ABSTRACT
Integrating linguistic features has been widely utilized in statistical machine translation (SMT) systems, resulting in improved translation quality. However, for low-resource languages such as Thai and Myanmar, the integration of linguistic features in neural machine translation (NMT) systems has yet to be implemented. In this study, we propose transformer-based NMT models (transformer, multi-source transformer, and shared-multi-source transformer models) using linguistic features for two-way translation of Thai-to-Myanmar, Myanmar-to-English, and Thai-to-English. Linguistic features such as part-of-speech (POS) tags or universal part-of-speech (UPOS) tags are added to each word on either the source or target side, or both the source and target sides, and the proposed models are conducted. The multi-source transformer and shared-multi-source transformer models take two inputs (i.e., string data and string data with POS tags) and produce string data or string data with POS tags. A transformer model that utilizes only word vectors was used as the first baseline model for comparison with the proposed models. The second baseline model, an Edit-Based Transformer with Repositioning (EDITOR) model, was also used to compare with our proposed models in addition to the baseline transformer model. The findings of the experiments show that adding linguistic features to the transformer-based models enhances the performance of a neural machine translation in low-resource language pairs. Moreover, the best translation results were yielded using shared-multi-source transformer models with linguistic features resulting in more significant Bilingual Evaluation Understudy (BLEU) scores and character n-gram F-score (chrF) scores than the baseline transformer and EDITOR models.
ABSTRACT
Research into semantic similarity has a long history in lexical semantics, and it has applications in many natural language processing (NLP) tasks like word sense disambiguation or machine translation. The task of calculating semantic similarity is usually presented in the form of datasets which contain word pairs and a human-assigned similarity score. Algorithms are then evaluated by their ability to approximate the gold standard similarity scores. Many such datasets, with different characteristics, have been created for English language. Recently, four of those were transformed to Thai language versions, namely WordSim-353, SimLex-999, SemEval-2017-500, and R&G-65. Given those four datasets, in this work we aim to improve the previous baseline evaluations for Thai semantic similarity and solve challenges of unsegmented Asian languages (particularly the high fraction of out-of-vocabulary (OOV) dataset terms). To this end we apply and integrate different strategies to compute similarity, including traditional word-level embeddings, subword-unit embeddings, and ontological or hybrid sources like WordNet and ConceptNet. With our best model, which combines self-trained fastText subword embeddings with ConceptNet Numberbatch, we managed to raise the state-of-the-art, measured with the harmonic mean of Pearson on Spearman ρ, by a large margin from 0.356 to 0.688 for TH-WordSim-353, from 0.286 to 0.769 for TH-SemEval-500, from 0.397 to 0.717 for TH-SimLex-999, and from 0.505 to 0.901 for TWS-65.
