ISCAS_NLP at SemEval-2016 Task 1: Sentence Similarity Based on Support Vector Regression using Multiple Features

Proceedings of SemEval-2016, pages 645–649, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ISCAS_NLP at SemEval-2016 Task 1: Sentence Similarity Based on Support Vector Regression using Multiple Features Cheng Fu, Bo An Institute of Software, Chinese Academy of Sciences, Beijing, China {fucheng,anbo}@nfs.iscas.ac.cn Xianpei Han, Le Sun State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China {xianpei,sunle}@nfs.iscas.ac.cn Abstract This paper describes our system developed for English Monolingual subtask (STS Core) of SemEval-2016 Task 1: “Semantic Textual Similarity: A Unified Framework for Seman- tic Processing and Evaluation”. We measure the similarity between two sentences using three different types of features, includ- ing word alignment-based similarity, sentence vector-based similarity and sentence constitu- ent similarity. The best performance of our submitted runs is a mean 0.69996 Pearson correlation which outperforms the median score from all participating systems. 1 Introduction Semantic Textual Similarity (STS) is the task of measuring the degree of semantic equivalence of a sentence pair (Agirre et al., 2012). STS was first held in SemEval 2012 and has drawn considerable attention in recent years. STS has been widely used in a lot of natural language processing tasks such as information retrieval, machine translation, ques- tion answering, text summarization, and so on. Previous methods for this task could be roughly divided into three categories: alignment approach- es, vector space approaches and machine learning approaches (Hänig et al., 2015). Alignment ap- proaches align words or phrases in a sentence pair, and then take the quality or coverage of alignments as similarity measure (Sultan et al., 2014). Vector space approaches represent sentences as bag-of- words vectors and take vector similarity as their similarity measure (Meadow et al., 1992). Machine learning approaches combine different similarity measures and features using supervised machine learning models (Bär et al., 2012). In our system, we measure semantic text simi- larity by combining evidence from all above three categories. Specifically, we extract align- ment-based similarity features, vector-based simi- larity features and sentence constituent similarity features from sentence pairs, and produce similari- ty scores between two sentences by combining these feature through a Support Vector Regression (SVR) model. 2 System Overview To measure the similarity between two sentences, we first extract a series of features, including alignment-based similarity features, vector-based similarity features and sentence constituent simi- larity features. Then we combine all these features and get an overall similarity using a Support Vec- tor Regression (SVR). In following we first de- scribe how to extract different types of features. Then we describe how to train the SVR model. 2.1 Alignment-based Similarity Features Sultan aligner is an open source unsupervised word alignment tool1 whose core algorithm is described in (Sultan et al., 2014). This aligner aligns related words in two sentenc- es based on the following two properties of the words: firstly, whether they are semantically simi- lar; secondly, whether they occur in similar seman- tic contexts in the respective sentences. As to the 1 https://github.com/ma-sultan/monolingual-word-aligner/ 645 former one, it mainly utilizes information provided …

Content truncated to 500 words