LLMs and RAG systems are now capable of handling millions of input tokens or more. However, evaluating the output quality of such systems on long-context tasks remains challenging, as tasks like Needle-in-a-Haystack lack complexity. In this work, we argue that summarization can play a central role in such evaluation. We design a procedure to synthesize Haystacks of documents, ensuring that specific insights repeat across documents. The “Summary of a Haystack” (SummHay) task then requires a system to process the Haystack and generate, given a query, a summary that identifies the relevant insights and precisely cites the source documents. Since we have precise knowledge of what insights should appear in a haystack summary and what documents should be cited, we implement a highly reproducible automatic evaluation that can score summaries on two aspects – Coverage and Citation. We generate Haystacks in two domains (conversation, news), and perform a large-scale evaluation of 10 LLMs and corresponding 50 RAG systems. Our findings indicate that SummHay is an open challenge for current systems, as even systems provided with an Oracle signal of document relevance lag our estimate of human performance (56%) by 10+ points on a Joint Score. Without a retriever, long-context LLMs like GPT-4o and Claude 3 Opus score below 20% on SummHay. We show SummHay can also be used to study enterprise RAG systems and position bias in long-context models. We hope future systems can equal and surpass human performance on SummHay.

Large language models (LLMs) have achieved remarkable performance on a variety of natural language understanding tasks. However, existing benchmarks are inadequate in measuring the complex logical reasoning capabilities of a model. We present FOLIO, a human-annotated, logically complex and diverse dataset for reasoning in natural language (NL), equipped with first-order logic (FOL) annotations. FOLIO consists of 1,430 examples (unique conclusions), each paired with one of 487 sets of premises used to deductively reason for the validity of each conclusion. The logical correctness of the premises and conclusions is ensured by their FOL annotations, which are automatically verified by an FOL inference engine. In addition to the main NL reasoning task, NL-FOL pairs in FOLIO constitute a new NL-FOL translation dataset. Our experiments on FOLIO systematically evaluate the FOL reasoning ability of supervised fine-tuning on medium-sized language models. For both NL reasoning and NL-FOL translation, we benchmark multiple state-of-the-art language models. Our results show that a subset of FOLIO remains a challenge for one of the most capable Large Language Model (LLM) publicly available, GPT-4.

Prompt leakage poses a compelling security and privacy threat in LLM applications. Leakage of system prompts may compromise intellectual property, and act as adversarial reconnaissance for an attacker. A systematic evaluation of prompt leakage threats and mitigation strategies is lacking, especially for multi-turn LLM interactions. In this paper, we systematically investigate LLM vulnerabilities against prompt leakage for 10 closed- and open-source LLMs, across four domains. We design a unique threat model which leverages the LLM sycophancy effect and elevates the average attack success rate (ASR) from 17.7% to 86.2% in a multi-turn setting. Our standardized setup further allows dissecting leakage of specific prompt contents such as task instructions and knowledge documents. We measure the mitigation effect of 7 black-box defense strategies, along with finetuning an open-source model to defend against leakage attempts. We present different combination of defenses against our threat model, including a cost analysis. Our study highlights key takeaways for building secure LLM applications and provides directions for research in multi-turn LLM interactions.

While large language models (LLMs) can already achieve strong performance on standard generic summarization benchmarks, their performance on more complex summarization task settings is less studied. Therefore, we benchmark LLMs on instruction controllable text summarization, where the model input consists of both a source article and a natural language requirement for desired summary characteristics. To this end, we curate an evaluation-only dataset for this task setting and conduct human evaluations of five LLM-based systems to assess their instruction-following capabilities in controllable summarization. We then benchmark LLM-based automatic evaluation for this task with 4 different evaluation protocols and 11 LLMs, resulting in 40 evaluation methods. Our study reveals that instruction controllable text summarization remains a challenging task for LLMs, since (1) all LLMs evaluated still make factual and other types of errors in their summaries; (2) no LLM-based evaluation methods can achieve a strong alignment with human annotators when judging the quality of candidate summaries; (3) different LLMs show large performance gaps in summary generation and evaluation capabilities. We make our collected benchmark InstruSum publicly available to facilitate future research in this direction.

