A Survey on Neural Machine Reading Comprehension

A Survey on Neural Machine Reading Comprehension

Boyu Qiu 111Co-first authours. B.Qiu. and X.Chen. contributed equally.    Xu Chen    Jungang Xu 222Corresponding authours.    Yingfei Sun \affiliationsUniversity of Chinese Academy of Sciences \emails{qiuboyu18, chenxu18}@mails.ucas.ac.cn
{xujg, yfsun}@ucas.ac.cn

Enabling a machine to read and comprehend the natural language documents so that it can answer some question remains an elusive challenge. In recent years, the popularity of deep learning and the establishment of large-scale datasets have both promoted the prosperity of Machine Reading Comprehension. This paper aims to present how to utilize the Neural Network to build a Reader and introduce some classic models, analyze what improvements they make. Further, we also point out the defects of existing models and future research directions.

A Survey on Neural Machine Reading Comprehension

Boyu Qiu thanks: Co-first authours. B.Qiu. and X.Chen. contributed equally., Xu Chen , Jungang Xu thanks: Corresponding authours., Yingfei Sun

University of Chinese Academy of Sciences

{qiuboyu18, chenxu18}@mails.ucas.ac.cn
{xujg, yfsun}@ucas.ac.cn

1 Introduction

As an important mean of measuring the machine’s ability to understand documents, Machine Reading Comprehension (MRC) has been the central goal of Natural Language Processing (NLP) research . MRC tasks usually require the machine to answer questions after reading one or more relevant passages. Traditional methods tend to use pattern matching techniques with additional automated linguistic processing like stemming, name identification, or analyzing the relationship between questions and answers, then use statistical or mathematical methods to calculate the similarity between the question and sentences to find the most likely answer. Such methods are more like based on retrieval of the relational database rather than the real understanding of given passages, so they lack of reasoning ability.

With the popularity of deep learning and the emergence of some brilliant architecture in NLP such as the sequence-to-sequence model and word vectors, MRC ushered in the stage of prosperity in 2015. In this year, [?] proposed two basic neural MRC models: attentive reader and impatient reader, which provide a lot of heuristic ideas, such as using bidirectional RNN/LSTM to capture the contextual semantics of the passage, using the attention mechanism to make the machine imitate human to read with questions. These ideas have become the basis of many later neural MRC models. In 2016, Stanford University released the Stanford Question Answering Dataset (SQuAD) [?], which contains more than 100,000 (questions, original, answer) triples. After that, a large number of representative models are proposed on this dataset. Large-scale datasets and end-to-end neural network models have jointly promoted the prosperity of MRC. Recent work such as GPT [?], BERT [?], etc. have surpassed human performance in several datasets.

The purpose of this paper is to analyze and summarize the classic models of neural MRC in recent years. In Section 2, we divide the existing dataset into four categories based on the question style and introduce them separately. In Section 3, we first extract the common architecture of existing models. Then for some efficient models, we present improvements what they make briefly. Besides, we also introduce how to utilize transfer learning to solve NLP tasks. In Section 4, we provide the application of machine reading comprehension, including conversational question answering and open-domain question answering, etc. The Sction 5 is discussion, presenting some experiments from the recent achievements and some inadequacies of the current MRC tasks.

2 Datasets

In the exploration of MRC, constructing a high-quality, large-scale dataset is as important as optimizing the model structure. According to the form of question and answer, the MRC datasets can be roughly divided into four categories.

Dataset Answer Type Data Source Question Document Metric
CNN/DM Cloze-style News report 1.4M 300k Accuracy
CBT Cloze-style Children’s book 688k 108 Accuracy
MCTest Multi-choice Fictional stories 2k 500 Accuracy
RACE Multi-choice English exams 97k 28k Accuracy
SQuAD Extractive Wikipedia 100k 536 EM, F1
SQuAD2.0 Extractive, Unanswerable Wikipedia 150k 505 EM, F1
TriviaQA Extractive Wikipedia,Web 40k 660k EM, F1
NarrativeQA Abstractive Wikipedia 46k 1.5K ROUGE-L
DuReader Abstractive, Unanswerable User logs 200k 1M BLEU, ROUGE-L
MC MARCO Abstractive, Unanswerable User logs 1M 3.2M Accuracy, ROUGE-L, BLEU
Table 1: Datasets

2.1 Cloze-Style

In this style MRC tasks, the question contains a placeholder and the machine must decide which word or entity is the most suitable option.During evaluation, accuracy is the only metric. Main datasets in this style are listed below.

