Design of traditional culture recommendation platform based on multimodal data mining and multivector lexical segmentation

With the acceleration of global informatization, the dissemination of traditional culture faces dual challenges: information overload and heterogeneity of multimodal data, making it difficult for recommender systems to capture user interests and cultural connotations accurately. To address issues of polysemy, ambiguity, and insufficient multimodal data fusion in traditional culture recommendations, this article proposes a recommendation platform design based on multivector lexical-sense disambiguation and multi-view text feature extraction (Multi-VT-RPM). By applying a lexical vector segmentation algorithm to resolve polysemy, enhancing text semantic representation with a multi-head self-attention network, and introducing a multimodal knowledge-sharing mechanism between images and text, we build an efficient cultural recommendation prediction model. Experiments indicate that the proposed lexical sense vector segmentation algorithm achieves superior accuracy in lexical sense disambiguation and semantic similarity tasks on the Senseval-3 and WordSim-240 datasets. Compared with baseline methods on the AION-5B dataset, Multi-VT-RPM improves key performance metrics, achieving an area under the receiver operating characteristic curve of 0.7754, a mean reciprocal rank of 0.3942, and a normalized discounted cumulative gain of 0.6018, respectively. The research results show that the proposed method significantly improves the accuracy of traditional culture recommendation and the efficiency of multimodal fusion, providing strong technical support for the inheritance and personalized promotion of traditional culture.