We introduce REPLUG, a retrieval-augmented language modeling framework that treats the language model (LM) as a black box and augments it with a tuneable retrieval model. Unlike prior retrieval-augmented LMs that train language models with special cross-attention mechanisms to encode the retrieved text, REPLUG simply prepends retrieved documents to the input for the frozen black-box LM. This simple design can be easily applied to any existing language models. Furthermore, we show that the LM can be used to supervise the retrieval model, which can then find documents that help the LM make better predictions. Our experiments demonstrate that REPLUG with the tuned retriever significantly improves the performance of GPT-3 (175B) on language modeling by 6.3%, as well as the performance of Codex on five-shot MMLU by 5.1%. Code is publicly released at github.com/swj0419/REPLUG.

Existing language models (LMs) predict tokens with a softmax over a finite vocabulary, which can make it difficult to predict rare tokens or phrases. We introduce NPM, the first nonparametric masked language model that replaces this softmax with a nonparametric distribution over every phrase in a reference corpus. NPM fills in the [MASK] solely from retrieving a token from a text corpus. We show that NPM can be efficiently trained with a contrastive objective and an in-batch approximation to full corpus retrieval. Zero-shot evaluation on 16 tasks including classification, fact probing and question answering demonstrates that NPM outperforms significantly larger parametric models, either with or without a retrieve-and-generate approach. It is particularly better at dealing with rare patterns (word senses or facts) and predicting rare or nearly unseen words (e.g., non-Latin script). We release the model and code at github.com/facebookresearch/NPM.

We investigate the ability of language models to perform compositional reasoning tasks where the overall solution depends on correctly composing the answers to sub-problems. We measure how often models can correctly answer all sub-problems but not generate the overall solution, a ratio we call the compositionality gap. We evaluate this ratio by asking multi-hop questions with answers that require composing multiple facts unlikely to have been observed together during pretraining. In the GPT-3 family of models, as model size increases we show that the single-hop question answering performance improves faster than the multi-hop performance does, therefore the compositionality gap does not decrease. This surprising result suggests that while more powerful models memorize and recall more factual knowledge, they show no corresponding improvement in their ability to perform this kind of compositional reasoning. We then demonstrate how elicitive prompting (such as chain of thought) narrows the compositionality gap by reasoning explicitly instead of implicitly. We present a new method, self-ask, that further improves on chain of thought. In our method, the model explicitly asks itself (and then answers) follow-up questions before answering the initial question. We finally show that self-ask’s structured prompting lets us easily plug in a search engine to answer the follow-up questions, which additionally improves accuracy.

In an information-seeking conversation, a user may ask questions that are under-specified or unanswerable. An ideal agent would interact by initiating different response types according to the available knowledge sources. However, most current studies either fail to or artificially incorporate such agent-side initiative. This work presents InSCIt, a dataset for Information-Seeking Conversations with mixed-initiative Interactions. It contains 4.7K user-agent turns from 805 human-human conversations where the agent searches over Wikipedia and either directly answers, asks for clarification, or provides relevant information to address user queries. The data supports two subtasks, evidence passage identification and response generation, as well as a human evaluation protocol to assess model performance. We report results of two systems based on state-of-the-art models of conversational knowledge identification and open-domain question answering. Both systems significantly underperform humans, suggesting ample room for improvement in future studies.1

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‘.

Although large language models can be prompted for both zero- and few-shot learning, performance drops significantly when no demonstrations are available. In this paper, we introduce Z-ICL, a new zero-shot method that closes the gap by constructing pseudo-demonstrations for a given test input using a raw text corpus. Concretely, pseudo-demonstrations are constructed by (1) finding the nearest neighbors to the test input from the corpus and pairing them with random task labels, and (2) applying a set of techniques to reduce the amount of direct copying the model does from the resulting demonstrations. Evaluation on nine classification datasets shows that Z-ICL outperforms previous zero-shot methods by a significant margin, and is on par with in-context learning with labeled training data in the few-shot setting. Overall, Z-ICL provides a significantly higher estimate of the zero-shot performance levels of a model, and supports future efforts to develop better pseudo-demonstrations that further improve zero-shot results.

