ConvNTM: Conversational Neural Topic Model

Authors: Hongda Sun, Quan Tu, Jinpeng Li, Rui Yan

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experimental results based on the benchmark datasets demonstrate that our proposed Conv NTM achieves the best performance both in topic modeling and in typical downstream tasks within conversational research (i.e., dialogue act classification and dialogue response generation).
Researcher Affiliation Academia 1 Gaoling School of Artificial Intelligence, Renmin University of China 2 Wangxuan Institute of Computer Technology, Peking University 3 Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education
Pseudocode No The paper describes a
Open Source Code Yes Our code and data are available at https://github.com/ssshddd/Conv NTM.
Open Datasets Yes We conduct the experiments on two widely used multi-turn dialogue datasets, Daily Dialog1 and Empathetic Dialogues2. Daily Dialog (Li et al. 2017) totally contains 13,118 high-quality open-domain daily conversations... Empathetic Dialogues (Rashkin et al. 2019) contains about 25k personal conversations...
Dataset Splits Yes We use the official splits, i.e., 11,118/1,000/1,000. Empathetic Dialogues... We also employ the official splits data, i.e. 19,533/2,770/2,547 for train/val/test respectively.
Hardware Specification Yes We implement the experiments on a Nvidia A40 GPU.
Software Dependencies No The paper mentions the
Experiment Setup Yes For the multi-role interaction graph, we set the window sizes Ks and Kc to 2. The Bo W dictionary size is set to 6,500 in Daily Dialog and 7,533 in Empathetic Dialogues. The embedding size and hidden size of the Transformer, LSTM and GCN are all set to 64. For the loss function, wkl and Wco are set to 0.01 and 0.05, while the value of dco is determined by the number of topics and the dataset. In our main results, dco is recommended to be set to 32 in Daily Dialog and 31.375 in Empathetic Dialogues. The training process has 100 epoches using the Adam optimizer with the base learning rate of 0.001.