Minimally Supervised Multilingual Taxonomy and Translation Lexicon Induction

Minimally Supervised Multilingual Taxonomy and Translation Lexicon Induction Nikesh Garera and David Yarowsky Department of Computer Science Center for Language and Speech Processing Johns Hopkins University Baltimore, MD 21218, USA {ngarera,yarowsky}@cs.jhu.edu Abstract We present a novel algorithm for the acqui- sition of multilingual lexical taxonomies (in- cluding hyponymy/hypernymy, meronymy and taxonomic cousinhood), from monolin- gual corpora with minimal supervision in the form of seed exemplars using discriminative learning across the major WordNet seman- tic relationships. This capability is also ex- tended robustly and effectively to a second language (Hindi) via cross-language projec- tion of the various seed exemplars. We also present a novel model of translation dic- tionary induction via multilingual transitive models of hypernymy and hyponymy, us- ing these induced taxonomies. Candidate lexical translation probabilities are based on the probability that their induced hyponyms and/or hypernyms are translations of one an- other. We evaluate all of the above models on English and Hindi. 1 Introduction Taxonomy resources such as WordNet are limited or non-existent for most of the world’s languages. Building a WordNet manually from scratch requires a huge amount of human effort and for rare lan- guages the required human and linguistic resources may simply not be available. Most of the automatic approaches for extracting semantic relations (such as hyponyms) have been demonstrated for English and some of them rely on various language-specific re- sources (such as supervised training data, language- specific lexicosyntactic patterns, shallow parsers, (grenade) haathagolaa (explosive) baaruuda (bomb) bama (gun) banduuka explosive grenade bomb gun weapon Induced Hindi Hypernymy (with glosses) Induced English Hypernymy hathiyaara (weapon) Figure 1: Goal: To induce multilingual taxonomy relation- ships in parallel in multiple languages (such as Hindi and En- glish) for information extraction and machine translation pur- poses. etc.). This paper presents a language independent approach for inducing taxonomies such as shown in Figure 1 using limited supervision and linguis- tic resources. We propose a seed learning based ap- proach for extracting semantic relations (hyponyms, meronyms and cousins) that improves upon existing induction frameworks by combining evidence from multiple semantic relation types. We show that us- ing a joint model for extracting different semantic relations helps to induce more relation-specific pat- terns and filter out the generic patterns1. The pat- 1By generic patterns, we mean patterns that cannot distin- guish between different semantic relations. For example, the 465 terns can then be used for extracting new wordpairs expressing the relation. Note that the only training data used in the algorithm are the few seed pairs re- quired to start the bootstrapping process, which are relatively easy to obtain. We evaluate the taxonomy induction algorithm on English and a second lan- guage (Hindi) and show that it can reliably and accu- rately induce taxonomies in two diverse languages. We further show how having induced parallel tax- onomies in two languages can be used for augment- ing a translation dictionary between those two lan- guages. We make use of the automatically induced hyponym/hypernym relations in each language to create a transitive …

Content truncated to 500 words