Word Alignment Based Parallel Corpora Evaluation and Cleaning Using Machine Learning Techniques

Word Alignment Based Parallel Corpora Evaluation and Cleaning Using Machine Learning Techniques Ieva Zariņa University of Latvia Pēteris Ņikiforovs Tilde Raivis Skadiņš Tilde {lu-ieva.zarina,peteris.nikiforovs,raivis.skadins}@tilde.lv Abstract This paper presents a method for cleaning and evaluating parallel corpora using word alignments and machine learning algorithms. It is based on the assumption that parallel sentences have many word alignments while non-parallel sentences have few or none. We show that it is possible to build an automatic classifier, which identifies most of non-parallel sentences in a parallel corpus. This method allows us to do (1) automatic quality evaluation of parallel corpus, and (2) automatic parallel corpus cleaning. The method allows us to get cleaner parallel corpora, smaller statistical models, and faster MT training, but this does not always guarantee higher BLEU scores. An open-source implementation of the tool described in this paper is available from https://github.com/tilde-nlp/c-eval. 1 Introduction In statistical machine translation, translation quality is largely dependent on the amount of parallel data available. In practice, a large chunk of data considered parallel might not be so, and it can interfere with good data and reduce translation quality. The problem of low quality parallel corpora is getting more and more important because it is becoming popular to build parallel corpora from web data using fully automatic methods. The quality of such corpora often is very low, especially in case of multilingual corpora, which are built by people who do not know the languages they are working with. As a result, we get corpora with broken encoding, many alignment errors and even texts in different languages. The problem can be mitigated by removing blatantly obvious non-parallel text that can be detected with handwritten rules. But that does not help in cases where there are alignment errors or two sentences are kind-of parallel but the translation is wrong or incomplete. The cleaning of such parallel text would require human involvement since devising rules for catching such errors would be nearly impossible. The idea presented in this work is to compare word alignments in a parallel text with those found in a non-parallel text. The intuition being that truly parallel text should have many alignments on word level while unrelated non- parallel text should have few to no alignments. Since word alignment computation is already a step in the training process of many phrase-based statistical machine translation systems, it can be used as input data for the corpus evaluation and cleaning method that we propose. Another benefit of cleaning a corpus is a reduced size, which leads to smaller storage and computational costs of statistical machine translation systems. 2 Related Work This paper is about evaluation and cleaning of parallel corpora, which has been researched from different aspects before. Typically corpus evaluation and cleaning are separate steps in the corpus development process, and corpus development goes through several cycles of evaluation and cleaning while corpus quality reaches acceptable level. Corpus quality is evaluated by both calculating quantitative measurements and assessing its suitability for the purpose. One of …

Content truncated to 500 words