[Objective] This paper summarizes various methods for interdisciplinary topic identification through a literature review and tries to find shortcomings with potential improvements. [Coverage] We retrieved 74 articles on the concepts and methods of interdisciplinary topic identification from the CNKI and Web of Science databases. [Methods] Based on clarifying the concepts of “interdisciplinarity” and related terms, this paper reviewed the method for interdisciplinary topic identification from three perspectives: recognition based on external characteristics, recognition based on internal features, and a combination of both. [Results] There are still some deficiencies in the existing methods, such as limited data source and identification corpus, insufficient semantics of identification method, coarse identification granularity, a lack of interdisciplinary measurement indicators at the subject level, as well as a lack of forward-looking and dynamic exploration in the identification results. [Limitations] We mainly selected representative literature and did not provide an in-depth exploration of the technical details of interdisciplinary topic identification. We did not review the study of interdisciplinary literature discovery. More research is needed to expand the application of trend tracking and subject clustering in interdisciplinary topic identification. [Conclusions] Future research should expand the identification methods based on multi-source data or full text, improve the semantic mining ability, conduct fine-grained identification, build multi-dimensional interdisciplinary topic measurement indices, and strengthen research on potential interdisciplinary topics and dynamic trends.

[Objective] This paper provides a new model to analyze public opinion from the perspective of sentiment divergence, aiming to address online public opinion events effectively. [Methods] First, we introduced the concept of sentiment disagreement and proposed a multi-level sentiment disagreement algorithm. Then, we constructed a multi-level sentiment disagreement analysis model for online opinion events. This model could calculate sentiment values and disagreement for the online opinion event, comment object, and user layers to perform correlation analysis among the three layers. [Results] Introducing sentiment disagreement can compensate for the lack of research on netizens’ opinion divergence in the original sentiment analysis. This model can identify the critical nodes of public opinion events and the comments generating significant controversy. It also evaluates the effectiveness of public opinion guidance and locates the causes of controversies. [Limitations] We only retrieved the needed data from Sina Weibo (Microblog). More research is needed to collect data from social platforms like Douban and Zhihu. [Conclusions] The proposed model can be applied to monitor the key nodes of public opinion events, select different public opinion guidance methods based on the reasons for controversies, and evaluate the effectiveness of public opinion guidance.

[Objective] This paper tries to efficiently and accurately extract sentiment information from Weibo texts and improve sentiment analysis performance. [Methods] First, we used WoBERT Plus and ALBERT to dynamically encode the character and word-level texts. Then, we extracted key local features with convolution operation. Next, we utilized cross-channel feature fusion and multi-head self-attention pooling operation to extract global semantic information and filter out critical data. Finally, we fused character-level and word-level semantic information using a multi-granularity feature interaction fusion operation and generated the classification results with the Softmax function. [Results] This model’s accuracy and F1 value were 98.51% and 98.53% on the weibo_senti_100k dataset and 80.11% and 75.62% on the SMP2020-EWECT dataset, respectively. Its performance was better than the advanced sentiment analysis models on each dataset. [Limitations] Our model does not include multimodal information such as video, image, and audio for sentiment classification. [Conclusions] The proposed model could effectively accomplish sentiment analysis of Weibo texts.

[Objective] This paper proposes a multimodal sentiment analysis model based on the bidirectional mask attention mechanism (BMAM) to utilize multimodal information and achieve more effective intermodal interaction. [Methods] First, we simultaneously modeled text and speech modalities. The mask attention dynamically adjusted the attention weight for each modality by introducing information from the other modality. Then, we obtained more accurate modality representations, which retained the inherent uniqueness of the modality. We also reduced the differences with the other modality and helped the model choose the optimal sentiment. [Results] The model was evaluated and validated on the general multimodal sentiment analysis dataset-IEMOCAP. The model’s sentiment analysis weighted accuracy rate reached 74.1%, significantly improving the existing mainstream methods. [Limitations] The model has a higher recognition effect on the Neutral and Anger emotional categories, accounting for a larger proportion in the data set. It has a poor recognition performance on the Happy and Sad emotional categories, which account for a smaller proportion in the data set. [Conclusions] The proposed BMAM model can effectively use the interaction between multiple modalities to adjust the attention weight between their emotional elements reasonably and decide sentiment accurately.

