Extracting Parallel Sentences from Nonparallel Corpora Using Parallel Hierarchical Attention Network.

Area of Science:

• Natural Language Processing
• Machine Translation
• Computational Linguistics

Background:

• Collecting parallel sentences from nonparallel data is crucial for machine translation.
• Existing methods struggle to learn diverse alignment weights within parallel sentences.

Purpose of the Study:

• To propose a novel parallel hierarchical attention neural network for extracting parallel sentences.
• To address the limitation of existing methods in learning alignment weights.

Main Methods:

• Developed a parallel hierarchical attention neural network.
• Encoded monolingual and bilingual sentences.
• Constructed a classifier to extract parallel sentences.
• Utilized an attention mechanism to learn word alignment weights.

Main Results:

• The proposed model achieved state-of-the-art performance on parallel sentence extraction.
• Demonstrated effectiveness on English-French, English-German, and English-Chinese datasets.
• Successfully learned different alignment weights of words in parallel sentences.

Conclusions:

• The novel attention neural network model significantly enhances parallel sentence extraction.
• The approach improves the quality of machine translation by providing better training data.
• The model's ability to learn alignment weights is a key advancement.