We propose a new paradigm to help Large Language Models (LLMs) generate more accurate factual knowledge without retrieving from an external corpus, called RECITation-augmented gEneration (RECITE). Different from retrieval-augmented language models that retrieve relevant documents before generating the outputs, given an input, RECITE first recites one or several relevant passages from LLMs' own memory via sampling, and then produces the final answers. We show that RECITE is a powerful paradigm for knowledge-intensive NLP tasks. Specifically, we show that by utilizing recitation as the intermediate step, a recite-and-answer scheme can achieve new state-of-the-art performance in various closed-book question answering (CBQA) tasks. In experiments, we verify the effectiveness of \method~on four pre-trained models (PaLM, UL2, OPT, and Codex) and three CBQA tasks (Natural Questions, TriviaQA, and HotpotQA). Our code is available at "https://github.com/Edward-Sun/RECITE".
Knowledge distillation is an effective way to transfer knowledge from a strong teacher to an efficient student model. Ideally, we expect the better the teacher is, the better the student. However, this expectation does not always come true. It is common that a better teacher model results in a bad student via distillation due to the nonnegligible gap between teacher and student. To bridge the gap, we propose PROD, a PROgressive Distillation method, for dense retrieval. PROD consists of a teacher progressive distillation and a data progressive distillation to gradually improve the student. We conduct extensive experiments on five widely-used benchmarks, MS MARCO Passage, TREC Passage 19, TREC Document 19, MS MARCO Document and Natural Questions, where PROD achieves the state-of-the-art within the distillation methods for dense retrieval. The code and models will be released.
Much recent research on information retrieval has focused on how to transfer from one task (typically with abundant supervised data) to various other tasks where supervision is limited, with the implicit assumption that it is possible to generalize from one task to all the rest. However, this overlooks the fact that there are many diverse and unique retrieval tasks, each targeting different search intents, queries, and search domains. In this paper, we suggest to work on Few-shot Dense Retrieval, a setting where each task comes with a short description and a few examples. To amplify the power of a few examples, we propose Prompt-base Query Generation for Retriever (Promptagator), which leverages large language models (LLM) as a few-shot query generator, and creates task-specific retrievers based on the generated data. Powered by LLM's generalization ability, Promptagator makes it possible to create task-specific end-to-end retrievers solely based on a few examples {without} using Natural Questions or MS MARCO to train %question generators or dual encoders. Surprisingly, LLM prompting with no more than 8 examples allows dual encoders to outperform heavily engineered models trained on MS MARCO like ColBERT v2 by more than 1.2 nDCG on average on 11 retrieval sets. Further training standard-size re-rankers using the same generated data yields another 5.0 point nDCG improvement. Our studies determine that query generation can be far more effective than previously observed, especially when a small amount of task-specific knowledge is given.
Deep prompt tuning (DPT) has gained great success in most natural language processing~(NLP) tasks. However, it is not well-investigated in dense retrieval where fine-tuning~(FT) still dominates. When deploying multiple retrieval tasks using the same backbone model~(e.g., RoBERTa), FT-based methods are unfriendly in terms of deployment cost: each new retrieval model needs to repeatedly deploy the backbone model without reuse. To reduce the deployment cost in such a scenario, this work investigates applying DPT in dense retrieval. The challenge is that directly applying DPT in dense retrieval largely underperforms FT methods. To compensate for the performance drop, we propose two model-agnostic and task-agnostic strategies for DPT-based retrievers, namely retrieval-oriented intermediate pretraining and unified negative mining, as a general approach that could be compatible with any pre-trained language model and retrieval task. The experimental results show that the proposed method (called DPTDR) outperforms previous state-of-the-art models on both MS-MARCO and Natural Questions. We also conduct ablation studies to examine the effectiveness of each strategy in DPTDR. We believe this work facilitates the industry, as it saves enormous efforts and costs of deployment and increases the utility of computing resources. Our code is available at https://github.com/tangzhy/DPTDR.
Large language models have shown impressive few-shot results on a wide range of tasks. However, when knowledge is key for such results, as is the case for tasks such as question answering and fact checking, massive parameter counts to store knowledge seem to be needed. Retrieval augmented models are known to excel at knowledge intensive tasks without the need for as many parameters, but it is unclear whether they work in few-shot settings. In this work we present Atlas, a carefully designed and pre-trained retrieval augmented language model able to learn knowledge intensive tasks with very few training examples. We perform evaluations on a wide range of tasks, including MMLU, KILT and NaturalQuestions, and study the impact of the content of the document index, showing that it can easily be updated. Notably, Atlas reaches over 42% accuracy on Natural Questions using only 64 examples, outperforming a 540B parameters model by 3% despite having 50x fewer parameters.
