A Review on Neural Network Question Answering Systems

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.8, No.2, March 2017
DOI : 10.5121/ijaia.2017.8206 59
A REVIEW ON NEURAL NETWORK QUESTION
ANSWERING SYSTEMS
Lorena Kodra and Elinda Kajo Meçe
Department of Computer Engineering, Polytechnic University of Tirana, Albania
ABSTRACT
In recent years neural networks (NN) are being used increasingly on Question Answering (QA) systems and
they seem to be successful in addressing different issues and challenges that these systems exhibit. This
paper presents a review to summarize the state of the art in question answering systems implemented using
neural networks. It identifies the main research topics and considers the most relevant research challenges.
Furthermore, it analyzes contributions, limitations, evaluation techniques, and directions proposed for
future research.
KEYWORDS
Information Retrieval, Question Answering Systems, Neural Networks
1. INTRODUCTION
In recent years neural networks are being used increasingly on QA systems and have been proven
effective in achieving good results [45]. In this paper we present a study of the latest research
being done on question answering systems based on neural networks. Our aim is to give an
answer to questions like: What are the research directions of Neural Network Question
Answering? What are the biggest challenges that researchers face in this area? What are the
solutions proposed? What criteria and datasets are used for evaluations? Which are the most
promising directions for future research? Our work can help researchers gain an insight on the
state of the art and future research directions of the research being done in the area of neural
network question answering systems. The rest of this paper is organized as follows: In section 2
we describe the methodology used in our study and define objectives and research questions.
Section 3 outlines the results of the review organized according to each research question defined
in Section 2. Section 4 discusses the results of our study. Section 5 contains conclusions and
future work.
2. METHODOLOGY
2.1. RESEARCH QUESTIONS
As a primary step in the investigation, retrieval and selection of the most accurate publications for
our review we have defined the following research questions:
RQ1: What are the research directions of neural network question answering?
RQ2: What are the challenges faced by researchers in this area?
RQ3: What solutions have been proposed?
RQ4: What criteria and datasets are used for evaluations?
RQ5: Which directions are the most promising for future research?

60
2.2. SEARCH KEYWORDS AND SOURCE SELECTION
In order to extract the most relevant information for our review we used the following keywords
and their combination and synonyms. The search string below was used as a query to search for
publications in different online digital libraries: (“question answering” OR “question answer” OR
“question answering system” OR “question answering systems”).
The search string contained only these keywords because not all the publications had on their title
the keywords “neural” or “neural network”. After retrieving the full publication another search
was made on the body of the publication with the keyword “neural”. If the search was successful
and neural networks were identified as an implementation mechanism the publication was
included in the list of publications.
Furthermore, we selected four scientific digital libraries that represent primary sources for
computer science research publications. We did not include online archives Google Scholar and
ArXiv because they index content from existing digital libraries. The sources are shown in Table
1.
Table 1. Selected sources for the search process.
Source URL
IEEExplore http://ieeexplore.ieee.org
ScienceDirect http://www.sciencedirect.com/
ACM Digital Library http://dl.acm.org
Springer Link http://link.springer.com
2.3. INCLUSION AND EXCLUSION CRITERIA
Table 2 lists the inclusion and exclusion criteria that we used to collect publications.
Table 2. Inclusion and exclusion criteria.
Inclusion criteria Exclusion criteria
Relevant to the topic of our review Review papers
Papers that have been published in the
last three years (2014 - 2016)
Reports
Papers from conferences and journals Papers reporting only abstract or presentation
slides because of the lack of information
describing the corresponding study.
Short and workshop papers that fulfil the
study criteria
Grey literature. PhD thesis, unpublished
studies and technical reports would not add
much more information against journal and
conference papers.
English language
We did not retrieve review papers and reports because our aim is to analyze the existing
implementations and latest developments of Question Answering Systems (QAS) based on neural
networks.

61
3. RESULTS
This section presents the relevant information found in the selected publications in order to
answer the proposed research questions. A further discussion and analysis of these results are
addressed in Section 4.
3.1. RESEARCH DIRECTIONS
We retrieved a total of 46 publications to include in our review. We identified three research
directions as shown in Table 3.
Table 1. Distribution of publications according to research directions
Research direction Number of publications
Knowledge Based Question Answering 18
Visual Question Answering 15
Community Question Answering 13
Knowledge Base Question Answering (KBQA) is an area that deals with answering questions
over a Knowledge Base. In such systems the information is organized in a form of structured
knowledge base (KB) where the data are linked by semantics. The information in the knowledge
base is organized in the form of triples composed of subject, predicate, and object. An example of
a triple is (Mount Everest, height, 8.848 m). The predicate is a characteristic of the subject, while
the object is the value of that characteristic. In question answering over a knowledge base, the
question is on subject and predicate, while the answer is from object.
Visual Question Answering (VQA) is an area that deals with answering questions about an
image. The information source is either a hybrid between text and image or entirely image-based.
Community Question Answering (CQA) Systems are modern QA systems where users rely on
expertize from the community to get an answer for their question. Examples of CQAs include
Yahoo! Answers1
and Quora2
.
For each of the research directions we identified the main research topics as shown in tables 4, 5
and 6.
Table 2. Distribution of publications according to research topics in Knowledge Based Question Answering
Research topic Number of publications
Entity linking 6
Answer selection 5
Question-answer matching 4
Question answering over paragraph questions 2
Learning to translate in multilingual QAS 1
Natural language answering 1
1
https://answers.yahoo.com/
2
https://www.quora.com/

