SXUCFN-Core: STS Models Integrating FrameNet Parsing Information

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 74–79, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics SXUCFN-Core: STS Models Integrating FrameNet Parsing Information Sai Wang, Ru Li, RuiboWang, ZhiqiangWang, Xia Zhang Shanxi University, Taiyuan, China enrique.s.wang@gmail.com {liru, wangruibo}@sxu.edu.cn {zhiq.wang, caesarzhangx}@163.com Abstract This paper describes our system submitted to *SEM 2013 Semantic Textual Similarity (STS) core task which aims to measure semantic si- milarity of two given text snippets. In this shared task, we propose an interpolation STS model named Model_LIM integrating Fra- meNet parsing information, which has a good performance with low time complexity com- pared with former submissions. 1 Introduction The goal of Semantic Textual Similarity (STS) is to measure semantic similarity of two given text snippets. STS has been recently proposed by Agirre et al. (2012) as a pilot task, which has close relationship with both tasks of Textual Entailment and Paraphrase, but not equivalent with them and it is more directly applicable to a number of NLP tasks such as Question Answering (Lin and Pantel, 2001), Text Summarization (Hatzivassiloglou et al., 1999), etc. And yet, the acquiring of sentence simi- larity has been the most important and basic task in STS. Therefore, the STS core task of *SEM 2013 conference, is formally defined as the degree of semantic equivalence between two sentences as follows:  5: completely equivalent, as they mean the same thing.  4: mostly equivalent, but some unimpor- tant details differ.  3: roughly equivalent, but some impor- tant information differs/missing.  2: not equivalent, but share some details.  1: not equivalent, but are on the same top- ic.  0: on different topics. In this paper, we attempt to integrate semantic information into STS task besides the lower-level word and syntactic information. Evaluation results show that our STS model could benefit from se- mantic parsing information of two text snippets. The rest of the paper is organized as follows: Sec- tion 2 reviews prior researches on STS. Section 3 illustrates three models measuring text similarity. Section 4 describes the linear interpolation model in detail. Section 5 provides the experimental re- sults on the development set as well as the official results on all published datasets. Finally, Section 6 summarizes our paper with direction for future works. 2 Related Work Several techniques have been developed for STS. The typical approach to finding the similarity be- tween two text segments is to use simple word matching method. In order to improve this simple method, Mihalcea et al. (2006) combine two cor- pus-based and six knowledge-based measures of word similarity, but the cost of their algorithm is expensive. In contrast, our method treats words and texts in essentially the same way. In 2012 STS task, 35 teams participate and sub- mit 88 runs. The two top scoring systems are UKP 74 and Takelab. The former system (Bär et al., 2012) uses a simple log-linear regression model to com- bine …

Content truncated to 500 words