Despite the progresses on pre-trained language models, there is a lack of unified frameworks for pre-trained sentence representation. As such, it calls for different pre-training methods for specific scenarios, and the pre-trained models are likely to be limited by their universality and representation quality. In this work, we extend the recently proposed MAE style pre-training strategy, RetroMAE, such that it may effectively support a wide variety of sentence representation tasks. The extended framework consists of two stages, with RetroMAE conducted throughout the process. The first stage performs RetroMAE over generic corpora, like Wikipedia, BookCorpus, etc., from which the base model is learned. The second stage takes place on domain-specific data, e.g., MS MARCO and NLI, where the base model is continuingly trained based on RetroMAE and contrastive learning. The pre-training outputs at the two stages may serve different applications, whose effectiveness are verified with comprehensive experiments. Concretely, the base model are proved to be effective for zero-shot retrieval, with remarkable performances achieved on BEIR benchmark. The continuingly pre-trained models further benefit more downstream tasks, including the domain-specific dense retrieval on MS MARCO, Natural Questions, and the sentence embeddings' quality for standard STS and transfer tasks in SentEval. The empirical insights of this work may inspire the future design of sentence representation pre-training. Our pre-trained models and source code will be released to the public communities.
Monumental advancements in artificial intelligence (AI) have lured the interest of doctors, lenders, judges, and other professionals. While these high-stakes decision-makers are optimistic about the technology, those familiar with AI systems are wary about the lack of transparency of its decision-making processes. Perturbation-based post hoc explainers offer a model agnostic means of interpreting these systems while only requiring query-level access. However, recent work demonstrates that these explainers can be fooled adversarially. This discovery has adverse implications for auditors, regulators, and other sentinels. With this in mind, several natural questions arise - how can we audit these black box systems? And how can we ascertain that the auditee is complying with the audit in good faith? In this work, we rigorously formalize this problem and devise a defense against adversarial attacks on perturbation-based explainers. We propose algorithms for the detection (CAD-Detect) and defense (CAD-Defend) of these attacks, which are aided by our novel conditional anomaly detection approach, KNN-CAD. We demonstrate that our approach successfully detects whether a black box system adversarially conceals its decision-making process and mitigates the adversarial attack on real-world data for the prevalent explainers, LIME and SHAP.
Existing benchmarks for open-domain question answering (ODQA) typically focus on questions whose answers can be extracted from a single paragraph. By contrast, many natural questions, such as "What players were drafted by the Brooklyn Nets?" have a list of answers. Answering such questions requires retrieving and reading from many passages, in a large corpus. We introduce QAMPARI, an ODQA benchmark, where question answers are lists of entities, spread across many paragraphs. We created QAMPARI by (a) generating questions with multiple answers from Wikipedia's knowledge graph and tables, (b) automatically pairing answers with supporting evidence in Wikipedia paragraphs, and (c) manually paraphrasing questions and validating each answer. We train ODQA models from the retrieve-and-read family and find that QAMPARI is challenging in terms of both passage retrieval and answer generation, reaching an F1 score of 32.8 at best. Our results highlight the need for developing ODQA models that handle a broad range of question types, including single and multi-answer questions.
Knowledge graph question answering (KGQA) facilitates information access by leveraging structured data without requiring formal query language expertise from the user. Instead, users can express their information needs by simply asking their questions in natural language (NL). Datasets used to train KGQA models that would provide such a service are expensive to construct, both in terms of expert and crowdsourced labor. Typically, crowdsourced labor is used to improve template-based pseudo-natural questions generated from formal queries. However, the resulting datasets often fall short of representing genuinely natural and fluent language. In the present work, we investigate ways to characterize and remedy these shortcomings. We create the IQN-KGQA test collection by sampling questions from existing KGQA datasets and evaluating them with regards to five different aspects of naturalness. Then, the questions are rewritten to improve their fluency. Finally, the performance of existing KGQA models is compared on the original and rewritten versions of the NL questions. We find that some KGQA systems fare worse when presented with more realistic formulations of NL questions. The IQN-KGQA test collection is a resource to help evaluate KGQA systems in a more realistic setting. The construction of this test collection also sheds light on the challenges of constructing large-scale KGQA datasets with genuinely NL questions.