Chain-of-Thought (CoT) prompting can dramatically improve the multi-step reasoning abilities of large language models (LLMs). CoT explicitly encourages the LLM to generate intermediate rationales for solving a problem, by providing a series of reasoning steps in the demonstrations. Despite its success, there is still little understanding of what makes CoT prompting effective and which aspects of the demonstrated reasoning steps contribute to its performance. In this paper, we show that CoT reasoning is possible even with invalid demonstrations - prompting with invalid reasoning steps can achieve over 80-90% of the performance obtained using CoT under various metrics, while still generating coherent lines of reasoning during inference. Further experiments show that other aspects of the rationales, such as being relevant to the query and correctly ordering the reasoning steps, are much more important for effective CoT reasoning. Overall, these findings both deepen our understanding of CoT prompting, and open up new questions regarding LLMs’ capability to learn to reason in context.

When asking about unfamiliar topics, information seeking users often pose questions with false presuppositions. Most existing question answering (QA) datasets, in contrast, assume all questions have well defined answers. We introduce CREPE, a QA dataset containing a natural distribution of presupposition failures from online information-seeking forums. We find that 25% of questions contain false presuppositions, and provide annotations for these presuppositions and their corrections. Through extensive baseline experiments, we show that adaptations of existing open-domain QA models can find presuppositions moderately well, but struggle when predicting whether a presupposition is factually correct. This is in large part due to difficulty in retrieving relevant evidence passages from a large text corpus. CREPE provides a benchmark to study question answering in the wild, and our analyses provide avenues for future work in better modeling and further studying the task.

Retrieval-based language models (LMs) have shown impressive performance on diverse NLP tasks. In this tutorial, we will provide a comprehensive and coherent overview of recent advances in retrieval-based LMs. We will start by providing preliminaries covering the foundation of LMs (e.g., masked LMs, autoregressive LMs) and retrieval systems (e.g., nearest-neighbor search). We will then detail recent progress in retrieval-based models, focusing on their model architectures and learning approaches. Finally, we will show how retrieval-based LMs are adapted to downstream applications, and extended to multilingual and multi-modal settings. Finally, we will use an exercise to showcase the effectiveness of retrieval-based LMs.

Large language models (LMs) are able to in-context learn—perform a new task via inference alone by conditioning on a few input-label pairs (demonstrations) and making predictions for new inputs. However, there has been little understanding of how the model learns and which aspects of the demonstrations contribute to end task performance. In this paper, we show that ground truth demonstrations are in fact not required—randomly replacing labels in the demonstrations barely hurts performance on a range of classification and multi-choce tasks, consistently over 12 different models including GPT-3. Instead, we find that other aspects of the demonstrations are the key drivers of endtask performance, including the fact that they provide a few examples of (1) the label space, (2) the distribution of the input text, and (3) the overall format of the sequence. Together, our analysis provides a new way of understanding how and why in-context learning works, while opening up new questions about how much can be learned from large language models through inference alone.

For users to trust model predictions, they need to understand model outputs, particularly their confidence — calibration aims to adjust (calibrate) models’ confidence to match expected accuracy. We argue that the traditional calibration evaluation does not promote effective calibrations: for example, it can encourage always assigning a mediocre confidence score to all predictions, which does not help users distinguish correct predictions from wrong ones. Building on those observations, we propose a new calibration metric, MacroCE, that better captures whether the model assigns low confidence to wrong predictions and high confidence to correct predictions. Focusing on the practical application of open-domain question answering, we examine conventional calibration methods applied on the widely-used retriever-reader pipeline, all of which do not bring significant gains under our new MacroCE metric. Toward better calibration, we propose a new calibration method (ConsCal) that uses not just final model predictions but whether multiple model checkpoints make consistent predictions. Altogether, we provide an alternative view of calibration along with a new metric, re-evaluation of existing calibration methods on our metric, and proposal of a more effective calibration method.

We conduct a large empirical evaluation to investigate the landscape of distributional robustness in question answering. Our investigation spans over 350 models and 16 question answering datasets, including a diverse set of architectures, model sizes, and adaptation methods (e.g., fine-tuning, adapter tuning, in-context learning, etc.). We find that, in many cases, model variations do not affect robustness and in-distribution performance alone determines out-of-distribution performance.Moreover, our findings indicate thati) zero-shot and in-context learning methods are more robust to distribution shifts than fully fine-tuned models;ii) few-shot prompt fine-tuned models exhibit better robustness than few-shot fine-tuned span prediction models;iii) parameter-efficient and robustness enhancing training methods provide no significant robustness improvements.In addition, we publicly release all evaluations to encourage researchers to further analyze robustness trends for question answering models.

