Methods For Solving Association Problems In High-school Geography Multiple Choice Questions

Answering questions in Gaokao(the national college entrance examination in China)is becoming a hot Question Answering Task in recent years.However,the task is also quiet challenging because of the difficult of question understanding and the lack of domain-specific knowledge base.This thesis is aimed at solving association prob-lems in high-school geography multiple choice questions,including the two types of conceptual association questions and causal association questions.The conceptual as-sociation questions describe geographic entities,concepts and the relationships among geographical concepts,however,there is still a lack of domain-specific knowledge base to describe the geographical concepts and their relationships.The causal association questions describe the causal relationship between geographical elements,there is also a lack of domain-specific knowledge base to fulfill their solution requirements.For answering the questions of conceptual association,a concept graph is semi-automatically constructed from textbook and Chinese wiki encyclopedia,to capture core concepts and relations in high-school geography.The concept graph depicts the attribute information of geographic concepts(such as alias,description text,etc.)and the relationships among geographic concepts(such as hypernymy,hyponym,disjoint,etc.).This thesis proposes the method for the questions of conceptual association based on concept graph(CGQA).Firstly,the natural language question is transformed into the conjunction of relation pairs by Entity Linking,Concept Matching and Description Mapping.Then,CGQA searchs the concept graph to get the inference path for each relation pair.The inference path can be divided into Protesting Path/Supporting Path,depending on whether the path contains "disjoint" relation.If the inference path of the relation pair is a Supporting Path,then the relation pair is true.Otherwise,it's false.Lastly,CGQA generates the judgment for each question by combining the judgment for each relation pair.Additionally,CGQA combines the inference path for each relation pair to generate the explanation for the whole question.For answering the questions of causal association,a causal knowledge base is automatically constructed by using pattern matching and statistics.This thesis pro-poses the method for the questions of causal association based on causal knowledge base(CKQA).Firstly,CKQA implement a causal association question identification method.And then,the identified causal association question is transformed into the form of template.After that,the template of question will be used as the query to re-trieve the causal knowledge base and return the candidate knowledge items.Lastly,the features of template about question,the features of candidate knowledge and the relevance features between question and candidate knowledge items are extracted.The solution of the causal association questions is modeled as binary classification task.Experiments indicate that CGQA can get a scoring rate of 34.64%on the dataset of all multiple-choice questions in Beijing Geography Gaokao from 2008 to 2018.And CGQA will give one or more inference paths for the each given question,the correctness of the inference paths is 81.35%.The score rate increases to 43.60%by combining CGQA with a neural network based method.Moreover,CKQA has a score rate of 30.25%on the dataset of causal association questions,which is better than a neural network based method.Meanwhile,the score rate increases to 33.12%by combining CKQA with a neural network based method.