Tables are an important form of structured data for both human and machine readers alike, providing answers to questions that cannot, or cannot easily, be found in texts. Recent work has designed special models and training paradigms for table-related tasks such as table-based question answering and table retrieval. Though effective, they add complexity in both modeling and data acquisition compared to generic text solutions and obscure which elements are truly beneficial. In this work, we focus on the task of table retrieval, and ask: "is table-specific model design necessary for table retrieval, or can a simpler text-based model be effectively used to achieve a similar result?" First, we perform an analysis on a table-based portion of the Natural Questions dataset (NQ-table), and find that structure plays a negligible role in more than 70% of the cases. Based on this, we experiment with a general Dense Passage Retriever (DPR) based on text and a specialized Dense Table Retriever (DTR) that uses table-specific model designs. We find that DPR performs well without any table-specific design and training, and even achieves superior results compared to DTR when fine-tuned on properly linearized tables. We then experiment with three modules to explicitly encode table structures, namely auxiliary row/column embeddings, hard attention masks, and soft relation-based attention biases. However, none of these yielded significant improvements, suggesting that table-specific model design may not be necessary for table retrieval.
Recently, retrieval models based on dense representations are dominant in passage retrieval tasks, due to their outstanding ability in terms of capturing semantics of input text compared to the traditional sparse vector space models. A common practice of dense retrieval models is to exploit a dual-encoder architecture to represent a query and a passage independently. Though efficient, such a structure loses interaction between the query-passage pair, resulting in inferior accuracy. To enhance the performance of dense retrieval models without loss of efficiency, we propose a GNN-encoder model in which query (passage) information is fused into passage (query) representations via graph neural networks that are constructed by queries and their top retrieved passages. By this means, we maintain a dual-encoder structure, and retain some interaction information between query-passage pairs in their representations, which enables us to achieve both efficiency and efficacy in passage retrieval. Evaluation results indicate that our method significantly outperforms the existing models on MSMARCO, Natural Questions and TriviaQA datasets, and achieves the new state-of-the-art on these datasets.
Long-form answers, consisting of multiple sentences, can provide nuanced and comprehensive answers to a broader set of questions. To better understand this complex and understudied task, we study the functional structure of long-form answers collected from three datasets, ELI5, WebGPT and Natural Questions. Our main goal is to understand how humans organize information to craft complex answers. We develop an ontology of six sentence-level functional roles for long-form answers, and annotate 3.9k sentences in 640 answer paragraphs. Different answer collection methods manifest in different discourse structures. We further analyze model-generated answers -- finding that annotators agree less with each other when annotating model-generated answers compared to annotating human-written answers. Our annotated data enables training a strong classifier that can be used for automatic analysis. We hope our work can inspire future research on discourse-level modeling and evaluation of long-form QA systems.
Recent large language models often answer factual questions correctly. But users can't trust any given claim a model makes without fact-checking, because language models can hallucinate convincing nonsense. In this work we use reinforcement learning from human preferences (RLHP) to train "open-book" QA models that generate answers whilst also citing specific evidence for their claims, which aids in the appraisal of correctness. Supporting evidence is drawn from multiple documents found via a search engine, or from a single user-provided document. Our 280 billion parameter model, GopherCite, is able to produce answers with high quality supporting evidence and abstain from answering when unsure. We measure the performance of GopherCite by conducting human evaluation of answers to questions in a subset of the NaturalQuestions and ELI5 datasets. The model's response is found to be high-quality 80\% of the time on this Natural Questions subset, and 67\% of the time on the ELI5 subset. Abstaining from the third of questions for which it is most unsure improves performance to 90\% and 80\% respectively, approaching human baselines. However, analysis on the adversarial TruthfulQA dataset shows why citation is only one part of an overall strategy for safety and trustworthiness: not all claims supported by evidence are true.
The process of model checkpoint validation refers to the evaluation of the performance of a model checkpoint executed on a held-out portion of the training data while learning the hyperparameters of the model, and is used to avoid over-fitting and determine when the model has converged so as to stop training. A simple and efficient strategy to validate deep learning checkpoints is the addition of validation loops to execute during training. However, the validation of dense retrievers (DR) checkpoints is not as trivial -- and the addition of validation loops is not efficient. This is because, in order to accurately evaluate the performance of a DR checkpoint, the whole document corpus needs to be encoded into vectors using the current checkpoint before any actual retrieval operation for checkpoint validation can be performed. This corpus encoding process can be very time-consuming if the document corpus contains millions of documents (e.g., 8.8m for MS MARCO and 21m for Natural Questions). Thus, a naive use of validation loops during training will significantly increase training time. To address this issue, in this demo paper, we propose Asyncval: a Python-based toolkit for efficiently validating DR checkpoints during training. Instead of pausing the training loop for validating DR checkpoints, Asyncval decouples the validation loop from the training loop, uses another GPU to automatically validate new DR checkpoints and thus permits to perform validation asynchronously from training. Asyncval also implements a range of different corpus subset sampling strategies for validating DR checkpoints; these strategies allow to further speed up the validation process. We provide an investigation of these methods in terms of their impact on validation time and validation fidelity. Asyncval is made available as an open-source project at https://github.com/ielab/asyncval.
