Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Knowledge-Bridged Causal Interaction Network for Causal Emotion Entailment
Authors: Weixiang Zhao, Yanyan Zhao, Zhuojun Li, Bing Qin
AAAI 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our model achieves better performance over most baseline models. |
| Researcher Affiliation | Academia | Research Center for Social Computing and Information Retrieval Harbin Insititute of Technology, China EMAIL |
| Pseudocode | No | The paper includes Figure 3, which is an overall architecture diagram, not structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Our source code is publicly available at https://github.com/circle-hit/KBCIN. |
| Open Datasets | Yes | We conduct experiments on the benchmark dataset RECCON-DD. It is collected from the popular dataset Daily Dialog (Li et al. 2017) with utterance-level emotion labels and the emotion cause labels are annotated by Poria et.al (2021). |
| Dataset Splits | Yes | Statistics of the processed RECCON-DD are shown in Table 1. Table 1: Dataset statistics: Train Valid Test Positive Causal Pairs 7,027 328 1,767 Negative Causal Pairs 20,646 838 5,330 Num. of Dialogue 834 47 225 Num. of Utterance 8,206 493 2,405 |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models or memory specifications used for experiments. |
| Software Dependencies | No | The paper mentions software like 'Transformer encoder (Vaswani et al. 2017)', 'COMET (Bosselut et al. 2019)', 'BART-based (Lewis et al. 2020)', and 'Adam W optimizer', but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | For utterance-level feature extraction, the dimension of hidden states in utterance encoder is 768, and the number of transformer encoder layer is 8 with 10 attention heads. Layers of emotion embedding and relative position embedding are randomly initialized and the dimension of both embedding layers are 300. Also, for all representations in the following parts of KBCIN, dh is set to 300. For causal utterance prediction, dimensions of MLP is set to [300, 300, 300, 1] and the dropout rate is set to 0.07. We utilize Adam W optimizer with learning rate of 4e-5 and L2 regularization of 3e-4 to train our model. And the batch size is 8. We pick the model which works best on the valid set, and then evaluate it on the test set. All of our results are averaged on 5 runs. |