Previous research in multi-document news summarization has typically concentrated on collating information that all sources agree upon. However, the summarization of diverse information dispersed across multiple articles about an event remains underexplored. In this paper, we propose a new task of summarizing diverse information encountered in multiple news articles encompassing the same event. To facilitate this task, we outlined a data collection schema for identifying diverse information and curated a dataset named DiverseSumm. The dataset includes 245 news stories, with each story comprising 10 news articles and paired with a human-validated reference. Next, to enable consistent automatic evaluation, we conducted a comprehensive analysis to pinpoint the position and verbosity biases when utilizing Large Language Model (LLM)-based metrics for evaluating the coverage and faithfulness of summaries. Through correlation analyses, we outline the best practices for effectively using automatic LLM-based metrics on the DiverseSumm dataset. Finally, we study how LLMs summarize multiple news articles by analyzing which type of diverse information LLMs are capable of identifying. Our analyses suggest that despite the extraordinary capabilities of LLMs in single-document summarization, the proposed task remains a complex challenge for them mainly due to their limited coverage, with GPT-4 only able to cover under 40% of the diverse information on average.

People from different social and demographic groups express diverse perspectives and conflicting opinions on a broad set of topics such as product reviews, healthcare, law, and politics. A fair summary should provide a comprehensive coverage of diverse perspectives without underrepresenting certain groups. However, current work in summarization metrics and Large Language Models (LLMs) evaluation has not explored fair abstractive summarization. In this paper, we systematically investigate fair abstractive summarization for user-generated data. We first formally define fairness in abstractive summarization as not underrepresenting perspectives of any groups of people, and we propose four reference-free automatic metrics by measuring the differences between target and source perspectives. We evaluate nine LLMs, including three GPT models, four LLaMA models, PaLM 2, and Claude, on six datasets collected from social media, online reviews, and recorded transcripts. Experiments show that both the model-generated and the human-written reference summaries suffer from low fairness. We conduct a comprehensive analysis of the common factors influencing fairness and propose three simple but effective methods to alleviate unfair summarization. Our dataset and code are available at https://github.com/psunlpgroup/FairSumm.

Recent studies have found that summaries generated by large language models (LLMs) are favored by human annotators over the original reference summaries in commonly used summarization datasets. Therefore, we study an LLM-as-reference learning setting for smaller text summarization models to investigate whether their performance can be substantially improved. To this end, we use LLMs as both oracle summary generators for standard supervised fine-tuning and oracle summary evaluators for efficient contrastive learning that leverages the LLMs’ supervision signals. We conduct comprehensive experiments with source news articles and find that (1) summarization models trained under the LLM-as-reference setting achieve significant performance improvement in both LLM and human evaluations; (2) contrastive learning outperforms standard supervised fine-tuning under both low and high resource settings. Our experimental results also enable a meta-analysis of LLMs’ summary evaluation capacities under a challenging setting, showing that LLMs are not well-aligned with human evaluators. Particularly, our expert human evaluation reveals remaining nuanced performance gaps between LLMs and our fine-tuned models, which LLMs fail to capture. Thus, we call for further studies into both the potential and challenges of using LLMs in summarization model development.

Ideal summarization models should generalize to novel summary-worthy content without remembering reference training summaries by rote. However, a single average performance score on the entire test set is inadequate in determining such model competencies. We propose a fine-grained evaluation protocol by partitioning a test set based on the lexical similarity of reference test summaries with training summaries. We observe up to a 5x (1.2x) difference in ROUGE-2 (entity recall) scores between the subsets with the lowest and highest similarity. Next, we show that such training repetitions also make a model vulnerable to rote learning, reproducing data artifacts such as factual errors, especially when reference test summaries are lexically close to training summaries. Consequently, we propose to limit lexical repetitions in training summaries during both supervised fine-tuning and likelihood calibration stages to improve the performance on novel test cases while retaining average performance. Our automatic and human evaluations on novel test subsets and recent news articles show that limiting lexical repetitions in training summaries can prevent rote learning and improve generalization.

