A Powerful And General Approach To Context Exploitation In Natural Language Processing

In natural language, the meaning of a lexeme often varies due to the specific surrounding context. Computational approaches to natural language processing can benefit from a reliable, long-range-context-dependent representation of the meaning of each lexeme that appears in a given sentence. We have developed a general new technique that produces a context-dependent ‘meaning’ representation for a lexeme in a specific surrounding context. The ‘meaning’ of a lexeme in a specific context is represented by a list of semantically replaceable elements the members of which are other lexemes from our experimental lexicon. We have performed experiments with a lexicon composed of individual English words and also with a lexicon of individual words and selected phrases. The resulting lists can be used to compare the ‘meaning’ of conceptual units (individual words or frequentlyoccurring phrases) in different contexts and also can serve as features for machine learning approaches to classify semantic roles and relationships.