UCSY's Research Repository

Syllable-Based Neural Named Entity Recognition for Myanmar Language

Show simple item record

dc.contributor.author Mo, Hsu Myat
dc.date.accessioned 2019-09-23T07:00:20Z
dc.date.available 2019-09-23T07:00:20Z
dc.date.issued 2019-08
dc.identifier.uri http://onlineresource.ucsy.edu.mm/handle/123456789/2258
dc.description.abstract More and more information is being created at online every day, and a lot of it is the natural language. Until recently, businesses have been unable to analyze this data. But advances in Natural Language Processing (NLP) make it possible to analyze and learn from a greater range of data sources. Additionally, NLP has many central implications on the ways that computers and humans network on our daily life. By promising a bridge between human and machine, and accessing stored information, NLP plays a vital role in the multilingual society. Technologies constructed on NLP are becoming increasingly widespread. Named Entity Recognition (NER), the task of recognizing names in text and assigning those recognized Named Entities (NEs) to particular NE types such as person name, location or organization, is a key component in many sophisticated systems, especially in information retrieval (IR) systems. NER for Myanmar language is essential for the development of Myanmar NLP and it is not an easy task for many reasons. This dissertation aims to develop Named Entity Recognition (NER) for Myanmar language as well as to promote Myanmar NLP research. Myanmar NLP is said to be still developing and has now been struggling to be developed. In the same situation, there are no publicly available resources that can be accessed freely or commercially for language computation so that Myanmar is being regarded as lowresourced language. For this reason, named entity (NE) tagged corpus for Myanmar NER research is manually annotated and constructed as part of this dissertation. The annotated NE corpus is essential for the development of Myanmar NER research. This NE tagged corpus is applied during all the conducted experiments for Myanmar NER and it will also be provided for future NER research. In written style of Myanmar language, there is no regular space between words or phrases. In Myanmar language, syllables are the basic units. Thus, all the experiments are conducted on syllable-level data instead of characters or words in this work. In this study, NER for Myanmar language is built by applying deep neural network architecture which can be said that Long Short-Term Memory (LSTM) - based network. The performance of neural model is also compared with baseline statistical Conditional Random Field (CRF) model. This statistical model totallyiv depends on feature engineering. As Myanmar language is low-resourced language, named dictionary or gazetteers are not available. If these external feature resources are available and feature engineering is carefully done based on knowledge to cover all situations, statistical methods provide a superior result. In this work, it has been proved that unless using additional features, deep neural networks work well on Myanmar NER and outperform baseline statistical CRF model. The best accuracy is achieved with bidirectional LSTM based network architecture. Therefore, this work eliminates the feature-engineering process and does not need to have language or domain knowledge. The proposed syllable-based neural architecture for Myanmar NER model has three main layers: a character sequence layer, a syllable sequence layer, and inference layer. For each input syllable sequence, syllables are represented with their syllable embeddings. The character sequence layer is used to automatically extract syllable level features by encoding the character sequence within the syllable. Convolutional Neural Network (CNN) is applied to learn character sequence feature within each input syllable at character sequence representation layer. The syllable sequence layer takes the syllable representations as input and extracts the sentence level features, which are fed into the inference layer. For the syllable sequence representation, bidirectional LSTM is utilized to learn sentence level feature, and then CRF inference layer is jointly added above the network to tag the name labels. This proposed CNN_BiLSTM_CRF neural model gives the best performance out of the conducted experiments for the Myanmar NER. en_US
dc.language.iso en_US en_US
dc.publisher University of Computer Studies, Yangon en_US
dc.title Syllable-Based Neural Named Entity Recognition for Myanmar Language en_US
dc.type Thesis en_US

Files in this item

This item appears in the following Collection(s)

Show simple item record

Search Repository


My Account