ReasoNet: Learning to Stop Reading in Machine Comprehension

Abstract

Teaching a computer to read a document and answer general questions pertaining to the document is a challenging yet unsolved problem. In this paper, we describe a novel neural network architecture called the Reasoning Network (ReasoNet) for machine comprehension tasks. ReasoNets make use of multiple turns to effectively exploit and then reason over the relation among queries, documents, and answers. Different from previous approaches using a fixed number of turns during inference, ReasoNets introduce a termination state to relax this constraint on the reasoning depth. With the use of reinforcement learning, ReasoNets can dynamically determine whether to continue the comprehension process after digesting intermediate results, or to terminate reading when it concludes that existing information is adequate to produce an answer. ReasoNets have achieved state-of-the-art performance in machine comprehension datasets, including unstructured CNN and Daily Mail datasets, and a structured Graph Reachability dataset.

Introduction

Teaching machines to read, process, and comprehend natural language documents is a coveted goal for artificial intelligence [ 2, 17, 6 ]. Genuine reading comprehension is extremely challenging, since effective comprehension involves thorough understanding of documents and performing sophisticated inference. Toward solving this machine reading comprehension problem, in recent years, several work has collected various datasets, in the form of question, passage, and answer, to test machine on answering a question based on the provided passage [ 17, 6, 7, 16 ]. Some large-scale cloze-style datasets [6, 7] have gained significant attention along with powerful deep learning models.

Recent approaches on cloze-style datasets can be separated into two categories: single-turn and multiturn reasoning. Single turn reasoning models utilize attention mechanisms [1] with deep learning models to emphasize specific parts of the document which are relevant to the query. These attention models subsequently calculate the relevance between a query and the corresponding weighted representations of document subunits (e.g. sentences or words) to score target candidates [ 7, 6, 8 ]. However, considering the sophistication of the problem, after a single-turn comprehension, readers often revisit some specific passage or the question to grasp a better understanding of the problem. With this motivation, recent advances in reading comprehension have made use of multiple turns to infer the relation between query, document and answer [ 7, 5, 22, 18 ]. By repeatedly processing the document and question after digesting intermediate information, multi-turn reasoning can generally produce a better answer and all existing work has demonstrated its superior performance consistently. Existing multi-turn models have a fixed number of hops or iterations in their inference, i.e., with predetermined reasoning depth, without regard to the complexity of each individual query or document. However, when a human reads a document with a question in mind, we often decide whether we want to stop reading if we believe the observed information is adequate already to answer the question, or continue reading after digesting intermediate information until we can answer the question with confidence. This behavior generally varies from document to document, or question to question Copyright © 2016 for this paper by its authors. Copying permitted for private and academic purposes. because it is related to the sophistication of the document or the difficulty of the question. Meanwhile, the analysis in [3] also illustrates the huge variations in the difficulty level with respect to questions in the CNN/Daily Mail datasets [6]. For a significant part of the datasets, this analysis shows that the problem cannot be solved without appropriate reasoning on both its query and document.

With this motivation, we propose a novel neural network architecture called Reasoning Network (ReasoNet). ReasoNets try to mimic the inference process of human readers. With a question in mind, ReasoNets read a document repeatedly, each time focusing on different parts of the document until a satisfying answer is found or formed. This reminds us of a Chinese proverb: “The meaning of a book will become clear if you read it hundreds of times.”. Moreover, unlike previous approaches using fixed number of hops or iterations, ReasoNets introduce a termination state in the inference. This state can decide whether to continue the inference to next turn after digesting intermediate information, or to terminate the whole inference when it concludes that existing information is sufficient to yield an answer. This number of turns in the inference is dynamically modeled by both the document and the query, and can be learned automatically according to the difficulty of the problem.

One of the significant challenges ReasoNets face is how to design an efficient training method, since the termination state is discrete and not connected to the final output. This prohibits canonical back-propagation method being directly applied to train ReasoNets. Inspired by [ 24, 13 ], we tackle this challenge by proposing a novel deep reinforcement learning method called Contrastive Reward (CR) to successfully train ReasoNets. Unlike traditional reinforcement learning optimization methods using a global variable to capture rewards, CR utilizes an instance-based reward baseline assignment. Experiments show the superiority of CR in both training speed and accuracy. Finally, by accounting for a dynamic termination state during inference and applying proposed deep reinforcement learning optimization method, ReasoNets achieve the state-of-the-art results in machine comprehension datasets when the paper is first publicly available in arXiv1, including unstructured CNN and Daily Mail datasets, and a proposed structured Graph Reachability dataset.

