Propositional Term Extraction over Short Text using Word Cohesiveness and Conditional Random Fields with Multi-Level Features

Propositional Term Extraction over Short Text using Word Cohesiveness and Conditional Random Fields with Multi-Level Features 張如瑩 Ru-Yng Chang 國立成功大學資訊工程學系 Department of Computer Science and Information Engineering National Cheng Kung University ruyng@csie.ncku.edu.tw 吳宗憲 Chung-Hsien Wu 國立成功大學資訊工程學系 Department of Computer Science and Information Engineering National Cheng Kung University chwu@csie.ncku.edu.tw Abstract Propositional terms in a research abstract (RA) generally convey the most important information for readers to quickly glean the contribution of a research article. This paper considers propositional term extraction from RAs as a sequence labeling task using the IOB (Inside, Outside, Beginning) encoding scheme. In this study, conditional random fields (CRFs) are used to initially detect the propositional terms, and the combined association measure (CAM) is applied to further adjust the term boundaries. This method can extract beyond simply NP-based propositional terms by combining multi-level features and inner lexical cohesion. Experimental results show that CRFs can significantly increase the recall rate of imperfect boundary term extraction and the CAM can further effectively improve the term boundaries. 摘要 命題術語(Propositional Term)表達文章中重要概念且引導讀者文章脈絡之發展。這篇論 文以學術論文摘要為實驗對象進行命題術語擷取，研究中整合條件隨機域(Conditional Random Fields, CRFs) 以及結合聯繫測量(Combined Association Measure, CAM) 兩種方 法，考量詞彙內部凝聚力和文脈兩大類訊息，截取出的命題術語不再侷限於名詞片語型 態，且可由單詞或多詞所構成。在命題術語擷取的過程中，將其視為一種序列資料標籤 的任務，並利用 IOB 編碼方式識別命題述語的邊界，CRF 考量多層次構成命題述語的 特徵，負責初步命題術語偵測，再利用 CAM 計算詞彙凝聚力，藉以加強確認命題術語 詞彙的邊界。實驗結果顯示 ，本研究所提出的方法比以往述語偵測方法在效能上有明 顯增進，其中，CRF 明顯增進非完美術語詞彙邊界辨識(Imperfect hits)的召回率，而 CAM 則有效修正術語詞彙邊界。 Keywords: Propositional Term Extraction, Conditional Random Fields, Combined Association Measure, Multi-Level Feature 關鍵詞：命題述語擷取，條件隨機域，結合聯繫測量，多層次特徵 1. Introduction Researchers generally review Research Abstracts (RAs) to quickly track recent research trends. However, many non-native speakers experience difficulties in writing and reading RAs [1]. The author-defined keywords and categories of the research articles currently utilized to provide researchers with access to content guiding information are cursory and general. Therefore, developing a propositional term extraction system is an attempt to exploit the linguistic evidence and other characteristics of RAs to achieve efficient paper comprehension. Other applications of the proposed method contain sentence extension, text generation, and content summarization. A term is a linguistic representation of a concept with a specific meaning in a particular field. It may be composed of a single word (called a simple term), or several words (a multiword term) [2]. A propositional term is a term that refers to the basic meaning of a sentence (the proposition) and helps to extend or control the development of ideas in a text. The main difference between a term and a propositional term is that a propositional term, which can guide the reader through the flow of the content, is determined by not only syntax or morphology but semantic information. Take RAs to illustrate the difference between a term and a propositional term. Cheng [3] indicted that a science RA is composed of background, manner, attribute, comparison and evaluation concepts. In Figure 1, the terms underlined are the propositional terms which convey the important information of the RA. In the clause “we present one of the first robust LVCSR systems that use a syllable-level acoustic unit for LVCSR, ＂ the terms “ LVCSR systems ＂ , “ syllable-level acoustic unit ＂ and “LVCSR＂ respectively represent the background, manner and background concepts of the research topic, and can thus be regarded as …

Content truncated to 500 words