Bilingual Parallel Sentence Extraction from Comparable Corpora

Bilingual Parallel Sentence Extraction from Comparable Corpora Chien-Yu Chien, Chin-Hua Chang, Chih-Ping Wei Department of Information Management National Taiwan University r07725026@ntu.edu.tw, r07725031@ntu.edu.tw, cpwei@ntu.edu.tw Abstract This research aims to develop a parallel sentence extraction method for automatically extracting parallel sentence pairs from bilingual comparable corpora based on cross-lingual word embeddings. Our task is to effectively identify matched sentence pairs from a Chinese- English corpus with the goal of maximizing F1 score. Our method employs pre-trained, task- specific, and hybrid (a combination of pre-trained and task-specific) monolingual word embeddings to construct a cross-lingual transformation matrix respectively to transform the word embeddings between the two languages, and develops two search strategies (sequential and exhaustive) for parallel sentence extraction. Our empirical evaluation results suggest that task-specific word embeddings (directly trained from a task-relevant corpus, i.e., 25,695 Chinese and English abstracts of theses) outperforms their counterparts. With respect to the two search strategies, our evaluation results suggest that the exhaustive search strategy attains a higher recall rate; the sequential search strategy is more efficient in time. Both strategies achieve a promising performance, with an F1 score up to 60.18%. 1. Introduction Recently, the tremendous development of neural network techniques for natural language processing has been introduced. Many studies have demonstrated promising results in many important applications, such as neural machine translation [1] and relation extraction [2]. One of the basic requirements for successful neural network model training is a sufficient number of qualified training data examples. In the case of neural machine translation, a bilingual parallel sentence corpus is a required data set. However, many low-resource language pairs (e.g., Chinese- English) or specific domains (e.g., biomedical research) have only limited bilingual parallel sentence corpora, which are not sufficient to support high- performance model construction. Generating parallel sentences by humans is both time consuming and resource intensive. Hence, recent research has focused on how to automatically extract parallel sentences from comparable corpora [3,4,5]. The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. �67-�81 ©The Association for Computational Linguistics and Chinese Language Processing 167 Comparable corpora include non-aligned sentences, phrases or documents that are not an exact translation of each other but share common features such as domain, genre, sampling period, etc. [6]. Compared to parallel corpora, there are more comparable corpora between languages, such as technical documents with bilingual abstracts. Therefore, once the accuracy of the parallel sentence pairs extracted from the comparable corpus can be realized, the problem of lack of a set of parallel sentences as a training corpus for neural machine translation can be effectively relaxed. Recent research on parallel sentence extraction from comparable corpora has shifted from the feature-based approach [6] to the word-embedding-based approach [3,4], due to the advances on cross-lingual word embedding. Several methods for building cross-lingual word embeddings have been proposed [7,8,9]; among them, a popular method is through transformation matrix [8,9]. Most of existing word-embedding-based parallel sentence extraction methods are conducted on European language corpora (such as English and French). Prior research pays less attention …

Content truncated to 500 words