[Objective] This paper aims to review the classical and cutting-edge research on the Regression Discontinuity Design(RDD) in the field of policy evaluation, and the prospect of its application in policy evaluation in China is prospected and discussed. [Coverage] We used “Regression Discontinuity” as a keyword to search the Web of Science and CNKI databases. The Chinese and English RDD literature databases were manually constructed for the period 2008-2022. [Methods] We used the bibliometric methods to analyze the application of RDD and conducted a systematic review of RDD in different policy areas. [Results] Our analysis shows that in addition to the dominat fields of education, public health, environment, and public administration, RDD has also appeared in science policy studies and in information and library sciences. The Chinese social science community has also witnessed a rapid growth of RDD research over the past decade. [Limitations] The literature on RDD research needs to be further expanded and comparative analysis with other policy evaluation methods can be further deepened. [Conclusions] RDD research has been widely used in policy evaluation research in the fields of education, public health, environment, and science and technology innovation, etc. In the future, the method can be enhanced with other research methods such as quasi-natural experiments, so as to expand the application of RDD in the analysis of China’s quantitative policy evaluation research and the international quantitative policy evaluation research.

[Objective] This paper aims to comprehensively review and summarize the current development status, mechanism principles, and application trends of question-answering techniques based on large language models. [Coverage] We retrieved a total of 73 relevant papers. [Methods] The study systematically reviews the development status of large language models and efficient parameter fine-tuning strategies. It analyzes the principles, mechanisms, application value, and existing issues of various techniques. It focuses on retrieval-enhanced generation question-answering inference for simple questions and prompt engineering question inference for complex questions. Through qualitative analysis, the research progress of question-answering techniques based on large language models is comprehensively summarized, and future research directions are proposed. [Results] Open-sourced pre-trained large language models continue to emerge, and efficient fine-tuning strategies can significantly improve model adaptability in vertical domains. Retrieval-augmented generation techniques, aided by text embeddings and approximate nearest neighbor retrieval technology, effectively enhance the interpretability and credibility of question-answering. With carefully crafted prompt engineering, the inference capabilities of large models for complex questions can be significantly expanded. [Limitations] The rapid development of research related to large models may result in incomplete coverage of relevant survey work. [Conclusions] Question-answering techniques based on large language models have made remarkable progress in semantic representation, complex reasoning, and other aspects. Retrieval-enhanced generation techniques and prompt engineering, which integrate external knowledge, are the main research hotspots in large models. Future research may focus on exploring aspects such as controllable and credible content generation.

[Objective] This paper aims to solve the problem that traditional automatic summarization technology cannot deeply integrate emotion and topic information synthetically, and cannot solve the lexical deficiency, a review summary generation model integrating emotion and topic information is proposed. [Methods] TextRank is used to dynamically extract the comment topic sentence, and PyABSA model is used to extract the aspect word-emotion word sequence in the topic sentence to concatenate the topic sentence to obtain the final topic information. The emotion sentence is obtained by constructing the emotion word set and Bi-LSTM emotion word extraction model integrating the topic, and the comment text and emotion sentence are concatenated to form dual-channel information with the topic sentence. The attention mechanism is used to obtain topic attention and emotion attention, respectively, and the superposition of them is deeply fused to obtain fusion attention. The single-channel attention of the pointer generation network is replaced, and the final comment summary is generated by the pointer network. [Results] Compared with the comparative experiment Topic+PNG, the proposed pointer generation network with dual-channel information improves the ROUGE-1, ROUGE-2 and ROUGE-L values by 2.87%, 6.14% and 2.64%, respectively. The ablation experiment showed that ROUGE-1, ROUGE-2 and ROUGE-L value of integrating dual-channel information were 4.49%, 3.66% and 4.16% higher than single-channel information. [Limitations] Because fine-grained attribute words may appear in comments, the integration of fine-grained attributes is not considered. [Conclusions] The model can effectively integrate the topic information and emotion information of the comments, improve the quality of the two-channel information fusion, and outperform the comparison model in the summary generation results. The generated summary can contain more emotion and topic information.

[Objective] This paper explores different DeepAL methods for identifying the structural function of scientific literature abstracts and their labeling costs. [Methods] Firstly, we constructed a SciBERT-BiLSTM-CRF model for the abstracts (SBCA), which utilized the contextual sequence information between sentences. Then, we developed an uncertainty active learning strategy for single sentences and full text of the abstracts. Finally, we conducted experiments on the PubMed 20K dataset. [Results] The SBCA model showed the best recognition performance and increased the F1 value by 11.93%, compared to the SciBERT model without sequence information. Using the Least Confidence strategy based on the abstracts, our SBCA model achieved its optimal F1 value with 60% of the experimental data. Using the Least Confidence strategy based on sentences, the SBCA model achieved optimal F1 value with 65% of the experimental data. [Limitations] In the future, we need to examine different active learning strategies in more fields or multi-language datasets. [Conclusions] The new model based on deep active learning could identify the structural function of scientific literature with a lower annotation cost.