62
Table 3. Distribution of publications according to research topics in Visual Question Answering
Answering free language questions about an image 7
Identifying clues in image and/or question to generate
answer
4
Detecting relevance of question to image. 1
Visual analogy 1
Situated visual question answering 1
Answering from a limited domain of questions 1
Table 4. Distribution of publications according to research topics in Community Question Answering
Retrieving similar questions 5
Question-answer matching 3
Expert finding 2
Answer ranking 2
Answer selection 2
3.2. RESEARCH CHALLENGES
In order to address RQ2 we summarize the main research challenges involved in the selected
publications. Tables 7 and 8 present these challenges for KBQAS and CQAS respectively. As
regards VQAs, we consider challenges to be the same with research topics.
Table 5. Distribution of publications according to research challenges in Knowledge Based Question
Answering
Research challenges Number of publications
Lexical gap between natural language and
structured semantics of the knowledge base
10
Entity identification and linking 5
Multiple entity question 1
Natural language answer generation 1
Questions involving multiple entities 1
Passage question answering 1
Table 6. Distribution of publications according to research challenges in Community Question Answering
Research challenges Number of publications
Lexical gap between questions 5
Lexical gap between questions and answers 5
Sparsity of CQA data 2
Deviation from question 2
3.2.1. RESEARCH CHALLENGES IN KNOWLEDGE BASE QUESTION ANSWERING SYSTEMS
We identified the following challenges for KBQA systems:
• Lexical gap between natural language and structured semantics of the knowledge
base: We identified it as the most frequent problem. It concerns differences in sentence

63
representations between the unstructured natural language question and the structured
knowledge base. It also concerns the many ways of expressing knowledge in a
knowledge base.
• Entity identification and linking: This was another prominent challenge. It concerns the
ability of the system to correctly identify the subject entity in question and link it to a
triple in the knowledge base.
• Natural language answer generation: It concerns the ability of a system to generate the
answer in a natural language form and not simply give the information the user asked for.
For example for the question “How many days are there in a leap year?” the system may
answer “There are 366 days in a leap year.” instead of just “366”. This feature is desirable
in making the system seem “smarter”.
• Questions involving multiple entities: It concerns the ability of the system to identify
and reason over multiple subject entities in question and link it to the relevant triple in the
knowledge base.
• Passage question answering: This is a challenge on non-factoid question answering
where the answer is in the form of a paragraph. Question-answer matching is a
challenging task as it requires effective representations that capture the complex semantic
relations between questions and answers.
3.2.2. RESEARCH CHALLENGES IN VISUAL QUESTION ANSWERING
We identified the following challenges for VQA systems:
• Answering free language questions about an image: We identified it as the most
frequent challenge. It concerns the ability of the system to answer questions posed in a
free language (i.e. without any lexical or semantic limitation). It is a complex research
problem since the system is required to answer different types of questions ranging from
simple questions like “What is the color of the flower?” to “Does it appear to be rainy?”
which requires extra knowledge or commonsense reasoning.
• Identifying clues in image and/or question to generate answer: It was another
frequent challenge. It concerns the ability of the system to reason about question and/or
image to identify possible words or image regions that can be helpful in answering the
question.
• Visual analogy: This topic deals with finding analogue images from a given set. The
questions take the form of “image A is to image B as image C is to what”. Answering
these questions requires discovering the mapping from image A to image B, extending
this mapping to image C and finding image D so that the relation from A to B holds for C
to D.
• Situated visual question answering: It concerns the problem of answering questions
about an environment such as an image or diagram. This problem requires jointly
reasoning about uncertain interpretations of both a question and an environment together
with background knowledge to select the correct answer.

