TL;DR
This paper presents two multilingual neural machine translation (NMT) systems for 10 Indic languages and English, leveraging romanized Indic data and a shared subword vocabulary. While the Indic-to-English system achieved strong results, the English-to-Indic system struggled, highlighting challenges with shared target-side vocabularies for diverse Indic languages.
What Problem Does This Paper Address?
Machine translation (MT) systems typically focus on specific language pairs. However, training individual models for every language pair is inefficient, especially for low-resource languages. Multilingual NMT (MNMT) aims to address this by training a single model capable of translating between multiple languages, leveraging transfer learning to improve performance, particularly for under-resourced languages. The paper specifically tackles the challenge of translating between English and a diverse set of Indic languages, which often have different scripts and varying amounts of available parallel data. A key question is how to effectively represent and process these diverse Indic languages within a single NMT framework.
Methods
The authors employ a Transformer-based NMT architecture, a standard choice for state-of-the-art MT. Their core innovation lies in the preprocessing and training strategy:
- Romanization: All Indic language data (both parallel and monolingual) is converted into a romanized script. This allows for a unified representation across different Indic languages, simplifying the input to the NMT model.
- Shared Subword Vocabulary: A single subword vocabulary is created for all romanized Indic languages. This is intended to facilitate knowledge transfer between languages.
- Back-Translation: To augment the limited parallel data, they use back-translation. This involves training an initial Indic-to-English model, using it to translate monolingual English data into synthetic Indic data, and then combining this synthetic data with the original parallel corpus to train the final models.
- Two MNMT Models: They train two distinct models:
- XX → EN: Translates from any of the 10 Indic languages to English.
- EN → XX: Translates from English to any of the 10 Indic languages.
- Language Tags: For the EN → XX model, language tags are prepended to the source English sentences to indicate the target Indic language.
Datasets
The primary datasets used are:
- MultiIndicMT Parallel Corpus: This corpus contains parallel data for 10 Indic languages (Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu) and English.
- PMI Monolingual Corpus: This corpus is used for generating synthetic data via back-translation.
The paper provides statistics on the size of these corpora for each language pair, noting that development and test sets contain 1,000 and 2,390 sentences, respectively.
Key Insights and Ideas
- Romanization as a Unifier: Romanizing Indic languages effectively creates a common ground for NMT, potentially enabling better transfer learning between languages that share linguistic similarities but differ in script.
- Shared Vocabulary's Double-Edged Sword: While a shared vocabulary can promote cross-lingual transfer, it can also hinder performance. The paper's results suggest that a shared romanized subword vocabulary might not be optimal for the English-to-Indic direction, especially for languages with distinct characteristics.
- Indic Languages' Diversity: The significant performance gap between the Indic-to-English and English-to-Indic models underscores the inherent diversity within Indic languages and the challenges of a one-size-fits-all approach for target-side generation.
Key Results
- Indic-to-English (XX → EN): This model performed consistently well across all language pairs. The Hindi-English pair achieved the highest BLEU score of 40.08. Even the language pair with the least data (Odia-English) yielded a respectable BLEU score of 31.19.
- English-to-Indic (EN → XX): This model showed inconsistent performance. While it achieved a high BLEU score of 34.48 for English-Hindi, it struggled significantly with other pairs, notably Tamil-English (8.51), Telugu-English (6.25), and Malayalam-English (3.79).
- RIBES and AMFM Scores: Similar trends were observed with RIBES scores, with the XX → EN model being more consistent. AMFM scores for the EN → XX model were more stable, despite lower BLEU and RIBES.
Limitations and Future Work
- Limitations: The authors explicitly acknowledge that the shared romanized subword vocabulary for the EN → XX model is likely the reason for its poor performance on certain language pairs. They hypothesize that this shared vocabulary might not be adequately helping the decoder during generation.
- Future Work: The paper suggests two potential avenues for improvement:
- Larger Target Vocabulary: Exploring a larger target subword vocabulary size than the current 100K might help.
- Language-Specific Vocabularies: Creating separate subword vocabularies for each Indic language, rather than a single combined one, could lead to better performance. This would involve a joint vocabulary creation process for the romanized data.
Summary updated