CNN/Daily Mail:

[?] collected 93k articles from the CNN and 220k articles form the Daily Mail (DM). The dataset extracts the entity of the critical sentence in the news story and replaces it with a placeholder. To remove the influence of external word knowledge, the named entities in the dataset are replaced by anonymous IDs, which are then further shuffled for each example. This processing makes the models to rely on text to answer questions.

Children’s Book Test:

[?] proposed Children’s Book Test dataset (CBT) using a different style. They take a sequence of 21 consecutive sentences from a children’s book. Then, the first 20 sentences are viewed as context and the query is to infer the missing word in the 21st sentence.

2.2 Multi-Choice

This style MRC tasks require the machine to find the only correct option in the k (usually 4) hypothesis based on the given passage. Similarly, the evaluation metric is accuracy. Main datasets in this style are listed below.


[?] constructed MCTest, the first comprehensive reading comprehension dataset since the wave of neural networks, which has 660 fictional stories, each with 4 questions and 4 candidate answers.


[?] collected more than 20,000 articles and 100,000 questions from Chinese middle and high school students’ English exams, covering a wide range of fields. These questions were proposed by experts, which was orig-inally to examine the level of human reading comprehension. Therefore, answering the question requires the machine to have some reasoning ability.

2.3 Span-Prediction

The machine needs to find the correct answer from one of the single spans in the passage. This category is also referred to as extractive question answering. There are two metrics to evaluate the task. EM (Exact Match) assigns credit 1.0 if the predicted answer is equal to the standard answer and 0.0 otherwise. F1 computes the average word overlap between the predicted and the standard answer. Main datasets in this style are listed below.


[?] proposed the reading comprehension dataset SQuAD, which contains more than 100,000 questions identified in 536 Wikipedia by manual workers. Each question corresponds to a specific passage, and the answer to the question is located at a span of the passage. The challenge based on the SQuAD greatly promoted the prosperity of MRC.


In 2018, the SQuAD 2.0 version was released by [?]. Unlike the previous version, more than 50,000 new, unanswerable questions have been added. These questions have plausible answers in the corresponding article, that is, the questions are of the same type, but not correct. The model needs to identify these questions to avoid answering them.


[?] TriviaQA, which contains over 650K question-answer-evidence triples.In comparison to the other datasets, TriviaQA has relatively considerable syntactic and lexical variability between questions and corresponding answer-evidence sentences and requires more cross sentence reasoning to find answers.

2.4 Free-Form

The last category allows the answer to be any free-text form. Because the standard answer is human-generated, there isn’t a consensus yet on the evaluation metric. A common way is to use standard evaluation metrics in natural language generation tasks such as BLEU (the metric of machine translate), GOUGE (the metric of automatic abstract).Main datasets in this style are listed below.


[?] proposed NarrativeQA, a more difficult data set, aimed to increase the difficulty of the question, making it not easy to locate the answer. The dataset contains 1567 complete stories from books and screenplays. Questions and answers are written by humans and are mostly more complex styles such as ”when/where/who/why.”


[?] proposed DuReader, a Chinese reading comprehension dataset, which collected more than one mil-lion documents and more than 200,000 different types of questions in Baidu Zhidao and Baidu Search, including Yes/No questions, descriptive questions, etc(up to 2018.6).

Ms Macro:

[?] introduced a large- scale machine reading comprehension dataset, which comprises of 8,841,823 pas-sages and 1,010,916 anonymized questions sampled from Bing’s search query logs(up to 2018.10).A question in the MS MARCO dataset may have multiple answers or no answers at all.

In addition, there have been some new forms of datasets in recent years, such as CoQA: a conversational QA dataset proposed by [?]. Each conversation has an average of 15 rounds. The answer to each round of dialogue is based on the questions in the article and the previous rounds of dialogue, and the answer is open-form.

