Adaptive and hybrid context-aware fine-grained word sense disambiguation in topic modeling based document representation

Abstract We propose a hybrid context based topic model with an adaptive window length for word sense disambiguation in document representation. Document representation is essential part of various tasks, and to capture the distinctions senses Traditional methods mainly rely on knowledge libraries data enrichment; however, semantics division may vary different domain-specific datasets. aim discover finer-grained semantic differences, such as entities or standpoints, handle problem without enrichment. There are two challenges this task: (1) dividing each polysemous word, (2) preserving differences between synonyms. Most existing models either separate clusters integrating auxiliary module specify senses. They can hardly achieve both since assumed be independent their intrinsic relationships ignored. To solve problem, we introduce “Bag-of-Senses” (BoS) assumption: multiset senses, generated instead words. The estimated by which it occurs contexts its other occurrences. Besides, scopes related occurrence, variable adjust adaptively. Our experiments three standard datasets show that our proposal outperforms state-of-the-art terms estimation, modeling, classification.