Hindi generation from interlingua.
Smriti Singh, Mrugank Dalal, Vishal Vachhani, Pushpak Bhattacharyya, Om Damani · 2007
This paper reports our work on generating Hindi sentences from an interlingua representation called Universal Networking Language (UNL). UNL represents knowledge in semantic net like graphs which contain disambiguated words, binary semantic relations, and speech act like attributes associated with the words. Assisted by a semantically rich lexicon, a priority-matrix of syntax plan, and elaborate morphology synthesis rules, we produce fluent Hindi sentences which also meet the adequacy requirement with respect to the reference sentences, and the faithfulness requirement with respect to the semantic graphs. The system has been tested on agricultural corpora, and the system generated sentences were scored by a team of evaluators. The BLEU scores against the reference sentences have been computed. The results show that our system is able to generate slightly flawed but easy to understand sentences that convey most of the meaning. We observe strong correlation between the fluency scores and the BLEU scores, as well as between fluency and the adequacy scores. Since fluency evaluation does not require reference translation, this correlation facilitates large scale evaluation of our system without translating large number of UNL sentences. This system is a step towards machine translation involving Hindi as the target language. Our approach is also adoptable to the generation of other languages, in particular Indian languages.