Despite significant interest in developing general purpose fact checking models, it is challenging to construct a large-scale fact verification dataset with realistic real-world claims. Existing claims are either authored by crowdworkers, thereby introducing subtle biases thatare difficult to control for, or manually verified by professional fact checkers, causing them to be expensive and limited in scale. In this paper, we construct a large-scale challenging fact verification dataset called FAVIQ, consisting of 188k claims derived from an existing corpus of ambiguous information-seeking questions. The ambiguities in the questions enable automatically constructing true and false claims that reflect user confusions (e.g., the year of the movie being filmed vs. being released). Claims in FAVIQ are verified to be natural, contain little lexical bias, and require a complete understanding of the evidence for verification. Our experiments show that the state-of-the-art models are far from solving our new task. Moreover, training on our data helps in professional fact-checking, outperforming models trained on the widely used dataset FEVER or in-domain data by up to 17% absolute. Altogether, our data will serve as a challenging benchmark for natural language understanding and support future progress in professional fact checking.

We introduce a noisy channel approach for language model prompting in few-shot text classification. Instead of computing the likelihood of the label given the input (referred as direct models), channel models compute the conditional probability of the input given the label, and are thereby required to explain every word in the input. We use channel models for recently proposed few-shot learning methods with no or very limited updates to the language model parameters, via either in-context demonstration or prompt tuning. Our experiments show that, for both methods, channel models significantly outperform their direct counterparts, which we attribute to their stability, i.e., lower variance and higher worst-case accuracy. We also present extensive ablations that provide recommendations for when to use channel prompt tuning instead of other competitive models (e.g., direct head tuning): channel prompt tuning is preferred when the number of training examples is small, labels in the training data are imbalanced, or generalization to unseen labels is required.

The ability to efficiently learn from little-to-no data is critical to applying NLP to tasks where data collection is costly or otherwise difficult. This is a challenging setting both academically and practically—particularly because training neutral models typically require large amount of labeled data. More recently, advances in pretraining on unlabelled data have brought up the potential of better zero-shot or few-shot learning (Devlin et al., 2019; Brown et al., 2020). In particular, over the past year, a great deal of research has been conducted to better learn from limited data using large-scale language models. In this tutorial, we aim at bringing interested NLP researchers up to speed about the recent and ongoing techniques for zero- and few-shot learning with pretrained language models. Additionally, our goal is to reveal new research opportunities to the audience, which will hopefully bring us closer to address existing challenges in this domain.

We introduce MetaICL (Meta-training for In-Context Learning), a new meta-training framework for few-shot learning where a pretrained language model is tuned to do in-context learning on a large set of training tasks. This meta-training enables the model to more effectively learn a new task in context at test time, by simply conditioning on a few training examples with no parameter updates or task-specific templates. We experiment on a large, diverse collection of tasks consisting of 142 NLP datasets including classification, question answering, natural language inference, paraphrase detection and more, across seven different meta-training/target splits. MetaICL outperforms a range of baselines including in-context learning without meta-training and multi-task learning followed by zero-shot transfer. We find that the gains are particularly significant for target tasks that have domain shifts from the meta-training tasks, and that using a diverse set of the meta-training tasks is key to improvements. We also show that MetaICL approaches (and sometimes beats) the performance of models fully finetuned on the target task training data, and outperforms much bigger models with nearly 8x parameters.

Fine-tuning continuous prompts for target tasks has recently emerged as a compact alternative to full model fine-tuning. Motivated by these promising results, we investigate the feasibility of extracting a discrete (textual) interpretation of continuous prompts that is faithful to the problem they solve. In practice, we observe a “wayward” behavior between the task solved by continuous prompts and their nearest neighbor discrete projections: We can find continuous prompts that solve a task while being projected to an arbitrary text (e.g., definition of a different or even a contradictory task), while being within a very small (2%) margin of the best continuous prompt of the same size for the task. We provide intuitions behind this odd and surprising behavior, as well as extensive empirical analyses quantifying the effect of various parameters. For instance, for larger model sizes we observe higher waywardness, i.e, we can find prompts that more closely map to any arbitrary text with a smaller drop in accuracy. These findings have important implications relating to the difficulty of faithfully interpreting continuous prompts and their generalization across models and tasks, providing guidance for future progress in prompting language models.

