Word Sense Disambiguation Based On Semantic Relatedness Computation

Determining the intended sense of words in text-word sense disambiguation(WSD)-is a long-standing problem in natural language processing(NLP)with broad applications.Supervised WSD systems perform best in public evaluations but they need large amounts of hand-tagged data.Semantic relatedness is one of the most important semantic features that influences the performance of WSD directly.The main difficulty lies in the study of semantic relatedness algrithom closer to human cognitive level.Considering the differences in existing resources of Chinese and English,two different semantic relatedness calculation approaches are presented.For Chinese,we present a semantic relatedness algorithm to disambiguate word senses which uses HowNet.This method uses HowNet Inference Machine and sememe provided by the HowNet to calculate semantic relatedness.We take full advantage of the structured definition of word sense and other resources in the HowNet.For English,we present a semantic relatedness algorithm to disambiguate word senses which uses context vectors that introduce part-of-speech.Context vectors are learned using deep learning model with additional part-of-speech features.Then we calculate semantic relatedness with context vectors of target sentences and example sentences.This modified semantic relatedness algorithm is shown to improve the accuracy of WSD.Based on the above algorithms,we propose a method of combining semantic relatedness with supervised WSD.For Chinese,the semantic relatedness calculated by HowNet is combined with supervised learning.For English,the semantic relatedness calculated by context vectors is combined with supervised learning.After combination,both Chinese and English disambiguation performance has been greatly improved.After evaluating in SemEval and Senseval public evaluations,this combining method achieves better disambiguation performance.