Latvian Twitter Sentiment

Repository for the paper Pretraining and Fine-Tuning Strategies for Sentiment Analysis of Latvian Tweets. Performs 3 class(Positive, Negative and Neutral) classification on Latvian tweets. The model is trained in tweets from the domain of politics.

Data

• Latvian Tweet Corpus. Since the twitter data cannot be shared directly due to twitter terms, kindly refer https://github.com/pmarcis/latvian-tweet-corpus for the data.
• Format - csv with label (0 - neutral; 1 - positive; 2 - negative) and text:

label,text
1,"@maljorka Hehe, man tad labāk garšo bez nekā, nevis šādi. :D Ai, gaumes ir tik atšķirīgas."
0,@IngaStirna Ābolu šarlote.
2,"Šodien bijām pie vecmāmiņas (malka + jāaizved lietas). Es gaidīju, ka paliks labāk, un man viņas pietrūks mazāk, bet mēs ar viņu varējām sarunāties tikai caur logu un pusdienu vietā apēdām bulciņas mašīnā. Nav ok."

How to run

• Train

  • Pre-training - run.sh present bert-twitter-language-pretraining/ folder. Update the paths to the train and eval split.
    • mBERT https://github.com/huggingface/transformers/blob/master/examples/language-modeling/
    • For pretraining ALBERT and ELECTRA https://github.com/shoarora/lmtuners/tree/master/experiments/disc_lm_small
  • Fine-tuning cd bert-sentiment/src;python train.py
    • Update the config.py file with the pretrained models link
    • Update train.py for the locations of the dataset
    • train.py expects pre-processed and pre-split data files( train, dev, test)

• Predict

  • Either train the model from scatch or download the pretrained mode from https://ffzghr-my.sharepoint.com/:f:/g/personal/gthakkar_m_ffzg_hr/EoCPu_Z4dhJPuRPGUipBra0BVLOe9e2a-kHuKGdtOjcoyA?e=7cjzIg
  • To perform a prediction on file, create a csv file with tweets with header text and label. The label column is blank.
  • python predict.py --test_file --model_path . where ignore the model_path flag if it is set in config.py. Pass the test_file created in previous step.

• DEMO

  • To run demo locally python app.py
  • Visit http://cleopatra.ijs.si/sentimentanalyzer/demo for the live demo

Performance Metrics

Accuracy score of around 76% on time-balanced dataset.

Publication

Gaurish Thakkar and Mārcis Pinnis. (2020). Pretraining and Fine-Tuning Strategies for Sentiment Analysis of Latvian Tweets. In Human Language Technologies – The Baltic Perspective - Proceedings of the Ninth International Conference Baltic HLT 2020. 55-61. IOS Press.

BibTeX

@inproceedings{thakkar2020sentiment,
  address = {Kaunas, Lithuania},
  author = {Thakkar, Gaurish and Pinnis, M\=arcis},
  booktitle = {Human Language Technologies – The Baltic Perspective - Proceedings of the Ninth International Conference Baltic HLT 2020},
  doi = {10.3233/FAIA200602},
  pages = {55--61},
  publisher = {IOS Press},
  title = {{Pretraining and Fine-Tuning Strategies for Sentiment Analysis of Latvian Tweets}},
  year = {2020}
}

Acknowledgement

The project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement no. 812997.

This work was done as a part of internship at TILDE.

License

This work is MIT licensed. See the LICENSE file for full details.