[Objective] This paper develops a multilingual sentence aligner for parallel corpora-based research in digital humanities and machine translation. [Methods] The system first encodes the bitext to be aligned in a shared vector space, and then calculates the semantic relationship between sentences based on modified cosine similarity. Finally, a two-stage dynamic programming algorithm is used to automatically extract parallel sentence pairs. [Results] We use both intrinsic and extrinsic evaluation to calculate the performance of the system. The intrinsic evaluation shows that the average accuracy, recall and F₁ values reached 0.950, 0.960 and 0.955. Furthermore, the chrF, chrF++ and COMET scores achieved in the extrinsic evaluation are 55.65, 55.85 and 87.31 respectively. [Limitations] A data capture platform that integrates document alignment and sentence alignment is yet to be developed. [Conclusions] The proposed approach outperforms existing methods in both intrinsic and extrinsic evaluation tasks, which may help to promote the construction of large and high quality multilingual parallel corpora.

[Objective] This study aims to solve problems such as text feature extraction bias and difficult separation of ambiguous semantics caused by the unique expressions and semantic drift phenomenon in Chinese. [Methods] This paper proposes a supervised contrastive learning semantic optimization method, which first uses a pre-trained model to generate semantic vectors, then designs a supervised joint self-supervised method to construct contrastive sample pairs, and finally constructs a supervised contrastive loss for semantic space measurement and optimization. [Results] On the ChnSentiCorp dataset, the five mainstream neural network models optimized by this method achieved F1 value improvements of 2.77%-3.82%. [Limitations] Due to limited hardware resources, a larger number of contrastive learning sample pairs were not constructed. [Conclusions] The semantic optimization method can effectively solve problems such as text feature extraction bias and difficult separation of ambiguous semantics, and provide new research ideas for text sentiment analysis tasks.

[Objective] This paper aims to construct a tweet interaction hypergraph-based rumor detection model that integrates user propagation bias information to improve the accuracy of rumor detection. [Methods] A rumor detection model named UPBI_HGRD is proposed. The model integrates the user propagation bias information when obtaining the tweet node embedding representation, and constructs hyperedges based on user IDs to form a hypergraph that can reflect the interactive relationship of tweets. In addition, this paper proposes a tweet node-user hyperedge level multi-layer dual-level multi-head attention mechanism to focus on important tweet relationships, so as to effectively learn the embedding representation of nodes, and finally input it into a classifier to judge whether it is a rumor or not. [Results] The experimental results on three publicly available datasets show that the accuracy of the model reaches 94.57%, 97.82% and 94.76%, respectively, which is better than the existing baseline model, and has an excellent performance in early detection of rumors, which proves the effectiveness of the model. [Limitations] The limitation of the model in this paper is that the process of obtaining the tweet embedding representation that integrates the user propagation bias information and constructing the hypergraph has a certain time overhead. In the future, further research will be done to improve the time efficiency of the model. [Conclusions] The proposed method effectively improves the accuracy of rumor detection and provides a novel approach to identifying online rumors.

[Objective] This paper aims to improve the accuracy of real-time rumor detection in social media and reduce the harm caused by rumors. [Methods] A real-time rumor detection method based on the graph attention neural network integrated with the knowledge graph is proposed. First, the background knowledge of the text is obtained from the external knowledge graph by knowledge distillation. Second, we transformed the text and background knowledge into a weighted graph structure representation by point mutual information, and a weighted graph attention neural network is used to learn the discontinuous semantic features of the text from the weighted graph. Then, the continuous semantic features of the text are learned by the pre-trained language model BERT, and the statistical features of users and content are converted into continuous vector representations using the embedding method. Finally, all the features are fused and input into the fully connected neural network for rumor detection. [Results] Experimental results on two public social media rumor datasets, PHEME and WEIBO, show that the method’s accuracy reaches 92.1% and 84.0%, respectively, higher than the state-of-the-art baseline methods. [Limitations] The method does not fuse the image or video information that may be attached to the post and cannot perform multi-modal fusion rumor detection. [Conclusions] Fusion of background knowledge can complement the semantic representation of short texts. Fusion of user and content statistical features can support semantic features in decision making and improve the accuracy of the model.

[Objective] This paper aims to achieve a trade-off between running efficiency and accuracy, this paper proposes a voting-based algorithm for identifying a set of influential nodes in social networks named KSEVoteRank. [Methods] Considering the node importance and the neighborhood information, the voting ability of a node is defined and a voting allocation strategy is designed. Meanwhile, an attenuation factor is introduced to discount the voting ability of neighbors. Finally, the node with the highest voting score is iteratively selected as the seed node. [Results] The experimental results show that the influence overlap of a set of influential nodes detected by the KSEVoteRank algorithm in the large social network Ca-AstroPh dataset is about 21% less than that of the VoteRank algorithm. [Limitations] During the repeated voting process, the voting allocation strategy of the neighbors is fixed, which may cause a slight deviation in the theoretical results. [Conclusions] The KSEVoteRank algorithm, based on a voting mechanism, selects a set of influential nodes in a distributed manner to achieve a widespread propagation of influence, which is applicable to large social networks.