64
• Answering from a limited domain of questions: This topic concerns systems that can
accept simple factoid question. The answer also may be in a form of a single word.
Sometimes the questions can be only from a single domain (e.g. health).
3.2.3. RESEARCH CHALLENGES IN COMMUNITY QUESTION ANSWERING
We identified the following challenges for CQA systems:
• Lexical gap between questions: It was one of the most frequent problems in the selected
publications. It concerns differences in natural language formulation of questions.
Different users ask for the same information but they formulate the question in different
ways. This results in many questions that are semantically equivalent but differ lexically.
• Lexical gap between questions and answers: This was another frequent problem.
Similar to the lexical gap between questions, sometimes question and answers can be
highly asymmetric in the information they contain. There is also a technical terminology
gap between questions and answers. Questions are posed by novices or non-experts who
use less technical terminology while experts who answer questions use the correct terms.
• Sparsity of CQA data: Another problem that we identified is the low participation of
answerers. Usually, each question is answered by a few users. As a consequence the
system cannot identify relevant answerers and most new questions cannot be routed to
appropriate users who can answer them.
• Deviation from question: It concerns the phenomenon of answer thread becoming
irrelevant to the question. Answers are given in the form of comments but sometimes
users engage in discussion and deviate from the original question.
3.3. PROPOSED SOLUTIONS
In order to address RQ3 we present and classify the solutions provided to research problems for
each research direction. In some cases there is only one publication dealing with a solution. Due
to the differences among solutions we cannot group or merge them into a single category.
3.3.1. PROPOSED SOLUTIONS IN KNOWLEDGE BASE QUESTION ANSWERING SYSTEMS
Lexical gap between natural language and structured semantics of the knowledge base:
Different solutions have been proposed to close this lexical gap. Some of them share the common
concept of using neural networks to obtain the semantic similarity by learning a similarity metric.
For example, for the task of answer selection a neural network is used to analyze question-answer
pairs from different aspects (i.e. answer path, answer context, and answer type) and then
computes a similarity score for the pair. The system uses probabilistic inference to select
candidates with the best scores [15]. Another proposed solution is a neural network that
sequentially reads the words from question and answer sentences and then outputs their relevance
scores. In order to identify the correct candidate answer sentences the system does keyword
matching between the words of the question and answer [18] [31].
For the task of question-answer matching the proposed solutions that we identified include a
neural network that uses graph-based semantic parsing to extract meaning, entities and relations
by leveraging the knowledge base more tightly when parsing input questions by linking entities in
the question to entities in the knowledge base [16], Or combination of semantic and structural
features to compute sentence similarity [42]. Another approach proposed is to look for answers

65
that are related to the question in a relevant way, according to the information need of the
question which may be determined through task-specific word embedding (i.e. different
embedding for different types of questions) [27]. A study using a factual memory network, which
learns to answer questions by extracting and reasoning over relevant facts from a knowledge base,
is also presented. Facts are visualized as a question-answer pair and are generated by matching all
n-grams of words of the question against knowledge base entities. To find the relevant answer,
the neural network calculates the similarity between the actual question and a hypothetical
question constructed from the fact [25].
For the task of entity linking, neural networks are used to measure the similarity of entity
mentions in the question with entities in the knowledge base. Triples in the knowledge base are
scored using these similarity measures and the top scoring triple is select to answer the question
[20], [30].
For multilingual systems, a proposed solution for the task of translating question into knowledge
base language or vice-versa is modeled by a neural network that learns the optimal translation of
question and/or candidate answer, based on how well it discriminates between good and bad
answers [28].
ENTITY IDENTIFICATION AND LINKING:
There were various solutions proposed to tackle this problem. One of them was to use a multistep
approach to “zooming in” a question to find more probable candidate subject mentions by using a
neural network to first identify focus segments in a question that are more probable of containing
the subject entity and then pruning the candidate knowledge base triples according to textual
similarities between question segment and knowledge base triple subject [9]. Another solution
was to use a neural network to compute character-level similarity score between the fact
candidate and entity in the question to determine the correct entity. Then use a word-level match
(pattern match) similarity score between the predicate in the fact and question words to find the
correct predicate [21]. Another proposed alternative was to use a neural network to tag whether
each question word is part of an entity, and optionally further divide into subject, object, and
predicate classes. Then converting the input question into a structured query, which can be ran
against the knowledge graph to retrieve the fact(s) that answer(s) the question [37]. In the case of
paragraph questions, the solutions proposed were to use a neural network to learn word and
phrase-level representations that combine across question sentences to capture the topic of the
paragraph and reason about entities [33], [43].
NATURAL LANGUAGE ANSWER GENERATION:
We identified only one publication trying to solve this problem and the proposed solution was a
generative neural network system that takes a sequence of words as input question and generates
another sequence of words as output answer. To provide right answers, the system is connected
with a knowledge base that contains facts. During the process of answering, the system queries
the knowledge base, retrieves a set of candidate facts and generates a correct answer to the
question using the right fact. The generated answer may contain two types of “words”: one is
common words for composing the answer (referred to as common word) and the other is
specialized words in the knowledge base denoting the answer (referred to as knowledge base
word). The generated answer is in the form of a whole sentence in natural language instead of a
simple fact [1].

