Constructing Parallel Corpora for Six Indian Languages via Crowdsourcing

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 401–409, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Constructing Parallel Corpora for Six Indian Languages via Crowdsourcing Matt Post⋆ and Chris Callison-Burch⋆† and Miles Osborne‡ ⋆Human Langage Technology Center of Excellence, Johns Hopkins University †Center for Language and Speech Processing, Johns Hopkins University ‡School of Informatics, University of Edinburgh Abstract Recent work has established the efficacy of Amazon’s Mechanical Turk for constructing parallel corpora for machine translation re- search. We apply this to building a collec- tion of parallel corpora between English and six languages from the Indian subcontinent: Bengali, Hindi, Malayalam, Tamil, Telugu, and Urdu. These languages are low-resource, under-studied, and exhibit linguistic phenom- ena that are difficult for machine translation. We conduct a variety of baseline experiments and analysis, and release the data to the com- munity. 1 Introduction The quality of statistical machine translation (MT) systems is strongly related to the amount of paral- lel text available for the language pairs. However, most language pairs have little or no readily available bilingual training data available. As a result, most contemporary MT research tends to opportunisti- cally focus on language pairs with large amounts of parallel data. A consequence of this bias is that language ex- hibiting certain linguistic phenomena are underrep- resented, including languages with complex mor- phology and languages with divergent word order- ings. In this paper, we describe our work gather- ing and refining document-level parallel corpora be- tween English and each of six verb-final languages spoken on the Indian subcontinent: Bengali, Hindi, Malayalam, Tamil, Telugu, and Urdu. This paper’s contributions are as follows: • We apply an established protocol for using Amazon’s Mechanical Turk (MTurk) to collect parallel data to train and evaluate translation systems for six Indian languages. • We investigate the relative performance of syn- tactic translation models over hierarchical ones, showing that syntax results in higher BLEU scores in most cases. • We explore the impact of training data quality on the quality of the resulting model. • We release the corpora to the research commu- nity under the Creative Commons Attribution- Sharealike 3.0 Unported License (CC BY-SA 3.0).1 2 Why Indian languages? Indian languages are important objects of study for a number of reasons. These languages are low- resource languages in terms of the availability of MT systems2 (and NLP tools in general) yet together they represent nearly half a billion native speakers (Table 1). Their speakers are well-educated, with many of them speaking English either natively or as a second language. Together with the degree of Inter- net penetration in India, it is reasonably straightfor- ward to find and hire non-expert translators through crowdsourcing services like Amazon’s Mechanical Turk. 1joshua-decoder.org/indian-parallel-corpora 2See sampark.iiit.ac.in/sampark/web/index.php/ content for a notable growing effort. 401 ெனட்டர் அவளை கருத்துக்கள் தயார் senator her remarks prepared Figure 1: An example of SOV word ordering in Tamil. Translation: The senator prepared her remarks. হাত কি ল আম walk CONT PAST 1p Figure 2: An example of …

Content truncated to 500 words