State-of-the-art Machine Reading Comprehension (MRC) models for Open-domain Question Answering (QA) are typically trained for span selection using distantly supervised positive examples and heuristically retrieved negative examples. This training scheme possibly explains empirical observations that these models achieve a high recall amongst their top few predictions, but a low overall accuracy, motivating the need for answer re-ranking. We develop a successful re-ranking approach (RECONSIDER) for span-extraction tasks that improves upon the performance of MRC models, even beyond large-scale pre-training. RECONSIDER is trained on positive and negative examples extracted from high confidence MRC model predictions, and uses in-passage span annotations to perform span-focused re-ranking over a smaller candidate set. As a result, RECONSIDER learns to eliminate close false positives, achieving a new extractive state of the art on four QA tasks, with 45.5% Exact Match accuracy on Natural Questions with real user questions, and 61.7% on TriviaQA. We will release all related data, models, and code.

In this tutorial, we aim at bringing interested NLP researchers up to speed about the recent and ongoing techniques for document-level representation learning. Additionally, our goal is to reveal new research opportunities to the audience, which will hopefully bring us closer to address existing challenges in this domain.

We study multi-answer retrieval, an under-explored problem that requires retrieving passages to cover multiple distinct answers for a given question. This task requires joint modeling of retrieved passages, as models should not repeatedly retrieve passages containing the same answer at the cost of missing a different valid answer. Prior work focusing on single-answer retrieval is limited as it cannot reason about the set of passages jointly. In this paper, we introduce JPR, a joint passage retrieval model focusing on reranking. To model the joint probability of the retrieved passages, JPR makes use of an autoregressive reranker that selects a sequence of passages, equipped with novel training and decoding algorithms. Compared to prior approaches, JPR achieves significantly better answer coverage on three multi-answer datasets. When combined with downstream question answering, the improved retrieval enables larger answer generation models since they need to consider fewer passages, establishing a new state-of-the-art.

Question answering (QA) tasks have been posed using a variety of formats, such as extractive span selection, multiple choice, etc. This has led to format-specialized models, and even to an implicit division in the QA community. We argue that such boundaries are artificial and perhaps unnecessary, given the reasoning abilities we seek to teach are not governed by the format. As evidence, we use the latest advances in language modeling to build a single pre-trained QA model, UNIFIEDQA, that performs well across 19 QA datasets spanning 4 diverse formats. UNIFIEDQA performs on par with 8 different models that were trained on individual datasets themselves. Even when faced with 12 unseen datasets of observed formats, UNIFIEDQA performs surprisingly well, showing strong generalization from its outof-format training data. Finally, simply finetuning this pre trained QA model into specialized models results in a new state of the art on 10 factoid and commonsense question answering datasets, establishing UNIFIEDQA as a strong starting point for building QA systems.

Ambiguity is inherent to open-domain question answering; especially when exploring new topics, it can be difficult to ask questions that have a single, unambiguous answer. In this paper, we introduce AmbigQA, a new open-domain question answering task which involves finding every plausible answer, and then rewriting the question for each one to resolve the ambiguity. To study this task, we construct AmbigNQ, a dataset covering 14,042 questions from NQ-open, an existing open-domain QA benchmark. We find that over half of the questions in NQ-open are ambiguous, with diverse sources of ambiguity such as event and entity references. We also present strong baseline models for AmbigQA which we show benefit from weakly supervised learning that incorporates NQ-open, strongly suggesting our new task and data will support significant future research effort. Our data and baselines are available at https://nlp.cs.washington.edu/ambigqa.

We present ELQ, a fast end-to-end entity linking model for questions, which uses a biencoder to jointly perform mention detection and linking in one pass. Evaluated on WebQSP and GraphQuestions with extended annotations that cover multiple entities per question, ELQ outperforms the previous state of the art by a large margin of +12.7% and +19.6% F1, respectively. With a very fast inference time (1.57 examples/s on a single CPU), ELQ can be useful for downstream question answering systems. In a proof-of-concept experiment, we demonstrate that using ELQ significantly improves the downstream QA performance of GraphRetriever.

Open-domain question answering relies on efficient passage retrieval to select candidate contexts, where traditional sparse vector space models, such as TF-IDF or BM25, are the de facto method. In this work, we show that retrieval can be practically implemented using dense representations alone, where embeddings are learned from a small number of questions and passages by a simple dual-encoder framework. When evaluated on a wide range of open-domain QA datasets, our dense retriever outperforms a strong Lucene-BM25 system greatly by 9%-19% absolute in terms of top-20 passage retrieval accuracy, and helps our end-to-end QA system establish new state-of-the-art on multiple open-domain QA benchmarks.

