A Semantically Enhanced Approach to Determine Textual Similarity

A Semantically Enhanced Approach to Determine Textual Similarity Eduardo Blanco and Dan Moldovan Lymba Corporation Richardson, TX 75080 USA {eduardo,moldovan}@lymba.com Abstract This paper presents a novel approach to deter- mine textual similarity. A layered methodol- ogy to transform text into logic forms is pro- posed, and semantic features are derived from a logic prover. Experimental results show that incorporating the semantic structure of sen- tences is beneficial. When training data is unavailable, scores obtained from the logic prover in an unsupervised manner outperform supervised methods. 1 Introduction The task of Semantic Textual Similarity (Agirre et al., 2012) measures the degree of semantic equiv- alence between two sentences. Unlike textual en- tailment (Giampiccolo et al., 2007), textual similar- ity is symmetric, and unlike both textual entailment and paraphrasing (Dolan and Brockett, 2005), tex- tual similarity is modeled using a graded score rather than a binary decision. For example, sentence pair (1) below is very similar [5 out of 5], (2) is some- what similar [3 out of 5] and (3) is not similar at all [0 out of 5]: 1. Someone is removing the scales from the fish. A person is descaling a fish. 2. A woman is chopping an herb. A man is finely chopping a green substance. 3. A cat is playing with a watermelon on a floor. A man is pouring oil into a pan. State-of-the-art systems to determine textual sim- ilarity (B¨ar et al., 2012; ˇSari´c et al., 2012; Banea et al., 2012) do not account for the semantic struc- ture of sentences, and mostly rely on word pair- ings and knowledge derived from large corpora, e.g., man holding AGENT � THEME �leaf (a) monkey fighting AGENT � THEME �man (b) Figure 1: Semantic representation of 1(a) A man is hold- ing a leaf and 1(b) A monkey is fighting a man. Wikipedia. Regardless of details, each word in sent1 is paired with the word in sent2 that is most simi- lar according to some similarity measure. Then, all similarities are added and normalized by the length of sent1 to obtain the similarity score from sent1 to sent2. The process is repeated to obtain the simi- larity score from sent2 to sent1, and both scores are then averaged to determine the overall textual sim- ilarity. Several word-to-word similarity measures are often combined with other shallow features, e.g., n-gram overlap, syntactic dependencies, to obtain the final similarity score. Consider sentences 1(a) A man is holding a leaf and 1(b) A monkey is fighting a man. These two sentences are very dissimilar, the only commonal- ity is the concept ‘man’. Any approach that blindly searches for the word in 1(b) that is the most similar to word ‘man’ in 1(a) will find ‘man’ from 1(b) to be a perfect match. One of three content words is a match and thus the estimated similarity will be much higher than it actually is. Consider now the semantic representations for sentences 1(a) and 1(b) in Figure 1. ‘man’ plays the role of …

Content truncated to 500 words