Ontology-based Geographic Knowledge Question Answering

In the field of artificial intelligence,it is a challenging task to automatically answer the college entrance examination(namely GaoKao)questions.Different from the questions of the general factoid question answering,the questions of GaoKao have a strong selection,the problem's forms are changeable,and the solution often cannot be obtained in one step,and it usually needs further knowledge inference.At present,there are two problems in assisting in solving the geography problem of GaoKao: the first problem is the lack of a well structured geographical core knowledge base(KB),and the second problem is that geographical problems are expressed in various forms,which makes it hard to understand.In view of the above two problems,this thesis has done the following three researches:(1)In order to solve the problem of the lack of a well structured geo-core knowledge base,this thesis constructs a Chinese Geographical Ontology(CGeoOnt).The ontology takes the version of the geography textbook published by people's education press as knowledge source,uses the World Wide Web Ontology language(OWL)as the knowledge representation language,takes the textbook chapter as the knowledge system structure,summarizes its core geographical concept,the geographical relation,the geographical test center,and expresses them in ontology form.At the same time,this thesis integrates the constructed CGeoOnt with ontology Clinga(chinese linked geographical dataset),and obtains a more large-scale chinese geographical ontology knowledge base.(2)In order to solve the problem of understanding geo-questions with various forms,this thesis employs the knowledge base question and answering model based on attention mechanism.The model is based on bidirectional long and short term memory(Bi-LSTM)network,combined with the attention mechanism to express the geographical problem and the answer,the vector generation of each word in the answer is combined with its attention weight distribution,so that the answer can better align the key information in the problem,weaken the interference of invalid information,which makes it easier to distinguish between correct answers and wrong answers.The experimental results show that this model has good reference and application value to assist in solving geography GaoKao questions.(3)In order to solve the problem of the lack of dataset in the training and testing procedures of Chinese geography question and answer model,this thesis collects a variety of chinese geography problem sets from the Internet.This thesis uses the Baidu question recommendation API as well as the baidu search API,takes the the high frequency core knowledge triples in ontology knowledge base as data source,then has access to 200,000 web geography questions,picks out the effective questions semi-automaticly and manually,then manually seeks the answers of the questions according to the knowledge base to obtain the final geography question and answering dataset.