Improved Transliteration Mining Using Graph Reinforcement

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1384–1393, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Improved Transliteration Mining Using Graph Reinforcement Ali El-Kahky1, Kareem Darwish1, Ahmed Saad Aldein2, Mohamed Abd El-Wahab3, Ahmed Hefny2, Waleed Ammar4 1 Qatar Computing Research Institute, Qatar Foundation, Doha, Qatar 2 Computer Engineering Department, Cairo University, Cairo, Egypt 3 Microsoft Research, Microsoft, Cairo, Egypt 4 Microsoft Research, Microsoft, Redmond, WA, US {aelkahky,kdarwish}@qf.org.qa1, asaadaldien@hotmail.com2, ahmed.s.hefny@gmail.com2, t-momah@microsoft.com3, i-waamma@microsoft.com4 Abstract Mining of transliterations from comparable or parallel text can enhance natural language processing applications such as machine translation and cross language information retrieval. This paper presents an enhanced transliteration mining technique that uses a generative graph reinforcement model to infer mappings between source and target character sequences. An initial set of mappings are learned through automatic alignment of transliteration pairs at character sequence level. Then, these mappings are modeled using a bipartite graph. A graph reinforcement algorithm is then used to enrich the graph by inferring additional mappings. During graph reinforcement, appropriate link reweighting is used to promote good mappings and to demote bad ones. The enhanced transliteration mining technique is tested in the context of mining transliterations from parallel Wikipedia titles in 4 alphabet-based languages pairs, namely English-Arabic, English-Russian, English-Hindi, and English-Tamil. The improvements in F1-measure over the baseline system were 18.7, 1.0, 4.5, and 32.5 basis points for the four language pairs respectively. The results herein outperform the best reported results in the literature by 2.6, 4.8, 0.8, and 4.1 basis points for the four language pairs respectively. Introduction Transliteration Mining (TM) is the process of finding transliterated word pairs in parallel or comparable corpora. TM has many potential applications such as mining training data for transliteration, improving lexical coverage for machine translation, and cross language retrieval via translation resource expansion. TM has been gaining some attention lately with a shared task in the ACL 2010 NEWS workshop (Kumaran, et al. 2010). One popular statistical TM approach is performed in two stages. First, a generative model is trained by performing automatic character level alignment of parallel transliterated word pairs to find character segment mappings between source and target languages. Second, given comparable or parallel text, the trained generative model is used to generate possible transliterations of a word in the source language while constraining the transliterations to words that exist in the target language. However, two problems arise in this approach: 1. Many possible character sequence mappings between source and target languages may not be observed in training data, particularly when limited training data is available – hurting recall. 2. Conditional probability estimates of obtained mappings may be inaccurate, because some mappings and some character sequences may not 1384 appear a sufficient number of times in training to properly estimate their probabilities – hurting precision. In this paper we focus on overcoming these two problems to improve overall TM. To address the first problem, we modeled the automatically obtained character sequence mappings (from alignment) as a bipartite …

Content truncated to 500 words