Character-based Thai Word Segmentation with Multiple Attentions

thwcc-attn

Yet another Thai Word Segmentation that employs multiple linguistic information, including characters, character-cluster, subwords, and words, with attention mechanisms.

Architecture

• Character-based word segmentation
• BiLSTM-CRF architecture
• Words and Character-cluster/subword attentions integrated to character representations
• BIES tagging scheme
  • B: beginning, I: inside, E: end, and S: single

Segmentation Performance (micro-averaged f1 score)

CWCC-WCON

• CoNLL (word-level): 97.67
  • conlleval
  • A simplified word-level evaluation script
    • https://github.com/tchayintr/word-level-eval.py
• BIES (character-level): 98.99
• Boundary (boundary-level): 99.38
  • http://www.lrec-conf.org/proceedings/lrec2020/pdf/2020.lrec-1.858.pdf

Datasets (based on BEST2010 corpus)

• train, valid, test
  • https://resources.aiat.or.th/thwcc-attn/
  • raw.sl: unsegmented-sentence line
  • seg.sl: segmented-sentence line

Requirements

• python3 >= 3.7.3
• torch >= 1.6.0+cu101 (original: 1.5.0+cu101)
• allennlp >= 1.1.0 (original: 1.0.0)
• numpy >= 1.19.2
• pathlib >= 1.0.1
• gensim >= 3.8.3
• pickle

Modes

• train: train and evaluate a model
• decode: decode an input file (unsegmented text) to a segmented words

Data format

• sl: sentence line

Usage

Modes can be specified by executing the following the sample scripts

Training models

• Character-BiLSTM-CRF(baseline): sample_scripts/sample_seg_ch.sh
  • tagging_unit: single
  • ./sample_scripts/sample_seg_ch.sh
• Character-Transformer-CRF: sample_scripts/sample_seg_tfm.sh
  • tagging_unit: transformer
  • ./sample_scripts/sample_seg_tfm_ch.sh
• W-WCON (strong-baseline): sample_scripts/sample_seg_w.sh
  • tagging_unit: hybrid
  • word-attention
  • ./sample_scripts/sample_seg_w.sh
• CCC-WCON (preliminary):
  • tagging_unit: mutant
  • cc-attention
  • ./sample_scripts/sample_seg_cc.sh
• CWSW-WCON (preliminary): sample_scripts/sample_seg_wsw.sh
  • tagging_unit: sub-combinative
  • word-attention and subword-attention
  • ./sample_scripts/sample_seg_wsw.sh
• CWCC-WCON (The best model): sample_scripts/sample_seg_wcc.sh
  • tagging_unit: combinative
  • word-attention and cc-attention
  • ./sample_scripts/sample_seg_wcc.sh

Logs

• A log file will be saved in log
  • training/evaluating scores
  • hyperparameters

Trained models

• Trained models will be saved in models/main
  • hyperparameters
  • dictionary
  • checkpoint for each break point

Acknowledgement

• Implementations based on modification of seikanlp

Citation

• https://aclanthology.org/2021.ranlp-1.31/