Figure 1: A generic architecture of MRC model

3 Models

3.1 Generic Architecture

We have extracted a generic architecture of existing MRC models. This architecture, which we can see as an extension of the Encoder-Decoder model proposed by [?], is shown in Figure 1. The original passage and the question are represented by the word vector through the Embed layer and the word vector is encoded by Encoder Layer to aggregate the context semantic information. The match layer is used to capture the interaction between passage and question, and a query-aware representation of the original passage is then obtained. The common idea here is attention mechanism, that is, the machine is better to pay more attention to the part of the passage that is most related to the question, which is the critical improvement part of most models. The Answer Layer is based on the passage representation from the match layer and predicts the answer, while the way of implementation varies according to the type of question.

Embed Layer

Word embedding, a pre-training process for NLP tasks, can project one-hot word representation into low dimensional space. In one-hot feature space, the dimension is as large as the size of vocabulary size (), which always incurs the dimension curse. While in the embedded space, the dimension () can be arbitrary (usually 256, 512). In addition, word embedding can capture some semantic features between words, such as similarity, analogy.

What the word embedding actually does is to find an embedded matrix, whose dimension is .There are three classic methods to do that. CBOW predicts a word using the context word of the passage, while Skim-Gram predicts context words of the passage using a given word. Both of them (word2vec) were proposed by [?].In addition, [?] proposed Glove, using global word frequency statistics to construct the embedded matrix. Some models not only use the word embedding, but also combine the character embedding, which is acquired by the character-level Convolution Neural Network (CNN) or Recurrent Neural Network (RNN), to process the corpus.

However, in a particular context, a word has only one semantic. The existence of polysemous words makes the word embedding a mixture of multiple semantics, which limits the representation performance of word embedding. To tackle the puzzle, [?] proposed ELMo (Embedding from Language Models), a dynamic word representation model. In the training process, ELMO not only learns a word embedding but also gains a two-layer, bidirectional LSTM structure. We can regard them as three semantically different word representations. When a word appears in the context, the weights of its three word representations are dynamically adjusted, and the word embedding generates by weighted summation.

Using pre-trained word embedding can improve the efficiency of training, some models use Glove ([?]) or word2vec ([?]) in embed layer. ELMo is also heavily used in the MRC task ([?]).

Encoder Layer

After the Embed layer, the embedded token in the passage and the question is fed to encoder Layer for modeling the sentence and passage by RNN, which is widely used in NLP because of its excellent performance in time and sequence feature tasks. Uni-directional RNN can capture the context information on the left side of a sequence, while bidirectional RNN can aggregate the context information on both the left and right sides of a sequence. However, RNN cannot capture long-distance dependencies, which is essential to comprehend the passage. Hence, two variants of RNN, Long Short Memory Network (LSTM) and Gated Recurrent Unit (GRU), with their bidirectional versions were proposed. The early models use RNN, while the recent models tend to use LSTM ([?]) or GRU ([?]).

Match Layer

At present, most models choose to use the attention mechanism (or its variants) in the match layer, which is very suitable to mine the association between passage and question. The essence of the attention mechanism can be described as a mapping process of the query to a series of key-value pairs (in NLP tasks, key = value). This process first use a function to measure the relevance between query and ıt key to gain the weight of value (Equation 1 gives four common calculation methods), then normalize these weights using a softmax function(Equation 2). The values and the corresponding weights are summed to get the final attention (Equation 3).


Answer Layer

Before entering the answer layer, the model has obtained the query-aware representation of passage. On this basis, the answer layer has different implementation methods to solve various tasks.

For the cloze-style question, the Reader calculate the similarity between every candidate answer with the passage, and choose the largest one. The conventional way of solving multi-choice question is almost the same as what cloze-style do: each option is encoded with LSTM or GRU, then compared with the passage, and choose the most likely answer.

For the span-prediction question, Pointer Networks [?] are quite suitable to determine the answer. [?] proposed two ways of using Pointer Net. One is the sequence model, which predicts all the words in the sentence, and the other is the boundary model, which only predicts the start and end points of the answer. To rank the candidate spans, feed-forward neural networks are always used to score the similarity.