This paper is organized as follows. In Section 2, we review and compare recent work on machine reading comprehension tasks. In Section 3, we introduce our proposed ReasoNet model architecture and training objectives. Section 4 presents the experimental setting and results on unstructured and structured machine reading comprehension tasks . 2

Related Work

Recently, with large-scale datasets available and the impressive advance of various statistical models, machine reading comprehension tasks have attracted much attention. Here we mainly focus on the related work in cloze-style datasets [6, 7]. Based on how they perform the inference, we can classify their models into two categories: single-turn and multi-turn reasoning.

Single-turn reasoning Single turn reasoning models utilize an attention mechanism to emphasis some sections of a document which are relevant to a query. This can be thought of as treating some parts unimportant while focusing on other important ones to find the most probable answer. [6] propose the attentive reader and the impatient reader models using neural networks with an attention over passages to predict candidates. [7] use attention over window-based memory, which encodes a window of words around entity candidates, by leveraging an end-to-end memory network [ 19 ]. Meanwhile, given the same entity candidate can appear multiple times in a passage, [ 8 ] propose the attention-sum reader to sum up all the attention scores for the same entity. This score captures the relevance between a query and a candidate. [3] propose using a bilinear term similarity function to calculate attention scores with pretrained word embedding. [ 22 ] propose the EpiReader which uses two neural network structures: one extracts candidates using the attention-sum reader; the other reranks candidates based on a bilinear term similarity score calculated from query and passage representations.

Multi-turn reasoning For complex passages and complex queries, human readers often revisit the given document in order to perform deeper inference after reading a document. Several recent studies try to simulate this revisit by combining the information in the query with the new information digested from previous iterations [ 7, 5, 18, 23, 12 ]. [7] use multiple hops memory network to augment the query with new information from the previous hop. Gated Attention reader [5] is an extension of Algorithm 1: Stochastic Inference in a ReasoNet the attention-sum reader with multiple iterations by pushing the query encoding into an attentionbased gate in each iteration. Iterative Alternative (IA) reader [ 18 ] produces a new query glimpse and document glimpse in each iteration and utilizes them alternatively in the next iteration. [4] further propose to extend the query-specific attention to both query-to-document attention and document-toquery attention, which is built from the intermediate results in the query-specific attention. By reading documents and enriching the query in an iterative fashion, multi-turn reasoning has demonstrated their superior performance consistently.

Our proposed approach explores the idea of using both attention-sum to aggregate candidate attention scores and multiple turns to attain a better reasoning capability. Unlike previous approaches using fixed number of hops or iterations, motivated by [ 14, 13 ], we propose a termination module in the inference. The termination module can decide whether to continue to infer the next turn after digesting intermediate information, or to terminate the whole inference process when it concludes existing information is sufficient to yield an answer. The number of turns in the inference is dynamically modeled by both a document and a query, and is generally related to the complexity of the document and the query. 3

Reasoning Networks

ReasoNets are devised to mimic the inference process of human readers. ReasoNets read a document repeatedly, with attention on different parts each time until a satisfying answer is found. As shown in Figure 1, a ReasoNet is composed of the following components: Internal State: The internal state is denoted as S which is a vector representation of the question state. Typically, the initial state s1 is the last-word vector representation of query by an RNN. The t-th time step of the internal state is represented by st. The sequence of internal state is modeled by an RNN: st+1 = RNN(st; xt; s); Memory: The external memory is denoted as M . It is a list of word vectors, M = fmigi=1::D, where mi is a fixed dimensional vector. In machine comprehensive tasks, mi is the vector representation of each word in the doc by a bidirectional-RNN.

Attention: Attention vector xt is generated based on the current internal state st and the external memory M : xt = fatt(st; M ; x); Termination Gate: Termination gate generates a stochastic random variable according to the current internal state; tt p( jftg(st; tg))). tt is a binary random variable. If tt is true, the ReasoNet stops, and the answer module executes at time step t; otherwise the ReasoNet generates an attention vector xt+1, and feed into the state network to update the next internal state st+1.

Answer: The action of answer module is triggered when the termination gate variable is true: at p( jfa(st; a)).

