While long-context large language models (LLMs) can technically summarize book-length documents (>100K tokens), the length and complexity of the documents have so far prohibited evaluations of input-dependent aspects like faithfulness. In this paper, we conduct the first large-scale human evaluation of faithfulness and content selection on LLM-generated summaries of fictional books. Our study mitigates the issue of data contamination by focusing on summaries of books published in 2023 or 2024, and we hire annotators who have fully read each book prior to the annotation task to minimize cost and cognitive burden. We collect FABLES, a dataset of annotations on 3,158 claims made in LLM-generated summaries of 26 books, at a cost of $5.2K USD, which allows us to rank LLM summarizers based on faithfulness: Claude-3-Opus significantly outperforms all closed-source LLMs, while the open-source Mixtral is on par with GPT-3.5-Turbo. An analysis of the annotations reveals that most unfaithful claims relate to events and character states, and they generally require indirect reasoning over the narrative to invalidate. While LLM-based auto-raters have proven reliable for factuality and coherence in other settings, we implement several LLM raters of faithfulness and find that none correlates strongly with human annotations, especially with regard to detecting unfaithful claims. Our experiments suggest that detecting unfaithful claims is an important future direction not only for summarization evaluation but also as a testbed for long-context understanding. Finally, we move beyond faithfulness by exploring content selection errors in book-length summarization: we develop a typology of omission errors related to crucial narrative elements and also identify a systematic over-emphasis on events occurring towards the end of the book.
As research in large language models (LLMs) continues to accelerate, LLM-based evaluation has emerged as a scalable and cost-effective alternative to human evaluations for comparing the ever increasing list of models. This paper investigates the efficacy of these "LLM evaluators", particularly in using them to assess instruction following, a metric that gauges how closely generated text adheres to the given instruction. We introduce a challenging meta-evaluation benchmark, LLMBar, designed to test the ability of an LLM evaluator in discerning instruction-following outputs. The authors manually curated 419 pairs of outputs, one adhering to instructions while the other diverging, yet may possess deceptive qualities that mislead an LLM evaluator, e.g., a more engaging tone. Contrary to existing meta-evaluation, we discover that different evaluators (i.e., combinations of LLMs and prompts) exhibit distinct performance on LLMBar and even the highest-scoring ones have substantial room for improvement. We also present a novel suite of prompting strategies that further close the gap between LLM and human evaluators. With LLMBar, we hope to offer more insight into LLM evaluators and foster future research in developing better instruction-following models.
Summarizing book-length documents (>100K tokens) that exceed the context window size of large language models (LLMs) requires first breaking the input document into smaller chunks and then prompting an LLM to merge, update, and compress chunk-level summaries. Despite the complexity and importance of this task, it has yet to be meaningfully studied due to the challenges of evaluation: existing book-length summarization datasets (e.g., BookSum) are in the pretraining data of most public LLMs, and existing evaluation methods struggle to capture errors made by modern LLM summarizers. In this paper, we present the first study of the coherence of LLM-based book-length summarizers implemented via two prompting workflows: (1) hierarchically merging chunk-level summaries, and (2) incrementally updating a running summary. We obtain 1193 fine-grained human annotations on GPT-4 generated summaries of 100 recently-published books and identify eight common types of coherence errors made by LLMs. Because human evaluation is expensive and time-consuming, we develop an automatic metric, BooookScore, that measures the proportion of sentences in a summary that do not contain any of the identified error types. BooookScore has high agreement with human annotations and allows us to systematically evaluate the impact of many other critical parameters (e.g., chunk size, base LLM) while saving $15K and 500 hours in human evaluation costs. We find that closed-source LLMs such as GPT-4 and Claude 2 produce summaries with higher BooookScore than the oft-repetitive ones generated by LLaMA 2. Incremental updating yields lower BooookScore but higher level of detail than hierarchical merging, a trade-off sometimes preferred by human annotators. We release code and annotations after blind review to spur more principled research on book-length summarization.
Great successes have been reported using Reinforcement Learning from Human Feedback (RLHF) to align large language models. Open-source preference datasets and reward models have enabled wider experimentation beyond generic chat settings, particularly to make systems more "helpful" for tasks like web question answering, summarization, and multi-turn dialogue. When optimizing for helpfulness, RLHF has been consistently observed to drive models to produce longer outputs. This paper demonstrates that optimizing for response length is a significant factor behind RLHF's reported improvements in these settings. First, we study the relationship between reward and length for reward models trained on three open-source preference datasets for helpfulness. Here, length correlates strongly with reward, and improvements in reward score are driven in large part by shifting the distribution over output lengths. We then explore interventions during both RL and reward model learning to see if we can achieve the same downstream improvements as RLHF without increasing length. While our interventions mitigate length increases, they aren't uniformly effective across settings. Furthermore, we find that even running RLHF with a reward based solely on length can reproduce most of the downstream improvements over the initial policy model, showing that reward models in these settings have a long way to go.
Models for textual entailment have increasingly been applied to settings like fact-checking, presupposition verification in question answering, and validating that generation models' outputs are faithful to a source. However, such applications are quite far from the settings that existing datasets are constructed in. We propose WiCE, a new textual entailment dataset centered around verifying claims in text, built on real-world claims and evidence in Wikipedia with fine-grained annotations. We collect sentences in Wikipedia that cite one or more webpages and annotate whether the content on those pages entails those sentences. Negative examples arise naturally, from slight misinterpretation of text to minor aspects of the sentence that are not attested in the evidence. Our annotations are over sub-sentence units of the hypothesis, decomposed automatically by GPT-3, each of which is labeled with a subset of evidence sentences from the source document. We show that real claims in our dataset involve challenging verification problems, and we benchmark existing approaches on this dataset. In addition, we show that reducing the complexity of claims by decomposing them by GPT-3 can improve entailment models' performance on various domains.