Free-form question is more difficult to handle than the previous ones because it does not limit the form of the answer. The usual processing method is the same as the sequence-to-sequence model, and LSTM is used to decode the passage of query-aware to get the final answer.

3.2 Classic Model

Many classic models are based on the previous architecture. Specially, [?] proposed two neural MRC models: AR (Attentive Reader) and IR (Impatient Reader), which are instructive for later researches. For example, using BiLSTM to encode the corpus and using attention mechanism in match layer, have been choices of most models. AR and IR also represent the two basic structures of the existing model and we will introduce them below.

Two Basic Structures

AR encodes the query using bidirectional LSTM and concatenates the results to get the representation of the query (Formula 4), which contains both forward semantics and reverse semantics. For the document, the composite output for each token at position t is Formula 5. Further, the representation of the document is formed by a weighted sum of these output vectors. The model is completed with the definition of the joint document and query embedding via a nonlinear combination (Formula 6):


The improvement of IR lies in the encoding of the query. Similar to document, the model computes a query representation (Formula 7) at each token i of the query. The result is an attention mechanism that allows the model to recurrently accumulate information from the document as it sees each query token, ultimately outputting a final joint document query representation for the answer prediction(Formula 8).


According to Figure 2, we can see that the fundamental difference between AR and IR is that the representation of the query is based on the single token or the whole sentence.

Figure 2: The left is the process of AR and the right is IR. In AR, the match layer, or the interaction(fusion) layer, is a vector. While in IR, a matrix W is obtained by the standard word-by-word attention mechanism, where indicates the degree of matching between the i-th token in the passage and the j-th token in the question.

Heuristic Improvements

Afterward, many excellent models appeared to improve or adjust the match layer of AR and IR. For example, the Attention Sum Reader (Kadlec et al. 2016), uses dot produce between the question embedding and the contextual embedding to select the answer. The Stanford Attentive Reader [?] uses a bilinear term, instead of a tanh layer to compute the relevance between question and contextual embeddings. Both models are the variants of AR.

Intuitively, IR can introduce more detailed information of question, which is critical in MRC tasks. Indeed, the performance of the relevant models is better, such as the Match-LSTM Reader [?], the BiDAF(Bi-Directional Attention flow) [?].

The Match-LSTM Reader also uses LSTM to preprocess the embedded passage and question. Then, the i-th row of the weight matrix is multiplied by the question representation, which means that the i-th word in the passage matches the question. Through the bidirectional Match-LSTM, the reader gets the overall semantics of the whole question. Further, the pointer net is use to look for the answer.

BiDAF is the first model to combine character embedding with word embedding in the embed layer. In addition, the other improvement of BiDAF is reflected in the introduction of a bidirectional attention mechanism, that is, first calculating the alignment matrix of the original passage and the question, and then calculating the Query2Context attention and Context2Query attention based on the matrix to get the query-aware passage representation. Finally, the reader uses BiLSTM to aggregate context information.

Innovation on Attention Mechanism

The excellent performance of BiDAF prompted more exploration of attention mechanism. Several models for improving attention gradually appears and achieves better performance. For example, the new feature extractor Transformer(see 3.3) [?], DCN(Dynamic Coattention Networks) [?], the FastQA [?], the R-net and so on.

DCN uses a coattention mechanism in the Encoder part, which generates weight attention for both context and question using a bidirectional attention mechanism. Besides, to recover from the local optimum (incorrect solution), DCN proposed the Dynamic Pointing Decoder, which dynamic iteratively predicts the start and the end point. Specifically, the Highway Maxout Network scores each word in every document as the start point (or end point) based on the historical prediction information and the previous prediction situation. It stores the historical prediction information with LSTM and takes the position with the highest score in the last iteration as the start position (or end position).

