Research On Chinese Word Sense Disambiguation Method Based On Deep Learning

Polysemy iscommonin natural languages. Word Sense Disambiguation(word sense disambiguation, WSD) task is aiming at determining the sense of a polysemy based on its context.The results of WSD task can directly affect machine translation(MT),information retrieval(IR) and other tasks.In this paper, we explore Chinese WSD task and out-of-vocabulary sense prediction task on the framework of machine learning. Specifically, the paper will expand the study of the following three aspects.(1) Chinese word sense disambiguationas sequence labeling. Weformalize WSD task as a sequence labeling problem. On the framework of machine learning, we use different sequence labeling models and different features to disambiguate, then analysis the result of WSD. The experiments show that CRFs with word and sense featuresis better than others.(2) Chinese word sense disambiguation based on the deep learning. We attempt to contain the semantic information into the WSD task. On the one hand, using co-occurrence word frequency anddependency structure to reduce the number of candidatesof ambiguous word, then calculate thesemantic similarity between sense vectors training from the neural network model; on the other hand, through semantic similarity the degree to expand the corpus, and we are using the maximum entropy model optimization to do the WSD task. Experimental results show that the extended corpus using optimization maximum entropy disambiguation model effects best.(3) Chinese out-of-vocabulary sense prediction based on deep learning. On the basis of vocabulary words WSD, weexplore the meaning of out-of-vocabularysense prediction, and it has two sub-problems: word sense construction and sense prediction. In construction step,we use internal morpheme feature?pos tag feature?cluster with similarity and merging methodto constructout-of-vocabulary word sense candidates. In prediction step, calculate the candidate's use of semantic similarity to predict meaning. The results showed that: the effect of the out-of-vocabulary sense prediction prefect best when we use the merging method to construct sense candidates.