In Algorithm 1, we describe the stochastic inference process of a ReasoNet. The process can be considered as a Partially Observable Markov Decision Process (POMDP) [ 9 ] in the reinforcement learning (RL) literature. The state sequence s1:T is hidden and dynamic, controlled by an RNN sequence model. The ReasoNet performs an answer action aT at the T -th step, which implies that the termination gate variables t1:T = (t1 = 0; t2 = 0; :::; tT 1 = 0; tT = 1). The ReasoNet learns a stochastic policy ((tt; at)jst; ) with parameters to get a distribution over termination actions, to RNN states

S1

Memory M

St fa(θa) at Query

Attention

Attention fatt(θx)

Xt fatt(θx)

Xt+1 ftg(θtg) False True

Tt Termination

St+1 fa(θa) ftg(θtg) False True

Tt+1 Termination

St+2 Answer

Answer

at+1 continue reading or to stop, and over answer actions if the model decides to stop at the current step. The termination step T varies from instance to instance.

The parameters of the ReasoNet are given by the parameters of the embedding matrices W , attention network x, the state RNN network s, the answer action network a, and the termination gate network tg. The parameters = fW; x; s; a; tgg are trained by maximizing the total expect reward. The expected reward for an instance is defined as:

J ( ) = E (t1:T ;aT ; ) " T

X rt

# t=1 The reward can only be received at the final termination step when an answer action aT is performed. We define rT = 1 if tT = 1 and the answer is correct, and rT = 0 otherwise. The rewards on intermediate steps are zeros, frt = 0gt=1:::T 1. J can be maximized by directly applying gradient based optimization methods. The gradient of J is given by:

r J ( ) = E (t1:T ;aT ; ) [r log (t1:T ; aT ; )rT ] We apply the REINFORCE algorithm [24] to compute r J ( ): E (t1:T ;aT ; ) [r log (t1:T ; aT ; )rT ] = where Ay is all the possible episodes, T; t1:T ; aT and rT are the termination step, termination action, answer action, and reward, respectively, for the (t1:T , aT ) episode. bT is called the reward baseline in the RL literature to lower variance [ 21 ]. It is common to select bT = E [rT ] [ 20 ], and can be updated via an online moving average approach : bT = bT + (1 )rT .

However, we empirically find that above approach leads to slow convergence in training ReasoNets. Intuitively, the average baselines fbT ; T = 1::Tmaxg are global variables independent of instances. It is hard for these baselines to capture the dynamic termination behavior of ReasoNets. In other words, ReasoNets may stop at different time steps for different instances. The adoption of a global variable without considering the dynamic variance in each instance is inappropriate. To resolve this weakness in traditional methods and account for the dynamic characteristic of ReasoNets, we propose an instance-based baseline method called “Contrastive Reward” (CR) to calculate r J ( ). The basic idea of CR is to utilize an instance-based baseline assignment. We will elaborate its implementation details in Section 3.1. Empirical results show that the proposed reward schema has produced better results compared to the baseline approach. 3.1

Training Details

In the machine reading comprehension tasks, a training dataset can be simplified as a collection of triplets of query q, passage p, and answer a. Say hqn; pn; ani is the n-th training instance. The first step is to extract memory M from pn by mapping each symbolic in the passage to a contextual representation given by the concatenation of forward and backward RNN hidden states, i.e., mk = [p!nk; pnjpnj k+1], and extract initial state s1 from qn by assigning s1 = [!qnjqnj; qn1]. Given M and s1 for the n-th training instance, a ReasoNet executes jAyj episodes, where all possible episodes Ay can be enumerated by setting a maximum step. Each episode generates actions and a reward from the last step: h(t1:T ; aT ), rT i(t1:T ;aT )2Ay .

Therefore, the gradient of J can be rewritten as: r J ( ) =

X where the baseline b = P(t1:T ;aT )2Ay (t1:T ; aT ; )rT is the average reward on the jAyj episodes for the n-th training instance. It allows different baselines for different training instances. This can be beneficial since the complexity of training instances varies significantly. Since the sum of the proposed rewards over jAyj episodes is zero, P(t1:T ;aT )2Ay (t1:T ; aT ; )(rT b) = 0, we call it Contrastive Reward in this work. In experiments, we empirically find using ( rbT 1) in replace of (rT b) can lead to a faster convergence. Therefore, we adopt this approach to train ReasoNets in the experiments. 4 4.1

Experiments

CNN and Daily Mail Datasets

We evaluate the performance of ReasoNets on CNN and Daily Mail datasets.2 The detailed settings of the ReasoNet model are as follows.

