Transferring Syntactic Relations from English to Hindi Using Alignments on Local Word Groups

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 447–455, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Transferring Syntactic Relations from English to Hindi Using Alignments on Local Word Groups Aswarth Dara, Prashanth Mannem, Hemanth Sagar Bayyarapu and Avinesh PVS Language Technologies Research Center International Institute of Information Technology Hyderabad, AP, India - 500032 {abhilash.d,prashanth,hemanth.sagar,avinesh}@research.iiit.ac.in Abstract Various works have used word alignments in parallel corpora to transfer information like POS tags, syntactic trees and word senses from source to target sentences. In this paper, we work on the problem of pro- jecting syntactic relations from English to morphologically rich Hindi parallel text. We show the effectiveness of Local Word Groups (LWGs) in simplifying alignments as well as in transferring syntactic depen- dencies by building an alignment model with LWGs as base units and training a de- pendency parser on the relations projected using these LWGs. The LWG alignment model using GIZA++ scores decreases the Alignment Error Rate by 1.16 points when compared to the best GIZA++ model trained on lemmas. We also show that a dependency parser trained on the syn- tactic relations projected using LWGs ob- tained statistical significant improvements over the relations projected using lemmas by a margin of 3.49%. 1 Introduction Data driven dependency parsers rely on the avail- ability of large amounts of annotated data and building them is a time consuming, labour inten- sive and expensive task. Recent efforts in tree- bank creation have looked at the concept of anno- tation projection from a resource-rich language to a resource-poor one in a parallel corpus (Yarowsky et al., 2001). The central idea is to transfer the relevant information from source to target text given the analysis on the source side and word- alignments in a parallel corpora. A projected de- pendency treebank is created by projecting source parse from an automatic parser on to a target sen- tence using word alignments between the source and target sentences (Hwa et al., 2005; Jiang and Liu, 2009; Spreyer and Kuhn, 2009; Spreyer et al., 2010). Difficulties in transferring syntactic relations from the source to target sentences arise mainly due to (i) errors in the source parse, (ii) errors in word alignments and (iii) differences in the depen- dency annotation schemes of the two languages (Ganchev et al., 2009). The third one can be han- dled by systematically identifying the differences in the annotation schemes and applying relevant transformations (Hwa et al., 2005). In this work, we minimize the scope of errors in (i) and (ii) and the effect they have on projection task by proposing a novel technique for projecting syntactic dependencies using alignments between local word groups instead of word forms. The aim is to make the projection task simpler by trans- ferring the relations in source LWGs to the corre- sponding target LWGs first and then get the align- ments and projections on the head words of these LWGs. We show how LWGs can effectively be used to handle the …

Content truncated to 500 words