66
QUESTIONS INVOLVING MULTIPLE ENTITIES:
The only publication that reported this research problem proposed a relation extractor neural
network that selects answer candidates from the knowledge base and then infers over external
unstructured sources of evidence like Wikipedia3
to choose the correct answer [11].
PASSAGE QUESTION ANSWERING:
The only publication that dealt with this research challenge proposed a non-factoid question
answering system that used neural networks to extract linguistic information from both question
and passage answer. Attention mechanisms where proposed to generate answer representations
according to the question, such that the embeddings do not overlook informative parts of the
answers [10].
Table 9 summarizes the proposed approaches mentioned above. Some systems use more than one
method.
Table 7. Distribution of publications according to methods used for Knowledge Base Question Answering
Method Number of publications
Similarity measures 8
Representation learning 6
Ranking mechanisms 3
Inference 3
3.3.2. PROPOSED SOLUTIONS IN VISUAL QUESTION ANSWERING SYSTEMS
ANSWERING FREE LANGUAGE QUESTIONS ABOUT AN IMAGE:
Various alternatives have been proposed regarding this research topic. Some of them use a
common idea such as attention mechanisms. The authors in [24] propose an attention model for
answering questions about a variety of world representations, including images and structured
knowledge bases. The model uses attention mechanisms to decompose the answering process into
modular sub-problems. It translates from questions to dynamically assembled neural networks
based on the type of question, then applies these networks to world representations (images or
knowledge bases) to produce answers. Another attentional system is proposed in [35]. This
system is composed of a stacked attention net (SAN) that uses semantic representation of a
question as a query to identify the regions of the image that are related to the answer. The SAN is
multi-layered and it queries the image multiple times to infer the answer progressively. The
authors state that the motivation behind this approach is the fact that image question answering
often requires multiple steps of reasoning. For example for the question: “What is the girl
holding? The system must first identify the girl. Then the system must identify her hands. Then it
needs to understand what the girl is holding.
The authors in [5] propose a reasoning model that is based on neural networks and is able to
update the question representation iteratively by inferring image information. The model updates
the question representation iteratively by selecting image regions relevant to the query and learns
to give the correct answer. A neural reasoned updates the question by interacting it with
supporting facts through multiple reasoning layers. With this technique, it is possible to make
questions more specific than the original ones focusing on important image information
automatically.
3
https://www.wikipedia.org/

67
Another proposed solution is a framework of neural networks for learning not only the image and
question representations, but also their inter-modal interactions to produce the answer [39].
Another alternative are multimodal systems composed of different neural network types to extract
question information, visual representation, store the linguistic con-text of the answer and
combine this information into generating a relevant answer to a free language question have also
been proposed by [7], [26] and [36].
IDENTIFYING CLUES IN IMAGE AND/OR QUESTION TO GENERATE ANSWER:
The authors in [4] use attention mechanisms to answer questions by finding both “where to look”
in an image as well as “what words to listen to” in a question. The mechanism proposed, jointly
reasons about visual attention, by identifying regions in an image that are relevant to the question,
and question attention, by identifying words in a question that might contain useful information
for answering the question, referred to as co-attention, which has a natural symmetry between the
image and question. It reasons about the questions (and consequently about the image via the co-
attention mechanism) in a hierarchical fashion by word level, phrase level and question level.
Other attention-based systems proposed in [44] and [47] use neural networks to store into
memory information about different regions of the image relevant for computing the answer.
Another proposed solution we identified is a dynamic parameter neural network whose
parameters are determined adaptively based on input questions [34]. In this way the system
reasons differently for each question. The authors stated that the motivation behind this approach
is the fact that different questions require different types and levels of understanding of an image
to find correct answers. For example, to answer the question “is this an apple?” we should decide
between yes and no, while for questions like “how is the weather?” we need to perform
classification on multiple choices related to weather. Hence, besides performance on a single
recognition task, the capability to select a proper task is equally important.
VISUAL ANALOGY:
We identified one publication dealing with this topic. In this study, the analogy problem is treated
as learning an embedding that makes pairs of analogous images with similar transformations to be
close together in the embedding space and pushing apart images with dissimilar transformations
[6].
SITUATED VISUAL QUESTION ANSWERING:
We identified one publication dealing with situated visual question answering. The authors
formulated situated question answering as semantic parsing with an execution model that is a
learned function of the environment that can use background knowledge and global features of
the question/environment interpretation while retaining efficient approximate inference to answer
a question [29].
ANSWERING FROM A LIMITED DOMAIN OF QUESTIONS:
The only publication that we identified for this research topic assumes that the answer consists of
only a single word and treats the answering process as a classification process. The system treats
the image as a word of a sentence. The authors propose the use of a neural network and visual
semantic embeddings without intermediate stages like object detection and image segmentation to
predict answers to simple questions about images [8].

