search

RaleLee / DialogueGCN

A preprocessing and training code for DialogueGCN on Dailydialogue and Mastodon dataset. Use Bert base to preprocess the sentences. Based on https://github.com/declare-lab/conv-emotion/tree/master/DialogueGCN
27 stars 3 forks source link

High Overfitting Probability Issue #6

Closed Salihcan29 closed 1 year ago

Salihcan29 commented 1 year ago

I've constructed a multilabel structure working with the DialogueGCN model. I've also tried connecting different encoders like DistilBERT and TinyBERT to the first layer of the model. Additionally, I experimented with the encoders in both trainable and non-trainable formats. I couldn't surpass a 99% train_f1 score and a 60% validation score. However, I need to point out that I am not using the model as a sentiment model but as a classification model. On average, my dialogues contain 6 utterances, and only 3 or 4 utterances are marked or multi-marked. I trained on 2400 dialogues and approximately 14,000 utterances, and despite playing with all the parameters, I'm encountering high overfitting. I'm wondering if this is due to a mistake in my model architecture or if the model's capacity is primarily sentiment-focused.

RaleLee commented 1 year ago

This repo is only a personal preprocess of DialogueGCN model. For further questions please contact the authors of DialogueGCN : )

Salihcan29 commented 1 year ago

Thank you for your answer.