Scale has opened new frontiers in natural language processing -- but at a high cost. In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models. But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning. Our work focuses on these issues and acts as a design guide. We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B). For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge), summarization (XSum, CNN-DM), closed book question answering (WebQA, Natural Questions), and adversarially constructed tasks (Winogrande, ANLI R3).
Text embeddings are useful features in many applications such as semantic search and computing text similarity. Previous work typically trains models customized for different use cases, varying in dataset choice, training objective and model architecture. In this work, we show that contrastive pre-training on unsupervised data at scale leads to high quality vector representations of text and code. The same unsupervised text embeddings that achieve new state-of-the-art results in linear-probe classification also display impressive semantic search capabilities and sometimes even perform competitively with fine-tuned models. On linear-probe classification accuracy averaging over 7 tasks, our best unsupervised model achieves a relative improvement of 4% and 1.8% over previous best unsupervised and supervised text embedding models respectively. The same text embeddings when evaluated on large-scale semantic search attains a relative improvement of 23.4%, 14.7%, and 10.6% over previous best unsupervised methods on MSMARCO, Natural Questions and TriviaQA benchmarks, respectively. Similarly to text embeddings, we train code embedding models on (text, code) pairs, obtaining a 20.8% relative improvement over prior best work on code search.
Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.
Commonsense question answering (CQA) aims to test if models can answer questions regarding commonsense knowledge that everyone knows. Prior works that incorporate external knowledge bases have shown promising results, but knowledge bases are expensive to construct and are often limited to a fixed set of relations. In this paper, we instead focus on better utilizing the \textit{implicit knowledge} stored in pre-trained language models. While researchers have found that the knowledge embedded in pre-trained language models can be extracted by having them fill in the blanks of carefully designed prompts for relation extraction and text classification, it remains unclear if we can adopt this paradigm in CQA where the inputs and outputs take much more flexible forms. To this end, we investigate four translation methods that can translate natural questions into cloze-style sentences to better solicit commonsense knowledge from language models, including a syntactic-based model, an unsupervised neural model, and two supervised neural models. In addition, to combine the different translation methods, we propose to encourage consistency among model predictions on different translated questions with unlabeled data. We demonstrate the effectiveness of our methods on three CQA datasets in zero-shot settings. We show that our methods are complementary to a knowledge base improved model, and combining them can lead to state-of-the-art zero-shot performance. Analyses also reveal distinct characteristics of the different cloze translation methods and provide insights on why combining them can lead to great improvements.
In this paper, we focus on studying robustness evaluation of Chinese question matching. Most of the previous work on analyzing robustness issue focus on just one or a few types of artificial adversarial examples. Instead, we argue that it is necessary to formulate a comprehensive evaluation about the linguistic capabilities of models on natural texts. For this purpose, we create a Chinese dataset namely DuQM which contains natural questions with linguistic perturbations to evaluate the robustness of question matching models. DuQM contains 3 categories and 13 subcategories with 32 linguistic perturbations. The extensive experiments demonstrate that DuQM has a better ability to distinguish different models. Importantly, the detailed breakdown of evaluation by linguistic phenomenon in DuQM helps us easily diagnose the strength and weakness of different models. Additionally, our experiment results show that the effect of artificial adversarial examples does not work on the natural texts.
This paper studies continual learning (CL) for sentiment classification (SC). In this setting, the CL system learns a sequence of SC tasks incrementally in a neural network, where each task builds a classifier to classify the sentiment of reviews of a particular product category or domain. Two natural questions are: Can the system transfer the knowledge learned in the past from the previous tasks to the new task to help it learn a better model for the new task? And, can old models for previous tasks be improved in the process as well? This paper proposes a novel technique called KAN to achieve these objectives. KAN can markedly improve the SC accuracy of both the new task and the old tasks via forward and backward knowledge transfer. The effectiveness of KAN is demonstrated through extensive experiments.