Computing Semantic Text Similarity Using Rich Features

Computing Semantic Text Similarity Using Rich Features Yang Liu1, Chengjie Sun1, Lei Lin1, Yuming Zhao2 and Xiaolong Wang1 1 Harbin Institute of Technology, Harbin Heilongjiang 150001, China 2 Northeast Forestry University, Harbin Heilongjiang 150040, China {yliu,cjsun,linl,ymzhao,wangxl}@insun.hit.edu.cn Abstract Semantic text similarity (STS) is an essential problem in many Natural Language Pro- cessing tasks, which has drawn a considerable amount of attention by research community in recent years. In this paper, our work focused on computing semantic similarity between texts of sentence length. We employed a Sup- port Vector Regression model with rich effec- tive features to predict the similarity scores between short English sentence pairs. Our model used WordNet-Based features, Corpus- Based features, Word2Vec-based features, Alignment-Based feature and Literal-Based features to cover various aspects of sentences. And the experiment conducted on SemEval 2015 task 2a shows that our method achieved a Pearson correlation: 80.486% which outper- formed the wining system (80.15%) by a small margin, the results indicated a high cor- relation with human judgments. Specially, among the five test sets which come from dif- ferent domains used in the estimation, our method got better results than the top team on two of them whose domain-related data is available for training, while comparable re- sults were achieved on the rest three unseen test sets. The experiments results indicated that our solution is more competitive when the domain-specific training data is available and our method still keeps good generalization ability on novel data. 1 Introduction Semantic text similarity is a fundamental challenge in many Natural Language Processing tasks, such as Machine Reading, Deep Question Answering (Narayanan & Harabagiu, 2004), Automatic Ma- chine Translation Evaluation (Papineni, Roukos et al., 2002), Automatic Text Summarization (Fattah & Ren, 2008) and Query Reformulation (Metzler, Dumais et al., 2007), etc. Previous researches on semantic text similarity have been focused on doc- uments and paragraphs, while comparison objects in many NLP tasks are texts of sentence length, such as Video descriptions, News headlines and beliefs, etc. In this paper, we study semantic simi- larity between sentences. Given two input text segments, we need to automatically determine a score that indicates their semantic similarity. The difficulties of this task lie on several aspects. First, there were no existing effective measures to repre- sent sentences which could be understood by com- puters without losing any information. Second, even with good representations, it’s very hard to find a metric which can fully compare the equiva- lence between two sentence representations. Third, similarity itself is a very complex concept, and se- mantic space is also hard to define and quantize. Given the same pair of sentences, different people may mark different similarity scores; this incon- sistency is derived from people’s judgments of dif- ference. Although with these difficulties ahead, a lot of methods have been proposed to handle this problem in recent years. And our efforts mainly focused on trying to combine different existing approaches to represent a sentence, and hope to cover as many aspects of sentence as possible on semantic …

Content truncated to 500 words