ValueNet: A New Dataset for Human Value Driven Dialogue System

Authors: Liang Qiu, Yizhou Zhao, Jinchao Li, Pan Lu, Baolin Peng, Jianfeng Gao, Song-Chun Zhu11183-11191

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive empirical results show that the learned value model could benefit a wide range of dialogue tasks.
Researcher Affiliation Collaboration 1UCLA Center for Vision, Cognition, Learning, and Autonomy 2Microsoft Research, Redmond
Pseudocode Yes Algorithm 1: Personalized Dialogue Value Matching
Open Source Code No The dataset is available at https://liang-qiu.github.io/Value Net/. The paper provides access to the dataset but does not explicitly state that the source code for their methodology is available or provide a link to it.
Open Datasets Yes The dataset is available at https://liang-qiu.github.io/Value Net/. The dataset we use for experiments is public available in Parl AI5. EMPATHETICDIALOGUES (Rashkin et al. 2019) provides 25k conversations grounded in emotional situations.
Dataset Splits Yes The data is split into the train (75%), valid (15%), and test (10%).
Hardware Specification Yes For an illustration of computational requirements, the training with MLE on 4 NVIDIA Tesla V100 takes 1 hours, and the reinforcement learning takes 30 minutes.
Software Dependencies No The paper mentions using Transformer variants like BERT, RoBERTa, DistilBERT, and BART, and fastText, but it does not provide specific version numbers for these software components or other dependencies like programming languages or libraries.
Experiment Setup Yes All Transformers are trained for 40 epochs with a learning rate of 5e 6. For the GPT2 (Radford et al. 2019) and Dialo GPT (Zhang et al. 2019) we have finetuned, we train them for 5k steps with a training batch size of 8. The learning rate is set to 2e 6.