Contrast and Generation Make BART a Good Dialogue Emotion Recognizer
Authors: Shimin Li, Hang Yan, Xipeng Qiu11002-11010
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments on four datasets demonstrate that our proposed model obtains significantly more favorable results than the state-of-the-art model in dialogue emotion recognition. The ablation study further demonstrates the effectiveness of supervised contrastive loss and generative loss. |
| Researcher Affiliation | Academia | 1 School of Computer Science, Fudan University 2 Peng Cheng Laboratory, Shenzhen, Guangdong, China 3 Shanghai Key Laboratory of Intelligent Information Processing, Fudan University |
| Pseudocode | No | The paper describes the model using mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Additionally, ablation experiments and case studies prove the effectiveness of the contrastive and generative losses in the ERC task1. 1https://github.com/whatissimondoing/Co G-BART. |
| Open Datasets | Yes | Datasets This section will introduce four benchmark datasets: MELD (Poria et al. 2019), Emory NLP (Zahiri and Choi 2018), Daily Dialog (Li et al. 2017), and IEMOCAP (Busso et al. 2008) for comparison with the baseline models. |
| Dataset Splits | Yes | Dataset DD MELD ENLP IEMOCAP #Dial Train 11118 1038 713 120 Dev 1000 114 99 120 Test 1000 280 85 31 #Utter Train 87170 9989 9934 5810 Dev 8069 1109 1344 5810 Test 7740 2610 1328 1623. We conducted a hyperparameter search for model training through the reserved validation set. The results on the test set come from the best checkpoint in the validation set, and we average the scores from five different random seeds. |
| Hardware Specification | Yes | All experiments are performed on Ge Force RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions 'The code framework and initial weight of BART come from Huggingface s Transformers (Wolf et al. 2020)' but does not provide specific version numbers for Python, PyTorch, or the Huggingface Transformers library itself. |
| Experiment Setup | No | The paper mentions 'The optimizer applied for model training is Adam W with a linearscheduled warm-up strategy. The parameters adjusted in this experiment include batch size, learning rate, warm-up ratio, α, and β. We conducted a hyperparameter search for model training through the reserved validation set.' However, it does not provide the specific hyperparameter values (e.g., exact learning rate, batch size, warm-up ratio, or the final chosen α and β values) used for the best results, only listing the types of parameters adjusted. |