The Web as an Implicit Training Set
Google Tech Talks
November, 5 2007
Speaker: Preslav Nakov
I will present Web-based approaches to the syntax and semantics of noun compounds (NCs), which can be used in query parsing, technical term understanding, etc. I will also describe an application to machine translation. First, I will present a highly accurate lightly supervised method based on surface features and paraphrases for making bracketing decisions for three-word noun compounds, e.g. "[[liver cell] antibody]" is left-bracketed, while "[liver [cell line]]" is right-bracketed. The enormous size of the Web makes such features frequent enough to be useful. Second, I will introduce an unsupervised method for discovering the implicit predicates characterizing the semantic relations that hold in noun-noun compounds. For example, "malaria mosquito" is a "mosquito that carries/spreads/causes/transmits/brings/infects with/... malaria". Finally, I will present a method for improving Machine Translation (SMT). Most modern SMT systems rely on aligned sentences of bilingual corpora for training. I will describe a method for expanding the training set with conceptually similar but syntactically differing paraphrases at the NP-level which involve NCs. The English to Spanish evaluation on the Europarl corpus shows an improvement equivalent to 33%-50% of that of doubling the amount of training data.