Machine Translation from an Intercomprehension Perspective

Yu Chen, Tania AvgustinovaACL 2019

TL;DR

This paper explores machine translation between closely related Slavic languages (Czech and Polish) by leveraging their mutual intelligibility. The authors propose methods that incorporate orthographic and phonological similarities, aiming to improve translation quality without relying on large parallel corpora, but found that current approaches did not outperform baselines.

What Problem Does This Paper Address?

Building high-quality machine translation (MT) systems for under-resourced language pairs is a significant challenge, often addressed by using English as a "pivot" language. However, when the source and target languages are typologically distant from English, this can lead to accumulated errors. This paper investigates an alternative: directly translating between closely related languages by exploiting their inherent similarities. Specifically, the authors focus on Czech and Polish, two West Slavic languages that exhibit a degree of mutual intelligibility. This intelligibility stems from shared linguistic features, making them candidates for intercomprehension-based MT, where speakers of different but related languages can understand each other to some extent. The core problem is how to effectively harness these linguistic similarities to improve direct MT performance, especially when data is scarce.

Methods

The authors propose two main avenues for incorporating intercomprehension principles into MT:

  • Orthographic Correspondences: They leverage the fact that Czech and Polish share the Latin alphabet but use different diacritics and letter combinations. Instead of a full manual rewriting of texts, they explore using correspondence rules to convert Czech letters not present in Polish into their Polish equivalents. This aims to create a more uniform input for the MT system. They also experimented with Byte Pair Encoding (BPE) to identify subword units that capture these orthographic similarities.
  • Phoneme Correspondences: Recognizing that pronunciation can aid comprehension, they propose a multi-source, multi-target NMT architecture. This system takes both the original text and its International Phonetic Alphabet (IPA) transcription as input for both Czech and Polish. The idea is to create "phoneme-infused" translation paths, treating IPA transcriptions as a "new" closely related language.

Datasets

The paper utilized the following datasets:

  • Parallel Corpora: Europarl, Wiki Titles, JRC-Acquis.
  • Monolingual Corpus: News Crawl 2018 for Polish.
  • Development and Test Sets: Disjoint subsets of the shared task's development set were used for training and blind testing.

No new benchmarks were introduced; the evaluation was performed on the official test set using BLEU scores.

Key Insights and Ideas

  • Mutual Intelligibility as a Resource: The core idea is that the inherent similarity between Czech and Polish, leading to mutual intelligibility, can be a valuable resource for MT, potentially reducing reliance on massive parallel data.
  • Orthographic Similarity is Detectable: The authors demonstrate that orthographic correspondences can be identified and potentially utilized, for instance, through BPE, which naturally groups similar subword units.
  • Phonetic Information for MT: The exploration of IPA transcriptions as an input for NMT suggests a novel way to inject phonetic information, aiming to improve translation by bridging pronunciation gaps.
  • Surprising Baseline Performance: The traditional phrase-based statistical machine translation (PBSMT) system surprisingly outperformed a simple recurrent neural network (RNN) based NMT system and was competitive with a Transformer-based ensemble, highlighting the robustness of older methods or the challenges in applying NMT effectively to this specific task with limited data.

Key Results

  • Orthographic Modifications: Applying character replacements based on orthographic correspondences did not improve the PBSMT system's performance on the test set. Joint BPE also showed minimal improvement.
  • Phoneme-Based NMT: The phoneme-based NMT system, while designed to leverage phonetic information, did not yield significant improvements and, in some configurations, seemed to degrade performance.
  • Overall Performance: None of the proposed methods, whether modifying PBSMT or introducing phoneme-based NMT, managed to outperform the established baselines on the final test set. The best Transformer-based NMT system achieved a BLEU score of 11.54, while the PBSMT baseline scored 11.58.

Limitations and Future Work

  • Preliminary Setups: The authors acknowledge that their current setups are preliminary and that the results are not yet competitive.
  • Limited Orthographic Rules: The fixed list of orthographic replacements used was a small subset of potential rules. Future work could involve integrating a more comprehensive set of orthographic correspondences, possibly combined with BPE.
  • G2P Tool Choice: For the phoneme-based system, the choice of grapheme-to-phoneme (g2p) tools and phoneme transcription sets needs further investigation to determine which best suits the task.
  • Grouping Similar Phonemes: Exploring methods to group similar phonemes could be a promising direction for the phoneme-based approach.
  • Architecture Design: The phoneme-based NMT architecture, while showing potential for producing different predictions, seemed to degrade performance due to deadline constraints preventing thorough experimentation. Further refinement of this architecture is needed.

Summary updated

BibTeX
@inproceedings{chen2019machine,
  author = {Yu Chen and Tania Avgustinova},
  title = {{Machine Translation from an Intercomprehension Perspective}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure