Abstract: Using modal representations like text to build universal session recommender systems has attracted the attention of the research community. However, these researches are based on the premise that the text information obtained by the recommender system is in the same language. This paper presents a cross-lingual session recommender system, named CLSRec. By using both the pivot language and the target language, the method uses the text representation as input, does not use the item ID, performs text whitening and linear pooling, and provides it to the session recommendation backbone, and finally obtains the score for recommendation. In the pre-training phase, contrastive learning technology is used to learn general knowledge of resource-rich language data. In the fine-tuning phase, user behavior patterns on the target language data set are learned to ensure the generalization of the method and handle cold start scenarios. While learning on the recommended target, the semantic modeling information is retained as much as possible. Experiments on real-world multilingual e-commerce datasets, the method is significantly higher than the current baseline on three different language data on NDCG@10, NDCG@50, Recall@10 and Recall@50. The paper also explores the factors that affect cross-lingual session recommender systems through comprehensive comparative experiments.

Key words: cross-lingual recommendation, session recommendation, large language models, transfer learning, contrastive learning

摘要： 利用文本等模态表征进行通用会话推荐系统的工作是学术界关注的焦点，在单语言环境下展现了优异性能，但其在跨语言环境下的研究仍属于空白区域。提出了一种跨语言会话推荐系统CLSRec，通过结合使用目标语言与大语言模型生成的中间语言，利用预训练模型生成的文本表征作为输入，摈弃传统物品ID依赖，将其进行文本白化后进行线性池化，提供给会话推荐骨架，最终得到用于推荐的评分。该方法分为预训练与微调两个学习阶段，在预训练阶段运用对比学习策略，从资源丰富语言数据提炼通用知识，在微调阶段学习目标语言数据集上的用户行为模式。采用两阶段学习策略，能够处理冷启动情景，在推荐目标上进行学习的同时，尽可能保留语义建模信息。在真实世界上的多语言电商数据集上进行实验，在三个不同语言数据集的NDCG@10、NDCG@50、Recall@10、Recall@50等指标上均实现了对目前基线的显著超越。还通过全面的对比实验，探究了影响多语言会话推荐系统的因素。

关键词: 跨语言推荐, 会话推荐, 大语言模型, 迁移学习, 对比学习