Multi-hop reading comprehension (RC) questions are challenging because they require reading and reasoning over multiple paragraphs. We argue that it can be difficult to construct large multi-hop RC datasets. For example, even highly compositional questions can be answered with a single hop if they target specific entity types, or the facts needed to answer them are redundant. Our analysis is centered on HotpotQA, where we show that single-hop reasoning can solve much more of the dataset than previously thought. We introduce a single-hop BERT-based RC model that achieves 67 F1—comparable to state-of-the-art multi-hop models. We also design an evaluation setting where humans are not shown all of the necessary paragraphs for the intended multi-hop reasoning but can still answer over 80% of questions. Together with detailed error analysis, these results suggest there should be an increasing focus on the role of evidence in multi-hop reasoning and possibly even a shift towards information retrieval style evaluations with large and diverse evidence collections.

Multi-hop Reading Comprehension (RC) requires reasoning and aggregation across several paragraphs. We propose a system for multi-hop RC that decomposes a compositional question into simpler sub-questions that can be answered by off-the-shelf single-hop RC models. Since annotations for such decomposition are expensive, we recast subquestion generation as a span prediction problem and show that our method, trained using only 400 labeled examples, generates sub-questions that are as effective as human-authored sub-questions. We also introduce a new global rescoring approach that considers each decomposition (i.e. the sub-questions and their answers) to select the best final answer, greatly improving overall performance. Our experiments on HotpotQA show that this approach achieves the state-of-the-art results, while providing explainable evidence for its decision making in the form of sub-questions.

Many question answering (QA) tasks only provide weak supervision for how the answer should be computed. For example, TriviaQA answers are entities that can be mentioned multiple times in supporting documents, while DROP answers can be computed by deriving many different equations from numbers in the reference text. In this paper, we show it is possible to convert such tasks into discrete latent variable learning problems with a precomputed, task-specific set of possible solutions (e.g. different mentions or equations) that contains one correct option. We then develop a hard EM learning scheme that computes gradients relative to the most likely solution at each update. Despite its simplicity, we show that this approach significantly outperforms previous methods on six QA tasks, including absolute gains of 2–10%, and achieves the state-of-the-art on five of them. Using hard updates instead of maximizing marginal likelihood is key to these results as it encourages the model to find the one correct answer, which we show through detailed qualitative analysis.

Machine reading comprehension, the task of evaluating a machine’s ability to comprehend a passage of text, has seen a surge in popularity in recent years. There are many datasets that are targeted at reading comprehension, and many systems that perform as well as humans on some of these datasets. Despite all of this interest, there is no work that systematically defines what reading comprehension is. In this work, we justify a question answering approach to reading comprehension and describe the various kinds of questions one might use to more fully test a system’s comprehension of a passage, moving beyond questions that only probe local predicate-argument structures. The main pitfall of this approach is that questions can easily have surface cues or other biases that allow a model to shortcut the intended reasoning process. We discuss ways proposed in current literature to mitigate these shortcuts, and we conclude with recommendations for future dataset collection efforts.

Neural models for question answering (QA) over documents have achieved significant performance improvements. Although effective, these models do not scale to large corpora due to their complex modeling of interactions between the document and the question. Moreover, recent work has shown that such models are sensitive to adversarial inputs. In this paper, we study the minimal context required to answer the question, and find that most questions in existing datasets can be answered with a small set of sentences. Inspired by this observation, we propose a simple sentence selector to select the minimal set of sentences to feed into the QA model. Our overall system achieves significant reductions in training (up to 15 times) and inference times (up to 13 times), with accuracy comparable to or better than the state-of-the-art on SQuAD, NewsQA, TriviaQA and SQuAD-Open. Furthermore, our experimental results and analyses show that our approach is more robust to adversarial inputs.

We show that the task of question answering (QA) can significantly benefit from the transfer learning of models trained on a different large, fine-grained QA dataset. We achieve the state of the art in two well-studied QA datasets, WikiQA and SemEval-2016 (Task 3A), through a basic transfer learning technique from SQuAD. For WikiQA, our model outperforms the previous best model by more than 8%. We demonstrate that finer supervision provides better guidance for learning lexical and syntactic information than coarser supervision, through quantitative results and visual analysis. We also show that a similar transfer learning procedure achieves the state of the art on an entailment task.