FlatTN

This repository contains code accompanying the paper "An End-to-End Chinese Text Normalization Model based on Rule-Guided Flat-Lattice Transformer" published on ICASSP 2022.

Paper

Requirement

Python: 3.7.3
PyTorch: 1.2.0
FastNLP: 0.5.0
Numpy: 1.16.4
fitlog

For more about FastNLP, please visit here. For Fitlog, please refer to this.

Dataset download

We release a large-scale Chinese Text Normalization (TN) Dataset in corporatioin with Databaker (Beijing) Technology Co., Ltd.

To download the dataset, please visit https://www.data-baker.com/en/#/data/index/TNtts.

(For Chinese version of the download page, please visit https://www.data-baker.com/data/index/TNtts.)

Data preprocessing

The raw dataset in jsonl format are saved at: dataset/processed/CN_TN_epoch-01-28645_2.jsonl

We preprocessed the data into the BMES format, and divided the data into train 、dev 、test by 8:1:1.

dataset/processed/shuffled_BMES
                      ├── train.char.bmes
                      ├── dev.char.bmes
                      └── test.char.bmes

An example of the processed data in BMES format is as follows:

2 B-DIGIT
0 M-DIGIT
1 M-DIGIT
5 E-DIGIT
年 S-SELF
， S-PUNC
只 S-SELF
剩 S-SELF
3 B-CARDINAL
9 E-CARDINAL
天 S-SELF
。 S-PUNC

You can re-run our code to preprocess and divide the raw dataset again:

cd dataset/processed
python preprocess.py

You can also used the following code to get statistics of all NSW categories of the data:

cd dataset/processed
python stat.py

Training

Our code are in version V1, run training code

cd V1
python flat_main.py --dataset databaker

Our proposed rule base are saved in a python file: V1/add_rule.py

Acknowledgement

Our code is based on Flat-Lattice-Transformer (FLAT) from LeeSureman.

For more information about FLAT, please refer to LeeSureman/Flat-Lattice-Transformer.

thuhcsi / FlatTN