With the recent appearance of LLMs in practical settings, having methods that can effectively detect factual inconsistencies is crucial to reduce the propagation of misinformation and improve trust in model outputs. When testing on existing factual consistency benchmarks, we find that a few large language models (LLMs) perform competitively on classification benchmarks for factual inconsistency detection compared to traditional non-LLM methods. However, a closer analysis reveals issues with existing evaluation benchmarks, affecting evaluation precision. To address this, we propose a new protocol for inconsistency detection benchmark creation and implement it in a 10-domain benchmark called SummEdits. This new benchmark is 20 times more cost-effective per sample than previous benchmarks and highly reproducible, as we estimate inter-annotator agreement at about 0.9. Most LLMs struggle on SummEdits, with performance close to random chance. The best-performing model, GPT-4, is still 8% below estimated human performance, highlighting the gaps in LLMs’ ability to reason about facts and detect inconsistencies when they occur.

Interpretability and efficiency are two important considerations for the adoption of neural automatic metrics. In this work, we develop strong-performing automatic metrics for reference-based summarization evaluation, based on a two-stage evaluation pipeline that first extracts basic information units from one text sequence and then checks the extracted units in another sequence. The metrics we developed include two-stage metrics that can provide high interpretability at both the fine-grained unit level and summary level, and one-stage metrics that achieve a balance between efficiency and interpretability. We make the developed tools publicly available at https://github.com/Yale-LILY/AutoACU.

Community Question Answering (CQA) fora such as Stack Overflow and Yahoo! Answers contain a rich resource of answers to a wide range of community-based questions. Each question thread can receive a large number of answers with different perspectives. One goal of answer summarization is to produce a summary that reflects the range of answer perspectives. A major obstacle for this task is the absence of a dataset to provide supervision for producing such summaries. Recent works propose heuristics to create such data, but these are often noisy and do not cover all answer perspectives present. This work introduces a novel dataset of 4,631 CQA threads for answer summarization curated by professional linguists. Our pipeline gathers annotations for all subtasks of answer summarization, including relevant answer sentence selection, grouping these sentences based on perspectives, summarizing each perspective, and producing an overall summary. We analyze and benchmark state-of-the-art models on these subtasks and introduce a novel unsupervised approach for multi-perspective data augmentation that boosts summarization performance according to automatic evaluation. Finally, we propose reinforcement learning rewards to improve factual consistency and answer coverage and analyze areas for improvement.

Factual consistency is an essential quality of text summarization models in practical settings. Existing work in evaluating this dimension can be broadly categorized into two lines of research, entailment-based and question answering (QA)-based metrics, and different experimental setups often lead to contrasting conclusions as to which paradigm performs the best. In this work, we conduct an extensive comparison of entailment and QA-based metrics, demonstrating that carefully choosing the components of a QA-based metric, especially question generation and answerability classification, is critical to performance. Building on those insights, we propose an optimized metric, which we call QAFactEval, that leads to a 14% average improvement over previous QA-based metrics on the SummaC factual consistency benchmark, and also outperforms the best-performing entailment-based metric. Moreover, we find that QA-based and entailment-based metrics can offer complementary signals and be combined into a single metric for a further performance boost.

Natural language processing researchers have identified limitations of evaluation methodology for generation tasks, with new questions raised about the validity of automatic metrics and of crowdworker judgments. Meanwhile, efforts to improve generation models tend to depend on simple n-gram overlap metrics (e.g., BLEU, ROUGE). We argue that new advances on models and metrics should each more directly benefit and inform the other. We therefore propose a generalization of leaderboards, bidimensional leaderboards (Billboards), that simultaneously tracks progress in language generation models and metrics for their evaluation. Unlike conventional unidimensional leaderboards that sort submitted systems by predetermined metrics, a Billboard accepts both generators and evaluation metrics as competing entries. A Billboard automatically creates an ensemble metric that selects and linearly combines a few metrics based on a global analysis across generators. Further, metrics are ranked based on their correlation with human judgments. We release four Billboards for machine translation, summarization, and image captioning. We demonstrate that a linear ensemble of a few diverse metrics sometimes substantially outperforms existing metrics in isolation. Our mixed-effects model analysis shows that most automatic metrics, especially the reference-based ones, overrate machine over human generation, demonstrating the importance of updating metrics as generation models become stronger (and perhaps more similar to humans) in the future.

