Measuring Term Informativeness in Context

Proceedings of NAACL-HLT 2013, pages 259–269, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Measuring Term Informativeness in Context Zhaohui Wu Computer Science and Engineering The Pennsylvania State University University Park, PA 16802, USA zzw109@psu.edu C. Lee Giles Information Sciences and Technology Computer Science and Engineering The Pennsylvania State University University Park, PA 16802, USA giles@ist.psu.edu Abstract Measuring term informativeness is a funda- mental NLP task. Existing methods, mostly based on statistical information in corpora, do not actually measure informativeness of a term with regard to its semantic context. This pa- per proposes a new lightweight feature-free approach to encode term informativeness in context by leveraging web knowledge. Given a term and its context, we model context- aware term informativeness based on semantic similarity between the context and the term’s most featured context in a knowledge base, Wikipedia. We apply our method to three ap- plications: core term extraction from snippets (text segment), scientific keywords extraction (paper), and back-of-the-book index genera- tion (book). The performance is state-of-the- art or close to it for each application, demon- strating its effectiveness and generality. 1 Introduction Computationally measuring importance of a word in text, or “term informativeness” (Kireyev, 2009; Rennie and Jaakkola, 2005), is fundamental to many NLP tasks such as keyword extraction, text catego- rization, clustering, and summarization, etc. Various features derived from statistical and linguistic infor- mation can be helpful in encoding term informative- ness, whereas practical feature definition and selec- tion are usually ad hoc, data-driven and application dependent. Statistical information based on term frequency (TF) and document frequency (DF) tend to be more effective in finding keywords in large corpora, but can have issues with small amounts of text or small corpora. Linguistic information such as POS tag patterns often require manual selection based on prior applications. We contend that few methods actually measure the informativeness of a term to the discourse unit it contains. For example, given a context such as “A graph comprises nodes (also called vertices) connected by links (also known as edges or arcs)”, it is difficult to measure the term informativeness of “graph”, “nodes”, or “links” based on any statistical or linguistic information. This raises many issues. Is there a fundamental and less ad hoc way to measure the term informa- tiveness of a word within a discourse unit? Can we actually find a general approach based on compre- hensive and high-level “knowledge” and not have to nitpick over features? Can this new metric be effectively applied to real world applications? To answer these questions, we develop a new term in- formativeness metric, motivated by query-document relevance in information retrieval. The higher the relevance score a query-document pair is, the more informative the query is to the document. If a sim- ilar principle also exists between word and con- text and there is an effective search engine return- ing ranked contexts for a given word, then we con- tend that word is more informative in the higher rank contexts. To see the …

Content truncated to 500 words