A Goal Interaction Graph Planning Framework for Conversational Recommendation

Authors: Xiaotong Zhang, Xuefang Jia, Han Liu, Xinyue Liu, Xianchao Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on two benchmark datasets show that our method achieves significant improvements in both the goal planning and response generation tasks.
Researcher Affiliation Academia Dalian University of Technology, Dalian, China zxt.dut@hotmail.com, jiaxuefang@hotmail.com, liu.han.dut@gmail.com, xyliu@dlut.edu.cn, xczhang@dlut.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/xfjdlut/GIGF.
Open Datasets Yes We conduct experiments on two multi-type conversational recommendation datasets Du Rec Dial and Du Rec Dial 2.0 (Liu et al. 2020), which are commonly used in conversational recommendation.
Dataset Splits Yes Following the previous works (Liu et al. 2020; Wang, Lin, and Li 2023b), we split the Du Rec Dial and Du Rec Dial 2.0 datasets into train/dev/test with 5,400/800/1,804 and 4,126/592/1,325 conversations, respectively.
Hardware Specification Yes The experiments are performed in Python 3.7 and with NVIDIA Ge Force RTX 3090.
Software Dependencies No The experiments are performed in Python 3.7 and with NVIDIA Ge Force RTX 3090. Only Python has a version number specified, and no other specific library versions are provided.
Experiment Setup Yes In our goal planning module, we set the embedding size to 768. The width of the hidden layer is as the same as the embedding size. We use Adam (Kingma and Ba 2015) to train the whole model. The learning rate and the weight decay rate are set to 2e-5 and 0.01. In the goal interaction graph learning module, we set the number of layers to 2 and the dimension of the edge embeddings to 64 with the dropout rate as 0.1. We set α = 0.42 when constructing the goal-entity adjacency matrix Ae.