This repository contains the implementation of the paper: "Span Classification with Structured Information for Disfluency Detection in Spoken Utterances".
Existing approaches in disfluency detection focus on solving a token-level classification task for identifying and removing disfluencies in text. Moreover, most works focus on leveraging only contextual information captured by the linear sequences in text, thus ignoring the structured information in text which is efficiently captured by dependency trees. In this paper, building on the span classification paradigm of entity recognition, we propose a novel architecture for detecting disfluencies in transcripts from spoken utterances, incorporating both contextual information through transformers and long-distance structured information captured by dependency trees, through graph convolutional networks (GCNs). Experimental results show that our proposed model achieves state-of-the-art results on the widely used English Switchboard for disfluency detection and outperforms prior-art by a significant margin.
All the dependencies are mentioned in requirements.txt file and can be installed using the following command:
pip install -r requirements.txt
For Data related information please refer the information provided in the data directory.
The dataset to be used for this model needs to be preprocessed before feeding it to the model. To do the same we provide a dataprocess/bio2spannerformat.py. First, you need to download datasets, and then convert them into BIO2 tagging format. We have used the switchboard dataset. Note that we use the dependency head index to incorporate the structured information. The data format can be understood by the dummy data provided in the data directory. english_bio is the data in the BIO2 format while english is the preprocessed data.
To run the experiment download and extract the pretrained model in the root directory. For our experiments we use the BERT-Large Architecture. You may need to change the DATA_DIR, PRETRAINED, dataname, n_class to your own dataset path, pre-trained model path, dataset name, and the number of labels in the dataset, respectively.
Note: We provide a dummy dataset in the data directory, as Switchboard dataset that we have used is not an opensource dataset. The dummy data is not a part of the switchboard dataset.
./run.sh
Pretrained models can be downloaded from this link
For running the token classification task use the flair_train.py by using the following command.
python flair_train.py --input input --output output --gpu cuda:1
The input data must be in BIO or IO format in the input directory to run this task.
To calculate the score, use the calculate_score.py which takes gold and prediction files as input.
Our code is inspired by SpanNER and SynLSTM-for-NER
