One Tree is not Enough: Cross-lingual Accumulative Structure Transfer for Semantic Indeterminacy

Proceedings of Recent Advances in Natural Language Processing, pages 739–746, Hissar, Bulgaria, Sep 7–9 2015. One Tree is not Enough: Cross-lingual Accumulative Structure Transfer for Semantic Indeterminacy Patrick Ziering Institute for Natural Language Processing University of Stuttgart, Germany Patrick.Ziering @ims.uni-stuttgart.de Lonneke van der Plas Institute of Linguistics University of Malta, Malta Lonneke.vanderPlas@um.edu.mt Abstract We address the task of parsing semanti- cally indeterminate expressions, for which several correct structures exist that do not lead to differences in meaning. We present a novel non-deterministic structure trans- fer method that accumulates all structural information based on cross-lingual word distance derived from parallel corpora. Our system’s output is a ranked list of trees. To evaluate our system, we adopted common IR metrics. We show that our system outperforms previous cross-lingual structure transfer methods significantly. In addition, we illustrate that tree accumula- tion can be used to combine partial evi- dence across languages to form a single structure, thereby making use of sparse parallel data in an optimal way. 1 Introduction Parsing linguistic expressions (e.g., noun phrases (NPs)) is a fundamental component in many natu- ral language processing (NLP) tasks like machine translation (MT) or information retrieval (IR) and indispensable for understanding the meaning of complex units. For example, while [natural lan- guage] processing means the (machine) process- ing of natural languages, natural [language pro- cessing] denotes the natural processing of (any) languages. As previous work has shown, multilingual data can help resolving various kinds of structural am- biguity such as prepositional phrase (PP) attach- ment (Schwartz et al., 2003; Fossum and Knight, 2008), subject/object distinction (Schwarck et al., 2010) or coordination ellipsis (Bergsma et al., 2011). Parallel sentences have been jointly parsed supported by word alignment features (Smith and Smith, 2004; Burkett and Klein, 2008). Yarowsky and Ngai (2001) project part-of-speech (PoS) tags and basic NP structures across languages. Hwa et al., (2005) use projected tree structures for bootstrapping new non-English parsers. Unsu- pervised multilingual grammar induction has been performed on parallel corpora (Snyder et al., 2009) and on non-parallel data (Berg-Kirkpatrick and Klein, 2010; Iwata et al., 2010). In addition to previous work focused on disam- biguation, we show that multilingual data can be used to point to semantic indeterminacy. Syntac- tic structures are usually understood deterministi- cally in that for every structure there exists con- ditions that can have no other structure. How- ever, previous work in NLP shows that such a deterministic take might not always be suitable. Hindle and Rooth (1993) were the first to discuss the phenomenon of semantic indeterminacy in PP attachment, e.g., in the sentence They mined the roads along the coast, the PP along the coast may be attached to both the verb or the object without changing the meaning. On the NP level, Lauer (1995) observed 12.54% semantically in- determinate three-noun compounds (3NCs) in his dataset, e.g., in ’Most advanced aircraft have pre- cision navigation systems’, both precision naviga- tion and navigation system can be bracketed lead- ing to the same meaning. We found more …

Content truncated to 500 words