Current pre-trained models applied for summarization are prone to factual inconsistencies that misrepresent the source text. Evaluating the factual consistency of summaries is thus necessary to develop better models. However, the human evaluation setup for evaluating factual consistency has not been standardized. To determine the factors that affect the reliability of the human evaluation, we crowdsource evaluations for factual consistency across state-of-the-art models on two news summarization datasets using the rating-based Likert Scale and ranking-based Best-Worst Scaling. Our analysis reveals that the ranking-based Best-Worst Scaling offers a more reliable measure of summary quality across datasets and that the reliability of Likert ratings highly depends on the target dataset and the evaluation design. To improve crowdsourcing reliability, we extend the scale of the Likert rating and present a scoring algorithm for Best-Worst Scaling that we call value learning. Our crowdsourcing guidelines will be publicly available to facilitate future work on factual consistency in summarization.

Query-focused summarization (QFS) aims to produce summaries that answer particular questions of interest, enabling greater user control and personalization. While recently released datasets, such as QMSum or AQuaMuSe, facilitate research efforts in QFS, the field lacks a comprehensive study of the broad space of applicable modeling methods. In this paper we conduct a systematic exploration of neural approaches to QFS, considering two general classes of methods: two-stage extractive-abstractive solutions and end-to-end models. Within those categories, we investigate existing models and explore strategies for transfer learning. We also present two modeling extensions that achieve state-of-the-art performance on the QMSum dataset, up to a margin of 3.38 ROUGE-1, 3.72 ROUGE2, and 3.28 ROUGE-L when combined with transfer learning strategies. Results from human evaluation suggest that the best models produce more comprehensive and factually consistent summaries compared to a baseline model. Code and checkpoints are made publicly available: https://github.com/salesforce/query-focused-sum.

This paper introduces the shared task of summrizing documents in several creative domains, namely literary texts, movie scripts, and television scripts. Summarizing these creative documents requires making complex literary interpretations, as well as understanding non-trivial temporal dependencies in texts containing varied styles of plot development and narrative structure. This poses unique challenges and is yet underexplored for text summarization systems. In this shared task, we introduce four sub-tasks and their corresponding datasets, focusing on summarizing books, movie scripts, primetime television scripts, and daytime soap opera scripts. We detail the process of curating these datasets for the task, as well as the metrics used for the evaluation of the submissions. As part of the CREATIVESUMM workshop at COLING 2022, the shared task attracted 18 submissions in total. We discuss the submissions and the baselines for each sub-task in this paper, along with directions for facilitating future work.

While online conversations can cover a vast amount of information in many different formats, abstractive text summarization has primarily focused on modeling solely news articles. This research gap is due, in part, to the lack of standardized datasets for summarizing online discussions. To address this gap, we design annotation protocols motivated by an issues–viewpoints–assertions framework to crowdsource four new datasets on diverse online conversation forms of news comments, discussion forums, community question answering forums, and email threads. We benchmark state-of-the-art models on our datasets and analyze characteristics associated with the data. To create a comprehensive benchmark, we also evaluate these models on widely-used conversation summarization datasets to establish strong baselines in this domain. Furthermore, we incorporate argument mining through graph construction to directly model the issues, viewpoints, and assertions present in a conversation and filter noisy input, showing comparable or improved results according to automatic and human evaluations.

The scarcity of comprehensive up-to-date studies on evaluation metrics for text summarization and the lack of consensus regarding evaluation protocols continue to inhibit progress. We address the existing shortcomings of summarization evaluation methods along five dimensions: 1) we re-evaluate 14 automatic evaluation metrics in a comprehensive and consistent fashion using neural summarization model outputs along with expert and crowd-sourced human annotations; 2) we consistently benchmark 23 recent summarization models using the aforementioned automatic evaluation metrics; 3) we assemble the largest collection of summaries generated by models trained on the CNN/DailyMail news dataset and share it in a unified format; 4) we implement and share a toolkit that provides an extensible and unified API for evaluating summarization models across a broad range of automatic metrics; and 5) we assemble and share the largest and most diverse, in terms of model types, collection of human judgments of model-generated summaries on the CNN/Daily Mail dataset annotated by both expert judges and crowd-source workers. We hope that this work will help promote a more complete evaluation protocol for text summarization as well as advance research in developing evaluation metrics that better correlate with human judgments.