68
DETECTING RELEVANCE OF QUESTION TO IMAGE:
We identified one study dealing with this research challenge. The proposed system uses a neural
network to determine the relevance of questions to images in a 2-step fashion. First it captures
visual-specific linguistic structure to determine if the question is visual or not. If the question is
visual, pre-trained captioning models generate relevant captions (or questions) for the given
image and then compare them to the given question to determine if the visual question is relevant
to the image [32]
Table 10 summarizes the proposed approaches mentioned above and illustrates the distribution of
publications in accordance with the proposed approaches.
Table 8. Distribution of publications according to approaches proposed for Visual Question Answering
Approach Number of publications
Attention mechanisms 6
Multimodal systems 3
Dynamic parameter NN 1
Other 5
3.3.3. PROPOSED SOLUTIONS IN VISUAL QUESTION ANSWERING SYSTEMS
LEXICAL GAP BETWEEN QUESTIONS:
Various solutions have been proposed to overcome this problem. They share the common concept
of using neural networks to obtain the semantic similarity by learning a similarity metric. The
solutions include capturing the semantic similarity be-tween the current and archived questions
and incorporating similarity scores to rank and select semantically equivalent questions [12], [17],
[22], also incorporating metadata information such as question category to learn the similarity
metric [14]. Another proposed solution was to use a neural network to first learn semantic
representation of question and answer pairs from a collection of question and answer pairs. Then
for each pair use a mixture score to compute its similarity with queried question [3].
LEXICAL GAP BETWEEN QUESTIONS AND ANSWERS:
This problem is relevant to the task of matching question and answer. The proposed solutions we
identified are similar to the problem of lexical gap between questions. They include using neural
networks to learn the similarity metric between questions and answers and incorporating
similarity scores to rank and select the best match [40], [48], and also enhance the quality of
question-answer matching by collaboratively utilizing the rich interaction among questions,
answers and answerers to learn the relative quality rank of different answers with a same question
[38].
SPARSITY OF CQA DATA:
The problem of data sparsity is relevant to the task of expert finding. A proposed solution to
tackle this problem was modeling the expertise of a user based on his/her answering history by
combining the questions answered by the user for which he or she has been selected as the best
answerer. A neural network is used to predict best answerers for new questions by classifying the
new questions and then obtaining the probabilities of every user to be the best answer [46].
Another proposed solution was using a neural network to learn the ranking metric by leveraging
both user answering history and their social relation with other experts in the community with the

69
idea that a social relation between two users provides a strong evidence for them to have common
background knowledge [2].
DEVIATION FROM QUESTION:
This problem is relevant to the task of answer selection and ranking. The proposed solutions
identified for this problem were borrowing ideas from Machine Translation Evaluation (MTE)
[41] by using a neural network to decide the quality of an answer comment by taking two
comments and use ranking mechanisms to filter out bad (irrelevant) comments [13], and using a
neural network to learn the joint semantic representation of a question-answer pair and use this
representation to predict the quality of each answer in the comments sequence [19].
Table 11 summarizes the proposed methods mentioned above. Some systems use more than one
method.
Table 9. Distribution of publications according to methods used to solve Community Question Answering
Methods Number of publications
Similarity metric learning 8
Probabilistic inference 6
Ranking mechanisms 6
3.4. EVALUATION TECHNIQUES
In order to address RQ4 we identified and collected data about the datasets used for evaluating
the proposed studies, as well as metrics used in the evaluation. We find worth mentioning that
some of the publications used multiple datasets and multiple metrics for evaluation. Table 12
presents the total distribution of publications according to the datasets they use for evaluation.
Table 10. Distribution of publications according to the dataset used for evaluation
Dataset Number of publications
Self-constructed 14
VQA 9
COCO-QA 8
DAQUAR 7
Yahoo! Answers 6
WebQuestions 4
Baidu Zhidao 3
SIMPLE-QUESTIONS 3
AskUbuntu 2
InsuranceQA 1
TREC-QA 1
SemEval 2015 CQA 1
PARALAX 1
GeoQA 1
Visual7W 1
InfoboxQA 1
BabI-10k English 1
StackOverflow 1

70
We notice that the most used datasets are self-constructed, accounting for 30% of total
publications. Among the publicly available datasets, the most used ones are visual datasets VQA,
COCO-QA, DAQUAR and Yahoo! Answers, WebQuestions, Baidu Zhidao and Simple-
Questions.
Table 13 presents the total distribution of publications according to the metrics they use for
evaluation.
Table 11. Distribution of studies according to evaluation metrics
Metric Number of publications
Accuracy 23
MRR 9
F1 7
MAP 6
P@K 6
WUPS 5
NDCG, DCG@K 4
Recall 3
Precision 2
Fluency 1
AvgRec 1
R-Prec 1
SVMRank 1
S@N 1
Visual Turing Test 1
3.5. FUTURE RESEARCH DIRECTIONS
To address RQ5 we analyzed the future work that the selected publications proposed in order to
improve or extend their approaches. We also analyzed the current system limitations. The
information collected is summarized in tables 14, 15 and 16.
Table 12. Future research directions for KBQA
Future research directions
Extend system to handle more complex questions
Include support for compositional questions or multiple entity
questions
Integrate external knowledge sources
Extend model for other tasks like CQA
Add support for KB with noisy data or unstructured KB
Include non-factoid questions
Include larger dataset for testing
Improve accuracy

