Learning to Memorize Entailment and Discourse Relations for Persona-Consistent Dialogues
Authors: Ruijun Chen, Jin Wang, Liang-Chih Yu, Xuejie Zhang
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on two large public datasets, Persona Chat and DSTC7-AVSD, demonstrated the effectiveness of the proposed method. Both automatic and human evaluations indicate that the proposed model outperforms several strong baselines in terms of both persona consistency and response coherence. |
| Researcher Affiliation | Academia | 1School of Information Science and Engineering, Yunnan University, Yunnan, China 2Department of Information Management, Yuan Ze University, Taiwan |
| Pseudocode | Yes | Algorithm 1: Latent memory learning |
| Open Source Code | Yes | Our source code is available at https://github.com/Chenrj233/LMEDR. |
| Open Datasets | Yes | Conv AI2 Persona Chat (Dinan et al. 2020) is a chitosan chat dataset based on Persona Chat (Zhang et al. 2018). DSTC7-AVSD (Alamri et al. 2019) provides a conversational question-answering dataset... DNLI (Welleck et al. 2020) is a dialogue inference dataset based on Persona Chat. MNLI (Williams, Nangia, and Bowman 2018) is a multigenre natural language inference corpus... |
| Dataset Splits | No | The paper mentions using the 'Persona Chat original mode' and 'original Persona Chat dev set for the human evaluation' but does not provide specific numerical details (percentages or counts) for the training, validation, or test splits. While these are standard datasets that often come with predefined splits, the paper does not explicitly state them. |
| Hardware Specification | Yes | The proposed model trained on one NVIDIA RTX 3090 with Py Torch framework. |
| Software Dependencies | No | The paper mentions using 'BART-large', 'Adam W', and 'Py Torch framework' but does not specify version numbers for these software dependencies. |
| Experiment Setup | Yes | The proposed model was initialized using BART-large. Adam W (Loshchilov and Hutter 2019) was applied to optimize the model, with an initial learning rate of 8e-6. We used DNLI on Persona Chat and MNLI on DSTC7-AVSD for the ERM learning. The batch size was 64 for training stage 1, and we used a batch size of two with a gradient accumulation of eight for training stage 2. The types of ERM and DDM were both set to 10 for Persona Chat and set to 20 and 5 for DSTC7-AVSD. For dialogue generation, we used a beam search, and the maximum sequence length was set to 50. |