Mohit Iyyer

Kalpesh Krishna, Yixiao Song, Marzena Karpinska et al. · deepmind

The rise in malicious usage of large language models, such as fake content creation and academic plagiarism, has motivated the development of approaches that identify AI-generated text, including those based on watermarking or outlier detection. However, the robustness of these detection algorithms to paraphrases of AI-generated text remains unclear. To stress test these detectors, we build a 11B parameter paraphrase generation model (DIPPER) that can paraphrase paragraphs, condition on surrounding context, and control lexical diversity and content reordering. Using DIPPER to paraphrase text generated by three large language models (including GPT3.5-davinci-003) successfully evades several detectors, including watermarking, GPTZero, DetectGPT, and OpenAI's text classifier. For example, DIPPER drops detection accuracy of DetectGPT from 70.3% to 4.6% (at a constant false positive rate of 1%), without appreciably modifying the input semantics. To increase the robustness of AI-generated text detection to paraphrase attacks, we introduce a simple defense that relies on retrieving semantically-similar generations and must be maintained by a language model API provider. Given a candidate text, our algorithm searches a database of sequences previously generated by the API, looking for sequences that match the candidate text within a certain threshold. We empirically verify our defense using a database of 15M generations from a fine-tuned T5-XXL model and find that it can detect 80% to 97% of paraphrased generations across different settings while only classifying 1% of human-written sequences as AI-generated. We open-source our models, code and data.

Katherine Thai, Marzena Karpinska, Kalpesh Krishna et al. · deepmind

Literary translation is a culturally significant task, but it is bottlenecked by the small number of qualified literary translators relative to the many untranslated works published around the world. Machine translation (MT) holds potential to complement the work of human translators by improving both training procedures and their overall efficiency. Literary translation is less constrained than more traditional MT settings since translators must balance meaning equivalence, readability, and critical interpretability in the target language. This property, along with the complex discourse-level context present in literary texts, also makes literary MT more challenging to computationally model and evaluate. To explore this task, we collect a dataset (Par3) of non-English language novels in the public domain, each aligned at the paragraph level to both human and automatic English translations. Using Par3, we discover that expert literary translators prefer reference human translations over machine-translated paragraphs at a rate of 84%, while state-of-the-art automatic MT metrics do not correlate with those preferences. The experts note that MT outputs contain not only mistranslations, but also discourse-disrupting errors and stylistic inconsistencies. To address these problems, we train a post-editing model whose output is preferred over normal MT output at a rate of 69% by experts. We publicly release Par3 at https://github.com/katherinethai/par3/ to spur future research into literary MT.

Yixiao Song, Kalpesh Krishna, Rajesh Bhatt et al. · deepmind

Grammatical error correction tools are effective at correcting grammatical errors in users' input sentences but do not provide users with \textit{natural language} explanations about their errors. Such explanations are essential for helping users learn the language by gaining a deeper understanding of its grammatical rules (DeKeyser, 2003; Ellis et al., 2006). To address this gap, we propose the task of grammar error explanation, where a system needs to provide one-sentence explanations for each grammatical error in a pair of erroneous and corrected sentences. We analyze the capability of GPT-4 in grammar error explanation, and find that it only produces explanations for 60.2% of the errors using one-shot prompting. To improve upon this performance, we develop a two-step pipeline that leverages fine-tuned and prompted large language models to perform structured atomic token edit extraction, followed by prompting GPT-4 to generate explanations. We evaluate our pipeline on German and Chinese grammar error correction data sampled from language learners with a wide range of proficiency levels. Human evaluation reveals that our pipeline produces 93.9% and 98.0% correct explanations for German and Chinese data, respectively. To encourage further research in this area, we will open-source our data and code.

Tu Vu, Mohit Iyyer, Xuezhi Wang et al.

Most large language models (LLMs) are trained once and never updated; thus, they lack the ability to dynamically adapt to our ever-changing world. In this work, we perform a detailed study of the factuality of LLM-generated text in the context of answering questions that test current world knowledge. Specifically, we introduce FreshQA, a novel dynamic QA benchmark encompassing a diverse range of question and answer types, including questions that require fast-changing world knowledge as well as questions with false premises that need to be debunked. We benchmark a diverse array of both closed and open-source LLMs under a two-mode evaluation procedure that allows us to measure both correctness and hallucination. Through human evaluations involving more than 50K judgments, we shed light on limitations of these models and demonstrate significant room for improvement: for instance, all models (regardless of model size) struggle on questions that involve fast-changing knowledge and false premises. Motivated by these results, we present FreshPrompt, a simple few-shot prompting method that substantially boosts the performance of an LLM on FreshQA by incorporating relevant and up-to-date information retrieved from a search engine into the prompt. Our experiments show that FreshPrompt outperforms both competing search engine-augmented prompting methods such as Self-Ask (Press et al., 2022) as well as commercial systems such as Perplexity.AI. Further analysis of FreshPrompt reveals that both the number of retrieved evidences and their order play a key role in influencing the correctness of LLM-generated answers. Additionally, instructing the LLM to generate concise and direct answers helps reduce hallucination compared to encouraging more verbose answers. To facilitate future work, we release FreshQA at github.com/freshllms/freshqa and commit to updating it at regular intervals.

Yapei Chang, Kyle Lo, Tanya Goyal et al.

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 USD 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 those generated by open-source models. While LLaMA 2 falls behind other models, Mixtral achieves performance on par with GPT-3.5-Turbo. Incremental updating yields lower BooookScore but higher level of detail than hierarchical merging, a trade-off sometimes preferred by annotators.

Kalpesh Krishna, Erin Bransom, Bailey Kuehl et al. · allen-ai, deepmind

While human evaluation remains best practice for accurately judging the faithfulness of automatically-generated summaries, few solutions exist to address the increased difficulty and workload when evaluating long-form summaries. Through a survey of 162 papers on long-form summarization, we first shed light on current human evaluation practices surrounding long-form summaries. We find that 73% of these papers do not perform any human evaluation on model-generated summaries, while other works face new difficulties that manifest when dealing with long documents (e.g., low inter-annotator agreement). Motivated by our survey, we present LongEval, a set of guidelines for human evaluation of faithfulness in long-form summaries that addresses the following challenges: (1) How can we achieve high inter-annotator agreement on faithfulness scores? (2) How can we minimize annotator workload while maintaining accurate faithfulness scores? and (3) Do humans benefit from automated alignment between summary and source snippets? We deploy LongEval in annotation studies on two long-form summarization datasets in different domains (SQuALITY and PubMed), and we find that switching to a finer granularity of judgment (e.g., clause-level) reduces inter-annotator variance in faithfulness scores (e.g., std-dev from 18.5 to 6.8). We also show that scores from a partial annotation of fine-grained units highly correlates with scores from a full annotation workload (0.89 Kendall's tau using 50% judgments). We release our human judgments, annotation templates, and our software as a Python library for future research.