[Objective] This paper explores the relationship between personalities and information needs. [Methods] First, we constructed a personality classification model based on linguistic and behavioural characteristics. Then, we utilized key phrase extraction method, Jaccard text clustering, and ERG theory for mining information needs. Finally, we used a one-way analysis of variance and logistic regression analysis to examine the relationship between the public’s personalities and information needs. [Results] Public information needs encompass nine main themes: medical resources, prevention, symptoms, diagnosis and treatment, public symptom emotion sharing, emotional support, disease awareness, social impact, and epidemic development. We categorized the themes into three types: survival information needs, relational information needs, and growth information needs. Survival information needs were positively correlated with conscientiousness; relational information needs were positively related to extraversion and negatively related to neuroticism; growth information needs were positively related to neuroticism and agreeableness. [Limitations] This study does not consider the dynamic nature of information needs. Future studies could analyze the information needs at different stages of public health emergencies. [Conclusions] This paper can help government departments understand the public’s information needs and conduct targeted intelligent information disclosure and risk management.

[Objective] Aimed at the peer review process of scientific research projects, this paper measures the coverage of reviewers’ knowledge on research project topics and constructs expert groups of maximum topic coverage. [Methods] We proposed three principles for recommending reviewer groups for research projects: the maximum topic coverage principle, the maximum knowledge matching principle, and the appropriate workload principle. Then, we developed a method for identifying the research topics of reviewers and projects using the Overlapping K-means. To achieve maximum topic coverage, we constructed a reviewer group recommendation model based on topic coverage, transforming the recommendation problem into an optimization problem. [Results] In two controlled experiments, the reviewer groups constructed by the proposed method increased the topic coverage by 32.38% and 29.01%, respectively. [Limitations] We need to quantitatively explore how to achieve multi-objective optimization for recommending reviewers for research projects according to the three principles. [Conclusions] This research took the reviewer group recommendation for the National Natural Science Foundation of China project application as a case study. It verified the feasibility and effectiveness of the proposed method through qualitative and quantitative analysis.

[Objective] To address the challenges of short texts, high similarity, and single and multiple entailments in the standardization of Chinese medical terminology, this paper proposes a research framework based on the fusion of multiple strategy comparison learning for recall-ranking-quantity prediction. [Methods] Firstly, we integrated text statistical and deep semantic features to retrieve candidate entities. Based on similarity scores, we obtained the candidate set. Secondly, we combined candidate ranking with original terms, standard entities, and candidate entities from recall by training vector representations with pre-trained models and contrastive learning strategies, followed by reordering based on cosine similarity. Next, we updated the vector representations of original terms through multi-head attention to predict the number of standard entities from the original terms. Finally, we selected the standard entities based on the quantity prediction results by integrating the similarity scores of candidate recall and ranking. [Results] We examined the new model on the Chinese medical terminology normalization dataset Yidu-N7k. Compared with statistical models and mainstream deep learning models, the proposed framework achieved an accuracy of 92.17%. This represents an improvement of at least 0.94% over the pre-trained binary classification baseline model. Additionally, on a dataset of 150 expert-labeled reports of mammography examinations for female breast cancer, the new framework’s accuracy reached 97.85%, achieving the best performance. [Limitations] The experiments are only conducted on medical datasets, and the effectiveness in other domains needs further exploration. [Conclusions] A multi-strategy candidate recall can comprehensively consider text information to address the challenge of short text. Contrastive learning candidate rank can capture subtle textual differences to address the challenge of high similarity. Quantity prediction with multi-head attention can enhance vector representation and address the challenges of single and multiple entailments. The proposed method provides the potential for promoting medical information mining and clinical research.

[Objective] This paper proposes a model for extracting patented technology triplets. It tries to improve the accuracy of personalization, fine-grained, multi-dimensional deep extraction, and semantic association. [Methods] We constructed an extraction method based on the WeakLabel-Bert-BiGRU-CRF model for four technical dimensions: problems, solutions, functions, and effects. We evaluated the model using indicators such as the macro average. [Results] We examined the new model with patents in graphene energy storage applications. Compared to the Bert-BiGRU-CRF model, the proposed method achieved a macro average of over 0.8 for triplet extraction and reduced the workload of data annotation. [Limitations] The proposed model requires domain experts and patent analysts in data annotation, and annotation quality affects application effectiveness. [Conclusions] The proposed model could effectively extract patent technology triplets, which has a broad application prospect in scientific literature knowledge mining.