71
Table 13. Future research directions for CQA
Enhance system to include meta-data information like user votes,
rating, user reputation, etc
Incorporate more similarity measures
Improve performance
Improve topic recognition
Apply more sophisticated NN models
Table 14. Future research directions for Visual QA
Improve object-counting ability
Improve word-region relation
Add support for longer answers
Include visual attention
Expand domain of question
Add support for non-visual questions by leveraging an external
knowledge
Improve accuracy
4. DISCUSSION
The purpose of this paper was to make a quantitative and topic-based analysis of the recent
developments of QAS implemented with neural networks. The scope of this study was not to give
a detailed comparison between different NN approaches in terms of performance or compare
them with traditional approaches.
To answer RQ1 we identified current research directions of Neural Network Question Answering.
The research directions are Knowledge Base Question Answering, Visual Question Answering
and Community Question Answering. Knowledge Base Question Answering has the most
number of contributions with 39.1% of the total. However there is no considerable difference in
the quantity of research contributions among the three directions. VQA has 32.7% of total
contributions and CQA 28.2%. Therefore we cannot say that KBQA is the prevalent research
direction.
Regarding RQ2 we can say that the main challenges faced today by researches are: 1) lexical gap
between natural language questions and answers, as well as structured semantics of the
knowledge base for text based question answering, 2) answering free language questions about an
image due to the fact that the system must reason differently for different kind of questions for
image based question answering. These challenges clearly have the most number of contributions.
To answer RQ3 we can say that for text based question answering the most used methods and
techniques are similarity measures, representation learning and inference while for image based
question answering the most used approaches are attention mechanisms and multimodal systems.
RQ4 regards datasets and metrics used to evaluate the proposed system. From the results
presented in Section 3 we noticed that a considerable number of studies used self-constructed
datasets to evaluate the proposed solutions. Among the publicly available datasets, the most used
ones are VQA, COCO-QA, DAQUAR and Yahoo! Answers, WebQuestions, Baidu Zhidao and
Simple-Questions. Regarding evaluation metrics we can say that the most used metrics are
Accuracy, Mean Reciprocal Rank (MRR) and F1.

72
Regarding RQ5, common future directions proposed for all types of systems include: improve
accuracy, increase the complexity of supported questions and answers and integrate external
knowledge sources. As QAS are considered to be the future of search engines, these
improvements are justified by the need to make QAS more interactive, more intelligent and closer
to humans.
5. CONCLUSIONS
In this paper we presented a study on the state of the art of the latest research being done on QAS
based on neural networks. We have identified the main research directions and answered the
defined research questions regarding them. We have presented the main challenges faced by
researchers in this area along with the proposed solutions and future research directions.
Evaluation criteria and datasets used for testing were also identified. From this study we can
conclude that NN are being used effectively and continuously improving their performance, but
still there are unsolved problems and new challenges arising. As a future work we plan to test
publicly available QAS on common datasets and metrics relevant to their type and evaluate them.
REFERENCES
[1] Yin, Jun & Jiang, Xin & Lu, Zhengdong & Shang, Lifeng & Li, Hang & Li, Xiaoming (2016) “Neural
Generative Question Answering”, IJCAI (International Joint Conference on AI)
[2] Zhao, Zhou & Yang, Qifan & Cai, Deng & He, Xiaofei & Zhuang, Yueting (2016) “Expert Finding
for Community-Based Question Answering via Ranking Metric Network Learning”, IJCAI
(International Joint Conference on AI)
[3] Qiu, Xipeng & Huang, Xuanjin (2015) “Convolutional Neural Tensor Network Architecture for
Community-Based Question Answering”, IJCAI (International Joint Conference on AI)
[4] Lu, Jiasen & Yang, Jianwei & Batra, Dhruv & Parikh, Devi (2016) “Hierarchical Question-Image Co-
Attention for Visual Question Answering”, NIPS (Neural Information Processing Systems
Conference)
[5] Li, Ruiyu & Jia, Jiaya (2016) “Visual Question Answering with Question Representation Update”,
NIPS (Neural Information Processing Systems Conference)
[6] Sadeghi, Fereshteh & Zitnick, C. Lawrence & Farhadi, Ali (2015) “Visalogy: Answering Visual
Analogy Questions”, NIPS (Neural Information Processing Systems Conference)
[7] Gao, Haoyuan & Mao, Junhua & Zhou, Jie & Huang, Zhiheng & Wang, Lei & Wei, Xu (2015)
“Are You Talking to a Machine? Dataset and Methods for Multilingual Image Question”, NIPS
(Neural Information Processing Systems Conference)
[8] Ren, Mengye & Kiros, Ryan & Zemel, Richard S. (2015) “Exploring Models and Data for Image
Question Answering”, NIPS (Neural Information Processing Systems Conference)
[9] Dai, Zihang & Li, Lei & Xu, Wei (2016) “CFO: Conditional Focused Neural Question Answering
with Large-scale Knowledge Bases”, ACL (Annual Meeting of the Association on Computational
Linguistics)
[10] Tan, Ming & dos Santos, Cicero & Xiang, Bing & Zhou, Bowen (2016) “Improved Representation
Learning for Question Answer Matching”, ACL (Annual Meeting of the Association on
Computational Linguistics)
[11] Xu, Kun & Reddy, Siva & Feng, Yansong & Huang, Songfang & Zhao, Dongyan (2016) “Question
Answering on Freebase via Relation Extraction and Textual Evidence”, ACL (Annual Meeting
of the Association on Computational Linguistics)
[12] Das, Arpita & Yenala, Harish & Chinnakotla, Manoj & Shrivastava, Manish (2016) “Together we
stand: Siamese Networks for Similar Question Retrieval”, ACL (Annual Meeting of the Association
on Computational Linguistics)
[13] Guzmán, Francisco & Màrquez, Lluı́s & Nakov, Preslav (2016) “Machine Translation Evaluation
Meets Community Question Answering”, ACL (Annual Meeting of the Association on Computational
Linguistics)