Kalpesh Krishna, Yapei Chang, John Wieting et al. · deepmind

Given an input sequence (or prefix), modern language models often assign high probabilities to output sequences that are repetitive, incoherent, or irrelevant to the prefix; as such, model-generated text also contains such artifacts. To address these issues we present RankGen, a 1.2B parameter encoder model for English that scores model generations given a prefix. RankGen can be flexibly incorporated as a scoring function in beam search and used to decode from any pretrained language model. We train RankGen using large-scale contrastive learning to map a prefix close to the ground-truth sequence that follows it and far away from two types of negatives: (1) random sequences from the same document as the prefix, and (2) sequences generated from a large language model conditioned on the prefix. Experiments across four different language models (345M-11B parameters) and two domains show that RankGen significantly outperforms decoding algorithms like nucleus, top-k, and typical sampling, as well as contrastive decoding and search, on both automatic metrics (85.0 vs 77.3 MAUVE over nucleus) as well as human evaluations with English writers (74.5% human preference over nucleus sampling). Analysis reveals that RankGen outputs are more relevant to the prefix and improve continuity and coherence compared to baselines. We release our model checkpoints, code, and human preference data with explanations to facilitate future research.

Tu Vu, Aditya Barua, Brian Lester et al. · berkeley, deepmind

In this paper, we explore the challenging problem of performing a generative task in a target language when labeled data is only available in English, using summarization as a case study. We assume a strict setting with no access to parallel data or machine translation and find that common transfer learning approaches struggle in this setting, as a generative multilingual model fine-tuned purely on English catastrophically forgets how to generate non-English. Given the recent rise of parameter-efficient adaptation techniques, we conduct the first investigation into how one such method, prompt tuning (Lester et al., 2021), can overcome catastrophic forgetting to enable zero-shot cross-lingual generation. Our experiments show that parameter-efficient prompt tuning provides gains over standard fine-tuning when transferring between less-related languages, e.g., from English to Thai. However, a significant gap still remains between these methods and fully-supervised baselines. To improve cross-lingual transfer further, we explore several approaches, including: (1) mixing in unlabeled multilingual data, and (2) explicitly factoring prompts into recombinable language and task components. Our approaches can provide further quality gains, suggesting that robust zero-shot cross-lingual generation is within reach.

Ali Naseh, Kalpesh Krishna, Mohit Iyyer et al. · deepmind

A key component of generating text from modern language models (LM) is the selection and tuning of decoding algorithms. These algorithms determine how to generate text from the internal probability distribution generated by the LM. The process of choosing a decoding algorithm and tuning its hyperparameters takes significant time, manual effort, and computation, and it also requires extensive human evaluation. Therefore, the identity and hyperparameters of such decoding algorithms are considered to be extremely valuable to their owners. In this work, we show, for the first time, that an adversary with typical API access to an LM can steal the type and hyperparameters of its decoding algorithms at very low monetary costs. Our attack is effective against popular LMs used in text generation APIs, including GPT-2, GPT-3 and GPT-Neo. We demonstrate the feasibility of stealing such information with only a few dollars, e.g., $\$0.8$, $\$1$, $\$4$, and $\$40$ for the four versions of GPT-3.

Simeng Sun, Yang Liu, Dan Iter et al. · stanford

Fine-tuning large language models is becoming ever more impractical due to their rapidly-growing scale. This motivates the use of parameter-efficient adaptation methods such as prompt tuning (PT), which adds a small number of tunable embeddings to an otherwise frozen model, and in-context learning (ICL), in which demonstrations of the task are provided to the model in natural language without any additional training. Recently, Singhal et al. (2022) propose ``instruction prompt tuning'' (IPT), which combines PT with ICL by concatenating a natural language demonstration with learned prompt embeddings. While all of these methods have proven effective on different tasks, how they interact with each other remains unexplored. In this paper, we empirically study when and how in-context examples improve prompt tuning by measuring the effectiveness of ICL, PT, and IPT on five text generation tasks with multiple base language models. We observe that (1) IPT does \emph{not} always outperform PT, and in fact requires the in-context demonstration to be semantically similar to the test input to yield improvements; (2) PT is unstable and exhibits high variance, but combining PT and ICL (into IPT) consistently reduces variance across all five tasks; and (3) prompts learned for a specific source task via PT exhibit positive transfer when paired with in-context examples of a different target task. Our results offer actionable insights on choosing a suitable parameter-efficient adaptation method for a given task.

Katherine Thai, Yapei Chang, Kalpesh Krishna et al. · deepmind

Humanities scholars commonly provide evidence for claims that they make about a work of literature (e.g., a novel) in the form of quotations from the work. We collect a large-scale dataset (RELiC) of 78K literary quotations and surrounding critical analysis and use it to formulate the novel task of literary evidence retrieval, in which models are given an excerpt of literary analysis surrounding a masked quotation and asked to retrieve the quoted passage from the set of all passages in the work. Solving this retrieval task requires a deep understanding of complex literary and linguistic phenomena, which proves challenging to methods that overwhelmingly rely on lexical and semantic similarity matching. We implement a RoBERTa-based dense passage retriever for this task that outperforms existing pretrained information retrieval baselines; however, experiments and analysis by human domain experts indicate that there is substantial room for improvement over our dense retriever.

Yixiao Song, Kalpesh Krishna, Rajesh Bhatt et al. · deepmind

To understand what kinds of linguistic knowledge are encoded by pretrained Chinese language models (LMs), we introduce the benchmark of Sino LINGuistics (SLING), which consists of 38K minimal sentence pairs in Mandarin Chinese grouped into 9 high-level linguistic phenomena. Each pair demonstrates the acceptability contrast of a specific syntactic or semantic phenomenon (e.g., The keys are lost vs. The keys is lost), and an LM should assign lower perplexity to the acceptable sentence. In contrast to the CLiMP dataset (Xiang et al., 2021), which also contains Chinese minimal pairs and was created by translating the vocabulary of the English BLiMP dataset, the minimal pairs in SLING are derived primarily by applying syntactic and lexical transformations to naturally-occurring, linguist-annotated sentences from the Chinese Treebank 9.0, thus addressing severe issues in CLiMP's data generation process. We test 18 publicly available pretrained monolingual (e.g., BERT-base-zh, CPM) and multi-lingual (e.g., mT5, XLM) language models on SLING. Our experiments show that the average accuracy for LMs is far below human performance (69.7% vs. 97.1%), while BERT-base-zh achieves the highest accuracy (84.8%) of all tested LMs, even much larger ones. Additionally, we find that most LMs have a strong gender and number (singular/plural) bias, and they perform better on local phenomena than hierarchical ones.

