Thai_NLP_platform

The project aims to develop Thai NLP Library based on Deep Learning techniques. With this library, users can train the model on their own dataset using the provided deep learning model structure and utilities. Moreover, Thai NLP Library also provides pre-trained models to process Thai text instantly. All pre-trained models was evaluated and compared across various deep learning techniques proposed in the previous researches in the experiments.

• Tokenization: Identify the boundaries between texts in natural languages to divide these texts into the meaningful unit as word.
• Word Embedding: Map a word from the vocabulary to vectors pf real numbers involving a mathematical embedding.
• Named Entity Recognition: Predict Named entity of each words in a sentence.
• Part-of-Speech Tagging: Predict Part-of-Speech of each words in a sentence.
• Sentiment Analysis: Predict the Sentiment of a document including Positive, Neutral and Negative sentiment.
• Text Categorization: Predict the pre-defined classes of sentences on the specific domain.
• Keyword Expansion: Find the related words from the vocabulary to the query word.

Short introduction for instant models

You could ues all instant models easily by its function. The models will be initialized when import instant_model so that means the import process execution time could be long.

>>> from bailarn.instant_model import instant_model
>>> text = "ฉันกินข้าว"
>>> instant_model.tokenize(text)
['ฉัน', 'กิน', 'ข้าว']
>>> instant_model.pos_tag(text)
['PPER', 'VV', 'NN']
>>> instant_model.ner(text)
['O', 'O', 'O']
>>> instant_model.sentiment("อาหารมื้อนี้อร่อย") # in Food domain
['POS']
>>> instant_model.categorzation("ใช้ไอโฟนของ AIS อยู่") # in Mobile domain
['iPhone', 'AIS', 'โทรศัพท์มือถือ']
>>> instant_model.keyword_expansion("ไอโฟน") in Mobile domain
[('iphone', 0.7972946763038635),
 ('6', 0.7576655149459839),
 ('แอนดรอย', 0.7489653825759888),
 ('โทรศัพท์', 0.7486546635627747),
 ('ios', 0.7471609115600586),
 ('6s', 0.744852602481842),
 ('รู้สึก', 0.7426965236663818),
 ('5s', 0.7402455806732178),
 ('ตก', 0.7397241592407227),
 ('ชอบ', 0.7394053339958191)]
 >>> instant_model.word_embedding("คำ")
 array([ 4.3708846e-02, -2.4344508e-01, -2.0937651e-01, ..., -8.1250496e-02,  3.1024747e-02,  4.8249800e-02], dtype=float32)

Additionally, other examples can be shown in introduction.ipynb

Performance on instant models

• Tokenization: f1-score 98.58%
• Named Entity Recognition: f1-macro 62.10%, f1-micro 79.59%
• Part-of-Speech Tagging: f1-macro 82.41%, f1-micro 95.48%
• Sentiment Analysis: f1-macro 87.47%, f1-micro 87.60%
• Text Categorization: f1-macro 10.73%, f1-micro 16.24%

Installation for instant models

Please make sure that you have keras-contrib, or you need to install keras-contrib by its github (also see https://github.com/keras-team/keras-contrib)

$ pip install git+https://www.github.com/keras-team/keras-contrib.git

Because many files are stored in Git Large File Storage (LFS), you should install Git LFS first by Install Git LFS.You'll suppose to fetch them while cloning and this should take time.

$ git lfs clone https://github.com/KawinL/Thai_NLP_platform.git

If you encounter any problems with the installation, make sure to install the correct versions of dependencies listed in requirements.txt file.

Authors

• Amarin Jettakul - Initial work - ammarinjtk
• Chavisa Thamjarat - Initial work - jijyisme
• Kawin Liaowongphuthorn - Initial work - KawinL

See also the list of contributors who participated in this project.

License

This project is licensed under the GNU LESSER GENERAL PUBLIC License

Acknownledgement

• This project is advised by Assistant Prof. Dr.Peerapon Vateekul, Department of Computer Engineering, Faculty of Engineering, Chulalongkorn University
• The BEST corpus is supported by Thailand's National Electronics and Computer Technology Center (also see https://www.nectec.or.th/en/)

References

• [1] Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space." arXiv preprint arXiv:1301.3781(2013).
• [2] Wutthiphat Phuriphatwatthana, Synthai: Thai Word Segmentation and Part-of-Speech Tagging with Deep Learning. (github)
• [3] Boonkwan, Prachya, and Thepchai Supnithi. "Bidirectional Deep Learning of Context Representation for Joint Word Segmentation and POS Tagging." In International Conference on Computer Science, Applied Mathematics and Applications, pp. 184-196. Springer, Cham, 2017.
• [4] Liu, Jingzhou, Wei-Cheng Chang, Yuexin Wu, and Yiming Yang. "Deep Learning for Extreme Multi-label Text Classification." In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 115-124. ACM, 2017.