Vocab Size: For training our ReasoNet, we keep the most frequent jV j = 101k words (not including 584 entities and 1 placeholder marker) in the CNN dataset, and jV j = 151k words (not including 530 entities and 1 placeholder marker) in the Daily Mail dataset.

Embedding Layer: We choose word embedding size d = 300, and use the 300 dimensional pretrained Glove word embeddings [ 15 ] for initialization. We also apply dropout with probability 0:2 to the embedding layer.

Bi-GRU Encoder: We apply bi-directional GRU for encoding query and passage into vector representations. We set the number of hidden units to be 256 and 384 for the CNN and Daily Mail datasets, respectively. The recurrent weights of GRUs are initialized with random orthogonal matrices. The other weights in GRU cell are initialized from a uniform distribution between 0:01 and 0:01. We use a shared GRU model for both query and passage.

Memory and Attention: The memory of the ReasoNet on CNN and Daily Mail dataset is composed of query memory and passage memory. M = (M query; M doc), where M query and M doc are extracted from query bidirectional-GRU encoder and passage bidirectional-GRU encoder respectively. We choose projected cosine similarity function as the attention module. The attention score atd;oic on memory midoc given the state st is computed as follows: atd;oic = softmaxi=1;:::;jMdocj cos(W1docmidoc; W2docst), where is set to 10. W1doc and W2doc are weight vectors associated with midoc and st, respectively, and are joint trained in the ReasoNet. Thus, attention vector on passage is given by xdoc = PjMj at;imidoc. The final attention vector is the t i concatenation of the query attention vector and the passage attention vector xt = (xtquery; xtdoc). The attention module is parameterized by x = (W1query; W2query; W1doc; W2doc); Internal State Controller: We choose GRU model as the internal state controller. The number of hidden units in the GRU state controller is 256 for CNN and 384 for Daily Mail. The initial state 2The CNN and Daily Mail datasets are available at https://github.com/deepmind/rc-data Query: passenger @placeholder1, 36 , died at the scene

11 Passage: ( @entity0 ) what was supposed to be a fantasy sports car ride at @entity3 turned deadly when a @entity4 crashed into a guardrail 1. the crash took place sunday at the @entity8 , which3bills itself as a chance to drive your dorfe@a menctaitry1o5n ,a@raecnettirtayc1k6.,tdhieed@aetntthiteys4c'esnpea,ss@enegnetirty, 1336s-ayidea.rth-eolddr@iveernotiftyth1e41 1

233

Step 1

Termination Module: We adopt a logistical regression to model the termination variable at each time step : ftg(st; tg) = sigmoid(Wtgst + btg); tg = (Wtg; btg) Answer Module: We apply a linear projection from GRU outputs and make predictions on the entity candidates. Following the settings in AS Reader [ 8 ], we sum up scores from the same candidate and make a prediction. Thus, AS Reader can be viewed as a special case of ReasoNets with Tmax = 1. Other Details: The maximum reasoning step, Tmax is set to 5 in experiments on both CNN and Daily Mail datasets. We use ADAM optimizer [ 10 ] for parameter optimization with an initial learning rate of 0:0005, 1 = 0:9 and 2 = 0:999; The absolute value of gradient on each parameter is clipped within 0:001. The batch size is 64 for both CNN and Daily Mail datasets. For each batch of the CNN and Daily Mail datasets we randomly reshuffle the assignment of named entities [6]. This forces the model to treat the named entities as semantically meaningless labels. In the prediction of test cases, we randomly reshuffle named entities up to 4 times, and report the averaged answer. Models are trained on GTX TitanX 12GB. It takes 7 hours per epoch to train on the Daily Mail dataset and 3 hours per epoch to train on the CNN dataset. The models are usually converged within 6 epochs on both CNN and Daily Mail datasets. Recent analysis and results [3] on the cloze-style machine comprehension tasks have suggested some simple models without multi-turn reasoning can achieve reasonable performance. Based on these results, we construct a synthetic structured Graph Reachability dataset3 to evaluate longer range machine inference and reasoning capability, since we expect ReasoNets have the capability to handle long range relationships.

