As a kind of important social network, online forums with rich and interactive usergenerated contents (UGCs) have shown an explosive rise year by year. Generally, an originally published article often gives rise to thousands of readers’ comments, which are related to specific points of the article or previous comments. This has formed the links of contents, which can provide a very good communication channel between publishers and their audience. Hence it has suggested the urgent need for automated methods to implement the content linking task, which can also help other related applications, such as information retrieval, summarization and content management. Up to now, most of the methods used for content linking are focused on similarity computing based on various traditional grammatical and semantic features. The major problem comes from the disadvantage that they mainly deal with the surface features of texts and words. In order to solve this problem, we propose to adopt deeper textual semantic analysis in this paper. Recently, the Word Embedding model based on deep learning has performed well in Natural Language Processing (NLP), especially in mining deep semantic information. Therefore, we study on the Word Embedding model trained by different neural network models from which we can learn the structure, principles and training ways of the neural network based language models in more depth to complete deep semantic feature extraction. We then put forward a new method for content linking between comments and their original articles for online forums, and verify the validity of the proposed method through experiments and comparison with traditional ways based on feature extraction using two realistic datasets.