NLPCC2018论文集

A Relateness-Based Ranking Method for

Knowledge-Based Question Answering

Han Ni

1(

B

)

, Liansheng Lin

2

,andGeXu

Abstract. In this paper, we report technique details of our approach for

the NLPCC 2018 shared task knowledge-based question answering. Our

system uses a word-based maximum matching method to find entity

candidates. Then, we combine editor distance, character overlap and

word2vec cosine similarity to rank SRO triples of each entity candidate.

Finally, the object of the top 1 score SRO is selected as the answer of

the question. The result of our system achieves 62.94% of answer exact

matching on the test set.

Keywords: Question answer

Question Answering (KBQA) is defined to retrieve a specific entity from knowl-

edge base as the answer to a given question.

The challenge of retrieval-based KBQA is how to match unstructured natural

language questions with structured data in knowledge base. To understand a

question, it is necessary to figure out the topic entity and relation chain inside the

question. Thus, topic entity linking and relation ranking are the most important

modules in our system.

Knowledge-based question answering is a challenging task in the field of NLP.

The mainstream approaches can be divided into three categories: semantic pars-

394 H. Ni et al.

The semantic parsing based approaches translate natural language questions

into a series of semantic representations in logic forms. They query the answer

in knowledge base through the corresponding query statement. Yih et al. [12]

present a semantic parsing method via staged query graph generation. Convo-

lution neural network is used to calculate the similarities between question and

relation chains.

The information extraction based approaches extract topic entities from ques-

tions and generate a knowledge base subgraph with the topic entity node as

the center. Each node in the subgraph can be used as a candidate answer. By

examining the questions and extracted information according to some rules or

The idea of retrieval-based method is similar to that of information extrac-

tion based methods. The question and candidate answers are mapped to dis-

tributed representation. The distributed representations are trained on labeled

data, aiming to optimize the matching function between the question and the

correct answer. Zhang et al. [14] combine bi-directional LSTM with an attention

mechanism to represent the questions dynamically according to diverse focuses

of various candidate answers.

These approaches work well on the English open dataset WebQuestion. How-

ever, their performances on a Chinese KBQA dataset have not been presented

triple is the answer of the question.

Since the entity in the knowledge base has various name, such as Chinese name,

English name, nick name, alias and so on, we build a Entity Map which maps

these names to the original entity. In order to detect the topic entity in the ques-

tion, we use a word-based maximum matching method to find entity candidates.

First of all, the question is segmented by ltp

1

[15] segmenter. Then, we join the

1

entity. After that, two sub-list of words are left, which are

and . Then, we do word-based maximum matching for each sub-list.

For sub-list,

and search it in the Entity Map. Apparently, it is not a entity. Then, we shorten

the string length by one. Now, we search

and in the Entity

triples. Each subject entity has dozens of Relation-Object pairs, each relation

corresponding to only one object entity. Therefore, finding the answer to the