We explore how continued pre-training on domain-specific corpora influences large language models, revealing that training on the raw corpora endows the model with domain knowledge, but drastically hurts its prompting ability for question answering. Taken inspiration from human learning via reading comprehension--practice after reading improves the ability to answer questions based on the learned knowledge--we propose a simple method for transforming raw corpora into reading comprehension texts. Each raw text is enriched with a series of tasks related to its content. Our method, highly scalable and applicable to any pre-training corpora, consistently enhances performance across various tasks in three different domains: biomedicine, finance, and law. Notably, our 7B language model achieves competitive performance with domain-specific models of much larger scales, such as BloombergGPT-50B. Furthermore, we demonstrate that domain-specific reading comprehension texts can improve the model's performance even on general benchmarks, showing the potential to develop a general model across even more domains. Our model, code, and data are available at https://github.com/microsoft/LMOps.
Robustness in Natural Language Processing continues to be a pertinent issue, where state of the art models under-perform under naturally shifted distributions. In the context of Question Answering, work on domain adaptation methods continues to be a growing body of research. However, very little attention has been given to the notion of domain generalization under natural distribution shifts, where the target domain is unknown. With drastic improvements in the quality and access to generative models, we answer the question: How do generated datasets influence the performance of QA models under natural distribution shifts? We perform experiments on 4 different datasets under varying amounts of distribution shift, and analyze how "in-the-wild" generation can help achieve domain generalization. We take a two-step generation approach, generating both contexts and QA pairs to augment existing datasets. Through our experiments, we demonstrate how augmenting reading comprehension datasets with generated data leads to better robustness towards natural distribution shifts.
Large language models (LLMs) trained on massive corpora demonstrate impressive capabilities in a wide range of tasks. While there are ongoing efforts to adapt these models to languages beyond English, the attention given to their evaluation methodologies remains limited. Current multilingual benchmarks often rely on back translations or re-implementations of English tests, limiting their capacity to capture unique cultural and linguistic nuances. To bridge this gap for the Korean language, we introduce the HAE-RAE Bench, a dataset curated to challenge models lacking Korean cultural and contextual depth. The dataset encompasses six downstream tasks across four domains: vocabulary, history, general knowledge, and reading comprehension. Unlike traditional evaluation suites focused on token and sequence classification or mathematical and logical reasoning, the HAE-RAE Bench emphasizes a model's aptitude for recalling Korean-specific knowledge and cultural contexts. Comparative analysis with prior Korean benchmarks indicates that the HAE-RAE Bench presents a greater challenge to non-Korean models by disturbing abilities and knowledge learned from English being transferred.
We present Belebele, a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants. Significantly expanding the language coverage of natural language understanding (NLU) benchmarks, this dataset enables the evaluation of text models in high-, medium-, and low-resource languages. Each question is based on a short passage from the Flores-200 dataset and has four multiple-choice answers. The questions were carefully curated to discriminate between models with different levels of general language comprehension. The English dataset on its own proves difficult enough to challenge state-of-the-art language models. Being fully parallel, this dataset enables direct comparison of model performance across all languages. We use this dataset to evaluate the capabilities of multilingual masked language models (MLMs) and large language models (LLMs). We present extensive results and find that despite significant cross-lingual transfer in English-centric LLMs, much smaller MLMs pretrained on balanced multilingual data still understand far more languages. We also observe that larger vocabulary size and conscious vocabulary construction correlate with better performance on low-resource languages. Overall, Belebele opens up new avenues for evaluating and analyzing the multilingual capabilities of NLP systems.
Although dominant in natural language processing, transformer-based models remain challenged by the task of long-sequence processing, because the computational cost of self-attention operations in transformers swells quadratically with the input sequence length. To alleviate the complexity of long-sequence processing, we propose a simple framework to enable the offthe-shelf pre-trained transformers to process much longer sequences, while the computation and memory costs remain growing linearly with the input sequence lengths. More specifically, our method divides each long-sequence input into a batch of chunks, then aligns the interchunk information during the encoding steps, and finally selects the most representative hidden states from the encoder for the decoding process. To extract inter-chunk semantic information, we align the start and end token embeddings among chunks in each encoding transformer block. To learn an effective hidden selection policy, we design a dual updating scheme inspired by reinforcement learning, which regards the decoders of transformers as environments, and the downstream performance metrics as the rewards to evaluate the hidden selection actions. Our empirical results on real-world long-text summarization and reading comprehension tasks demonstrate effective improvements compared to prior longsequence processing baselines.
In this paper, we create NaijaRC: a new multi-choice Reading Comprehension dataset for three native Nigeria languages that is based on high-school reading comprehension examination. We provide baseline results by performing cross-lingual transfer using existing English RACE and Belebele training dataset based on a pre-trained encoder-only model. Additionally, we provide results by prompting large language models (LLMs) like GPT-4.
