發展少量特徵擷取方法之問題分類技術; A Method to Extract Fewer Features for Question Classification

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/13433

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/13433

Title:	發展少量特徵擷取方法之問題分類技術;A Method to Extract Fewer Features for Question Classification
Authors:	曾增仁;Tseng-Jen Tseng
Contributors:	資訊管理研究所
Keywords:	文件分類;問題分類;問答系統;特徵擷取;機器學習;text classification;question classification;question answering system;machine learning;feature extraction
Date:	2008-06-26
Issue Date:	2009-09-22 15:31:46 (UTC+8)
Publisher:	國立中央大學圖書館
Abstract:	現今使用者利用問題回答系統 (question answering system) 進行資訊檢索時，通常期望在其查詢問題的當中得到一個確切的答案；而非傳統的檢索系統一般，回應一連串相關的文件列表。在問題回答系統的架構之中，系統在回答問題之前必須先進行問題的分類，以便了解問題的義涵。而問題分類也是問題回答系統裡處理程序之中最易出現錯誤的模組。以機器學習導向來說，問題分類與文件分類是兩個相似的程序。因此，特徵擷取在問題分類的處理之中是相當重要的任務。傳統特徵擷取的方法是依賴成百上千甚至更多的特徵，研究者在處理大量的特徵面臨了許多的問題。因此，本篇研究發展一個新的特徵擷取方法，試圖以少量的特徵擷取用於機器學習的分類器。在實驗當中，我們使用統計顯著性檢定來判別每一種不同特徵對於分類器效能的影響。實驗發現我們所擷取的特徵與一般常使用的bag-of-words 特徵表現一樣好。而在小型訓練資料集當中，我們所擷取的特徵也跟bag-of-ngrams 特徵的表現一樣好。 Today, some users usually prefer to receive answers in response to their questions by a question answering (QA) system, as opposed to the document lists returned by information retrieval (IR) system. In the architecture of a QA system, question classification is needed to extract the meaning of a question for answering the question. It causes most errors in the procedure of QA system. And question classification is very similar to text classification in machine learning approach. Therefore, the one of its important issues is to extract effective features. Traditional feature extraction depends on thousands or more features. Researches have problems in handling a large-dimension feature vectors. In view of this, this study is aimed to define a small number of features for machine learning classifiers. In our experiment, we test the efficacy of each feature with statistical significant test. We discover that our features are as good as bag-of-words feature. In small training dataset, our features are as good as bag-of-ngrams feature.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Size	Format

社群 sharing

Loading...