This repository contains data and code for our Interspeech 2024 paper Identifying Speakers in Dialogue Transcripts: A Text-based Approach Using Pretrained Language Models.
End-to-end speaker identification is done in two stages:
- Person name recognition: given a meeting transcript, find spans for person names for each sentence in the transcript.
- Speaker identification: given a meeting transcript and person names, find names for speaker IDs.
The following documentation provides instructions to run the code for models of each of the two subtasks.
Before following the instructions below, make sure that you create a conda environment as below:
conda create -n speaker-id python=3.8
conda activate speaker-id
pip install transformers==3.5.1
pip install torch==1.7.1
pip install trankit==1.1.1
pip install nltk==3.6.7
pip install spacy==3.3.0
pip install thefuzz==0.19.0
pip install six==1.16.0
pip install protobuf==3.20.0
pip install xlsxwriter==3.0.3
We developed a model for detecting person names in text. The model follows a standard NER model architecture: a BERT-based encoder for producing contextualized representations, a Conditional Random Field layer for tagging start and end words of person names based on the BERT representations. The implementation of the model can be found in the class TranscriptNER defined in the file ./individual_model.py.
To train the custom NER model, please run the command:
python train.py --dataset CoNLL03-English-synthetic
Here, we use CoNLL03-English-synthetic, which is a special version containing both the original version of the well-known CoNLL03-English dataset and its lower-cased version, to train the model. This helps improve the robustness of the custom NER model in predicting person names in transcript data, where names could be incorrectly lowercased by an ASR system. The trained model will be stored at logs/CoNLL03-English-synthetic/best-model.mdl. A trained custom NER model can be found at: Google Drive link (mirror).
For speaker identification, we developed two main models:
- Individual Model: given an inidividual person name, the model tries to decide if the name belongs to any speaker around.
- Joint Model: given a sentence containing multiple names, the model tries to decide if each of the given names belongs to any speaker around.
We use the trained NER model to make predictions on Mediasum Corpus to obtain person names on transcripts. Afterwards, text matching is performed to obtain training data for speaker identification. To run this process, use the following command:
python util.py
The preprocessed dataset will be stored at datasets/mediasum. The preprocessed dataset can be downloaded at: Google Drive link (mirror).
To train the models, please use the following command:
python train.py --dataset mediasum --model_type [joint|individual]
In this command:
- [REQUIRED] --model_type [joint|individual] specifies what type of models that we use for training. joint means the Joint Model while individual means the Individual Model.
Trained models can be found at: Google Drive link (mirror).
If you use this work in your research, please cite our paper Identifying Speakers in Dialogue Transcripts: A Text-based Approach Using Pretrained Language Models as follows:
@inproceedings{nguyen2024identifyingspeakersdialoguetranscripts,
title={Identifying Speakers in Dialogue Transcripts: A Text-based Approach Using Pretrained Language Models},
author={Minh Nguyen and Franck Dernoncourt and Seunghyun Yoon and Hanieh Deilamsalehy and Hao Tan and Ryan Rossi and Quan Hung Tran and Trung Bui and Thien Huu Nguyen},
year={2024},
booktitle={Interspeech 2024},
}
The code and model are licensed under the Adobe Research License. The license prohibits commercial use and allows non-commercial research use.
