This is a reproduction on DialogueEIN: Emotion Interaction Network for Dialogue Affective Analysis-COLING22.

Project structure refers to @shenwzh3/DAG-ERC. Features and Dataset can be found there.

• V1: the inputs of the encoder in the semantic interaction network are featrues have been extracted by roberta-large in advance, fixed.

• SSC: remove DialogueEIN structure, substitute with a three layers MLP, hidden size the same with BERT config.hidden_size.

• V2: nearly the same as DialogueEIN raised in original paper, except separate learning rate and the linear transform layer after roberta.

• V3: add linear transform compared to v2

• V4: add separate learning rate compared to v3

• V5: add linear learning rate decay compared to v4

• V6: fix a fatal bug: extended_mask should be -10000 or 0

These experiments trained on JDDC dataset. Params setting is nearly the same as original paper(bsz=8, lr settings are exactly two groups refered in paper, local_att_window_size=7)

Expr1 and Expr3 are feature fixed(using features extracted from chinese-roberta-base-wwm) and all use last four layers cls as utterance feature.

feature-fixed method boost the model performance showing my reproduction is relatively believable.

Some model structure modification and ablation, which are implemented on JDDC Chinese Dataset.

on the reproduction on MELD, I wrongly choose the 12th layer cls repr as the utterance feature using roberta-large, so the result-v4 on MELD is little worse than original paper.

the training process on IEMOCAP isn't that satisfactory...