- 學位論文
中文開放領域自動問答系統之研究
A Study on Chinese Open-Domain Question Answering Systems
摘要
本論文主旨為研究如何建立一個開放領域的中文自動問答系統，目的在於自動回答非限定領域的自然語言問句，並研究在中文環境中所會面對的問題及所需建立的資源為何。論文中提出了新的中文問句類型分類規則、候選答案分數的多層次排名策略、以及短答問題及長答問題的可能候選答案為何等等。系統中各模組的效能及影響也將有所討論。NTU QA System現在已經開放在網際網路上，供使用者可以線上試用這套系統。 中文問句類型定義有十一類，包括有是非 (YESNO)、選擇 (SELECTION)、人物 (PERSON)、地點 (LOCATION)、時間 (TIME)、數量 (QUANTITY)、物品 (OBJECT)、方法 (METHOD)、定義 (DEFINITION)、人物描述 (PERSONDEF) 及原因 (REASON) 各類問題。問句類型分類規則共有136條，集內測試的正確率為72.7%。在決定問句類型時也會同時決定問句中心語。 針對不同的問句類型，論文中提出不同的候選答案找尋方式。短答問題的候選詞多半來自於具名實體辨識系統的輸出，或是在語意辭典中的下位詞等。長答問題的候選答案則是藉由特定句型所抽取出的詞組，並可針對特別問問類型限制組中所帶的語意等。 不同的問句類型亦有著不同的候選答案排名策略。各種分數計算、權重設定及排名策略等都將由實驗結果來決定最好的組合。各問句類型的效能如下：短答問題：答題率52.9%、MRR 0.446；定義類問題：答題率56.3%、MRR 0.443；人物描述類問題：答題率60.7%、MRR 0.418。 NTU QA系統已經實際在網際網路上運作，除了提供自然語言的問句輸入外，系統也會立即在網上搜尋可能的答案回應給使用者，並且提供來源網頁以及網頁中提到答案成立的時間點等資訊。
並列摘要
Development of a question answering system focusing on open-domain knowledge in Chinese environment is studied in this dissertation. New question type categories, ranking strategies, and identification of answer candidates for short-answer questions and long-answer questions are proposed. Performances of modules in the QA system are evaluated, and the effects of different factors are studied. The NTU QA System is now working online, which receives questions and finds answers immediately from the Internet. Questions are classified into eleven question types, including YESNO, SELECTION, PERSON, LOCATION, TIME, QUANTITY, OBJECT, METHOD, DEFINITION, PERSONDEF, and REASON questions. A classification rule set is constructed, including 136 rules with a correct rate of 72.7% in an inside test. Question cores can also be decided at the same time. Answer candidate extractions of different question types are proposed. The candidates of short-answer questions are terms extracted by a named entity identifier or descendents in a thesaurus. Candidates of long-answer questions are phrases extracted by patterns or noun phrases with head nouns denoting persons. Different ranking scores are proposed for different types of questions. Different Ranking strategies and weight assignments are experimented. The performances of different question types are: Coverage 52.9%, MRR 0.446 for short-answer questions; Coverage 56.3%, MRR 0.443 for DEFINITION questions; and Coverage 60.7%, MRR 0.418 for PERSONDEF questions. An online system has been built and working on the Internet. Besides receiving questions in natural languages and returning answers found on the Internet immediately, the QA system also provides the time point when an answer took place.