Model description

The RST-LSTM module is used to refine predictions of a high-performance sequential text classifier (BERT) on documents with rhetorical structure.

Training pipeline

1. The first stage involves fine-tuning the sequential model on the dataset including texts of different lengths and complexity.
2. In the second stage, we freeze the base model and then train a discourse-aware neural module on top of it for the classification of texts with discourse structure.

Prediction pipeline

1. The text is parsed with end-to-end RST parser
2. Predictions are obtained on each discourse unit in the structure with the BERT
3. Non-elementary discourse structures with assigned BERT predictions go through the trained RST-LSTM

RuARG-2022

This repository is for applying this method on RuARG-2022 argument mining shared task.

Requirements

• AllenNLP == 2.9.3
• IsaNLP RST parser for Russian

Code

• *.ipynb - Data analysis, scripts for training and evaluation.
• models_scripts/ - BERT-based and RST-LSTM-based classifiers scripts for AllenNLP.
  • Both classifiers predict two labels (Stance and Premise) jointly.
  • RST-LSTM includes both Child-sum and Binary options for Tree LSTM (no significant difference was found for the current task, Binary by default).

Reference

Further information and examples can be found in our paper:

@INPROCEEDINGS{chistova2022dialogue,
      author = {Chistova, E. and Smirnov, I.},
      title = {Discourse-aware text classification for argument mining},
      booktitle = {Computational Linguistics and Intellectual Technologies. Papers from the Annual International Conference "Dialogue" (2022)},
      year = {2022},
      number = {21},
      pages = {93--105}
}