Ankita Gupta, Marzena Karpinska, Wenlong Zhao et al. · deepmind

Large-scale, high-quality corpora are critical for advancing research in coreference resolution. However, existing datasets vary in their definition of coreferences and have been collected via complex and lengthy guidelines that are curated for linguistic experts. These concerns have sparked a growing interest among researchers to curate a unified set of guidelines suitable for annotators with various backgrounds. In this work, we develop a crowdsourcing-friendly coreference annotation methodology, ezCoref, consisting of an annotation tool and an interactive tutorial. We use ezCoref to re-annotate 240 passages from seven existing English coreference datasets (spanning fiction, news, and multiple other domains) while teaching annotators only cases that are treated similarly across these datasets. Surprisingly, we find that reasonable quality annotations were already achievable (>90% agreement between the crowd and expert annotations) even without extensive training. On carefully analyzing the remaining disagreements, we identify the presence of linguistic cases that our annotators unanimously agree upon but lack unified treatments (e.g., generic pronouns, appositives) in existing datasets. We propose the research community should revisit these phenomena when curating future unified annotation guidelines.

Shufan Wang, Fangyuan Xu, Laure Thompson et al. · princeton

Exemplification is a process by which writers explain or clarify a concept by providing an example. While common in all forms of writing, exemplification is particularly useful in the task of long-form question answering (LFQA), where a complicated answer can be made more understandable through simple examples. In this paper, we provide the first computational study of exemplification in QA, performing a fine-grained annotation of different types of examples (e.g., hypotheticals, anecdotes) in three corpora. We show that not only do state-of-the-art LFQA models struggle to generate relevant examples, but also that standard evaluation metrics such as ROUGE are insufficient to judge exemplification quality. We propose to treat exemplification as a \emph{retrieval} problem in which a partially-written answer is used to query a large set of human-written examples extracted from a corpus. Our approach allows a reliable ranking-type automatic metrics that correlates well with human evaluation. A human evaluation shows that our model's retrieved examples are more relevant than examples generated from a state-of-the-art LFQA model.

Andrew Drozdov, Shufan Wang, Razieh Rahimi et al.

Retrieval-enhanced language models (LMs), which condition their predictions on text retrieved from large external datastores, have recently shown significant perplexity improvements compared to standard LMs. One such approach, the $k$NN-LM, interpolates any existing LM's predictions with the output of a $k$-nearest neighbors model and requires no additional training. In this paper, we explore the importance of lexical and semantic matching in the context of items retrieved by $k$NN-LM. We find two trends: (1) the presence of large overlapping $n$-grams between the datastore and evaluation set plays an important factor in strong performance, even when the datastore is derived from the training data; and (2) the $k$NN-LM is most beneficial when retrieved items have high semantic similarity with the query. Based on our analysis, we define a new formulation of the $k$NN-LM that uses retrieval quality to assign the interpolation coefficient. We empirically measure the effectiveness of our approach on two English language modeling datasets, Wikitext-103 and PG-19. Our re-formulation of the $k$NN-LM is beneficial in both cases, and leads to nearly 4% improvement in perplexity on the Wikitext-103 test set.

Marzena Karpinska, Mohit Iyyer

Large language models (LLMs) are competitive with the state of the art on a wide range of sentence-level translation datasets. However, their ability to translate paragraphs and documents remains unexplored because evaluation in these settings is costly and difficult. We show through a rigorous human evaluation that asking the Gpt-3.5 (text-davinci-003) LLM to translate an entire literary paragraph (e.g., from a novel) at once results in higher-quality translations than standard sentence-by-sentence translation across 18 linguistically-diverse language pairs (e.g., translating into and out of Japanese, Polish, and English). Our evaluation, which took approximately 350 hours of effort for annotation and analysis, is conducted by hiring translators fluent in both the source and target language and asking them to provide both span-level error annotations as well as preference judgments of which system's translations are better. We observe that discourse-level LLM translators commit fewer mistranslations, grammar errors, and stylistic inconsistencies than sentence-level approaches. With that said, critical errors still abound, including occasional content omissions, and a human translator's intervention remains necessary to ensure that the author's voice remains intact. We publicly release our dataset and error annotations to spur future research on evaluation of document-level literary translation.

Marzena Karpinska, Nishant Raj, Katherine Thai et al.

While machine translation evaluation metrics based on string overlap (e.g., BLEU) have their limitations, their computations are transparent: the BLEU score assigned to a particular candidate translation can be traced back to the presence or absence of certain words. The operations of newer learned metrics (e.g., BLEURT, COMET), which leverage pretrained language models to achieve higher correlations with human quality judgments than BLEU, are opaque in comparison. In this paper, we shed light on the behavior of these learned metrics by creating DEMETR, a diagnostic dataset with 31K English examples (translated from 10 source languages) for evaluating the sensitivity of MT evaluation metrics to 35 different linguistic perturbations spanning semantic, syntactic, and morphological error categories. All perturbations were carefully designed to form minimal pairs with the actual translation (i.e., differ in only one aspect). We find that learned metrics perform substantially better than string-based metrics on DEMETR. Additionally, learned metrics differ in their sensitivity to various phenomena (e.g., BERTScore is sensitive to untranslated words but relatively insensitive to gender manipulation, while COMET is much more sensitive to word repetition than to aspectual changes). We publicly release DEMETR to spur more informed future development of machine translation evaluation metrics

Simeng Sun, Dhawal Gupta, Mohit Iyyer

During the last stage of RLHF, a large language model is aligned to human intents via PPO training, a process that generally requires large-scale computational resources. In this technical report, we empirically investigate an efficient implementation of RLHF using low-rank adaptation (LoRA), which allows us to align the LLaMA 7B checkpoint on the Alpaca dataset using only two A100 GPUs instead of the eight required for full model fine-tuning. Despite tuning only 0.2% of LLaMA 7B's parameters, our implementation achieves better performance than the publicly-released AlpacaFarm checkpoint with full model fine-tuning. Next, we analyze several configurations of our LoRA-based PPO implementation, varying the form of the KL regularization term in the training objective. We find that (1) removing this penalty term does not harm performance on the AlpacaFarm evaluation set under our LoRA setup; (2) other regularizers, such as Jensen-Shannon divergence, lead to improved performance; and (3) while PPO training negatively impacts the factuality of model-generated responses, training with LoRA largely mitigates this effect. We release our code and pretrained checkpoints to facilitate future research on more efficient RLHF.

