Combining Supervised and Unsupervised Parsing for Distributional Similarity

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1435–1446, Dublin, Ireland, August 23-29 2014. Combining Supervised and Unsupervised Parsing for Distributional Similarity Martin Riedl, Irina Alles and Chris Biemann FG Language Technology Computer Science Department, Technische Universit¨at Darmstadt Hochschulstrasse 10, D-64289 Darmstadt, Germany {riedl,biem}@cs.tu-darmstadt.de, ialles@gmx.de Abstract In this paper, we address the role of syntactic parsing for distributional similarity. On the one hand, we are exploring distributional similarities as an extrinsic test bed for unsupervised parsers. On the other hand, we explore whether single unsupervised parsers, or their combination, can contribute to better distributional similarities, or even replace supervised parsing as a prepro- cessing step for word similarity. We evaluate distributional thesauri against manually created taxonomies both for English and German for five unsupervised parsers. While for English, a supervised parser is the best single parser in this evaluation, we find an unsupervised parser to work best for German. For both languages, we show significant improvements in word similarity when combining features from supervised and unsupervised parsers. To our knowledge, this is the first work where unsupervised parsers are systematically evaluated extrinsically in a seman- tic task, and the first work to show that unsupervised parsing can complement and even replace supervised parsing, when used as a pre-processing feature. 1 Introduction While the field has seen increased interest in automatically inducing syntactic structures from raw or part- of-speech (POS) tagged text, the evaluation of unsupervised data-driven parsers has almost exclusively been conducted either by introspection or by automatic comparison to treebanks. It might be due to comparatively low scores on reproducing a treebank’s syntactic annotation that hardly anyone has yet attempted to use the output of unsupervised parsers for an NLP task other than parsing itself. A further complication with unsupervised parsers – be it dependency parsers, constituency parsers or combinatory categorial grammar parsers – is that the categories induced by such parsers cannot be straightforwardly mapped to linguistically-inspired categories as defined in a treebank. But also when considering only unlabeled syntactic annotations, an unsupervised parser is hardly to blame if it does not adhere to sometimes arbitrary conventions: e.g. for dependencies, it is not a priori clear how to connect auxiliary and main verbs, where to attach the complementizer of subordinate clauses, how to represent a conjunction and its conjuncts, how to relate the preposition and the nominal in prepositional phrases, and how to handle punctuation, cf. Nivre and K¨ubler (2006), Schwartz et al. (2011). When it comes to utilizing syntactic structures, however, it is more important that they are consistent across different sentences than that they adhere to specific syntactic theories and conventions. Here, we choose a task that makes only intermediary use of syntactic structures: we employ unsupervised parsing for preprocessing corpora for the purpose of computing distributional similarities. Since it is generally accepted (e.g. (Lin, 1997; Curran and Moens, 2002)), that syntactic preprocessing plays an important role for the quality of distributional thesauri, and comparing words along their syntactic contexts …

Content truncated to 500 words