Deciphering Compatibility Relationships with Textual Descriptions via Extraction and Explanation
Authors: Yu Wang, Zexue He, Zhankui He, Hao Xu, Julian McAuley
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments showcase the model s potential in crafting descriptions that are knowledgeable, aligned with ground-truth matching correlations, and that produce understandable and informative descriptions, as assessed by both automatic metrics and human evaluation. Experiments Experimental Setup Datasets for Stage I To train models fθ and hϕ in accordance with Eq.(3), we construct a composite dataset Combined from Amazon Reviews(Ni, Li, and Mc Auley 2019) and Fashion VC(Song et al. 2017). |
| Researcher Affiliation | Academia | Yu Wang, Zexue He, Zhankui He, Hao Xu, Julian Mc Auley University of California, San Diego {yuw164, zehe, zhh004, hax019, jmcauley}@ucsd.edu |
| Pseudocode | No | The paper does not contain explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code and data are released at https://github.com/wangyuustc/Pair Fashion Explanation. |
| Open Datasets | Yes | Our code and data are released at https://github.com/wangyuustc/Pair Fashion Explanation. We construct a composite dataset Combined from Amazon Reviews(Ni, Li, and Mc Auley 2019) and Fashion VC(Song et al. 2017). |
| Dataset Splits | Yes | We partition the Combined dataset into training, validation, and test sets at a ratio of 8:1:1. The whole dataset is split into training set and testing set with the ratio of 9:1. The held-out testing set is also used for the evaluation in Table 1. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used, such as GPU or CPU models, for running the experiments. |
| Software Dependencies | No | The paper mentions software like 'spacy', 'peft' (with citation), 'Lo RA' (with citation), 'GPT-2', and 'Flan-T5-large/xl', but does not provide specific version numbers for these software packages or libraries. |
| Experiment Setup | Yes | Stage I Details We partition the Combined dataset into training, validation, and test sets at a ratio of 8:1:1. We jointly train the extractor fθ and the classifier hϕ as depicted in Eq.(3) to acquire the extractor fθ. For Cross-Attn, we apply lasso regularization with a weight value of 0.01. For Rationale Extraction, we fix the selection ratio at 0.3, indicating we select 30% of text from ti and tj as the rationale. Stage II Details We employ prompts fi ci and fj cj match because and Generate the reason why fi ci and fj cj match: for GPT2 and Flan-T5, respectively. For GPT2, we set the batch size as 5 to train the whole model. While for Flan-T5-large and Flan-T5-xl, we use the package peft (Mangrulkar et al. 2022) with the method Lo RA (Hu et al. 2022) to finetune. For Flan-T5-large, we set the batch size as 5. For Flan-T5-xl, we set the batch size as 1 but accumulate the gradient over 5 iterations. The learning rate for all models is set to 2e-4. |