We generate two synthetic datasets: a small graph dataset and a large graph dataset. In the small graph dataset, it contains 500K small graphs, where each graph contains 9 nodes, and 16 direct edges 3The dataset is available at https://github.com/MSRDL/graph_reachability_dataset to randomly connect pairs of nodes. The large graph dataset contains 500K graphs, where each graph contains 18 nodes, and 32 random direct edges. Duplicated edges are removed. Table 2 shows the graph reachability statistics on the two datasets.

In Table 3, we show examples of a small graph and a large graph in the synthetic dataset. Both graph and query are represented by a sequence of symbols. In the experiment, we use a 100-dimensional embedding vector for each symbol, and bidirectional-LSTM with 128 and 256 cells for query and graph embedding in the small and the large graph datasets, respectively. The last states of bidirectionalLSTM on query are concatenated to be the initial internal state s1 = [ !qjqj; q 1] in the ReasoNet. Another bidirectional-LSTM on graph description maps each symbol gi to a contextual representation given by the concatenation of forward and backward LSTM hidden states mi = [ !gi; g jgj i+1]. The final answer is either “Yes” or “No” and hence logistical regression is used as the answer module: at = (Wast + ba); a = (Wa; ba). We apply another logistical regression as the termination gate module: tt = (Wtgst + btg). The maximum reasoning step Tmax is set to 15 and 25 for the small graph and large graph dataset, respectively.

We study the effectiveness of the termination gate in ReasoNets. We denote “ReasoNet” as a standard ReasoNet with termination gate, as described in Section 3.1. If we remove the termination gate, and just simply use the last state answer action as the final answer, say a^ = aTmax (Tmax is the maximum reasoning step), denoted as “ReasoNet-Last ”. To study the effectiveness of multi-turn reasoning, we choose “ReasoNet-Tmax = 2”, which only has single-turn reasoning, as a baseline. In Table 4, we report the performance of ReasoNet, ReasoNet-Last and ReasoNet-Tmax = 2 models on the Graph Reachability dataset. The ReasoNet-Last model performs well on the small graph dataset, and it obtains 100% accuracy. However, the ReasoNet-Last model fails to learn on the large graph dataset, as the task becomes much more challenging. Meanwhile, the ReasoNet model converges faster than the ReasoNet-Last model. The ReasoNet model converges in 20 epochs in the small graph dataset, and 40 epochs in the large graph dataset, while the ReasoNet-Last model converges around 40 epochs in the small graph dataset, and 70 epochs in the large graph dataset. The results suggest that the termination gate variable in the ReasoNet is helpful when training with sophisticated examples, and makes models converge faster. Both the ReasoNet and ReasoNet-Last models perform better than the ReasoNet-Tmax = 2 model, which demonstrates the importance of multi-turn reasoning. To further understand the inference process in ReasoNets, we present two examples of the graph reachability results in appendix A.

Acknowledgments

In this paper, we propose ReasoNets that dynamically decide whether to continue or to terminate the inference process in machine comprehension tasks. Using reinforcement learning with the proposed contractive reward, our proposed model achieves the start-of-the-art results in machine comprehension datasets, including unstructured CNN and Daily Mail datasets, and a proposed structured Graph Reachability dataset. For future work, ReasoNets can be generalized to other tasks that requires reasoning capability, such as question answering and knowledge graph inference. We thank Ming-Wei Chang, Li Deng, Lihong Li, and Xiaodong Liu for their thoughtful feedback and discussions. [1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In Proceedings of the International Conference on Learning Representations, 2015. [2] Léon Bottou. From machine learning to machine reasoning. Machine Learning, 94(2):133–149, 2014. [3] Danqi Chen, Jason Bolton, and Christopher D Manning. A thorough examination of the CNN / Daily Mail reading comprehension task. In ACL, 2016. [4] Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. Attention-over-attention neural networks for reading comprehension. CoRR, abs/1607.04423, 2016. [5] Bhuwan Dhingra, Hanxiao Liu, William W. Cohen, and Ruslan Salakhutdinov. Gated-attention readers for text comprehension. CoRR, abs/1606.01549, 2016. [6] Karm Moritz Hermann, Tomáš Kocˇiský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems, pp. 1693–1701, 2015. [7] Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. The Goldilocks principle: Reading children’s books with explicit memory representations. In Proceedings of the International Conference on Learning Representations, 2016. [23] Dirk Weissenborn.

abs/1607.03316, 2016.