Simeng Sun, Katherine Thai, Mohit Iyyer

While numerous architectures for long-range language models (LRLMs) have recently been proposed, a meaningful evaluation of their discourse-level language understanding capabilities has not yet followed. To this end, we introduce ChapterBreak, a challenge dataset that provides an LRLM with a long segment from a narrative that ends at a chapter boundary and asks it to distinguish the beginning of the ground-truth next chapter from a set of negative segments sampled from the same narrative. A fine-grained human annotation reveals that our dataset contains many complex types of chapter transitions (e.g., parallel narratives, cliffhanger endings) that require processing global context to comprehend. Experiments on ChapterBreak show that existing LRLMs fail to effectively leverage long-range context, substantially underperforming a segment-level model trained directly for this task. We publicly release our ChapterBreak dataset to spur more principled future research into LRLMs.

Jiachen Zhao, Wenlong Zhao, Andrew Drozdov et al.

We study semi-supervised sequence generation tasks, where the few labeled examples are too scarce to finetune a model, and meanwhile, few-shot prompted large language models (LLMs) exhibit room for improvement. In this paper, we present the discovery that a student model distilled from a few-shot prompted LLM can commonly generalize better than its teacher to unseen examples on such tasks. We find that the student is able to learn a general pattern from the high-quality pseudolabels produced by the teacher during knowledge distillation (KD), and favorably not a general pattern from the low-quality pseudolables. Leveraging this discovery, we propose a new method, Multistage Collaborative Knowledge Distillation from an LLM (MCKD), for these tasks. MCKD first few-shot prompts an LLM to produce pseudolabels for unlabeled data. Then at each stage of an iterative KD process, a new pair of students is trained on disjoint partitions of the pseudolabeled data, and produces new and improved pseudolabels for their unseen partitions. We conduct extensive experiments on four syntactic and semantic parsing datasets and show the effectiveness of MCKD for low-resource semi-supervised sequence generation. On CRAFT biomedical parsing, for example, 3-stage MCKD with 50 labeled examples outperforms an LLM teacher and vanilla KD by 7.5% and 3.7% parsing F1, respectively, and matches the performance of supervised finetuning with 500 labeled examples.

Rachneet Sachdeva, Yixiao Song, Mohit Iyyer et al.

Long-form question answering (LFQA) aims to provide thorough and in-depth answers to complex questions, enhancing comprehension. However, such detailed responses are prone to hallucinations and factual inconsistencies, challenging their faithful evaluation. This work introduces HaluQuestQA, the first hallucination dataset with localized error annotations for human-written and model-generated LFQA answers. HaluQuestQA comprises 698 QA pairs with 1.8k span-level error annotations for five different error types by expert annotators, along with preference judgments. Using our collected data, we thoroughly analyze the shortcomings of long-form answers and find that they lack comprehensiveness and provide unhelpful references. We train an automatic feedback model on this dataset that predicts error spans with incomplete information and provides associated explanations. Finally, we propose a prompt-based approach, Error-informed refinement, that uses signals from the learned feedback model to refine generated answers, which we show reduces errors and improves answer quality across multiple models. Furthermore, humans find answers generated by our approach comprehensive and highly prefer them (84%) over the baseline answers.

Yekyung Kim, Yapei Chang, Marzena Karpinska et al.

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.

Jenna Russell, Marzena Karpinska, Mohit Iyyer

In this paper, we study how well humans can detect text generated by commercial LLMs (GPT-4o, Claude, o1). We hire annotators to read 300 non-fiction English articles, label them as either human-written or AI-generated, and provide paragraph-length explanations for their decisions. Our experiments show that annotators who frequently use LLMs for writing tasks excel at detecting AI-generated text, even without any specialized training or feedback. In fact, the majority vote among five such "expert" annotators misclassifies only 1 of 300 articles, significantly outperforming most commercial and open-source detectors we evaluated even in the presence of evasion tactics like paraphrasing and humanization. Qualitative analysis of the experts' free-form explanations shows that while they rely heavily on specific lexical clues ('AI vocabulary'), they also pick up on more complex phenomena within the text (e.g., formality, originality, clarity) that are challenging to assess for automatic detectors. We release our annotated dataset and code to spur future research into both human and automated detection of AI-generated text.

Chau Minh Pham, Simeng Sun, Mohit Iyyer

Existing research on instruction following largely focuses on tasks with simple instructions and short responses. In this work, we explore multi-constraint instruction following for generating long-form text. We create Suri, a dataset with 20K human-written long-form texts paired with LLM-generated backtranslated instructions that contain multiple complex constraints. Because of prohibitive challenges associated with collecting human preference judgments on long-form texts, preference-tuning algorithms such as DPO are infeasible in our setting; thus, we propose Instructional ORPO (I-ORPO), an alignment method based on the ORPO algorithm. Instead of receiving negative feedback from dispreferred responses, I-ORPO obtains negative feedback from synthetically corrupted instructions generated by an LLM. Using Suri, we perform supervised and I-ORPO fine-tuning on Mistral-7b-Instruct-v0.2. The resulting models, Suri-SFT and Suri-I-ORPO, generate significantly longer texts (~5K tokens) than base models without significant quality deterioration. Our human evaluation shows that while both SFT and I-ORPO models satisfy most constraints, Suri-I-ORPO generations are generally preferred for their coherent and informative incorporation of the constraints. We release our code at https://github.com/chtmp223/suri.

Yapei Chang, Kalpesh Krishna, Amir Houmansadr et al.

The most effective techniques to detect LLM-generated text rely on inserting a detectable signature -- or watermark -- during the model's decoding process. Most existing watermarking methods require access to the underlying LLM's logits, which LLM API providers are loath to share due to fears of model distillation. As such, these watermarks must be implemented independently by each LLM provider. In this paper, we develop PostMark, a modular post-hoc watermarking procedure in which an input-dependent set of words (determined via a semantic embedding) is inserted into the text after the decoding process has completed. Critically, PostMark does not require logit access, which means it can be implemented by a third party. We also show that PostMark is more robust to paraphrasing attacks than existing watermarking methods: our experiments cover eight baseline algorithms, five base LLMs, and three datasets. Finally, we evaluate the impact of PostMark on text quality using both automated and human assessments, highlighting the trade-off between quality and robustness to paraphrasing. We release our code, outputs, and annotations at https://github.com/lilakk/PostMark.

