---
language: th
license: cc-by-sa-4.0
tags:
- word segmentation
datasets:
- best2010
- lst20
- tlc
- vistec-tp-th-2021
- wisesight_sentiment
pipeline_tag: token-classification
---

# Multi-criteria BERT base Thai with Lattice for Word Segmentation

This is a variant of the pre-trained model [BERT](https://github.com/google-research/bert) model.
The model was pre-trained on texts in the Thai language and fine-tuned for word segmentation based on [bert-base-multilingual-cased](https://huggingface.co/bert-base-multilingual-cased).
This version of the model processes input texts with character-level with word-level incorporated with a lattice structure.

The scripts for the pre-training are available at [tchayintr/latte-ptm-ws](https://github.com/tchayintr/latte-ptm-ws).

The LATTE scripts are available at [tchayintr/latte-ws](https://github.com/tchayintr/latte-ws).

## Model architecture

The model architecture is described in this [paper](https://www.jstage.jst.go.jp/article/jnlp/30/2/30_456/_article/-char/ja).

## Training Data

The model is trained on multiple Thai word segmented datasets, including best2010, lst20, tlc (tnhc), vistec-tp-th-2021 (vistec2021) and wisesight_sentiment (ws160).
The datasets can be accessed as follows: 
- [best2010](https://thailang.nectec.or.th)
- [lst20](https://huggingface.co/datasets/lst20)
- [tlc](https://huggingface.co/datasets/tlc)
- [vistec-tp-th-2021](https://github.com/mrpeerat/OSKut/tree/main/VISTEC-TP-TH-2021)
- [wisesight_sentiment](https://huggingface.co/datasets/wisesight_sentiment).

## Licenses

The pre-trained model is distributed under the terms of the [Creative Commons Attribution-ShareAlike 4.0](https://creativecommons.org/licenses/by-sa/4.0/).

## Acknowledgments

This model was trained with GPU servers provided by [Okumura-Funakoshi NLP Group](https://lr-www.pi.titech.ac.jp).