FastQA uses a lightweight architecture with no interaction layer. In the Embed layer: in addition to word and char embedding as input, for each word in the context, two word-in-question features are used to replace the attention mechanism: one is the binary feature, indicating whether the word appears in the question; the other is the weighted feature, which combines the token’s term frequency in the question with the inverse term frequency in the context. These two features are also used in the question words. In the answer Layer, FastQA uses Beam-search to determine the answer range. FastQAExt (the extended version of FastQA) has an interaction layer: two lightweight information fusion strategies are used: Intra-Fusion and Inter-Fusion, the former calculates the similarity between each word in the original text, and the latter calculates the similarity between each original word and the question word. Then send it to Highway Networks to get the query-aware passage representation.

R-NET introduces a self-attention mechanism into the model. The Reader chooses Glove and char-embedding to pre-train the question and passage and uses BiRNN to get representation answer. Then, a Gated attention-based recurrent network incorporates the question into the passage to reach the question-aware passage representation. Further, self-matching attention layer is used to capture the knowledge of the context. Ultimately, R-net uses pointer networks to predict the start and end position of the answer. In addition, an attention-pooling over the question representation is used to generate the initial hidden vector for the pointer network.

There are also many other brilliant innovationas, such as the forthputting of reinforcement learning. ReasoNet does not have a fiexed approach of turns during inference, but utilizes reinforcement learning to dynamically determine whether to continue the comprehension results [?]. The Mnemonic Reader encourages to predict a more acceptable answer to address the convergence suppresion by dynamic-critical reinforcement learning. Besides, its reattention mechanism is effecive to avoid attention redundancy and attention deficency [?].

3.3 Transfer Learning Model

All models have been improving performance by attempting to mine the information contained in the context. But there is a fact that when people do reading-comprehension tasks, they often use a lot of knowledge that does not appear in the context. For example, we always default everyone knows that Roosevelt is from the United States while Churchill is from the United Kingdom. For the phenomenon, transfer learning is a heuristic thought.

Recently, the pre-training model has achieved excellent results in NLP tasks. The basic structure of such a model is to conduct semi-supervised pre-training on the large corpus and supervised fine-tuning on specific tasks (the process is called transfer learning). These models achieved the best performance in several datasets. Up to now, successful pre-training models mainly include: GPT, BERT, and ELMO (details in 3.1). Before introducing them, it is essential to know what the Transformer is.


Previously, RNN, LSTM or GRU is be seen as the best choice in sequence modeling or language models, but all these recurrent architectures have a fatal flaw, which is hard to parallel in the training process, limiting the computational efficiency. [?] proposed Transformer, based entirely on self-attention rather than RNN or Convolution.

The Encoder part consists of six identical network layers. Each layer has two sub-layers, the first is the multi-head self-attention mechanism, and the second is point-wise fully connected feed-forward network. Layer normalization and residual connection are applied to every sublayer. The decoder part has the almost same architecture with the encoder, except that it adds encoder-decoder attention layer between the self-attention layer and feed-forward neural network.

Transformer’s most notable improvement is the application of self-attention and multi-headed attention: Self-attention, also known as Intra attention, captures attention between words in a sequence. The attention calculation method here uses Scaled Dot-Product Attention in formula 9, where indicates the dimension of keys.


Multi-headed attention means that the original sequence is mapped by several mapping functions to obtain several sets of Q, K, V. Each set of Q, K, V values can be used to compute attention in parallel and the resulting values are then concatenated. The advantage of this approach is that it extends the ability of the model to focus on different locations, giving multiple representation subspaces of the attention layer.


The transformer can not only achieve parallel calculations but also capture the semantic correlation of any span. Therefore, more and more language models tend to choose it to be the feature extractor.

OpenAI proposed a generative pre-training model based on the transformer, using a multi-layer transformer decoder for language modeling. Unlike the vanilla transformer, it does not contain the encoder-decoder attention layer. This model achieved an absolute improvement of 8.9% on commonsense reasoning and optimal results on RACE dataset. However, OpenAI transformer has one glaring flaw: it only trains forward language models.

Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations by jointly conditioning on both left and right context in all layers. After training on the BooksCorpus (800M words) and Wik-ipedia (2,500M words) to learn some expressions of the language, BERT can achieve the best performance on SQuAD dataset with just one additional fine-tuned output layer. In addition to large amounts of data and calculations, BERT’s superior performance is based on Masked Language Model (MLM) and Next Sentence Prediction (NSP) during the training process. The former is to capture semantics between words and words (similar to CBOW) and the latter is to discover the logic between sentences and sentences.