Abhinav Kumar, Jaechul Roh, Ali Naseh et al.

We increase overhead for applications that rely on reasoning LLMs-we force models to spend an amplified number of reasoning tokens, i.e., "overthink", to respond to the user query while providing contextually correct answers. The adversary performs an OVERTHINK attack by injecting decoy reasoning problems into the public content that is used by the reasoning LLM (e.g., for RAG applications) during inference time. Due to the nature of our decoy problems (e.g., a Markov Decision Process), modified texts do not violate safety guardrails. We evaluated our attack across closed-(OpenAI o1, o1-mini, o3-mini) and open-(DeepSeek R1) weights reasoning models on the FreshQA and SQuAD datasets. Our results show up to 18x slowdown on FreshQA dataset and 46x slowdown on SQuAD dataset. The attack also shows high transferability across models. To protect applications, we discuss and implement defenses leveraging LLM-based and system design approaches. Finally, we discuss societal, financial, and energy impacts of OVERTHINK attack which could amplify the costs for third-party applications operating reasoning models.

Ali Naseh, Katherine Thai, Mohit Iyyer et al.

With the digital imagery landscape rapidly evolving, image stocks and AI-generated image marketplaces have become central to visual media. Traditional stock images now exist alongside innovative platforms that trade in prompts for AI-generated visuals, driven by sophisticated APIs like DALL-E 3 and Midjourney. This paper studies the possibility of employing multi-modal models with enhanced visual understanding to mimic the outputs of these platforms, introducing an original attack strategy. Our method leverages fine-tuned CLIP models, a multi-label classifier, and the descriptive capabilities of GPT-4V to create prompts that generate images similar to those available in marketplaces and from premium stock image providers, yet at a markedly lower expense. In presenting this strategy, we aim to spotlight a new class of economic and security considerations within the realm of digital imagery. Our findings, supported by both automated metrics and human assessment, reveal that comparable visual content can be produced for a fraction of the prevailing market prices ($0.23 - $0.27 per image), emphasizing the need for awareness and strategic discussions about the integrity of digital media in an increasingly AI-integrated landscape. Our work also contributes to the field by assembling a dataset consisting of approximately 19 million prompt-image pairs generated by the popular Midjourney platform, which we plan to release publicly.

Yekyung Kim, Jenna Russell, Marzena Karpinska et al.

We present ONERULER, a multilingual benchmark designed to evaluate long-context language models across 26 languages. ONERULER adapts the English-only RULER benchmark (Hsieh et al., 2024) by including seven synthetic tasks that test both retrieval and aggregation, including new variations of the "needle-in-a-haystack" task that allow for the possibility of a nonexistent needle. We create ONERULER through a two-step process, first writing English instructions for each task and then collaborating with native speakers to translate them into 25 additional languages. Experiments with both open-weight and closed LLMs reveal a widening performance gap between low- and high-resource languages as context length increases from 8K to 128K tokens. Surprisingly, English is not the top-performing language on long-context tasks (ranked 6th out of 26), with Polish emerging as the top language. Our experiments also show that many LLMs (particularly OpenAI's o3-mini-high) incorrectly predict the absence of an answer, even in high-resource languages. Finally, in cross-lingual scenarios where instructions and context appear in different languages, performance can fluctuate by up to 20% depending on the instruction language. We hope the release of ONERULER will facilitate future research into improving multilingual and cross-lingual long-context training pipelines.

Chaitanya Malaviya, Joseph Chee Chang, Dan Roth et al.

Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping benchmark rankings between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure suggests a potential bias towards WEIRD (Western, Educated, Industrialized, Rich and Democratic) contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts.

Rishanth Rajendhran, Amir Zadeh, Matthew Sarte et al.

Metrics like FactScore and VeriScore that evaluate long-form factuality operate by decomposing an input response into atomic claims and then individually verifying each claim. While effective and interpretable, these methods incur numerous LLM calls and can take upwards of 100 seconds to evaluate a single response, limiting their practicality in large-scale evaluation and training scenarios. To address this, we propose VeriFastScore, which leverages synthetic data to fine-tune Llama3.1 8B for simultaneously extracting and verifying all verifiable claims within a given text based on evidence from Google Search. We show that this task cannot be solved via few-shot prompting with closed LLMs due to its complexity: the model receives ~4K tokens of evidence on average and needs to concurrently decompose claims, judge their verifiability, and verify them against noisy evidence. However, our fine-tuned VeriFastScore model demonstrates strong correlation with the original VeriScore pipeline at both the example level (r=0.80) and system level (r=0.94) while achieving an overall speedup of 6.6x (9.9x excluding evidence retrieval) over VeriScore. To facilitate future factuality research, we publicly release our VeriFastScore model and synthetic datasets.

Katherine Thai, Bradley Emi, Elyas Masrour et al.

A significant proportion of queries to large language models ask them to edit user-provided text, rather than generate new text from scratch. While previous work focuses on detecting fully AI-generated text, we demonstrate that AI-edited text is distinguishable from human-written and AI-generated text. First, we propose using lightweight similarity metrics to quantify the magnitude of AI editing present in a text given the original human-written text and validate these metrics with human annotators. Using these similarity metrics as intermediate supervision, we then train EditLens, a regression model that predicts the amount of AI editing present within a text. Our model achieves state-of-the-art performance on both binary (F1=94.7%) and ternary (F1=90.4%) classification tasks in distinguishing human, AI, and mixed writing. Not only do we show that AI-edited text can be detected, but also that the degree of change made by AI to human writing can be detected, which has implications for authorship attribution, education, and policy. Finally, as a case study, we use our model to analyze the effects of AI-edits applied by Grammarly, a popular writing assistance tool. To encourage further research, we commit to publicly releasing our models and dataset.

Alisha Srivastava, Emir Korukluoglu, Minh Nhat Le et al.

