COLA: Improving Conversational Recommender Systems by Collaborative Augmentation
Authors: Dongding Lin, Jian Wang, Wenjie Li
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two benchmark datasets demonstrate the effectiveness of our method. |
| Researcher Affiliation | Academia | The Hong Kong Polytechnic University {dongding88.lin, jian-dylan.wang}@connect.polyu.hk, cswjli@comp.polyu.edu.hk |
| Pseudocode | No | The paper describes its method using equations and textual explanations, but does not include a structured pseudocode or algorithm block. |
| Open Source Code | Yes | Our code and data are available at https://github.com/Dongding Lin/COLA. |
| Open Datasets | Yes | We conduct experiments on two widely-used CRS datasets, namely REDIAL (Li et al. 2018) and TG-Re Dial (Zhou et al. 2020c). |
| Dataset Splits | Yes | Both two datasets are split into training/validation/testing sets with a ratio of 8:1:1. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used for experiments, such as GPU/CPU models or memory specifications. |
| Software Dependencies | No | We implement our approach with Pytorch. We use the pretrained 300-d word2vec (Mikolov et al. 2013) embeddings1 for Transformer during response generation. During training, we use the Adam (Kingma and Ba 2015) optimizer. |
| Experiment Setup | Yes | The hidden dimension is set to 128 and 300 for the recommendation module and response generation module, respectively. For the BM25 algorithm, the number of the top conversations n is set to 1. The max length of conversation history is limited to 256. For both R-GCN and GCN, the number of layers is set to 2 in consideration of efficacy and efficiency. The embeddings of both RCGN and GCN are randomly initialized. The normalization factor of R-GCN is set to 1.0 by default. We use the Adam (Kingma and Ba 2015) optimizer with an initial learning rate of 0.001 and a gradient clipping strategy to restrict the gradients within [0, 0.1]. The batch size is set to 256. We train our model with 30 epochs for both recommendation and response generation. During testing, we use the greedy search algorithm to generate responses, where the max decoding length is set to 30. |