DiscoFlan: Instruction Fine-tuning and Refined Text Generation for Discourse Relation Label Classification
Kaveri Anuranjana · 2023
This paper introduces DiscoFlan, our system for the DISRPT 2023 shared task on discourse relation classification.We leverage recent advances in NLP finetuning and use Flan-T5 as a multilingual discourse relation classifier.Our model uses multilingual instructional prompts to finetune on datasets from different languages and generate relation labels as classification outputs.The model's hyperparameters are tuned to enable efficient label generation by finetuning on low-resource datasets.Moreover, we introduce a post-processing step to tackle the problem of label mismatches caused by the generative nature of a seq2seq model by using the label distribution.In contrast to the previous state-of-the-art model, our approach eliminates the need for hand-crafted features in computing the discourse relation classes.Overall, DiscoFlan showcases how instruction finetuning can perform multilingual discourse relation classification for the DISRPT 2023 discourse relation classification shared task.