Large language models (LLMs) are known to memorize and recall English text from their pretraining data. However, the extent to which this ability generalizes to non-English languages or transfers across languages remains unclear. This paper investigates multilingual and cross-lingual memorization in LLMs, probing if memorized content in one language (e.g., English) can be recalled when presented in translation. To do so, we introduce OWL, a dataset of 31.5K aligned excerpts from 20 books in ten languages, including English originals, official translations (Vietnamese, Spanish, Turkish), and new translations in six low-resource languages (Sesotho, Yoruba, Maithili, Malagasy, Setswana, Tahitian). We evaluate memorization across model families and sizes through three tasks: (1) direct probing, which asks the model to identify a book's title and author; (2) name cloze, which requires predicting masked character names; and (3) prefix probing, which involves generating continuations. We find that LLMs consistently recall content across languages, even for texts without direct translation in pretraining data. GPT-4o, for example, identifies authors and titles 69% of the time and masked entities 6% of the time in newly translated excerpts. Perturbations (e.g., masking characters, shuffling words) modestly reduce direct probing accuracy (7% drop for shuffled official translations). Our results highlight the extent of cross-lingual memorization and provide insights on the differences between the models.

Chau Minh Pham, Yapei Chang, Mohit Iyyer

LLM developers are increasingly reliant on synthetic data, but generating high-quality data for complex long-context reasoning tasks remains challenging. We introduce CLIPPER, a compression-based approach for generating synthetic data tailored to narrative claim verification - a task that requires reasoning over a book to verify a given claim. Instead of generating claims directly from the raw text of the book, which results in artifact-riddled claims, CLIPPER first compresses the book into chapter outlines and book summaries and then uses these intermediate representations to generate complex claims and corresponding chain-of-thoughts. Compared to naive approaches, CLIPPER produces claims that are more valid, grounded, and complex. Using CLIPPER, we construct a dataset of 19K synthetic book claims paired with their source texts and chain-of-thought reasoning, and use it to fine-tune three open-weight models. Our best model achieves breakthrough results on narrative claim verification (from 28% to 76% accuracy on our test set) and sets a new state-of-the-art for sub-10B models on the NoCha leaderboard. Further analysis shows that our models generate more detailed and grounded chain-of-thought reasoning while also improving performance on other narrative understanding tasks (e.g., NarrativeQA).

Jenna Russell, Marzena Karpinska, Destiny Akinode et al.

AI is rapidly transforming journalism, but the extent of its use in published newspaper articles remains unclear. We address this gap by auditing a large-scale dataset of 186K articles from online editions of 1.5K American newspapers published in the summer of 2025. Using Pangram, a state-of-the-art AI detector, we discover that approximately 9% of newly-published articles are either partially or fully AI-generated. This AI use is unevenly distributed, appearing more frequently in smaller, local outlets, in specific topics such as weather and technology, and within certain ownership groups. We also analyze 45K opinion pieces from Washington Post, New York Times, and Wall Street Journal, finding that they are 6.4 times more likely to contain AI-generated content than news articles from the same publications, with many AI-flagged op-eds authored by prominent public figures. Despite this prevalence, we find that AI use is rarely disclosed: a manual audit of 100 AI-flagged articles found only five disclosures of AI use. Overall, our audit highlights the immediate need for greater transparency and updated editorial standards regarding the use of AI in journalism to maintain public trust.

Katherine Thai, Mohit Iyyer

How well do modern long-context language models understand literary fiction? We explore this question via the task of literary evidence retrieval, repurposing the RELiC dataset of That et al. (2022) to construct a benchmark where the entire text of a primary source (e.g., The Great Gatsby) is provided to an LLM alongside literary criticism with a missing quotation from that work. This setting, in which the model must generate the missing quotation, mirrors the human process of literary analysis by requiring models to perform both global narrative reasoning and close textual examination. We curate a high-quality subset of 292 examples through extensive filtering and human verification. Our experiments show that recent reasoning models, such as Gemini Pro 2.5 can exceed human expert performance (62.5% vs. 50% accuracy). In contrast, the best open-weight model achieves only 29.1% accuracy, highlighting a wide gap in interpretive reasoning between open and closed-weight models. Despite their speed and apparent accuracy, even the strongest models struggle with nuanced literary signals and overgeneration, signaling open challenges for applying LLMs to literary analysis. We release our dataset and evaluation code to encourage future work in this direction.

Garima Dhanania, Sheshera Mysore, Chau Minh Pham et al.

Topic models are widely used to analyze document collections. While they are valuable for discovering latent topics in a corpus when analysts are unfamiliar with the corpus, analysts also commonly start with an understanding of the content present in a corpus. This may be through categories obtained from an initial pass over the corpus or a desire to analyze the corpus through a predefined set of categories derived from a high level theoretical framework (e.g. political ideology). In these scenarios analysts desire a topic modeling approach which incorporates their understanding of the corpus while supporting various forms of interaction with the model. In this work, we present EdTM, as an approach for label name supervised topic modeling. EdTM models topic modeling as an assignment problem while leveraging LM/LLM based document-topic affinities and using optimal transport for making globally coherent topic-assignments. In experiments, we show the efficacy of our framework compared to few-shot LLM classifiers, and topic models based on clustering and LDA. Further, we show EdTM's ability to incorporate various forms of analyst feedback and while remaining robust to noisy analyst inputs.

Yixiao Song, Yekyung Kim, Mohit Iyyer

Existing metrics for evaluating the factuality of long-form text, such as FACTSCORE (Min et al., 2023) and SAFE (Wei et al., 2024), decompose an input text into "atomic claims" and verify each against a knowledge base like Wikipedia. These metrics are not suitable for most generation tasks because they assume that every claim is verifiable (i.e., can plausibly be proven true or false). We address this issue with VERISCORE, a metric for diverse long-form generation tasks that contain both verifiable and unverifiable content. VERISCORE can be effectively implemented with either closed or fine-tuned open-weight language models, and human evaluation confirms that VERISCORE's extracted claims are more sensible than those from competing methods across eight different long-form tasks. We use VERISCORE to evaluate generations from 16 different models across multiple long-form tasks and find that while GPT-4o is the best-performing model overall, open-weight models such as Mixtral-8x22 are closing the gap. We show that an LM's VERISCORE on one task (e.g., biography generation) does not necessarily correlate to its VERISCORE on a different task (e.g., long-form QA), highlighting the need for expanding factuality evaluation across tasks with varying fact density.

Shane Arora, Marzena Karpinska, Hung-Ting Chen et al.