4 Applications

The application of MRC has two general directions: one is Conversational Question Answering, where a machine has to understand passages and answer questions that appear in a conversation, the other is Open-domain Question Answering, where a machine needs to answer questions without passage.

The conversational question answering is directly related to dialogue, which is the core of building dialogue systems to converse with humans in natural language. Usually, the dialogue systems are designed for a particular task and set up to have short conversations,(e.g.booking a flight). When dealing with such tasks, the Reader often has to consider the sequence of questions. For example, the first question is “How many flights to Paris today?”, the second question is “Which is the earliest?”. [?] gave a comprehensive survey on how to use deep neural networks to build dialogue systems.

Unlike the previous tasks, the open-domain question answering does not specify the passage of the given question. One solution is to build large-scale structured Knowledge Bases, such as DBpedia, and then retrieve the answers in the KB. This method is very intuitive, but because external knowledge is always growing, it’s costly to construct and maintain KB. Another way is to combine Information Retrieval (IR) and MRC. This method is generally divided into two stages. First, the passages are sorted according to the degree of relevance to the question from the Knowledge Source (Wikipedia and so on), and the first n articles are taken as the base of the reading comprehension process. With these articles, the reading comprehension model mentioned earlier can be applied to get the answer.

5 Discussion

Since 2015, MRC has made very significant achievements, which have attracted the attention of many researchers.

The construction of large-scale corpus is a necessary prerequisite for promoting the development of MRC. From the CNN/DM, to the recent SQuAD dataset, and then to the MS MARCO, DuReader, each new dataset has presented new problems which current methods cannot effectively solve . New issues have prompted researchers to continually explore new models and promote the development of the field.

The match layer built on attention mechanism is essential for the model to understand the original passage and the question. The more complex bidirectional interaction layer design, which enables more semantic information to flow in the original passage, thus partially solving the long-distance dependencies in the passage, is undoubtedly better than the previous single interaction layer design. In particular, self-attention mechanism, which can capture more context information, will be the main research direction of the future.

Nevertheless, it’s still too early to be optimistic. [?] tested whether systems can answer questions about paragraphs that contain adversarially inserted sentences, which are automatically generated to distract computer systems without changing the correct answer or misleading humans. The result shows that the accuracy of sixteen published models drops from an average of 75% F1 score to 36%. [?] also employed simple heuristics to split each dataset into easy and hard subsets and the experiment showed that the existing models perform worse in the latter, where the questions need more reasoning skills. Therefore, the efficient and accurate language models (such as BERT, GPT2.0) are necessary to help the machine learn reasoning. Although BERT has a visible improvement for NLP tasks, but it requires a lot of computing resources and training time, which is not something that the average team can afford. Whether it is possible to find a more efficient method and to train an efficient language model with a small amount of resources are essential to further study.

6 Conclusion

We have surveyed the main datasets and models in the MRC task in recent years. In terms of datasets, the span-prediction and open-form styles have been more popular. As for models, the improvements are mainly made on the Match Layer. The proposed pre-training model recently has achieved State-Of-The-Art performance on several datasets and is the key research direction in the future MRC field.