Separating answers from queries for neural reading comprehension.

CoRR, [24] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229–256, 1992.

A

Examples of Graph Reachability Results in ReasoNets

Step 9

Step 10

Termination Probability Prediction 1.00E-06 0.172 1.00E-06 0.625 1.00E-06 0.752 1.00E-06 0.202 1.00E-06 0.065 1.00E-06 0.041 2.30E-06 0.137 0.0017 0.136 0.49 0.761 0.99 0.927 1 -> 16 # 1 -> 12 # 1 -> 14 # 1 -> 7 # 2 3 > 17 # 3 -> 1 #2 4 -> 0 # 4 -> 1 #1 4 -> 12 1 # 4 -> 6 #26 -> 0 # 6 -> 3 # 6 -> 7 # 8 -> 2 # 8 -> 4 # 8 -> 13 # 8 -> 14 # 9 -> 16 # 10 -> 0 # 10 -> 6 # 11 -> 10 # 11 -> 2 # 12 -> 2 # 13 -> 2 # 13 -> 6 #314 -> 2 # 14 -> 7 # 16 -> 13 # 16 -> 14 # 17 -> 0 # 17 -> 13 #

Step 2

Termination Probability 1.40E-05 0.999

Prediction 4.49E-04 1.40E-05

References

[8] Rudolf Kadlec , Martin Schmid, Ondrej Bajgar, and Jan Kleindienst . Text understanding with the attention sum reader network . arXiv:1603.01547v1 [cs.CL] , 2016 .

[9] Leslie Pack Kaelbling , Michael L. Littman , and Anthony R. Cassandra . Planning and acting in partially observable stochastic domains . Artificial Intelligence , 101 : 99 - 134 , 1998 .

[10] Diederik P. Kingma and Jimmy Ba . Adam: A method for stochastic optimization . In Proceedings of the International Conference on Learning Representations , 2015 .

[11] Sosuke Kobayashi , Ran Tian, Naoaki Okazaki, and Kentaro Inui . Dynamic entity representation with max-pooling improves machine reading . In Proceedings of the North American Chapter of the Association for Computational Linguistics and Human Language Technologies (NAACL-HLT) , 2016 .

[12] Ankit Kumar , Ozan Irsoy, Peter Ondruska , Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor Zhong, Romain Paulus, and Richard Socher. Ask me anything: Dynamic memory networks for natural language processing . In Proceedings of the International Conference on Machine Learning , 2016 .

[13] Volodymyr Mnih , Nicolas Heess, Alex Graves , et al. Recurrent models of visual attention . In Advances in Neural Information Processing Systems , pp. 2204 - 2212 , 2014 .

[14] Rodrigo Nogueira and Kyunghyun Cho . Webnav: A new large-scale task for natural language based sequential decision making . In Advances in Neural Information Processing Systems , 2016 .

[15] Jeffrey Pennington , Richard Socher, and Christopher D. Manning . Glove: Global vectors for word representation . In EMNLP , 2014 .

[16] Pranav Rajpurkar , Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQuAD: 100 , 000+ questions for machine comprehension of text . In EMNLP , 2016 .

[17] Matthew Richardson , Christopher JC Burges, and Erin Renshaw. MCTest: A challenge dataset for the open-domain machine comprehension of text . In EMNLP , 2013 .

[18] Alessandro Sordoni , Phillip Bachman, and Yoshua Bengio . Iterative alternating neural attention for machine reading . CoRR, abs/1606.02245, 2016 .

[19] Sainbayar Sukhbaatar , Jason Weston, Rob Fergus , et al. End-to-end memory networks . In Advances in neural information processing systems , pp. 2440 - 2448 , 2015 .

[20] Richard S. Sutton , David McAllester , Satinder Singh , and Yishay Mansour . Policy gradient methods for reinforcement learning with function approximation . In Advances in Neural Information Processing Systems , 1999 .

[21] Richard Stuart Sutton. Temporal Credit Assignment in Reinforcement Learning . PhD thesis , 1984 .

[22] Adam Trischler , Zheng Ye, Xingdi Yuan, and Kaheer Suleman . Natural language comprehension with the EpiReader . In EMNLP, 2016 .