation. In scenarios with limited resources, our pixel2emb demonstrates superior performance compared to existing state-of-the-art (SOTA) approaches in both the location input and output tasks under fair comparison. Leveraging the proposed pixel2emb method, we train an LMM named NExT-Chat and demonstrate its capability of handling multiple tasks like visual grounding, region caption, and grounded reasoning.
Authors: Shuo Yang, Wei-Lin Chiang, Lianmin Zheng, Joseph E. Gonzalez, Ion Stoica
Large language models are increasingly trained on all the data ever produced by humans. Many have raised concerns about the trustworthiness of public benchmarks due to potential contamination in pre-training or fine-tuning datasets. While most data decontamination efforts apply string matching (e.g., n-gram overlap) to remove benchmark data, we show that these methods are insufficient, and simple variations of test data (e.g., paraphrasing, translation) can easily bypass these decontamination measures. Furthermore, we demonstrate that if such variation of test data is not eliminated, a 13B model can easily overfit a test benchmark and achieve drastically high performance, on par with GPT-4. We validate such observations in widely used benchmarks such as MMLU, GSK8k, and HumanEval. To address this growing risk, we propose a stronger LLM-based decontamination method and apply it to widely used pre-training and fine-tuning datasets, revealing significant previously unknown test overlap. For example, in pre-training sets such as RedPajama-Data-1T and StarCoder-Data, we identified that 8-18\% of the HumanEval benchmark overlaps. Interestingly, we also find such contamination in synthetic dataset generated by GPT-3.5/4, suggesting a potential risk of unintentional contamination. We urge the community to adopt stronger decontamination approaches when using public benchmarks. Moreover, we call for the community to actively develop fresh one-time exams to evaluate models accurately. Our decontamination tool is publicly available at https://github.com/lm-sys/llm-decontaminator.
Authors: Suhaima Jamal, Hayden Wimmer
Phishing and spam detection is long standing challenge that has been the subject of much academic research. Large Language Models (LLM) have vast potential to transform society and provide new and innovative approaches to solve well-established challenges. Phishing and spam have caused financial hardships and lost time and resources to email users all over the world and frequently serve as an entry point for ransomware threat actors. While detection approaches exist, especially heuristic-based approaches, LLMs offer the potential to venture into a new unexplored area for understanding and solving this challenge. LLMs have rapidly altered the landscape from business, consumers, and throughout academia and demonstrate transformational potential for the potential of society. Based on this, applying these new and innovative approaches to email detection is a rational next step in academic research. In this work, we present IPSDM, our model based on fine-tuning the BERT family of models to specifically detect phishing and spam email. We demonstrate our fine-tuned version, IPSDM, is able to better classify emails in both unbalanced and balanced datasets. This work serves as an important first step towards employing LLMs to improve the security of our information systems.
Authors: Qiusi Zhan, Richard Fang, Rohan Bindu, Akul Gupta, Tatsunori Hashimoto, Daniel Kang
As large language models (LLMs) have increased in their capabilities, so does their potential for dual use. To reduce harmful outputs, produces and vendors of LLMs have used reinforcement learning with human feedback (RLHF). In tandem, LLM vendors have been increasingly enabling fine-tuning of their most powerful models. However, concurrent work has shown that fine-tuning can remove RLHF protections. We may expect that the most powerful models currently available (GPT-4) are less susceptible to fine-tuning attacks.
In this work, we show the contrary: fine-tuning allows attackers to remove RLHF protections with as few as 340 examples and a 95% success rate. These training examples can be automatically generated with weaker models. We further show that removing RLHF protections does not decrease usefulness on non-censored outputs, providing evidence that our fine-tuning strategy does not decrease usefulness despite using weaker models to generate training data. Our results show the need for further research on protections on LLMs.
Authors: Nanna Inie, Jonathan Stray, Leon Derczynski
Engaging in the deliberate generation of abnormal outputs from large language models (LLMs) by attacking them is a novel human activity. This paper presents a thorough exposition of how and why people perform such attacks. Using a formal qualitative methodology, we interviewed dozens of practitioners from a broad range of backgrounds, all contributors to this novel work of attempting to cause LLMs to fail. We relate and connect this activity between its practitioners' motivations and goals; the strategies and techniques they deploy; and the crucial role the community plays. As a result, this paper presents a grounded theory of how and why people attack large language models: LLM red teaming in the wild.
Authors: Shahriar Golchin, Mihai Surdeanu
Data contamination, i.e., the presence of test data from downstream tasks in the training data of large language models (LLMs), is a potential major issue in measuring LLMs' real effectiveness on other tasks. We propose a straightforward yet effective method for identifying data contamination within LLMs. At its core, our approach starts by identifying potential contamination at the instance level; using this information, our approach then assesses wider contamination at the partition level. To estimate contamination of individual instances, we employ "guided instruction:" a prompt consisting of the dataset name, partition type, and the random-length initial segment of a reference instance, asking the LLM to complete it. An instance is flagged as contaminated if the LLM's output either exactly or nearly matches the latter segment of the reference. To understand if an entire partition is contaminated, we propose two ideas. The first idea marks a dataset partition as contaminated if the average overlap score with the reference instances (as measured by ROUGE-L or BLEURT) is statistically significantly better with the completions from guided instruction compared to a "general instruction" that does not include the dataset and partition name. The second idea marks a dataset partition as contaminated if a classifier based on GPT-4 with few-shot in-context learning prompt marks multiple generated completions as exact/near-exact matches of the corresponding reference instances. Our best method achieves an accuracy between 92% and 100% in detecting if an LLM is contaminated with seven datasets, containing train and test/validation partitions, when contrasted with manual evaluation by human experts. Further, our findings indicate that GPT-4 is contaminated with AG News, WNLI, and XSum datasets.