Using Web-scale N-grams to Improve Base NP Parsing Performance

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 886–894, Beijing, August 2010 Using Web-scale N-grams to Improve Base NP Parsing Performance Emily Pitler Computer and Information Science University of Pennsylvania epitler@seas.upenn.edu Shane Bergsma Department of Computing Science University of Alberta sbergsma@ualberta.ca Dekang Lin Google, Inc. lindek@google.com Kenneth Church Human Language Technology Center of Excellence Johns Hopkins University kenneth.church@jhu.edu Abstract We use web-scale N-grams in a base NP parser that correctly analyzes 95.4% of the base NPs in natural text. Web-scale data improves performance. That is, there is no data like more data. Performance scales log-linearly with the number of parame- ters in the model (the number of unique N-grams). The web-scale N-grams are particularly helpful in harder cases, such as NPs that contain conjunctions. 1 Introduction Noun phrases (NPs) provide an index to the world’s information. About 70% of web queries are NPs (Barr et al., 2008). A robust NP parser could help search engines improve retrieval per- formance on multi-word NP queries (Zhai, 1997). For example, by knowing the correct parse of “washed (baby carrots),” a search engine could ensure that returned pages (and advertisements) concern clean carrots rather than clean babies. NP structure is also helpful for query expansion and substitution (Jones et al., 2006). ing.This paper is concerned with base NP pars- We are given a base NP string as input, and the task is to produce a parse tree as output. Base NPs are NPs that do not contain embedded noun phrases. These are sometimes called NP chunks, or core/non-recursive NPs (Church, 1988; Ramshaw and Marcus, 1995). Correctly parsing (or, equivalently, bracketing) base NPs is chal- lenging because the same part-of-speech (POS) sequence can be parsed differently depending on the specific words involved. For example, “retired (science teacher)” and “(social science) teacher” have different structures even though they have identical POS sequences. Lexical statistics are therefore needed in order to parse the above examples, and they must be computed over a lot of text to avoid sparsity. All of our lexical statistics are derived from a new and improved web-scale N-gram corpus (Lin et al., 2010), which we call Google V2. Despite the importance of base NPs, most sentence parsers do not parse base NPs, since the main training corpus for parsers, the Penn Treebank (PTB) (Marcus et al., 1994), leaves a flat structure for base NPs. Recent annotations by Vadas and Curran (2007a) added NP structure to the PTB. We use these annotations (described in Section 3) for our experiments. NP parsers usually focus on bracketing three- word noun compounds. Parsing three-word noun compounds is a fairly artificial task; we show that sequences of three nouns make up less than 1% of the three-word-or-longer base NPs in natural text. As the NP length increases, the number of possible binary trees (parses) increases with the Catalan numbers (Church and Patil, 1982). NPs of length three have just two possible parses (chance is 50%), while NPs of length six already have forty-two possible …

Content truncated to 500 words