A Multi-answer Multi-task Framework for Real-world Machine Reading Comprehension

BackgroundExtractive methods for machine reading comprehension (MRC) tasks have achieved comparable or better accuracy than human performance on benchmark data sets. However, such models are not as successful when adapted to complex domains such as health care. One of the main reasons is that the context that the MRC model needs to process when operating in a complex domain can be much larger compared with an average open-domain context. This causes the MRC model to make less accurate and slower predictions. A potential solution to this problem is to reduce the input context of the MRC model by extracting only the necessary parts from the original context.ObjectiveThis study aims to develop a method for extracting useful contexts from long articles as an additional component to the question answering task, enabling the MRC model to work more efficiently and accurately.MethodsExisting approaches to context extraction in MRC are based on sentence selection strategies, in which the models are trained to find the sentences containing the answer. We found that using only the sentences containing the answer was insufficient for the MRC model to predict correctly. We conducted a series of empirical studies and observed a strong relationship between the usefulness of the context and the confidence score output of the MRC model. Our investigation showed that a precise input context can boost the prediction correctness of the MRC and greatly reduce inference time. We proposed a method to estimate the utility of each sentence in a context in answering the question and then extract a new, shorter context according to these estimations. We generated a data set to train 2 models for estimating sentence utility, based on which we selected more precise contexts that improved the MRC model’s performance.ResultsWe demonstrated our approach on the Question Answering Data Set for COVID-19 and Biomedical Semantic Indexing and Question Answering data sets and showed that the approach benefits the downstream MRC model. First, the method substantially reduced the inference time of the entire question answering system by 6 to 7 times. Second, our approach helped the MRC model predict the answer more correctly compared with using the original context (F1-score increased from 0.724 to 0.744 for the Question Answering Data Set for COVID-19 and from 0.651 to 0.704 for the Biomedical Semantic Indexing and Question Answering). We also found a potential problem where extractive transformer MRC models predict poorly despite being given a more precise context in some cases.ConclusionsThe proposed context extraction method allows the MRC model to achieve improved prediction correctness and a significantly reduced MRC inference time. This approach works technically with any MRC model and has potential in tasks involving processing long texts.

With the rapid growth of legal cases, professionals are under pressure to go through lengthy documents and grasp informative pieces of text in the limited time. Most of the existing techniques focus on simple legal information retrieval task, such as name or address of the prosecutor or the defendant, which can be easily accomplished with the help of handcrafted patterns or sequence labeling methods. Yet complicated texts always challenge such pattern-based methods and sequence labeling approaches. These texts state the same facts or describe the same events, but they do not share common or similar patterns. In this paper, we design a unified framework to extract legal information in various formats, including directly extracted information (a piece of span) and information that needs to be deduced. The framework follows the methodology to answer questions in machine reading comprehension (MRC) tasks. We treat the extraction fact labels as the counterpart of questions in MRC task and propose several strategies to represent them. We construct several datasets regarding different cases for training and testing. Our best strategy achieves up to 4% enhancement in F1 score on each dataset compared to the MRC baseline.

We report two experiments that investigated the regulation of memory accuracy with a new regulatory mechanism: the plurality option. This mechanism is closely related to the grain-size option but involves control over the number of alternatives contained in an answer rather than the quantitative boundaries of a single answer. Participants were presented with a slideshow depicting a robbery (Experiment 1) or a murder (Experiment 2), and their memory was tested with five-alternative multiple-choice questions. For each question, participants were asked to generate two answers: a single answer consisting of one alternative and a plural answer consisting of the single answer and two other alternatives. Each answer was rated for confidence (Experiment 1) or for the likelihood of being correct (Experiment 2), and one of the answers was selected for reporting. Results showed that participants used the plurality option to regulate accuracy, selecting single answers when their accuracy and confidence were high, but opting for plural answers when they were low. Although accuracy was higher for selected plural than for selected single answers, the opposite pattern was evident for confidence or likelihood ratings. This dissociation between confidence and accuracy for selected answers was the result of marked overconfidence in single answers coupled with underconfidence in plural answers. We hypothesize that these results can be attributed to overly dichotomous metacognitive beliefs about personal knowledge states that cause subjective confidence to be extreme.

Medical Named Entity Recognition (NER) is a fundamental but challenging task due to the lack of specialized entity datasets like tumor entities, which are often overlapped and discontinuous. In this paper, we propose a novel Prompt Enhanced Generative Machine Reading Comprehension Framework (PGMRC) to improve the overlapped and discontinuous NER performance. Specifically, we formulate NER as a Machine Reading Comprehension (MRC) task and employ a pre-trained encoder-decoder module to generate entity span sequences according to their entity query. In this way, we adopt query to guide the model to focus on answer entities in context, which can naturally solve entity overlap and alleviate the exposure bias of the generative model. Then, we introduce continuous prompts to the self-attention mechanism in Transformer to reduce the dependence on manually constructed queries. In addition, we annotate 875 pathological documents of pancreatic cancer and construct a Chinese pathological NER dataset (PAN) containing overlapped and discontinuous entities. Finally, we conduct our experiments on three widely used benchmarks (GENIA, ACE04, ACE05) and our dataset PAN. Experiments have demonstrated its effectiveness and better performance than state-of-the-art methods.