Catastrophic forgetting (CF) is a phenomenon that occurs in machine learning when a model forgets previously learned information while acquiring new knowledge for achieving a satisfactory performance in downstream tasks. As large language models (LLMs) have demonstrated remarkable performance, it is intriguing to investigate whether CF exists during the continual instruction tuning of LLMs. This study empirically evaluates the forgetting phenomenon in LLMs' knowledge during continual instruction tuning from the perspectives of domain knowledge, reasoning, and reading comprehension. The experiments reveal that catastrophic forgetting is generally observed in LLMs ranging from 1b to 7b parameters. Surprisingly, as the model scale increases, the severity of forgetting intensifies in such a model sale range which may result from the much significant initial performance in the larger LLM. Comparing the decoder-only model BLOOMZ with the encoder-decoder model mT0, BLOOMZ exhibits less forgetting and retains more knowledge. Interestingly, we also observe that LLMs can mitigate language biases, such as gender bias, during continual fine-tuning. Furthermore, our findings indicate that general instruction tuning can help alleviate the forgetting phenomenon in LLMs during subsequent fine-tuning.
Although deep learning techniques have shown significant achievements, they frequently depend on extensive amounts of hand-labeled data and tend to perform inadequately in few-shot scenarios. The objective of this study is to devise a strategy that can improve the model's capability to recognize biomedical entities in scenarios of few-shot learning. By redefining biomedical named entity recognition (BioNER) as a machine reading comprehension (MRC) problem, we propose a demonstration-based learning method to address few-shot BioNER, which involves constructing appropriate task demonstrations. In assessing our proposed method, we compared the proposed method with existing advanced methods using six benchmark datasets, including BC4CHEMD, BC5CDR-Chemical, BC5CDR-Disease, NCBI-Disease, BC2GM, and JNLPBA. We examined the models' efficacy by reporting F1 scores from both the 25-shot and 50-shot learning experiments. In 25-shot learning, we observed 1.1% improvements in the average F1 scores compared to the baseline method, reaching 61.7%, 84.1%, 69.1%, 70.1%, 50.6%, and 59.9% on six datasets, respectively. In 50-shot learning, we further improved the average F1 scores by 1.0% compared to the baseline method, reaching 73.1%, 86.8%, 76.1%, 75.6%, 61.7%, and 65.4%, respectively. We reported that in the realm of few-shot learning BioNER, MRC-based language models are much more proficient in recognizing biomedical entities compared to the sequence labeling approach. Furthermore, our MRC-language models can compete successfully with fully-supervised learning methodologies that rely heavily on the availability of abundant annotated data. These results highlight possible pathways for future advancements in few-shot BioNER methodologies.
Text matching is the task of matching two texts and determining the relationship between them, which has extensive applications in natural language processing tasks such as reading comprehension, and Question-Answering systems. The mainstream approach is to compute text representations or to interact with the text through attention mechanism, which is effective in text matching tasks. However, the performance of these models is insufficient for texts that require commonsense knowledge-based reasoning. To this end, in this paper, We introduce a new model for text matching called the Knowledge Enhanced Text Matching model (KETM), to enrich contextual representations with real-world common-sense knowledge from external knowledge sources to enhance our model understanding and reasoning. First, we use Wiktionary to retrieve the text word definitions as our external knowledge. Secondly, we feed text and knowledge to the text matching module to extract their feature vectors. The text matching module is used as an interaction module by integrating the encoder layer, the co-attention layer, and the aggregation layer. Specifically, the interaction process is iterated several times to obtain in-depth interaction information and extract the feature vectors of text and knowledge by multi-angle pooling. Then, we fuse text and knowledge using a gating mechanism to learn the ratio of text and knowledge fusion by a neural network that prevents noise generated by knowledge. After that, experimental validation on four datasets are carried out, and the experimental results show that our proposed model performs well on all four datasets, and the performance of our method is improved compared to the base model without adding external knowledge, which validates the effectiveness of our proposed method. The code is available at https://github.com/1094701018/KETM
Machine Reading Comprehension (MRC) models tend to take advantage of spurious correlations (also known as dataset bias or annotation artifacts in the research community). Consequently, these models may perform the MRC task without fully comprehending the given context and question, which is undesirable since it may result in low robustness against distribution shift. The main focus of this paper is answer-position bias, where a significant percentage of training questions have answers located solely in the first sentence of the context. We propose a Single-Sentence Reader as a new approach for addressing answer position bias in MRC. Remarkably, in our experiments with six different models, our proposed Single-Sentence Readers trained on biased dataset achieve results that nearly match those of models trained on normal dataset, proving their effectiveness in addressing the answer position bias. Our study also discusses several challenges our Single-Sentence Readers encounter and proposes a potential solution.