Despite rising global usage of large language models (LLMs), their ability to generate long-form answers to culturally specific questions remains unexplored in many languages. To fill this gap, we perform the first study of textual multilingual long-form QA by creating CaLMQA, a dataset of 51.7K culturally specific questions across 23 different languages. We define culturally specific questions as those that refer to concepts unique to one or a few cultures, or have different answers depending on the cultural or regional context. We obtain these questions by crawling naturally-occurring questions from community web forums in high-resource languages, and by hiring native speakers to write questions in under-resourced, rarely-studied languages such as Fijian and Kirundi. Our data collection methodologies are translation-free, enabling the collection of culturally unique questions like "Kuber iki umwami wa mbere w'uburundi yitwa Ntare?" (Kirundi; English translation: "Why was the first king of Burundi called Ntare (Lion)?"). We evaluate factuality, relevance and surface-level quality of LLM-generated long-form answers, finding that (1) for many languages, even the best models make critical surface-level errors (e.g., answering in the wrong language, repetition), especially for low-resource languages; and (2) answers to culturally specific questions contain more factual errors than answers to culturally agnostic questions -- questions that have consistent meaning and answer across many cultures. We release CaLMQA to facilitate future research in cultural and multilingual long-form QA.

Marzena Karpinska, Katherine Thai, Kyle Lo et al.

Synthetic long-context LLM benchmarks (e.g., "needle-in-the-haystack") test only surface-level retrieval capabilities, but how well can long-context LLMs retrieve, synthesize, and reason over information across book-length inputs? We address this question by creating NoCha, a dataset of 1,001 minimally different pairs of true and false claims about 67 recently-published English fictional books, written by human readers of those books. In contrast to existing long-context benchmarks, our annotators confirm that the largest share of pairs in NoCha require global reasoning over the entire book to verify. Our experiments show that while human readers easily perform this task, it is enormously challenging for all ten long-context LLMs that we evaluate: no open-weight model performs above random chance (despite their strong performance on synthetic benchmarks), while GPT-4o achieves the highest accuracy at 55.8%. Further analysis reveals that (1) on average, models perform much better on pairs that require only sentence-level retrieval vs. global reasoning; (2) model-generated explanations for their decisions are often inaccurate even for correctly-labeled claims; and (3) models perform substantially worse on speculative fiction books that contain extensive world-building. The methodology proposed in NoCha allows for the evolution of the benchmark dataset and the easy analysis of future models.

Fangyuan Xu, Yixiao Song, Mohit Iyyer et al.

Long-form question answering (LFQA) enables answering a wide range of questions, but its flexibility poses enormous challenges for evaluation. We perform the first targeted study of the evaluation of long-form answers, covering both human and automatic evaluation practices. We hire domain experts in seven areas to provide preference judgments over pairs of answers, along with free-form justifications for their choices. We present a careful analysis of experts' evaluation, which focuses on new aspects such as the comprehensiveness of the answer. Next, we examine automatic text generation metrics, finding that no existing metrics are predictive of human preference judgments. However, some metrics correlate with fine-grained aspects of answers (e.g., coherence). We encourage future work to move away from a single "overall score" of the answer and adopt a multi-faceted evaluation, targeting aspects such as factuality and completeness. We publicly release all of our annotations and code to spur future work into LFQA evaluation.

Shufan Wang, Yixiao Song, Andrew Drozdov et al.

In this paper, we study the generation quality of interpolation-based retrieval-augmented language models (LMs). These methods, best exemplified by the KNN-LM, interpolate the LM's predicted distribution of the next word with a distribution formed from the most relevant retrievals for a given prefix. While the KNN-LM and related methods yield impressive decreases in perplexity, we discover that they do not exhibit corresponding improvements in open-ended generation quality, as measured by both automatic evaluation metrics (e.g., MAUVE) and human evaluations. Digging deeper, we find that interpolating with a retrieval distribution actually increases perplexity compared to a baseline Transformer LM for the majority of tokens in the WikiText-103 test set, even though the overall perplexity is lower due to a smaller number of tokens for which perplexity dramatically decreases after interpolation. However, when decoding a long sequence at inference time, significant improvements on this smaller subset of tokens are washed out by slightly worse predictions on most tokens. Furthermore, we discover that the entropy of the retrieval distribution increases faster than that of the base LM as the generated sequence becomes longer, which indicates that retrieval is less reliable when using model-generated text as queries (i.e., is subject to exposure bias). We hope that our analysis spurs future work on improved decoding algorithms and interpolation strategies for retrieval-augmented language models.

Simeng Sun, Yang Liu, Shuohang Wang et al.

Strategies such as chain-of-thought prompting improve the performance of large language models (LLMs) on complex reasoning tasks by decomposing input examples into intermediate steps. However, it remains unclear how to apply such methods to reason over long input documents, in which both the decomposition and the output of each intermediate step are non-trivial to obtain. In this work, we propose PEARL, a prompting framework to improve reasoning over long documents, which consists of three stages: action mining, plan formulation, and plan execution. More specifically, given a question about a long document, PEARL decomposes the question into a sequence of actions (e.g., SUMMARIZE, FIND_EVENT, FIND_RELATION) and then executes them over the document to obtain the answer. Each stage of PEARL is implemented via zero-shot or few-shot prompting of LLMs (in our work, GPT-4) with minimal human input. We evaluate PEARL on a challenging subset of the QuALITY dataset, which contains questions that require complex reasoning over long narrative texts. PEARL outperforms zero-shot and chain-of-thought prompting on this dataset, and ablation experiments show that each stage of PEARL is critical to its performance. Overall, PEARL is a first step towards leveraging LLMs to reason over long documents.

Sewon Min, Kalpesh Krishna, Xinxi Lyu et al.

Evaluating the factuality of long-form text generated by large language models (LMs) is non-trivial because (1) generations often contain a mixture of supported and unsupported pieces of information, making binary judgments of quality inadequate, and (2) human evaluation is time-consuming and costly. In this paper, we introduce FACTSCORE, a new evaluation that breaks a generation into a series of atomic facts and computes the percentage of atomic facts supported by a reliable knowledge source. We conduct an extensive human evaluation to obtain FACTSCOREs of people biographies generated by several state-of-the-art commercial LMs -- InstructGPT, ChatGPT, and the retrieval-augmented PerplexityAI -- and report new analysis demonstrating the need for such a fine-grained score (e.g., ChatGPT only achieves 58%). Since human evaluation is costly, we also introduce an automated model that estimates FACTSCORE using retrieval and a strong language model, with less than a 2% error rate. Finally, we use this automated metric to evaluate 6,500 generations from a new set of 13 recent LMs that would have cost $26K if evaluated by humans, with various findings: GPT-4 and ChatGPT are more factual than public models, and Vicuna and Alpaca are some of the best public models. FACTSCORE is available for public use via `pip install factscore`.

Simeng Sun, Kalpesh Krishna, Andrew Mattarella-Micke et al.

