Minimally Supervised Question Classification and Answering based on WordNet and Wikipedia

Minimally Supervised Question Classification and Answering based on WordNet and Wikipedia 張至 Joseph Chang 顏孜羲 Tzu-Hsi Yen 蔡宗翰 Richard Tzong-Han Tsai* 元智大學資訊工程學系 Department of Computer Science and Engineering, Yuan Ze University, Taiwan {s951533, s940635}@mail.yzu.edu.tw, thtsai@saturn.cse.yzu.edu.tw *corresponding author 摘要 在此篇論文中，我們提出一個自動將問題分類至現有詞網 (WordNet) 中之細分類方 法。為此，我們利用維基百科之特性以及其文章標題，建立大規模語意實體分類表， 包含了1,581,865個實體。為了表現我們研究之效用，我們建構了一個基於冗餘原則的 自動問題回答系統，並透過所提出的問題分類方法來增進其效能。實驗結果顯示所提 出的方法能夠有效地提升問題分類與回答的精確率。 Abstract In this paper, we introduce an automatic method for classifying a given question using broad semantic categories in an existing lexical database (i.e., WordNet) as the class tagset. For this, we also constructed a large scale entity supersense database that contains over 1.5 million entities to the 25 WordNet lexicographer’s files (supersenses) from titles of Wikipedia entry. To show the usefulness of our work, we implement a simple redundancy-based system that takes the advantage of the large scale semantic database to perform question classification and named entity classification for open domain question answering. Experimental results show that the proposed method outperform the baseline of not using question classification. 關鍵詞： 自動問題回答，問題分類，辭彙語意資料庫，辭網，維基百科 Keywords: question answering, question classification, semantic category, WordNet, Wikipedia. 1. Introduction Question classification is considered crucial to the question answering task due to its ability 209 to eliminating answer candidates irrelevant to the question. For example, answers to person- questions (e.g., Who wrote Hamlet?) should always be a person (e.g., William Shakespeare). Common classification strategies includes semantic categorization and surface patterns identification. In order to fully benefit from question classification techniques, answer candidates should be classified the same way as questions. Surface patterns identification methods classifies questions to sets of word-based patterns. Answers are then extracted from retrieved documents using these patterns. Without the help of external knowledge, surface pattern methods suffer from limited ability to exclude answers that are in irrelevant semantic classes, especially when using smaller or heterogeneous corpora. An other common approach uses external knowledge to classify questions to semantic types. In some previous QA systems that deploy question classification, named entity recognition (NER) techniques are used for selecting answers from classified candidates. State-of-the-art NER systems produce near human performances. Good results are often achieved by handcrafted complex grammar models or large amount of hand annotated training data. However, most high performance NER systems deal with a specific domain, focus on homogeneous corpora, and support a small set of NE types. For example, in the Message Understanding Conference 7 (MUC-7) NER task, the domain is “Airplane crashes, and Rocket/Missile Launches” using news reports as the corpus. There are only three NE classes containing seven sub classes: ORG, PERSON, LOCATION, DATE, TIME, MONEY, PERCENT. Notice that in the seven subclasses, only three of them are NEs of physical objects, others are number based entities. This is apparently insufficient for candidates filtering for general question answering. Owing to the need of wider range NE types, some of the later proposed NE classes construct of up to 200 sub classes, but NER systems targeting these types of fine-grained NE classes may not be precise enough to achieve high performance. The amount of supported classification types greatly influences the performance of QA systems. A coarse-grained …

Content truncated to 500 words