[Objective] This paper identifies the literature subjects according to their contents, aiming to meet the needs of interdisciplinary measurement based on the discipline classification of a single paper. [Methods] With the help of the Leuven-Budapest subject classification system, we used machine learning, deep learning, and pre-training language models to classify abstracts from 15 primary disciplines. Then, we used the improved SCIBERT model to conduct interdisciplinary measurement analysis. [Results] The improved SCIBERT model had the best automatic classification performance, with an average F1 score of 81.45%. Some individual categories achieved a classification performance of over 90%. The highest interdisciplinary degree among the 15 primary disciplines was 0.38 for biomedical research, while the lowest was 0.08 for physics. [Limitations] This paper measures the interdisciplinary from the perspective of text content and does not consider multi-dimensional methods for interdisciplinary measurement. [Conclusions] The pre-training model performs best in automatically classifying journal articles, followed by deep learning models. In contrast, machine learning models had the worst performance. Using automatic classification for interdisciplinary measurement based on literature content expanded the current research system and is helpful for a multi-angle and deep understanding of interdisciplinary research.

[Objective] To promote collaboration and academic community building among researchers, this paper proposes a hypergraph-based recommendation algorithm, SCRH. [Methods] Firstly, we constructed a scientific collaboration hyper-network based on hypergraph structure. Then, we created the hypergraph’s structural similarity index based on common neighbors and resource allocation. Next, we built the attribute similarity index of the hypergraph using the author topic model and deep autoencoder. Finally, the two measurement indices were linearly fused to achieve scientific collaboration recommendations. [Results] In the collaboration recommendation task, the AUC and MR index values of SCRH reached 0.88 and 2.35, which were 0.11 and 0.79 better than the optimal metrics of the comparison algorithms. [Limitations] SCRH only considers the author’s text attributes in the node attribute similarity measurement. It needs to fully utilize the author’s citation information, institution information, and publication levels. [Conclusions] SCRH considers the hypergraph’s structural and attribute features. It can effectively accomplish the research collaboration recommendation tasks in stem cells field.

[Objective] This paper tries to alleviate the problems of information overload and low processing efficiency caused by user-generated content in the open innovation of virtual communities. It also aims to optimize the quality of user-generated content, effectively identify and analyze user-generated content, and improve the collaborative innovation performance of virtual communities. [Methods] We proposed a method for identifying the innovation value of user-generated content based on the interactive and collaborative perspective. First, in terms of innovation element features, we introduced the heterogeneity attributes of innovation elements based on user and content attributes. Second, regarding innovation process features, we focused on interactive content’s temporal and collaborative nature. Third, we established a Conv-BiLSTM model incorporating element and process characteristics to identify the value of user-generated content. [Results] We examined the new model with virtual community data. This empirical study showed that the new model’s accuracy reached 88.65%. The introduction of process characteristics increased the model’s accuracy by 14.22%, and the introduction of heterogeneous attributes of collaborative elements increased the accuracy by 6.48%, higher than other baseline or combination models. [Limitations] The model identified innovative content in virtual communities well. The model’s generalization ability needs to be improved for other types of collaborative innovation recognition. [Conclusions] Introducing process characteristics and the heterogenous collaborative innovation element attributes in the user-generated content recognition model of virtual communities improves the recognition accuracy and provides some reference for community open innovation management.

[Objective] This study aims to improve the performance of session-based recommendation models in repeat consumption scenarios and reduce the negative impact of information overload. [Methods] First, we improved the Repeat-Explore Mechanism suitable for repeat consumption scenarios. Then, based on Self-Attention Mechanism, we fused the position information in a non-invasive approach to optimize the utilization of side information. The performance of the new model was validated on public datasets. [Results] Compared to the suboptimal values, the Recall and Mean Reciprocal Rank of the new model on the Yoochoose 1/64 dataset increased by 0.71% and 1.69%, respectively. On the Diginetica dataset, the Recall and Mean Reciprocal Rank were improved by 3.08% and 5.72%. [Limitations] Our experiment only used position information as side information, and the datasets used for verification were limited. [Conclusions] The experimental results verify the effectiveness of the proposed model, which could optimize recommendation systems and improve personalized information services.