Language models are generally trained on short, truncated input sequences, which limits their ability to use discourse-level information present in long-range context to improve their predictions. Recent efforts to improve the efficiency of self-attention have led to a proliferation of long-range Transformer language models, which can process much longer sequences than models of the past. However, the ways in which such models take advantage of the long-range context remain unclear. In this paper, we perform a fine-grained analysis of two long-range Transformer language models (including the \emph{Routing Transformer}, which achieves state-of-the-art perplexity on the PG-19 long-sequence LM benchmark dataset) that accept input sequences of up to 8K tokens. Our results reveal that providing long-range context (i.e., beyond the previous 2K tokens) to these models only improves their predictions on a small set of tokens (e.g., those that can be copied from the distant context) and does not help at all for sentence-level prediction tasks. Finally, we discover that PG-19 contains a variety of different document types and domains, and that long-range context helps most for literary novels (as opposed to textbooks or magazines).

Marzena Karpinska, Nader Akoury, Mohit Iyyer

Recent text generation research has increasingly focused on open-ended domains such as story and poetry generation. Because models built for such tasks are difficult to evaluate automatically, most researchers in the space justify their modeling choices by collecting crowdsourced human judgments of text quality (e.g., Likert scores of coherence or grammaticality) from Amazon Mechanical Turk (AMT). In this paper, we first conduct a survey of 45 open-ended text generation papers and find that the vast majority of them fail to report crucial details about their AMT tasks, hindering reproducibility. We then run a series of story evaluation experiments with both AMT workers and English teachers and discover that even with strict qualification filters, AMT workers (unlike teachers) fail to distinguish between model-generated text and human-generated references. We show that AMT worker judgments improve when they are shown model-generated output alongside human-generated references, which enables the workers to better calibrate their ratings. Finally, interviews with the English teachers provide deeper insights into the challenges of the evaluation process, particularly when rating model-generated text.

Shufan Wang, Laure Thompson, Mohit Iyyer

Phrase representations derived from BERT often do not exhibit complex phrasal compositionality, as the model relies instead on lexical similarity to determine semantic relatedness. In this paper, we propose a contrastive fine-tuning objective that enables BERT to produce more powerful phrase embeddings. Our approach (Phrase-BERT) relies on a dataset of diverse phrasal paraphrases, which is automatically generated using a paraphrase generation model, as well as a large-scale dataset of phrases in context mined from the Books3 corpus. Phrase-BERT outperforms baselines across a variety of phrase-level similarity tasks, while also demonstrating increased lexical diversity between nearest neighbors in the vector space. Finally, as a case study, we show that Phrase-BERT embeddings can be easily integrated with a simple autoencoder to build a phrase-based neural topic model that interprets topics as mixtures of words and phrases by performing a nearest neighbor search in the embedding space. Crowdsourced evaluations demonstrate that this phrase-based topic model produces more coherent and meaningful topics than baseline word and phrase-level topic models, further validating the utility of Phrase-BERT.

Tu Vu, Minh-Thang Luong, Quoc V. Le et al.

Despite their recent successes in tackling many NLP tasks, large-scale pre-trained language models do not perform as well in few-shot settings where only a handful of training examples are available. To address this shortcoming, we propose STraTA, which stands for Self-Training with Task Augmentation, an approach that builds on two key ideas for effective leverage of unlabeled data. First, STraTA uses task augmentation, a novel technique that synthesizes a large amount of data for auxiliary-task fine-tuning from target-task unlabeled texts. Second, STraTA performs self-training by further fine-tuning the strong base model created by task augmentation on a broad distribution of pseudo-labeled data. Our experiments demonstrate that STraTA can substantially improve sample efficiency across 12 few-shot benchmarks. Remarkably, on the SST-2 sentiment dataset, STraTA, with only 8 training examples per class, achieves comparable results to standard fine-tuning with 67K training examples. Our analyses reveal that task augmentation and self-training are both complementary and independently effective.

Zhiyang Xu, Andrew Drozdov, Jay Yoon Lee et al.

For over thirty years, researchers have developed and analyzed methods for latent tree induction as an approach for unsupervised syntactic parsing. Nonetheless, modern systems still do not perform well enough compared to their supervised counterparts to have any practical use as structural annotation of text. In this work, we present a technique that uses distant supervision in the form of span constraints (i.e. phrase bracketing) to improve performance in unsupervised constituency parsing. Using a relatively small number of span constraints we can substantially improve the output from DIORA, an already competitive unsupervised parsing system. Compared with full parse tree annotation, span constraints can be acquired with minimal effort, such as with a lexicon derived from Wikipedia, to find exact text matches. Our experiments show span constraints based on entities improves constituency parsing on English WSJ Penn Treebank by more than 5 F1. Furthermore, our method extends to any domain where span constraints are easily attainable, and as a case study we demonstrate its effectiveness by parsing biomedical text from the CRAFT dataset.

Hiroshi Iida, Dung Thai, Varun Manjunatha et al.

Existing work on tabular representation learning jointly models tables and associated text using self-supervised objective functions derived from pretrained language models such as BERT. While this joint pretraining improves tasks involving paired tables and text (e.g., answering questions about tables), we show that it underperforms on tasks that operate over tables without any associated text (e.g., populating missing cells). We devise a simple pretraining objective (corrupt cell detection) that learns exclusively from tabular data and reaches the state-of-the-art on a suite of table based prediction tasks. Unlike competing approaches, our model (TABBIE) provides embeddings of all table substructures (cells, rows, and columns), and it also requires far less compute to train. A qualitative analysis of our model's learned cell, column, and row representations shows that it understands complex table semantics and numerical trends.

Simeng Sun, Wenlong Zhao, Varun Manjunatha et al.

While large-scale pretrained language models have significantly improved writing assistance functionalities such as autocomplete, more complex and controllable writing assistants have yet to be explored. We leverage advances in language modeling to build an interactive writing assistant that generates and rephrases text according to fine-grained author specifications. Users provide input to our Intent-Guided Assistant (IGA) in the form of text interspersed with tags that correspond to specific rhetorical directives (e.g., adding description or contrast, or rephrasing a particular sentence). We fine-tune a language model on a dataset heuristically-labeled with author intent, which allows IGA to fill in these tags with generated text that users can subsequently edit to their liking. A series of automatic and crowdsourced evaluations confirm the quality of IGA's generated outputs, while a small-scale user study demonstrates author preference for IGA over baseline methods in a creative writing task. We release our dataset, code, and demo to spur further research into AI-assisted writing.