  • [Chen et al., 2016] Danqi Chen, Jason Bolton, and Christopher D. Manning. A thorough examination of the cnn/daily mail reading comprehension task. In ACL, 2016.
  • [Devlin et al., 2018] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018.
  • [Guo et al., 2018] Daya Guo, Duyu Tang, Nan Duan, Ming Zhou, and Jian Yin. Dialog-to-action: Conversational question answering over a large-scale knowledge base. In NeurIPS, pages 2946–2955, 2018.
  • [He et al., 2018] Wei He, Kai Liu, Jing Liu, Yajuan Lyu, Shiqi Zhao, Xinyan Xiao, Yuan Liu, Yizhong Wang, Hua Wu, Qiaoqiao She, Xuan Liu, Tian Wu, and Haifeng Wang. Dureader: a chinese machine reading comprehension dataset from real-world applications. In ACL, pages 37–46, 2018.
  • [Hermann et al., 2015] Karl Moritz Hermann, Tomás Kociský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In NIPS, pages 1693–1701, 2015.
  • [Hill et al., 2016] Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The goldilocks principle: Reading children’s books with explicit memory representations. In ICLR, 2016.
  • [Hu et al., 2018] Minghao Hu, Yuxing Peng, Zhen Huang, Xipeng Qiu, Furu Wei, and Ming Zhou. Reinforced mnemonic reader for machine reading comprehension. In IJCAI, pages 4099–4106, 2018.
  • [Jia and Liang, 2017] Robin Jia and Percy Liang. Adversarial examples for evaluating reading comprehension systems. In EMNLP, pages 2021–2031, 2017.
  • [Joshi et al., 2017] Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In ACL, pages 1601–1611, 2017.
  • [Kociský et al., 2018] Tomás Kociský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. The narrativeqa reading comprehension challenge. TACL, 6:317–328, 2018.
  • [Lai et al., 2017] Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard H. Hovy. RACE: large-scale reading comprehension dataset from examinations. In EMNLP, pages 785–794, 2017.
  • [Mikolov et al., 2013] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013.
  • [Nguyen et al., 2016] Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. MS MARCO: A human generated machine reading comprehension dataset. In NIPS, 2016.
  • [Pan et al., 2017] Boyuan Pan, Hao Li, Zhou Zhao, Bin Cao, Deng Cai, and Xiaofei He. MEMEN: multi-layer embedding with memory networks for machine comprehension. CoRR, abs/1707.09098, 2017.
  • [Pennington et al., 2014] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532–1543, 2014.
  • [Peters et al., 2018] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In NAACL, pages 2227–2237, 2018.
  • [Radford et al., ] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving Language Understanding by Generative Pre-Training. page 12.
  • [Rajpurkar et al., 2016] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100, 000+ questions for machine comprehension of text. In EMNLP, pages 2383–2392, 2016.
  • [Rajpurkar et al., 2018] Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don’t know: Unanswerable questions for squad. In ACL, pages 784–789, 2018.
  • [Reddy et al., 2018] Siva Reddy, Danqi Chen, and Christopher D. Manning. Coqa: A conversational question answering challenge. CoRR, abs/1808.07042, 2018.
  • [Richardson et al., 2013] Matthew Richardson, Christopher J. C. Burges, and Erin Renshaw. Mctest: A challenge dataset for the open-domain machine comprehension of text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 193–203, 2013.
  • [Seo et al., 2016] Min Joon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. Bidirectional attention flow for machine comprehension. CoRR, abs/1611.01603, 2016.
  • [Shen et al., 2017] Yelong Shen, Po-Sen Huang, Jianfeng Gao, and Weizhu Chen. Reasonet: Learning to stop reading in machine comprehension. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada, August 13 - 17, 2017, pages 1047–1055, 2017.
  • [Sugawara et al., 2018] Saku Sugawara, Kentaro Inui, Satoshi Sekine, and Akiko Aizawa. What makes reading comprehension questions easier? In EMNLP, pages 4208–4219, 2018.
  • [Sutskever et al., 2014] Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural networks. In NIPS, pages 3104–3112, 2014.
  • [Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 6000–6010, 2017.
  • [Vinyals et al., 2015] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NIPS, pages 2692–2700, 2015.
  • [Wang and Jiang, 2016] Shuohang Wang and Jing Jiang. Machine comprehension using match-lstm and answer pointer. CoRR, abs/1608.07905, 2016.
  • [Wang et al., 2017] Wenhui Wang, Nan Yang, Furu Wei, Baobao Chang, and Ming Zhou. Gated self-matching networks for reading comprehension and question answering. In ACL, pages 189–198, 2017.
  • [Weissenborn et al., 2017] Dirk Weissenborn, Georg Wiese, and Laura Seiffe. Making neural QA as simple as possible but not simpler. In CoNLL, pages 271–280, 2017.
  • [Xiong et al., 2016] Caiming Xiong, Victor Zhong, and Richard Socher. Dynamic coattention networks for question answering. CoRR, abs/1611.01604, 2016.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description