Jeff Siskind's Lexicon Learning Research

The most closely related previous research into automated lexicon acquisition is that of Siskind (1996), itself inspired by work by Rayner (1988). As we will be comparing our system to his in Section 5, we describe the main features of his research in this section. His goal is one of cognitive modeling of children's acquisition of the lexicon, where that lexicon can be used for both comprehension and generation. Our goal is a machine learning and engineering one, and focuses on a lexicon for comprehension and use in parsing, using a learning process that does not claim any cognitive plausibility, and with the goal of learning a lexicon that generalizes well from a small number of training examples. His system takes an incremental approach to acquiring a lexicon. Learning proceeds in two stages. The first stage learns which symbols in the representation are to be used in the final ``conceptual expression'' that represents the meaning of a word, by using a version-space approach. The second stage learns how these symbols are put together to form the final representation. For example, when learning the meaning of the word ``raise'', the algorithm may learn the set {CAUSE, GO, UP} during the first stage and put them together to form the expression CAUSE(x, GO(y, UP)) during the second stage. Siskind (1996) shows the effectiveness of his approach on a series of artificial corpora. The system handles noise, lexical ambiguity, referential uncertainty, and very large corpora, but the usefulness of lexicons learned is only compared to the ``correct,'' artificial lexicon. The goal of the experiments presented there was to evaluate the correctness and completeness of learned lexicons. Earlier work [Siskind1992] also evaluated versions of his technique on a quite small corpus of real English and Japanese sentences. We extend that evaluation to a demonstration of the system's usefulness in performing real world natural language processing tasks, using a larger corpus of real sentences.