Patent

[authorization] A BERT and MLM based noise reduction method for remote supervised relation extraction. ZL202110525465.9. 2023.07.21. Tielin Shen, Daling Wang, Shi Feng, Yifei Zhang

The present invention discloses a noise reduction method based on BERT and MLM for remote supervised relationship extraction, which relates to the technical field of remote supervised relationship extraction. The entity pair is defined as source entity and target entity, monadic word entities are screened from the target questions corresponding to each category, and these sentences are sorted according to the possibility that their corresponding sentences are not noisy sentences and a sentence set is formed. Positive and negative samples are selected from each sentence set to establish data sets for each category. Based on the hidden state output by the encoder at the last layer of BERT model, the MASK-lhs features of sentences are extracted from the entities of sentences in the data set. The MASK-lhs features of noisy sentences and non-noisy sentences are captured by the full connection layer to train the binary classifier. The trained binary classifier is used as a special noise reduction device for all classes. All sentences in the training set are sent to the corresponding noise eliminator according to the categories to find and eliminate the noisy sentences. Noise canceller is plug-and-play, practical.

[authorization] An approach of building user profile and its application for scholars. ZL201910976349.1. 2023.04.07. Daling Wang, Yinghao Chen, Shi Feng, Yifei Zhang

The present invention provides a scholar - oriented user profile construction and application method. Firstly, the scholar basic information is obtained from the personal home pages of domestic scholars, and the scholar research information is obtained from famous academic websites at home and abroad. On this basis, the above information is preprocessed, so as to obtain the corpus required for the construction of scholars' profile. Then, by mining the basic attributes and research attributes of scholars, the scholar profile is constructed. Finally, based on the scholar profile, the application of scholar search and expert discovery, paper reviewer recommendation, cooperative scholar recommendation and so on can be implemented. The invention supports and assists scholars looking for collaborators on projects and papers, conference and journal sponsors looking for paper reviewers, and senior scholars looking for starters in a new field of research.

[authorization] A multi-labeled emotion intensity prediction method based on hierarchical convolutional neural network. ZL201910751989.2. 2022.11.29. Shi Feng, Hongliang Xie, Daling Wang, Yifei Zhang

The present invention proposes a multi-labeled emotion intensity prediction method based on hierarchical convolutional neural network, which includes: (1) Dividing the original multi-labeled social media essays into a training set and a test set. (2) The data of a raw multi-labeled social media short text in the training set were preprocessed to obtain the basic single-tagged emotional data of the training set. (3) A single-labeled emotion classification model based on hierarchical convolutional neural network was constructed. (4) The emotional intensity value model was constructed based on the attention convolutional neural network. (5) For the test data of the multi-labeled social media short text, the single-labeled emotion classification model of the hierarchical convolutional neural network was used for prediction, and the optimized multi-labeled emotion intensity vector was obtained. The multi-label emotional intensity prediction method based on the hierarchical convolutional neural network of the present invention can further improve the accuracy of emotional intensity prediction in social media texts, which is especially suitable for situations in which multiple basic emotions exist simultaneously in texts.

[published] An approach of constructing and displaying multimodal emotion knowledge graph. ZL202011319237.8. 2021.09.07. Daling Wang, Jiufeng Li, Shi Feng, Yifei Zhang

The present invention provides a method for constructing and displaying a multimodal emotional knowledge graph, which relates to the technical field of multimodal knowledge management. The method firstly collects and downloads the ontology or dictionary of emotional vocabulary, images of facial and animal expressions, and emoticons commonly used on social media platforms. The obtained text and image data resources are preprocessed respectively to make these data resources meet the requirements of the unified abstract ontology model. Then, based on the preprocessing results of text emotion words, facial expression images and social media emoji, a multi-modal emotional knowledge graph was constructed to express different modal entities and their relationships. Finally, the nodes, edges and attributes of the multi-modal emotional knowledge graph are displayed and rendered by using the front-end text markup language and the front-end page element rendering language. On this basis, the user interaction function of the graph is realized. At the same time, the dynamic effect of knowledge graph is increased through the graph visualization library.

[authorization] A local adaptive optimization method based on transitive relation for knowledge graph. ZL201910154252.2. 2020.09.01. Daling Wang, Hongchen Liu, Shi Feng, Yifei Zhang

The present invention proposes a local adaptive knowledge graph optimization method based on transfer relation, which includes: (1) Set training sample set. (2) Set that any r_i and e_i initially belong to a certain distribution. (3) Normalized. (4) Constitute a new training sample set. (5) Initializes the set of triples to be empty. (6) Set the correct triplet, which corresponds to the wrong triplet, replace the correct triplet with the wrong triplet head entity or tail entity, form the wrong training sample set, and merge into the triplet set. (7) Obtain the edge parameters of its entity. (8) Obtain edge parameters of the relationship. (9) Calculate that parameters of edge parameters vary with entities and relationships. (10) Obtain a new loss function based on transitive relation. (11) Judge and optimize each entity or relationship vector using SGD. The invention can make up for the incompleteness of data, and can better express the potential different semantics between relations and entities. The new knowledge graph constructed after optimization has higher accuracy.

[authorization] A personalized clothing recommendation method based on content analysis of clothing image and label text. ZL201710947454.3. 2020.06.09. Daling Wang, Dandan Sha, Shi Feng, Yifei Zhang, Ge Yu

The present invention puts forward a recommended method of personalized clothing based on clothing image and the label text content analysis of the, including the following steps: (1) Through clothing in the image of shopping website details properties were analyzed, and establish a clothing detail parts and features form the image of clothing goods model and user preferences model. (2) Through the analysis of the text of clothing labels on shopping websites, establish the clothing commodity model described in text form and user preference model. (3) Combine the clothing commodity model based on clothing image established in step (1) with the user preference model established in step (2) to produce the recommendation results. The invention combines the image and text information of clothing, and makes personalized clothing recommendation for users based on the fusion of the two models mentioned above.