Machine reading comprehension (MRC) is a challenging natural language processing (NLP) task. It has a wide application potential in the fields of question answering robots, human‐computer interactions in mobile virtual reality systems, etc. Recently, the emergence of pretrained models (PTMs) has brought this research field into a new era, in which the training objective plays a key role. The masked language model (MLM) is a self‐supervised training objective widely used in various PTMs. With the development of training objectives, many variants of MLM have been proposed, such as whole word masking, entity masking, phrase masking, and span masking. In different MLMs, the length of the masked tokens is different. Similarly, in different machine reading comprehension tasks, the length of the answer is also different, and the answer is often a word, phrase, or sentence. Thus, in MRC tasks with different answer lengths, whether the length of MLM is related to performance is a question worth studying. If this hypothesis is true, it can guide us on how to pretrain the MLM with a relatively suitable mask length distribution for MRC tasks. In this paper, we try to uncover how much of MLM’s success in the machine reading comprehension tasks comes from the correlation between masking length distribution and answer length in the MRC dataset. In order to address this issue, herein, (1) we propose four MRC tasks with different answer length distributions, namely, the short span extraction task, long span extraction task, short multiple‐choice cloze task, and long multiple‐choice cloze task; (2) four Chinese MRC datasets are created for these tasks; (3) we also have pretrained four masked language models according to the answer length distributions of these datasets; and (4) ablation experiments are conducted on the datasets to verify our hypothesis. The experimental results demonstrate that our hypothesis is true. On four different machine reading comprehension datasets, the performance of the model with correlation length distribution surpasses the model without correlation.

This paper presents the development process of a Vietnamese spoken language corpus for machine reading comprehension (MRC) tasks and provides insights into the challenges and opportunities associated with using realworld data for machine reading comprehension tasks.The existing MRC corpora in Vietnamese mainly focus on formal written documents such as Wikipedia articles, online newspapers, or textbooks.In contrast, the VlogQA consists of 10,076 question-answer pairs based on 1,230 transcript documents sourced from YouTube -an extensive source of user-uploaded content, covering the topics of food and travel.By capturing the spoken language of native Vietnamese speakers in natural settings, an obscure corner overlooked in Vietnamese research, the corpus provides a valuable resource for future research in reading comprehension tasks for the Vietnamese language.Regarding performance evaluation, our deep-learning models achieved the highest F1 score of 75.34% on the test set, indicating significant progress in machine reading comprehension for Vietnamese spoken language data.In terms of EM, the highest score we accomplished is 53.97%, which reflects the challenge in processing spoken-based content and highlights the need for further improvement.

Pre-trained language models have achieved human-level performance on many Machine Reading Comprehension (MRC) tasks, but it remains unclear whether these models truly understand language or answer questions by exploiting statistical biases in datasets. Here, we demonstrate a simple yet effective method to attack MRC models and reveal the statistical biases in these models. We apply the method to the RACE dataset, for which the answer to each MRC question is selected from 4 options. It is found that several pre-trained language models, including BERT, ALBERT, and RoBERTa, show consistent preference to some options, even when these options are irrelevant to the question. When interfered by these irrelevant options, the performance of MRC models can be reduced from human-level performance to the chance-level performance. Human readers, however, are not clearly affected by these irrelevant options. Finally, we propose an augmented training method that can greatly reduce models' statistical biases.

Pretrained language models (PrLMs) have been widely used in machine reading comprehension (MRC) tasks. Although PrLMs can learn powerful representations from large-scale corpora, it introduces the problem of model redundancy, which severely restricts the application of mobile devices. Recent studies have suggested the use of early exit (EE) methods for the fast inference of PrLMs. However, there are two limitations to the application of existing EE methods. First, different Transformer layers have different capabilities for learning semantic features. Thus, the lower layers may behave differently from the higher layers. As a result, the earlier the exit from the model, the lower its performance. Second, a greedy search is applied to the classifiers exiting either the start or the end position. However, the start and end positions may also be related. Choosing only one best candidate might be suitable for the current step, but it may be a suboptimal choice in global inference. To address these issues, this study proposes an on-device MRC model with fast inference based on an EE strategy. To ensure that the lower layers have similar abilities as the higher layers, we propose the use of self-distillation to transfer knowledge from the higher layers to the lower ones. Furthermore, a dynamic programming strategy was applied to choose the optimal confidence combination of both the start and end classifiers to better select the exit layers in the inference stage. Experimental results on both SQuAD 1.1 and 2.0 show that the inference of the proposed model is accelerated at the minimum cost of performance loss, thereby outperforming previous EE models.KeywordsMachine reading comprehensionDynamic early exitSelf-distillationAdaptive inference

