Context-driven lexicon extension
Speaker:Emily Thomforde
Institution: University of Edinburgh
Abstract:
One of the many limitations for state-of-the-art parsing systems is the sparsity of existing lexicons. However, vast amounts of noisy, raw text are readily available. What is clearly needed is an automatic method for using the patterns already present in the lexicon to discover new entries from unlabeled data with high precision. Combinatory Categorial Grammar (CCG) is particularly suited to this task because it is entirely lexicalised, such that word/category pairs are the only target for learning. This talk will outline a first attempt to extend the existing CCG parsing lexicon by exploiting context cues over a large amount of unlabeled data. Further directions and implications of the model will also be discussed.