73
[14] Zhou, Guangyou & He, Tingting & Zhao, Jun & Hu, Po (2015) “Learning Continuous Word
Embedding with Metadata for Question Retrieval in Community Question Answering”, ACL (Annual
Meeting of the Association on Computational Linguistics)
[15] Dong, Li & Wei, Furu & Zhou, Ming & Xu, Ke (2015) “Question Answering over Freebase with
Multi-Column Convolutional Neural Networks”, ACL (Annual Meeting of the Association on
Computational Linguistics)
[16] Yih, Wen-tau & Chang, Ming-Wei & He, Xiaodong & Gao, Jianfeng (2015) “Semantic Parsing via
Staged Query Graph Generation: Question Answering with Knowledge Base”, ACL (Annual Meeting
of the Association on Computational Linguistics)
[17] Dos Santos, Cı́cero & Barbosa, Luciano & Bogdanova, Dasha & Zadrozny, Bianca (2015) “Learning
Hybrid Representations to Retrieve Semantically Equivalent Questions”, ACL (Annual Meeting of
the Association on Computational Linguistics)
[18] Wang, Di & Nyberg, Eric (2015) “A Long Short-Term Memory Model for Answer Sentence
Selection in Question Answering”, ACL (Annual Meeting of the Association on Computational
Linguistics)
[19] Zhou, Xiaoqiang & Hu, Baotian & Chen, Qingcai & Tang, Buzhou & Wang, Xiaolong (2015)
“Answer Sequence Learning with Neural Networks for Answer Selection in Community Question
Answering”, ACL (Annual Meeting of the Association on Computational Linguistics)
[20] Yih, Wen-tau & He, Xiaodong & Meek, Christopher (2014) “Semantic Parsing for Single-Relation
Question Answering”, ACL (Annual Meeting of the Association on Computational Linguistics)
[21] Yin, Wenpeng & Yu, Mo & Xiang, Bing & Zhou, Bowen & Schütze, Hinrich (2016) “Simple
Question Answering by Attentive Convolutional Neural Network”, COLING (International
Conference on Computational Linguistics)
[22] Bogdanova, Dasha & dos Santos, Cı́cero & Barbosa, Luciano & Zadrozny, Bianca (2015) “Detecting
Semantically Equivalent Questions in Online User Forums”, CONLL (Conference on Natural
Language Learning)
[23] Bogdanova, Dasha & Foster, Jennifer (2016) “This is how we do it: Answer Reranking for Open-
domain How Questions with Paragraph Vectors and Minimal Feature Engineering”, NAACL (North
American chapter of ACL)
[24] Andreas, Jacob & Rohrbach, Marcus & Darrell, Trevor & Klein, Dan (2016) “Learning to Compose
Neural Networks for Question Answering”, NAACL (North American chapter of ACL)
[25] Jain, Sarthak (2016) “Question Answering over Knowledge Base using Factual Memory Networks”,
NAACL (North American chapter of ACL)
[26] Fukui, Akira & Huk Park, Dong & Yang, Daylen & Rohrbach, Anna & Darrell, Trevor & Rohrbach,
Marcus (2016) “Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual
Grounding”, EMNLP (Conference on Empirical Methods on Natural Language Processing)
[27] Sharp, Rebecca & Surdeanu, Mihai & Jansen, Peter & Clark, Peter & Hammond, Michael (2016)
“Creating Causal Embeddings for Question Answering with Minimal Supervision”, EMNLP
(Conference on Empirical Methods on Natural Language Processing)
[28] Ture, Ferhan & Boschee, Elizabeth (2016) “Learning to Translate for Multilingual Question
Answering”, EMNLP (Conference on Empirical Methods on Natural Language Processing)
[29] Krishnamurthy, Jayant & Tafjord, Oyvind & Kembhavi, Aniruddha (2016) “Semantic Parsing to
Probabilistic Programs for Situated Question Answering”, EMNLP (Conference on Empirical
Methods on Natural Language Processing)
[30] Golub, David & He, Xiaodong (2016) “Character-Level Question Answering with Attention”,
EMNLP (Conference on Empirical Methods on Natural Language Processing)
[31] Morales, Alvaro & Premtoon, Varot & Avery, Cordelia & Felshin, Sue & Katz, Boris (2016)
“Learning to Answer Questions from Wikipedia Infoboxes”, EMNLP (Conference on Empirical
Methods on Natural Language Processing)
[32] Ray, Arijit & Christie, Gordon & Bansal, Mohit & Batra, Dhruv & Parikh, Devi (2016) “Question
Relevance in VQA: Identifying Non-Visual And False-Premise Questions”, EMNLP (Conference on
Empirical Methods on Natural Language Processing)
[33] Iyyer, Mohit & Boyd-Graber, Jordan & Claudino, Leonardo & Socher, Richard & Daumé III, Hal
(2014) “A Neural Network for Factoid Question Answering over Paragraphs”, EMNLP (Conference
on Empirical Methods on Natural Language Processing)
[34] Noh, Hyeonwoo & Seo, Paul Hongsuck & Han, Bohyung (2016) “Image Question Answering Using
Convolutional Neural Network With Dynamic Parameter Prediction”, Conference on Computer

