Interpreting Compound Noun Phrases Using Web Search Queries

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 335–344, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Interpreting Compound Noun Phrases Using Web Search Queries Marius Pas¸ca Google Inc. 1600 Amphitheatre Parkway Mountain View, California 94043 mars@google.com Abstract A weakly-supervised method is applied to anonymized queries to extract lexical interpre- tations of compound noun phrases (e.g., “for- tune 500 companies”). The interpretations explain the subsuming role (“listed in”) that modifiers (fortune 500) play relative to heads (companies) within the noun phrases. Exper- imental results over evaluation sets of noun phrases from multiple sources demonstrate that interpretations extracted from queries have encouraging coverage and precision. The top interpretation extracted is deemed relevant for more than 70% of the noun phrases. 1 Introduction Motivation: Semantic classes of interest to Web users are often expressed as lexical class labels (e.g., “fortune 500 companies”, “italian composers”, “victorinox knives”). Each class label hints at the implicit properties shared among its instances (e.g., general electric, gaetano donizetti, swiss army jet- setter respectively). Class labels allow for the orga- nization of instances into hierarchies, which in turn allows for the systematic development of knowl- edge repositories. This motivates research efforts to acquire as many relevant class labels of instances as possible, which have received particular empha- sis (Wang and Cohen, 2009; Dalvi et al., 2012; Flati et al., 2014). The efforts are part of the larger area of extracting open-domain facts and relations (Banko et al., 2007; Hoffart et al., 2013; Yao and Van Durme, 2014), ultimately delivering richer results in Web search. Different methods can associate instances (gen- eral electric) with both class labels (“fortune 500 companies”) and facts (<general electric, founded in, 1892>) extracted from text. But the class labels tend to be extracted, maintained and used separately from facts. Beyond organizing the class labels hi- erarchically (Kozareva and Hovy, 2010), the mean- ing of a class label is rarely explored (Nastase and Strube, 2008), nor is it made available downstream to applications using the extracted data. Contributions: The method introduced in this paper is the first to exploit Web search queries to uncover the semantics of open-domain class labels in par- ticular; and of compound noun phrases in general. The method extracts candidate, lexical interpreta- tions of compound noun phrases from queries. The interpretations turn implicit properties or subsuming roles (“listed in”, “from”, “made by”) that mod- ifiers (fortune 500, italian, victorinox) play within longer noun phrases (“fortune 500 companies”, “italian composers”, “victorinox knives”) into ex- plicit strings. The roles of modifiers relative to heads of noun phrase compounds cannot be characterized in terms of a finite list of possible compounding re- lationships (Downing, 1977). Hence, the interpreta- tions are not restricted to a closed, pre-defined set. Experimental results over evaluation sets of noun phrases from multiple sources demonstrate that in- terpretations can be extracted from queries for a sig- nificant fraction of the input noun phrases. Without relying on syntactic …

Content truncated to 500 words