With the success of large language models, generative retrieval has emerged as a new retrieval technique for recommendation. It can be divided into two stages: the first stage involves constructing discrete Codes (i.e., codes), and the second stage involves decoding the code sequentially via the transformer architecture. Current methods often construct item semantic codes by reconstructing based quantization on item textual representation, but they fail to capture item discrepancy that is essential in modeling item relationships in recommendation sytems. In this paper, we propose to construct the code representation of items by simultaneously considering both item relationships and semantic information. Specifically, we employ a pre-trained language model to extract item's textual description and translate it into item's embedding. Then we propose to enhance the encoder-decoder based RQVAE model with contrastive objectives to learn item code. To be specific, we employ the embeddings generated by the decoder from the samples themselves as positive instances and those from other samples as negative instances. Thus we effectively enhance the item discrepancy across all items, better preserving the item neighbourhood. Finally, we train and test semantic code with with generative retrieval on a sequential recommendation model. Our experiments demonstrate that our method improves NDCG@5 with 43.76% on the MIND dataset and Recall@10 with 80.95% on the Office dataset compared to the previous baselines.

翻译：随着大型语言模型的发展，生成式检索已成为推荐系统中一种新的检索技术。该方法分为两个阶段：第一阶段构建离散编码（即代码），第二阶段通过Transformer架构顺序解码。现有方法通常基于物品文本表征的量化重构来构建语义编码，但未能捕捉到推荐系统中建模物品关系至关重要的物品差异性。本文提出通过同时考虑物品关系与语义信息来构建物品的编码表征。具体而言，我们利用预训练语言模型提取物品文本描述并将其转化为物品嵌入，随后提出在编码器-解码器结构的RQVAE模型中融入对比学习目标以学习物品编码。具体地，我们将解码器从样本自身生成的嵌入视为正样本，将来自其他样本的嵌入视为负样本，从而有效增强所有物品间的差异性，更好地保留物品邻域结构。最后，我们在序列推荐模型上对语义编码进行生成式检索训练与测试。实验结果表明，相比现有基线方法，本方法在MIND数据集上NDCG@5提升43.76%，在Office数据集上Recall@10提升80.95%。