74
Vision and Pattern Recognition, CVPR (29th IEEE Conference on Computer Vision and Pattern
Recognition)
[35] Yang, Zichao & He, Xiaodong & Gao, Jianfeng & Deng, Li & Smola, Alex (2016) “Stacked
Attention Networks for Image Question Answering”, CVPR (29th IEEE Conference on Computer
Vision and Pattern Recognition)
[36] Malinowski, Mateusz & Rohrbach, Marcus & Fritz, Mario (2015) “Ask Your Neurons: A Neural-
Based Approach to Answering Questions about Images”, ICCV (International Conference on
Computer Vision)
[37] Ture, Ferhan & Jojic, Oliver (2016) “Ask Your TV: Real-Time Question Answering with Recurrent
Neural Networks”, SIGIR (Special Interest Group on Information Retrieval)
[38] Fang, Hanyin & Wu, Fei & Zhao, Zhou & Duan, Xinyu & Zhuang, Yueting (2016) “Community-
Based Question Answering via Heterogeneous Social Network Learning”, AAAI Conference on
Artificial Intelligence
[39] Ma, Lin & Lu, Zhengdong & Li, Hang (2016) “Learning to Answer Questions from Image Using
Convolutional Neural Network”, AAAI Conference on Artificial Intelligence
[40] Shen, Yikang & Rong, Wenge & Sun, Zhiwei & Ouyang, Yuanxin & Xiong, Zhang (2015)
“Question/Answer Matching for CQA System via Combining Lexical and Sequential
Information”, AAAI Conference on Artificial Intelligence
[41] Sun, Huan & Ma, Hao & He, Xiaodong & Yih, Wen-tau & Su, Yu & Yan, Xifeng (2016) “Table Cell
Search for Question Answering”, WWW (World Wide Web Conference)
[42] Ruan, Haipeng & Li, Yuan & Wang, Qinglin & Liu, Yu (2016) “A Research on Sentence
Similarity for Question Answering System Based on Multi-feature Fusion”, International Conference
on Web Intelligence, WI (International Conference on Web Intelligence)
[43] Zheng, Suncong & Bao, Hongyun & Zhao, Jun & Zhang, Jie & Qi, Zhenyu & Hao, Hongwei (2015)
“A Novel Hierarchical Convolutional Neural Network for Question Answering over Paragraphs”, WI
(International Conference on Web Intelligence)
[44] Xiong, Caiming & Merity, Stephen & Socher, Richard (2016) “Dynamic Memory Networks for
Visual and Textual Question Answering”, ICML (International Conference on Machine Learning)
[45] Yang, Liu & Ai, Qingyao & Spina, Damiano & Chen, Ruey-Cheng & Pang, Liang & Croft, W. Bruce
& Guo, Jiafeng & Scholer, Falk (2016) “Beyond Factoid QA: Effective Methods for Non-factoid
Answer Sentence Retrieval”, ECIR (European Conference on Information Retrieval)
[46] Wang, Jian & Sun, Jiqing & Lin, Hongfei & Dong, Hualei & Zhang, Shaowu (2016) “Predicting Best
Answerers for New Questions: An Approach Leveraging Convolution Neural Networks in
Community Question Answering”, Chinese National Conference on Social Media Processing, SMP
2016: Social Media Processing pp 29-41
[47] Xu, Huijuan & Saenko, Kate (2016) “Ask, Attend and Answer: Exploring Question-Guided Spatial
Attention for Visual Question Answering”, European Conference on Computer Vision, ECCV 2016:
Computer Vision – ECCV 2016 pp 451-466
[48] Zhou, Guangyou & Zhou, Yin & He, Tingting & Wu, Wensheng (2015) “Learning semantic
representation with neural networks for community question answering retrieval”, Knowledge-Based
Systems, Volume 93, 1 February 2016, Pages 75–83

A Review on Neural Network Question Answering Systems

More Related Content

What's hot (15)

Similar to A Review on Neural Network Question Answering Systems (20)

Recently uploaded (20)

A Review on Neural Network Question Answering Systems