Conversion can require a reading decision

Software can transliterate Punjabi usefully, but a plausible output is not a guarantee of the intended reading. Saini and Lehal’s 2011 study identifies missing short vowels and characters with several possible mappings as causes of ambiguity in Shahmukhi-to-Gurmukhi conversion. Some input words therefore produce more than one possible target form. Selecting among them requires contextual evidence, rather than a simple replacement of one letter by another.[1]

Their system tested statistical methods for choosing between these alternatives. This matters for a historian using a converted passage: the program can be making an interpretive choice even when the output looks fluent. The original text should remain available beside the converted version so that a disputed reading can be checked.[1]

Ask what an accuracy percentage measures

A different task appears in Kaur and Singh’s 2014 English-to-Punjabi study. It reports approximately 95 per cent word accuracy on 10,000 news-domain units. That result describes a particular direction, dataset and evaluation. It does not mean that every passage was completely correct or that the same system could reliably convert Shahmukhi manuscripts.[2]

The paper distinguishes transliteration from translation: representing an English word’s sound in Punjabi script does not supply its Punjabi meaning. It also uses different rules for different sound and spelling patterns. Before comparing accuracy claims, establish whether two tools are performing the same conversion and whether the reported measure counts words, characters or complete texts.[2]

Keep the output traceable and test the hard cases

The 2011 paper reports word-disambiguation performance above 92 per cent, while still documenting failures, including named-entity abbreviations and cases poorly represented in training data. This is a different measure from complete end-to-end correctness of a historical document. Names, abbreviations and unusual vocabulary deserve particular attention when a converted text is used to support a claim.[1]

The practical inference from the 2014 evaluation is to test material from the intended task, including its vocabulary, spelling patterns and direction of conversion. Preserve the source and identify the tool. Keep automatic output distinct from a verified transcription or a translation. Software can reduce a reading barrier while leaving particular readings and meanings open to correction.[2]

Sources & notes

Source records include access dates and, where known, limits on what was examined. Checks used AI assistance; independent human or specialist review has not been completed.

  1. Tejinder Singh Saini and Gurpreet Singh Lehal / Proceedings of the 9th Workshop on Asian Language Resources. Word Disambiguation in Shahmukhi to Gurmukhi Transliteration ↗primary computational-linguistics research · Source date: 2011-11 · Locator: Printed pp.79–80,85–86 / PDF pp.1–2,7–8: ambiguity, evaluation tables and failure cases. Accessed 2026-11-01.Actual abstract, ambiguity discussion, experimental tables and failure cases inspected. Word-disambiguation accuracy is not relabelled whole-document accuracy. No new native-script example is reproduced.Return to Conversion can require a reading decision, paragraph 1 ↑Return to Conversion can require a reading decision, paragraph 2 ↑Return to Keep the output traceable and test the hard cases, paragraph 1 ↑
  2. Kamaljeet Kaur and Parminder Singh / Proceedings of the 11th International Conference on Natural Language Processing. English to Punjabi Transliteration using Orthographic and Phonetic Information ↗primary computational-linguistics research · Source date: 2014-12 · Locator: Printed pp.315–316,319–320 / PDF pp.1–2,5–6: task definition, rules, evaluation table 8 and conclusion. Accessed 2026-11-01.Actual abstract, introduction, design, evaluation table and conclusion inspected. The 2014 English-to-Punjabi and 2011 Shahmukhi-to-Gurmukhi papers concern different tasks; their results are not directly ranked. No current product is certified.Return to Ask what an accuracy percentage measures, paragraph 1 ↑Return to Ask what an accuracy percentage measures, paragraph 2 ↑Return to Keep the output traceable and test the hard cases, paragraph 2 ↑
Help improve this entry. Suggest a correction with a source.
Record PH-JOU-0008 · English edition · human review not performed