engleski

ConGraCNet 0.3: Corpus-based graph syntactic- semantic relations analysis

The paper will demonstrate the ConGraCNet application for distinguishing word senses and identifying semantically related lexemes in a corpus by using the syntactic-semantic patterns of language usage. This unsupervised tagged corpus graph analysis method is based on the construction grammar approach to syntactic dependencies. ConGraCNet relies explicitly on the coordinated [x and|or y] and [x is_a y] syntactic grammatical relations between the lexical co- occurrences for the construction of the network representation. For a given source lexeme in a corpus, the method yields associated communities of collocation lexemes that represent the sense structure and different meanings based on the context of its usage. By projecting semantic value to a coordinated syntactical relation [x and|or y], we can filter the lexical collocates with high conceptual similarity from a corpus and construct clustered lexical networks that reveal ambiguous referential meanings of a source lexeme. The members of a cluster are processed with an iterative graph function that finds best candidates for abstracted class label using [x is_a y] syntactic-semantic construction. We will explain the impact of the modulation of the linguistic and graph parameters, exemplify the application of the procedure on several lexemes in different languages and corpora and present the implementation of the WordNet external knowledge databases for further refinement of the results.

Natural Language Processing ; Conceptual Similarity ; Word Sense Induction ; Corpus

nije evidentirano