Question answering is a task that answers factoid questions using a large collection of documents. It aims to provide precise answers in response to the user's questions in natural language. 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. On the web, there is no single article that could provide all the possible answers available on the internet to the question of the problem asked by the user. The existing Dense Passage Retrieval model has been trained on Wikipedia dump from Dec. 20, 2018, as the source documents for answering questions. Question answering (QA) has made big strides with several open-domain and machine comprehension systems built using large-scale annotated datasets. However, in the clinical domain, this problem remains relatively unexplored. According to multiple surveys, Biomedical Questions cannot be answered correctly from Wikipedia Articles. In this work, we work on the existing DPR framework for the biomedical domain and retrieve answers from the Pubmed articles which is a reliable source to answer medical questions. When evaluated on a BioASQ QA dataset, our fine-tuned dense retriever results in a 0.81 F1 score.
As the popularity of voice assistants continues to surge, conversational search has gained increased attention in Information Retrieval. However, data sparsity issues in conversational search significantly hinder the progress of supervised conversational search methods. Consequently, researchers are focusing more on zero-shot conversational search approaches. Nevertheless, existing zero-shot methods face three primary limitations: they are not universally applicable to all retrievers, their effectiveness lacks sufficient explainability, and they struggle to resolve common conversational ambiguities caused by omission. To address these limitations, we introduce a novel Zero-shot Query Reformulation (or Query Rewriting) (ZeQR) framework that reformulates queries based on previous dialogue contexts without requiring supervision from conversational search data. Specifically, our framework utilizes language models designed for machine reading comprehension tasks to explicitly resolve two common ambiguities: coreference and omission, in raw queries. In comparison to existing zero-shot methods, our approach is universally applicable to any retriever without additional adaptation or indexing. It also provides greater explainability and effectively enhances query intent understanding because ambiguities are explicitly and proactively resolved. Through extensive experiments on four TREC conversational datasets, we demonstrate the effectiveness of our method, which consistently outperforms state-of-the-art baselines.
Deep text understanding, which requires the connections between a given document and prior knowledge beyond its text, has been highlighted by many benchmarks in recent years. However, these benchmarks have encountered two major limitations. On the one hand, most of them require human annotation of knowledge, which leads to limited knowledge coverage. On the other hand, they usually use choices or spans in the texts as the answers, which results in narrow answer space. To overcome these limitations, we build a new challenging benchmark named KoRc in this paper. Compared with previous benchmarks, KoRC has two advantages, i.e., broad knowledge coverage and flexible answer format. Specifically, we utilize massive knowledge bases to guide annotators or large language models (LLMs) to construct knowledgable questions. Moreover, we use labels in knowledge bases rather than spans or choices as the final answers. We test state-of-the-art models on KoRC and the experimental results show that the strongest baseline only achieves 68.3% and 30.0% F1 measure in the in-distribution and out-of-distribution test set, respectively. These results indicate that deep text understanding is still an unsolved challenge. The benchmark dataset, leaderboard, and baseline methods are released in https://github.com/THU-KEG/KoRC.
Text is by far the most ubiquitous source of knowledge and information and should be made easily accessible to as many people as possible; however, texts often contain complex words that hinder reading comprehension and accessibility. Therefore, suggesting simpler alternatives for complex words without compromising meaning would help convey the information to a broader audience. This paper proposes mTLS, a multilingual controllable Transformer-based Lexical Simplification (LS) system fined-tuned with the T5 model. The novelty of this work lies in the use of language-specific prefixes, control tokens, and candidates extracted from pre-trained masked language models to learn simpler alternatives for complex words. The evaluation results on three well-known LS datasets -- LexMTurk, BenchLS, and NNSEval -- show that our model outperforms the previous state-of-the-art models like LSBert and ConLS. Moreover, further evaluation of our approach on the part of the recent TSAR-2022 multilingual LS shared-task dataset shows that our model performs competitively when compared with the participating systems for English LS and even outperforms the GPT-3 model on several metrics. Moreover, our model obtains performance gains also for Spanish and Portuguese.
