An Urdu semantic tagger - lexicons, corpora, methods and tools

Shafi, Jawad and Rayson, Paul and Adeel Nawab, Rao Muhammad (2019) An Urdu semantic tagger - lexicons, corpora, methods and tools. PhD thesis, SCC (Data Science).

Text (2019jawadshafiphd)
2019jawadshafiphd.pdf - Published Version
Available under License Creative Commons Attribution.
Download (1MB)

Abstract

Extracting and analysing meaning-related information from natural language data has attracted the attention of researchers in various fields, such as Natural Language Processing (NLP), corpus linguistics, data sciences, etc. An important aspect of such automatic information extraction and analysis is the semantic annotation of language data using semantic annotation tool (a.k.a semantic tagger). Generally, different semantic annotation tools have been designed to carry out various levels of semantic annotations, for instance, sentiment analysis, word sense disambiguation, content analysis, semantic role labelling, etc. These semantic annotation tools identify or tag partial core semantic information of language data, moreover, they tend to be applicable only for English and other European languages. A semantic annotation tool that can annotate semantic senses of all lexical units (words) is still desirable for the Urdu language based on USAS (the UCREL Semantic Analysis System) semantic taxonomy, in order to provide comprehensive semantic analysis of Urdu language text. This research work report on the development of an Urdu semantic tagging tool and discuss challenging issues which have been faced in this Ph.D. research work. Since standard NLP pipeline tools are not widely available for Urdu, alongside the Urdu semantic tagger a suite of newly developed tools have been created: sentence tokenizer, word tokenizer and part-of-speech tagger. Results for these proposed tools are as follows: word tokenizer reports $F_1$ of 94.01\%, and accuracy of 97.21\%, sentence tokenizer shows F$_1$ of 92.59\%, and accuracy of 93.15\%, whereas, POS tagger shows an accuracy of 95.14\%. The Urdu semantic tagger incorporates semantic resources (lexicon and corpora) as well as semantic field disambiguation methods. In terms of novelty, the NLP pre-processing tools are developed either using rule-based, statistical, or hybrid techniques. Furthermore, all semantic lexicons have been developed using a novel combination of automatic or semi-automatic approaches: mapping, crowdsourcing, statistical machine translation, GIZA++, word embeddings, and named entity. A large multi-target annotated corpus is also constructed using a semi-automatic approach to test accuracy of the Urdu semantic tagger, proposed corpus is also used to train and test supervised multi-target Machine Learning classifiers. The results show that Random k-labEL Disjoint Pruned Sets and Classifier Chain multi-target classifiers outperform all other classifiers on the proposed corpus with a Hamming Loss of 0.06\% and Accuracy of 0.94\%. The best lexical coverage of 88.59\%, 99.63\%, 96.71\% and 89.63\% are obtained on several test corpora. The developed Urdu semantic tagger shows encouraging precision on the proposed test corpus of 79.47\%.