Models pretrained with self-supervised objectives on large text corpora achieve state-of-the-art performance on English text summarization tasks. However, these models are typically fine-tuned on hundreds of thousands of data points, an infeasible requirement when applying summarization to new, niche domains. In this work, we introduce a novel and generalizable method, called WikiTransfer, for fine-tuning pretrained models for summarization in an unsupervised, dataset-specific manner. WikiTransfer fine-tunes pretrained models on pseudo-summaries, produced from generic Wikipedia data, which contain characteristics of the target dataset, such as the length and level of abstraction of the desired summaries. WikiTransfer models achieve state-of-the-art, zero-shot abstractive summarization performance on the CNN-DailyMail dataset and demonstrate the effectiveness of our approach on three additional diverse datasets. These models are more robust to noisy data and also achieve better or comparable few-shot performance using 10 and 100 training examples when compared to few-shot transfer from other summarization datasets. To further boost performance, we employ data augmentation via round-trip translation as well as introduce a regularization term for improved few-shot transfer. To understand the role of dataset aspects in transfer performance and the quality of the resulting output summaries, we further study the effect of the components of our unsupervised fine-tuning data and analyze few-shot performance using both automatic and human evaluation.

The task of concept prerequisite chain learning is to automatically determine the existence of prerequisite relationships among concept pairs. In this paper, we frame learning prerequisite relationships among concepts as an unsupervised task with no access to labeled concept pairs during training. We propose a model called the Relational-Variational Graph AutoEncoder (R-VGAE) to predict concept relations within a graph consisting of concept and resource nodes. Results show that our unsupervised approach outperforms graph-based semi-supervised methods and other baseline methods by up to 9.77% and 10.47% in terms of prerequisite relation prediction accuracy and F1 score. Our method is notably the first graph-based model that attempts to make use of deep learning representations for the task of unsupervised prerequisite learning. We also expand an existing corpus which totals 1,717 English Natural Language Processing (NLP)-related lecture slide files and manual concept pair annotations over 322 topics.

Question Answering (QA) is in increasing demand as the amount of information available online and the desire for quick access to this content grows. A common approach to QA has been to fine-tune a pretrained language model on a task-specific labeled dataset. This paradigm, however, relies on scarce, and costly to obtain, large-scale human-labeled data. We propose an unsupervised approach to training QA models with generated pseudo-training data. We show that generating questions for QA training by applying a simple template on a related, retrieved sentence rather than the original context sentence improves downstream QA performance by allowing the model to learn more complex context-question relationships. Training a QA model on this data gives a relative improvement over a previous unsupervised model in F1 score on the SQuAD dataset by about 14%, and 20% when the answer is a named entity, achieving state-of-the-art performance on SQuAD for unsupervised QA.

Automatic generation of summaries from multiple news articles is a valuable tool as the number of online publications grows rapidly. Single document summarization (SDS) systems have benefited from advances in neural encoder-decoder model thanks to the availability of large datasets. However, multi-document summarization (MDS) of news articles has been limited to datasets of a couple of hundred examples. In this paper, we introduce Multi-News, the first large-scale MDS news dataset. Additionally, we propose an end-to-end model which incorporates a traditional extractive summarization model with a standard SDS model and achieves competitive results on MDS datasets. We benchmark several methods on Multi-News and hope that this work will promote advances in summarization in the multi-document setting.

In this paper, we propose to boost low-resource cross-lingual document retrieval performance with deep bilingual query-document representations. We match queries and documents in both source and target languages with four components, each of which is implemented as a term interaction-based deep neural network with cross-lingual word embeddings as input. By including query likelihood scores as extra features, our model effectively learns to rerank the retrieved documents by using a small number of relevance labels for low-resource language pairs. Due to the shared cross-lingual word embedding space, the model can also be directly applied to another language pair without any training label. Experimental results on the Material dataset show that our model outperforms the competitive translation-based baselines on English-Swahili, English-Tagalog, and English-Somali cross-lingual information retrieval tasks.

