A Punjabi To Hindi Machine Translation System

Gurpreet Singh Josan, Gurpreet Singh LehalCOLING 2008

TL;DR

This paper presents a direct machine translation (MT) system for the Punjabi to Hindi language pair, leveraging their syntactic similarities. The system achieves a notable accuracy of 90.67% by employing a combination of lexical resources, rule-based processing, and statistical methods for ambiguity resolution, offering a practical solution for bridging communication gaps between these closely related languages.

What Problem Does This Paper Address?

India's rich linguistic diversity presents a significant challenge for seamless communication. While many Indian languages share common origins and syntactic structures, they are often not mutually intelligible, especially in written form. This necessitates effective machine translation systems to facilitate information exchange. Specifically, Punjabi and Hindi, despite their shared Indo-Iranian roots, have enough differences to hinder direct understanding. The paper addresses the need for a robust and accurate MT system for this pair, aiming to improve communication in government, education, and everyday life.

Methods

The authors developed a direct MT system based on exploiting syntactic similarities between Punjabi and Hindi. The system architecture involves several key stages:

  • Text Normalization: Source Punjabi text is converted to Unicode to handle font variations and simplify processing.
  • Tokenization: The text is broken down into individual words (tokens) for processing.
  • Lexical Resources: A suite of custom-built resources is central to the system:
  • Root word Lexicon: Contains Punjabi root words, their lexical categories, and Hindi equivalents, including gender and verb type information (approx. 33,000 entries).
  • Inflectional form lexicon: Maps inflected Punjabi words to their root forms and Hindi equivalents (approx. 90,000 entries).
  • Ambiguous word lexicon: Lists ambiguous words and their most frequent meanings.
  • Bigram and Trigram Tables: Used for ambiguity resolution, storing contextual word pairs (bigrams) and triplets (trigrams) with their Hindi meanings, derived from a 7-million-word corpus.
  • Translation Engine: This core component handles the translation of each token. It includes:
  • Named Entity Recognition: Identifies proper names for transliteration.
  • Repetitive Construct Handling: Detects and handles repeated words or phrases (e.g., "gharōṁ ghar").
  • Ambiguity Resolution: Employs n-gram language modeling (trigram and bigram tables) to resolve word sense ambiguities. If these fail, the most frequent meaning is used.
  • Transliteration: Handles words not found in the lexicon, including foreign words, through a three-stage process (direct mapping, rule-based improvement, and Soundex-based improvement).
  • Target Language Generation: A rule-based engine refines the translated text for grammatical correctness.

The system was implemented using ASP.net for the front-end and MSAccess for the back-end.

Datasets

The paper mentions the use of a 7-million-word corpus for creating the Bigram and Trigram tables. No other specific datasets for training or evaluation are detailed, beyond the internal lexical resources developed by the authors.

Key Insights and Ideas

  • Direct MT for Related Languages: The paper reinforces the idea that direct MT is a viable and often simpler approach for closely related language pairs, especially when syntactic similarities are high.
  • Importance of Lexical Resources: The success of the system hinges on the comprehensive and carefully constructed lexical resources, particularly the root word and inflectional form lexicons, which are crucial for accurate word mapping.
  • Contextual Ambiguity Resolution: The use of bigram and trigram tables demonstrates a practical approach to resolving word sense ambiguity by leveraging local word context, a common challenge in MT.
  • Handling Repetitive Constructs: Explicitly addressing repetitive word patterns is a nuanced aspect that contributes to more natural-sounding translations.

Key Results

  • Accuracy: The system achieved an accuracy score of 90.67%.
  • Subjective Analysis:
    • Intelligibility rating: 2.76 on a 3-point scale (94% of sentences were intelligible).
  • Fidelity rating: 2.72 on a 3-point scale.
  • Error Analysis:
    • Word Error Rate (WER): 2.34%, which is noted as comparably lower than general MT systems.
  • Sentence Error Rate: 24.26%.
  • Comparative Analysis: The system's accuracy (90.67%) is shown to be comparable to or better than other MT systems for different language pairs (e.g., RUSLAN, CESILKO, Czech-to-Polish, Czech-to-Slovak, Czech-to-Lithuanian).

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the direct translation methodology has inherent deficiencies, though it remains useful for related languages.
  • The output may still require post-editing for perfect grammatical correctness and stylistic nuance.
  • The system's accuracy is comparable to other systems, suggesting room for improvement.
  • Future Work:
    • The authors suggest that the simple word-for-word translation, combined with statistical and rule-based approaches, provides a high accuracy and simple solution. This implies that further refinement of these components could lead to even better results.
  • The paper mentions the development of web services, indicating a direction towards making the MT module more accessible for integration into other applications.

Summary updated

BibTeX
@inproceedings{josan2008punjabi,
  author = {Gurpreet Singh Josan and Gurpreet Singh Lehal},
  title = {{A Punjabi To Hindi Machine Translation System}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2008},
}
Figure