Despite recent progress in abstractive summarization, models often generate summaries with factual errors. Numerous approaches to detect these errors have been proposed, the most popular of which are question answering (QA)-based factuality metrics. These have been shown to work well at predicting summary-level factuality and have potential to localize errors within summaries, but this latter capability has not been systematically evaluated in past research. In this paper, we conduct the first such analysis and find that, contrary to our expectations, QA-based frameworks fail to correctly identify error spans in generated summaries and are outperformed by trivial exact match baselines. Our analysis reveals a major reason for such poor localization: questions generated by the QG module often inherit errors from non-factual summaries which are then propagated further into downstream modules. Moreover, even human-in-the-loop question generation cannot easily offset these problems. Our experiments conclusively show that there exist fundamental issues with localization using the QA framework which cannot be fixed solely by stronger QA and QG models.
The recent success of zero- and few-shot prompting with models like GPT-3 has led to a paradigm shift in NLP research. In this paper, we study its impact on text summarization, focusing on the classic benchmark domain of news summarization. First, we investigate how zero-shot GPT-3 compares against fine-tuned models trained on large summarization datasets. We show that not only do humans overwhelmingly prefer GPT-3 summaries, but these also do not suffer from common dataset-specific issues such as poor factuality. Next, we study what this means for evaluation, particularly the role of gold standard test sets. Our experiments show that both reference-based and reference-free automatic metrics, e.g. recently proposed QA- or entailment-based factuality approaches, cannot reliably evaluate zero-shot summaries. Finally, we discuss future research challenges beyond generic summarization, specifically, keyword- and aspect-based summarization, showing how dominant fine-tuning approaches compare to zero-shot prompting. To support further research, we release: (a) a corpus of 10K generated summaries from fine-tuned and zero-shot models across 4 standard summarization benchmarks, (b) 1K human preference judgments and rationales comparing different systems for generic- and keyword-based summarization.
The propensity of abstractive summarization systems to make factual errors has been the subject of significant study, including work on models to detect factual errors and annotation of errors in current systems' outputs. However, the ever-evolving nature of summarization systems, error detectors, and annotated benchmarks make factuality evaluation a moving target; it is hard to get a clear picture of how techniques compare. In this work, we collect labeled factuality errors from across nine datasets of annotated summary outputs and stratify them in a new way, focusing on what kind of base summarization model was used. To support finer-grained analysis, we unify the labeled error types into a single taxonomy and project each of the datasets' errors into this shared labeled space. We then contrast five state-of-the-art error detection methods on this benchmark. Our findings show that benchmarks built on modern summary outputs (those from pre-trained models) show significantly different results than benchmarks using pre-Transformer models. Furthermore, no one factuality technique is superior in all settings or for all error types, suggesting that system developers should take care to choose the right system for their task at hand.
Progress in summarizing long texts is inhibited by the lack of appropriate evaluation frameworks. When a long summary must be produced to appropriately cover the facets of that text, that summary needs to present a coherent narrative to be understandable by a reader, but current automatic and human evaluation methods fail to identify gaps in coherence. In this work, we introduce SNaC, a narrative coherence evaluation framework rooted in fine-grained annotations for long summaries. We develop a taxonomy of coherence errors in generated narrative summaries and collect span-level annotations for 6.6k sentences across 150 book and movie screenplay summaries. Our work provides the first characterization of coherence errors generated by state-of-the-art summarization models and a protocol for eliciting coherence judgments from crowd annotators. Furthermore, we show that the collected annotations allow us to train a strong classifier for automatically localizing coherence errors in generated summaries as well as benchmarking past work in coherence modeling. Finally, our SNaC framework can support future work in long document summarization and coherence evaluation, including improved summarization modeling and post-hoc summary correction.
Existing abstractive summarization models lack explicit control mechanisms that would allow users to influence the stylistic features of the model outputs. This results in generating generic summaries that do not cater to the users needs or preferences. To address this issue we introduce HydraSum, a new summarization architecture that extends the single decoder framework of current models, e.g. BART, to a mixture-of-experts version consisting of multiple decoders. Our proposed model encourages each expert, i.e. decoder, to learn and generate stylistically-distinct summaries along dimensions such as abstractiveness, length, specificity, and others. At each time step, HydraSum employs a gating mechanism that decides the contribution of each individual decoder to the next token's output probability distribution. Through experiments on three summarization datasets (CNN, Newsroom, XSum), we demonstrate that this gating mechanism automatically learns to assign contrasting summary styles to different HydraSum decoders under the standard training objective without the need for additional supervision. We further show that a guided version of the training process can explicitly govern which summary style is partitioned between decoders, e.g. high abstractiveness vs. low abstractiveness or high specificity vs. low specificity, and also increase the stylistic-difference between individual decoders. Finally, our experiments demonstrate that our decoder framework is highly flexible: during inference, we can sample from individual decoders or mixtures of different subsets of the decoders to yield a diverse set of summaries and enforce single- and multi-style control over summary generation.