We release UniPunc model code under fairseq_code folder.
The code is implemented based on fairseq.
You can download MUST-C data here, where we use release v1.0.
We also use mTEDx for construct English-mixed subset.
The data split is in data/ folder.
The case study instance and multilingual performance of our ICASSP submission. We provide those results for reviewers' convenience.
This repo does not contain an analysis, which is provided in section 5 of the paper.
We will release the code upon paper acceptance.
Please refer to case.tsv
We also compare UniPunc and other baseline on multilingual sentences from mTEDx, where we select 6 languages, namely English, German, French, Spanish, Portuguese, Italian.
| Overall | Comma | Full Stop | Question Mark | |
|---|---|---|---|---|
| Att-GRU | 54.8 | 48.2 | 65.6 | 32.1 |
| BiLSTM | 53.6 | 46 | 65.2 | 30.1 |
| BERT | 74.7 | 71.5 | 80.1 | 61.1 |
| UniPunc-Mix | 75.4 | 72.1 | 80.8 | 71.3 |
https://ieeexplore.ieee.org/document/9747131
@INPROCEEDINGS{9747131,
author={Zhu, Yaoming and Wu, Liwei and Cheng, Shanbo and Wang, Mingxuan},
booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={Unified Multimodal Punctuation Restoration Framework for Mixed-Modality Corpus},
year={2022},
volume={},
number={},
pages={7272-7276},
doi={10.1109/ICASSP43922.2022.9747131}}