That’ll Do Fine!: A Coarse Lexical Resource for English-Hindi MT, Using Polylingual Topic Models

That’ll do Fine!: A Coarse Lexical Resource for English-Hindi MT, using Polylingual Topic Models Diptesh Kanojia2, Aditya Joshi1,2,3, Pushpak Bhattacharyya2, Mark James Carman3 1 IITB-Monash Research Academy, Mumbai, India, 2 IIT Bombay, Mumbai, India, 3 Monash University, Melbourne, Australia diptesh@cse.iitb.ac.in, adityaj@cse.iitb.ac.in, pb@cse.iitb.ac.in, mark.carman@monash.edu Abstract Parallel corpora are often injected with bilingual lexical resources for improved Indian language machine translation (MT). In absence of such lexical resources, multilingual topic models have been used to create coarse lexical resources in the past, using a Cartesian product approach. Our results show that for morphologically rich languages like Hindi, the Cartesian product approach is detrimental for MT. We then present a novel ‘sentential’ approach to use this coarse lexical resource from a multilingual topic model. Our coarse lexical resource when injected with a parallel corpus outperforms a system trained using parallel corpus and a good quality lexical resource. As demonstrated by the quality of our coarse lexical resource and its benefit to MT, we believe that our sentential approach to create such a resource will help MT for resource-constrained languages. Keywords: Topic models, Machine Translation, Statistical Machine Translation, Coarse Dictionary 1. Introduction Parallel corpora are often injected with bilingual lexi- cal resources for improved machine translation (MT). These lexical resources are often manually created dic- tionaries. Creation of such a resource is a time and effort intensive task. This paper presents a novel approach to use a coarse lexical resource using par- allel topics obtained from multilingual topic models. We observe that for a machine translation system for English-Hindi, these coarse alignments do fine! In a country like India where more than 22 official lan- guages are spoken across 29 states, the task of trans- lation becomes immensely important. A statistical machine translation (SMT) system typically uses two modules: alignment and reordering. The quality of an SMT system is dependent on the alignments discov- ered. The initial quality of word alignment is known to impact the quality of SMT (Och and Ney, 2003; Ganchev et al., 2008). Many SMT based systems are evaluated in terms of the information gained from the word alignment results. However, there is not a lot of parallel data available for these languages making it necessary for specialized techniques that improve alignment quality has been felt (Sanchis and Sánchez, 2008; Lee et al., 2006; Koehn et al., 2007). The ex- isting baseline approach is called Cartesian product Approach. This approach was used by Mimno et al. (2009). In their work, they analyzed the character- istics of MLTM in comparison to monolingual LDA, and demonstrated that it is possible to discover aligned topics. They also demonstrated that relatively small numbers of topically comparable document tuples are sufficient to align topics between languages in non- comparable corpora. They then use MLTM to create bilingual lexicons for low resource language pairs, and provided candidate translations for more computation- Figure 1: Our Sentential Approach to create pseudo- parallel data ally intense alignment processes without the sentence- aligned translations. They conduct experiments for Spanish, English, …

Content truncated to 500 words