Document Sub-structure in Neural Machine Translation

TL;DR

This paper explores incorporating document sub-structure, specifically section topics, into neural machine translation to improve translation quality by leveraging document organization.

Contribution

It introduces methods to include section topic information in neural MT, adapting ideas from statistical MT and applying them to neural models with new datasets.

Findings

Section-aware models outperform baseline in translation quality.

Side constraints and cache-based methods effectively incorporate sub-structure.

New Wikipedia biography datasets for three language pairs are released.

Abstract

Current approaches to machine translation (MT) either translate sentences in isolation, disregarding the context they appear in, or model context at the level of the full document, without a notion of any internal structure the document may have. In this work we consider the fact that documents are rarely homogeneous blocks of text, but rather consist of parts covering different topics. Some documents, such as biographies and encyclopedia entries, have highly predictable, regular structures in which sections are characterised by different topics. We draw inspiration from Louis and Webber (2014) who use this information to improve statistical MT and transfer their proposal into the framework of neural MT. We compare two different methods of including information about the topic of the section within which each sentence is found: one using side constraints and the other using a cache-based model. We create and release the data on which we run our experiments - parallel corpora for three language pairs (Chinese-English, French-English, Bulgarian-English) from Wikipedia biographies, which we extract automatically, preserving the boundaries of sections within the articles.