[Objective] This paper proposes a Temporal Sequence Generative Adversarial Network (T-SeqGAN) automatically generating online comments, aiming to counteract malicious information on social networks and guide the correct direction of public opinion. [Methods] First, we modified the Sequence Generative Adversarial Network (SeqGAN) generator to a Seq2Seq structure. Then, we used the bidirectional gated recurrent unit (BiGRU) and the sequential convolutional neural network (TCN) as the skeleton network of the encoder and decoder, respectively. Next, we improved the similarity of the syntactic structure and semantic features between the generated posts and the real online comments. Finally, we modified the discriminator of SeqGAN to a model combing TCN and attention mechanism layers to improve the fluency of generated posts. [Results] Compared with the baseline model, the comments generated by the proposed model have significantly higher BLEU-2 (0.799 35), BLEU-3(0.603 96), BLEU-4(0.476 42), and KenLM (-27.670 29)metrics, as well as lower PPL(0.752 47) metrics. [Limitations] The vocabulary and language style of the generated posts are limited by actual posts, and the applicability of our method is limited. [Conclusions] The comments generated by the proposed model have higher syntactic and grammatical correctness and higher similarity to the real-world ones, which can guide the correct direction of public opinion on social networks.

[Objective] This paper addresses data sparsity and dynamic changes in user interests with multimodal feature fusion and deep reinforcement learning. [Methods] First, we used a pre-trained model and attention mechanism to achieve intra-modal representation and fusion of three modalities. Then, we created a model for user-item interactions. Finally, we utilized the deep reinforcement learning algorithm to capture user interest drift and long and short-term rewards in real time to achieve personalized recommendations. [Results] Compared with the highest value in the baseline models, the proposed model improved precision@5 by 11.8%, 16.5%, 11.4%, and NDCG@5 by 5.3%, 8.0%, 6.4%, on the MovieLens-1M, MovieLens-100K, and Douban datasets, respectively. [Limitations] The user interaction history in the Douban dataset is relatively small, and the proposed model cannot learn more accurate user preferences during training. Compared with the experiments on the MovieLens dataset, we received limited recommendation results. [Conclusions] The proposed model integrates multimodal information to reconstruct the state representation network of deep reinforcement learning, improving the recommendation effect.

[Objective] This paper proposes a knowledge fusion method based on fuzzy ontologies, aiming to address the issues of representing and storing uncertain or inaccurate information in large-group emergency decision-making. [Methods] First, we used the multi-granular hesitant fuzzy language to construct fuzzy ontologies. Then, we implemented expert clustering based on K-Means and defined the value measure to determine cluster weights and realize knowledge fusion. Finally, we built an emergency knowledge base for the large group to find the optimal solutions. [Results] The proposed method could represent and store expert knowledge and utilize them in the emergency decision-making of a large group. The case analysis shows that our new method constructed an emergency knowledge base, improved the efficiency of knowledge fusion, and handled multi-stage emergency decision-making. [Limitations] The proposed model did not consider complex relationships among experts and only included the similarity of opinions in expert clustering. The attribute information can also be determined from other dimensions. [Conclusions] This study enriches the method of decision knowledge fusion and provides new directions for multi-stage emergency decision-making of large groups.

[Objective] To solve the problem of low accuracy in classification models for Traditional Chinese Medicine (TCM) patents due to the complexity of TCM and insufficient extracted information on the characteristics of TCM patents. [Methods] We proposed a classification model for TCM patents called MPMFC (Medicine Patent Multi-feature Fusion Classifier). Firstly, we constructed a TCM patent similarity network based on the similarity information of the patent core fields. Then, we used the Node2Vec algorithm to capture the neighborhood structure information of potential patents from the global structure of the TCM patent similarity network, which was mapped to low-dimensional vectors as additional features. Finally, the attention mechanism was utilized to fuse the patent semantic feature vector pre-trained by RoBERTa-Tiny with their corresponding supplementary features to classify TCM patents automatically. [Results] We examined the MPMFC model on a corpus of 7,000 TCM patents. It achieved the accuracy, recall, and F1 values of 0.8436, 0.8017, and 0.822 1, respectively, which were 1.58%, 2.59%, and 2.11% higher than the baseline classification model. [Limitations] The weight allocation when constructing the similarity network of TCM patents has subjectivity issues. There may be some classification errors when Non-TCM researchers label patents. [Conclusions] The MPMFC model can acquire and learn more comprehensive feature representations from multiple perspectives during TCM patent classification, improving classification accuracy.