MotivationCurrent studies in extractive question answering (EQA) have modeled the single-span extraction setting, where a single answer span is a label to predict for a given question-passage pair. This setting is natural for general domain EQA as the majority of the questions in the general domain can be answered with a single span. Following general domain EQA models, current biomedical EQA (BioEQA) models utilize the single-span extraction setting with post-processing steps.ResultsIn this article, we investigate the question distribution across the general and biomedical domains and discover biomedical questions are more likely to require list-type answers (multiple answers) than factoid-type answers (single answer). This necessitates the models capable of producing multiple answers for a question. Based on this preliminary study, we propose a sequence tagging approach for BioEQA, which is a multi-span extraction setting. Our approach directly tackles questions with a variable number of phrases as their answer and can learn to decide the number of answers for a question from training data. Our experimental results on the BioASQ 7b and 8b list-type questions outperformed the best-performing existing models without requiring post-processing steps.Availability and implementationSource codes and resources are freely available for download at https://github.com/dmis-lab/SeqTagQA.Supplementary informationSupplementary data are available at Bioinformatics online.

Factoid question answering (QA) has recently benefited from the development of deep learning (DL) systems. Neural network models outperform traditional approaches in domains where large datasets exist, such as SQuAD (ca. 100,000 questions) for Wikipedia articles. However, these systems have not yet been applied to QA in more specific domains, such as biomedicine, because datasets are generally too small to train a DL system from scratch. For example, the BioASQ dataset for biomedical QA comprises less then 900 factoid (single answer) and list (multiple answers) QA instances. In this work, we adapt a neural QA system trained on a large open-domain dataset (SQuAD, source) to a biomedical dataset (BioASQ, target) by employing various transfer learning techniques. Our network architecture is based on a state-of-the-art QA system, extended with biomedical word embeddings and a novel mechanism to answer list questions. In contrast to existing biomedical QA systems, our system does not rely on domain-specific ontologies, parsers or entity taggers, which are expensive to create. Despite this fact, our systems achieve state-of-the-art results on factoid questions and competitive results on list questions.

Machine reading comprehension is a classic issue artificial intelligence. It is a key technology in the next generation search engine and intelligent interactive service. The traditional methods usually work in a small scale of data sets. The traditional system cannot meet the emerging demand. Deep learning and cloud computation have ability to deal with the large scale data sets. In real scene, the parameters affect the performance of machine reading comprehension task. In this paper, we analyze how the parameters of deep neural network affect the machine reading comprehension. The experiment results show that the performance is only sensitive to a few parameters which should be key point for engineers.

Chinese electronic medical records named entity recognition (NER) is a core task in medical knowledge mining, which is usually viewed as a sequence labeling problem. Recent works introduce the machine reading comprehension (MRC) framework into this task and extract named entities in a question-answering manner, resulting in state-of-the-art performance. However, they extract entities of different types independently, ignoring the fact that entities presented in the context might highly correlate with each other. To address this issue, we extend the MRC-based model and propose Fusion Label Relations with MRC (FLR-MRC). The method implicitly models the relations between labels through graph attention networks and fuse label information with text for named entity recognition. Experimental results on the benchmark datasets CMeEE and CCKS2017-CNER demonstrate FLR-MRC outperform existing clinical NER methods, with F1-score reaching 0.6652 and 0.9101, respectively.KeywordsMedical named entity recognitionLabel relationsMachine reading comprehension

Development of an Arabic HQAS-based ASAG to consider an ignored knowledge in misspelled multiple words short answers

We propose a global optimization method under length constraint (GOLC) for neural text summarization models. GOLC increases the probabilities of generating summaries that have high evaluation scores, ROUGE in this paper, within a desired length. We compared GOLC with two optimization methods, a maximum log-likelihood and a minimum risk training, on CNN/Daily Mail and a Japanese single document summarization data set of The Mainichi Shimbun Newspapers. The experimental results show that a state-of-the-art neural summarization model optimized with GOLC generates fewer overlength summaries while maintaining the fastest processing speed; only 6.70% overlength summaries on CNN/Daily and 7.8% on long summary of Mainichi, compared to the approximately 20% to 50% on CNN/Daily Mail and 10% to 30% on Mainichi with the other optimization methods. We also demonstrate the importance of the generation of in-length summaries for post-editing with the dataset Mainich that is created with strict length constraints. The ex- perimental results show approximately 30% to 40% improved post-editing time by use of in-length summaries.

Log-linear model combination is the standard approach in LVCSR to combine several knowledge sources, usually an acoustic and a language model.Instead of using a single scaling factor per knowledge source, we make the scaling factor wordand pronunciation-dependent.In this work, we combine three acoustic models, a pronunciation model, and a language model for a Mandarin BN/BC task.The achieved error rate reduction of 2% relative is small but consistent for two test sets.An analysis of the results shows that the major contribution comes from the improved interdependency of language and acoustic model.