We present CoSQL, a corpus for building cross-domain, general-purpose database (DB) querying dialogue systems. It consists of 30k+ turns plus 10k+ annotated SQL queries, obtained from a Wizard-of-Oz (WOZ) collection of 3k dialogues querying 200 complex DBs spanning 138 domains. Each dialogue simulates a real-world DB query scenario with a crowd worker as a user exploring the DB and a SQL expert retrieving answers with SQL, clarifying ambiguous questions, or otherwise informing of unanswerable questions. When user questions are answerable by SQL, the expert describes the SQL and execution results to the user, hence maintaining a natural interaction flow. CoSQL introduces new challenges compared to existing task-oriented dialogue datasets: (1) the dialogue states are grounded in SQL, a domain-independent executable representation, instead of domain-specific slot value pairs, and (2) because testing is done on unseen databases, success requires generalizing to new domains. CoSQL includes three tasks: SQL-grounded dialogue state tracking, response generation from query results, and user dialogue act prediction. We evaluate a set of strong baselines for each task and show that CoSQL presents significant challenges for future research. The dataset, baselines, and leaderboard will be released at https://yale-lily.github.io/cosql.

The field of Natural Language Processing (NLP) is growing rapidly, with new research published daily along with an abundance of tutorials, codebases and other online resources. In order to learn this dynamic field or stay up-to-date on the latest research, students as well as educators and researchers must constantly sift through multiple sources to find valuable, relevant information. To address this situation, we introduce TutorialBank, a new, publicly available dataset which aims to facilitate NLP education and research. We have manually collected and categorized over 5,600 resources on NLP as well as the related fields of Artificial Intelligence (AI), Machine Learning (ML) and Information Retrieval (IR). Our dataset is notably the largest manually-picked corpus of resources intended for NLP education which does not include only academic papers. Additionally, we have created both a search engine and a command-line tool for the resources and have annotated the corpus to include lists of research topics, relevant resources for each topic, prerequisite relations among topics, relevant sub-parts of individual resources, among other annotations. We are releasing the dataset and present several avenues for further research.

Computational models for sarcasm detection have often relied on the content of utterances in isolation. However, the speaker’s sarcastic intent is not always apparent without additional context. Focusing on social media discussions, we investigate three issues: (1) does modeling conversation context help in sarcasm detection? (2) can we identify what part of conversation context triggered the sarcastic reply? and (3) given a sarcastic post that contains multiple sentences, can we identify the specific sentence that is sarcastic? To address the first issue, we investigate several types of Long Short-Term Memory (LSTM) networks that can model both the conversation context and the current turn. We show that LSTM networks with sentence-level attention on context and current turn, as well as the conditional LSTM network, outperform the LSTM model that reads only the current turn. As conversation context, we consider the prior turn, the succeeding turn, or both. Our computational models are tested on two types of social media platforms: Twitter and discussion forums. We discuss several differences between these data sets, ranging from their size to the nature of the gold-label annotations. To address the latter two issues, we present a qualitative analysis of the attention weights produced by the LSTM models (with attention) and discuss the results compared with human performance on the two tasks.

Computational models for sarcasm detection have often relied on the content of utterances in isolation. However, speaker’s sarcastic intent is not always obvious without additional context. Focusing on social media discussions, we investigate two issues: (1) does modeling of conversation context help in sarcasm detection and (2) can we understand what part of conversation context triggered the sarcastic reply. To address the first issue, we investigate several types of Long Short-Term Memory (LSTM) networks that can model both the conversation context and the sarcastic response. We show that the conditional LSTM network (Rocktäschel et al. 2015) and LSTM networks with sentence level attention on context and response outperform the LSTM model that reads only the response. To address the second issue, we present a qualitative analysis of attention weights produced by the LSTM models with attention and discuss the results compared with human performance on the task.