Improving Open-Domain Dialogue Response Generation with Multi-Source Multilingual Commonsense Knowledge

Authors: Sixing Wu, Jiong Yu, Jiahao Chen, Xiaofan Deng, Wei Zhou

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments have verified the effectiveness of our dataset and approach in monolingual, cross-lingual, and multilingual scenarios.
Researcher Affiliation Academia 1National Pilot School of Software, Yunnan University, Kunming, China 2Engineering Research Center of Cyberspace, Yunnan University, Kunming, China
Pseudocode No The paper describes the Estimate-Cluster-Penalize mechanism and its steps with equations, but it is presented as a textual description within paragraphs and mathematical formulas, not as a structured pseudocode block or algorithm figure.
Open Source Code Yes More details of our LLM instruction prompt pattern and some sampled dialogue cases can be found in our Git Hub project https://github.com/YNLP/Chatbots/tree/main/AAAI2024 MMK-BART.
Open Datasets No The paper states it constructs MMK-Daily Dialog by extending XDaily Dialog (Liu et al. 2023) and aligning Concept Net (Speer, Chin, and Havasi 2016), both of which are cited. However, it does not provide a specific link, DOI, or repository for the constructed MMK-Daily Dialog dataset itself.
Dataset Splits Yes Table 1: The statistics of our MMK-Daily Dialog. #Training 10.5K Sessions and 39.7K Dialogues #Valid/Test 995/996 Sessions and 3.83K/3.69K Dialogues
Hardware Specification Yes Depending on the 24GB V-GRAM of the Nvidia RTX-3090 GPU and the input length, the gradient acclimation step is set to either 4 (when there are 20*4 facts) or 2 (in other scenarios).
Software Dependencies No All methods are implemented with Py Torch and Huggingface library. While the software names are mentioned, specific version numbers are not provided.
Experiment Setup Yes we use the mini-batch of 32 in fine-tuning. Depending on the 24GB V-GRAM of the Nvidia RTX-3090 GPU and the input length, the gradient acclimation step is set to either 4 (when there are 20*4 facts) or 2 (in other scenarios). We use the Adam optimizer and 500 warming steps. We search the learning rate and the epoch number based on the English subset. For m T5, we set the learning rate to 3e-4 and train 5 epochs. For the m BART(MMKBART), we set the learning rate to 2.5e-5 and train 3 epochs. In the inference, we select the last epoch and use the beam width of 5. For our MMK-BART, ECP clusters facts into 10 groups and sets the penalty factor λ to 0.99 by default.