Large Language Models (LLMs) show promising results in language generation and instruction following but frequently "hallucinate", making their outputs less reliable. Despite Uncertainty Quantification's (UQ) potential solutions, implementing it accurately within LLMs is challenging. Our research introduces a simple heuristic: not all tokens in auto-regressive LLM text equally represent the underlying meaning, as "linguistic redundancy" often allows a few keywords to convey the essence of long sentences. However, current methods underestimate this inequality when assessing uncertainty, causing tokens with limited semantics to be equally or excessively weighted in UQ. To correct this, we propose Shifting Attention to more Relevant (SAR) components at both token- and sentence-levels for better UQ. We conduct extensive experiments involving a range of popular "off-the-shelf" LLMs, such as Vicuna, WizardLM, and LLaMA-2-chat, with model sizes extending up to 33B parameters. We evaluate various free-form question-answering tasks, encompassing domains such as reading comprehension, science Q&A, and medical Q&A. Our experimental results, coupled with a comprehensive demographic analysis, demonstrate the superior performance of SAR. The code is available at https://github.com/jinhaoduan/SAR.
In the field of machine reading comprehension (MRC), existing systems have surpassed the average performance of human beings in many tasks like SQuAD. However, there is still a long way to go when it comes to logical reasoning. Although some methods for it have been put forward, they either are designed in a quite complicated way or rely too much on external structures. In this paper, we proposed IDOL (InDicator-Oriented Logic Pre-training), an easy-to-understand but highly effective further pre-training task which logically strengthens the pre-trained models with the help of 6 types of logical indicators and a logically rich dataset LGP (LoGic Pre-training). IDOL achieves state-of-the-art performance on ReClor and LogiQA, the two most representative benchmarks in logical reasoning MRC, and is proven to be capable of generalizing to different pre-trained models and other types of MRC benchmarks like RACE and SQuAD 2.0 while keeping competitive general language understanding ability through testing on tasks in GLUE. Besides, at the beginning of the era of large language models, we take several of them like ChatGPT into comparison and find that IDOL still shows its advantage.
The traditional way of sentence-level event detection involves two important subtasks: trigger identification and trigger classifications, where the identified event trigger words are used to classify event types from sentences. However, trigger classification highly depends on abundant annotated trigger words and the accuracy of trigger identification. In a real scenario, annotating trigger words is time-consuming and laborious. For this reason, we propose a trigger-free event detection model, which transforms event detection into a two-tower model based on machine reading comprehension and prompt learning. Compared to existing trigger-based and trigger-free methods, experimental studies on two event detection benchmark datasets (ACE2005 and MAVEN) have shown that the proposed approach can achieve competitive performance.
Entity Linking (EL) is a fundamental task for Information Extraction and Knowledge Graphs. The general form of EL (i.e., end-to-end EL) aims to first find mentions in the given input document and then link the mentions to corresponding entities in a specific knowledge base. Recently, the paradigm of retriever-reader promotes the progress of end-to-end EL, benefiting from the advantages of dense entity retrieval and machine reading comprehension. However, the existing study only trains the retriever and the reader separately in a pipeline manner, which ignores the benefit that the interaction between the retriever and the reader can bring to the task. To advance the retriever-reader paradigm to perform more perfectly on end-to-end EL, we propose BEER$^2$, a Bidirectional End-to-End training framework for Retriever and Reader. Through our designed bidirectional end-to-end training, BEER$^2$ guides the retriever and the reader to learn from each other, make progress together, and ultimately improve EL performance. Extensive experiments on benchmarks of multiple domains demonstrate the effectiveness of our proposed BEER$^2$.
Machine reading comprehension (MRC) poses new challenges over logical reasoning, which aims to understand the implicit logical relations entailed in the given contexts and perform inference over them. Due to the complexity of logic, logical relations exist at different granularity levels. However, most existing methods of logical reasoning individually focus on either entity-aware or discourse-based information but ignore the hierarchical relations that may even have mutual effects. In this paper, we propose a holistic graph network (HGN) which deals with context at both discourse level and word level, as the basis for logical reasoning, to provide a more fine-grained relation extraction. Specifically, node-level and type-level relations, which can be interpreted as bridges in the reasoning process, are modeled by a hierarchical interaction mechanism to improve the interpretation of MRC systems. Experimental results on logical reasoning QA datasets (ReClor and LogiQA) and natural language inference datasets (SNLI and ANLI) show the effectiveness and generalization of our method, and in-depth analysis verifies its capability to understand complex logical relations.
Conventional knowledge distillation (KD) methods require access to the internal information of teachers, e.g., logits. However, such information may not always be accessible for large pre-trained language models (PLMs). In this work, we focus on decision-based KD for PLMs, where only teacher decisions (i.e., top-1 labels) are accessible. Considering the information gap between logits and decisions, we propose a novel method to estimate logits from the decision distributions. Specifically, decision distributions can be both derived as a function of logits theoretically and estimated with test-time data augmentation empirically. By combining the theoretical and empirical estimations of the decision distributions together, the estimation of logits can be successfully reduced to a simple root-finding problem. Extensive experiments show that our method significantly outperforms strong baselines on both natural language understanding and machine reading comprehension datasets.