[Objective] This paper summarizes and comments on the research methods of text semantic parsing with neural networks in the past decade. [Coverage] With Google Scholar and CNKI as the data retrieval platforms, and “Neural Semantic Parsing” as the keywords, all relevant papers and their important citations from 2010 to 2022 were retrieved for analysis. [Methods] The paper classified the existing neural semantic parsing methods according to the technical path, explained the basic ideas of each technical path, compared and analyzed the similarities and differences of each technology method in data, performance, application goals, etc., and summarized the existing problems and development tendency of text neural semantic parsing technology. [Results] Neural semantic parsing methods could be summarized into three types, sequence to sequence method, intermediate form based method, and semantic unit decomposition and combination method. The latter two methods are improvements to the first method. At present, intermediate representations such as semantic sketch, canonical utterance and few-shot neural semantic parsing are the main research focuses. [Limitations] The paper mainly summarized and analyzed the existing research ideas from the methodology, but does not elaborate the internal implementation mechanism of the neural semantic parsing models. [Conclusions] The neural semantic parsing method gains the best performance in text semantic parsing at present. The current popular practice is to design targeted neural network models for specific applications. But the effect of semantic parsing is still far from the practical application.

[Objective] This paper analyzes and summarizes the current situation of the development of text sentiment in financial markets, and provides reference for subsequent related research. [Coverage] We used “financial market”, “text sentiment analysis”, “text sentiment” and “investor sentiment” as keywords to search on academic platforms such as CNKI, Web of Science and Google Academic, and extended the search for relevant literatures. A total of 115 papers were reviewed. [Methods] We classified the extracted text sentiment according to the type of the source financial text, then introduced the framework of text sentiment analysis, and finally sorted out the relevant research results on the impact of text sentiment on the financial markets. [Results] Text sentiment in financial markets can be divided into information reporting sentiment, news media sentiment and social media sentiment. In the construction of sentiment indicators, dictionary-based methods and machine learning-based methods are widely used. The above three text sentiments have a certain impact on the financial markets. [Limitations] Due to the universality of text analysis methods in various fields, the selected literature on the framework of text sentiment analysis is not entirely focused on the financial markets. [Conclusions] When constructing financial text sentiment indicators, we should choose the appropriate sentiment analysis method according to the text characteristics, research conditions and research objectives.

[Objective] This paper aims to identify breakthrough innovation inventors through their collaboration and citation features. [Methods] First, we defined the metrics of breakthrough innovation inventors. Then, we examined the features of cooperation and citation relationship of inventors. Third, we established a statistical learning model to predict their future innovations based on the dynamic learning of the patent knowledge graph. Finally, we analyzed the characteristics of breakthrough innovation inventors. [Results] We examined our model with patent data and found its overall prediction accuracy reached 83.51%. The model’s accuracy for predicting breakthrough and continuation innovation inventors reached 85.99% and 81.40%, respectively. While predicting the inventors, their collaboration and citation-related features were ranked high. [Limitations] The ambiguity of the technological innovation metric of patents around the value of 0 was not fully resolved. We filtered inventors with unidentified categories due to this problem. [Conclusions] The proposed model could discover breakthrough innovation inventors earlier.

[Objective] This paper proposes a technical novelty assessment method based on knowledge reorganization and variation as well as the source and formation mechanism of technological novelty. It addresses the problem of using substitute indicators and ignoring the connotation of novelty in technical novelty assessment. [Methods] First, we analyzed the sources of technological novelty at the micro level, which helps us clarify the internal relationship from knowledge units to knowledge reorganization and from variation to technological novelty. Then, we constructed the technical novelty evaluation indexes at three levels: knowledge source diversity, reorganization novelty, and knowledge variation breakthrough degree around two main lines of knowledge reorganization and variation. Finally, we verified the feasibility and effectiveness of the method with digital medical technology. [Results] We identified the technologies with high novelty and their scores. The recall values of this proposed method were about 23.19%, 5.24%, and 9.69% higher than the commonly used citation, cosine similarity, and knowledge diversity methods. [Limitations] More research is needed to explore categorizing knowledge units under different classification schemes. [Conclusions] Knowledge reorganization and variation are two leading reasons for technology innovation. The proposed method can effectively identify technologies of high novelty.

[Objective] This paper aims to improve the accuracy of event detection tasks by fusing semantic information. [Methods] First, we stored the non-relational semantic information with an initial word vector and encoded them with the Bi-LSTM model to aggregate their contexts. Then, we developed a relation graph based on relational semantic information. Third, we used a multi-scale convolutional neural network to capture the spatial information from the adjacency matrix and fuse it with the word vector. Finally, we built a Gate-GCN model to aggregate relational semantic information between adjacent word vectors to enhance their representation ability. [Results] We examined the new model with the ACE05 benchmark dataset. Our method’s F1 value reached 76.3%, which was 1.2% higher than the existing mainstream models. [Limitations] The proposed model needs to be validated with general datasets. [Conclusions] Fusion of multiple types of semantic information can effectively improve the event detection performance.

[Objective] This study aims to introduce cross-market and cross-source sentiment analysis into the RMB exchange rate forecasting model to improve the performance. [Methods] We built a CCSA-DL model for fusing cross-market and cross-source sentiment analysis. First, we used a BERT-TextCNN model to extract deep sentiment features from China and the United States respectively. Then, we shared them with LSTM-based deep features of exchange rate time series to achieve deep fusion, based on which exchange rate forecasting is realized with the help of SVM model. [Results] Compared with the baseline model, the CCSA-DL model achieved optimal performance in predicting indicators and economic returns. Especially compared with the LSTM prediction model, there was an average improvement of about 16.77% in the three evaluation indicators. [Limitations] The source of sentiment analysis data needs to be further expanded and optimized. [Conclusions] The CCSA-DL model with cross-market and cross-source sentiment analysis has better exchange rate forecasting performance and economic returns.

[Objective] This paper constructs a clique recognition model for Chinese ceramic images. It aims to automatically classify and recognize the clique of ceramic images and provide technical support for the research and digital protection of ceramic culture. [Methods] We adopted the “end-to-end learning” paradigm to build the new model. It applied transfer learning and ensemble learning technology to ceramic cliff identification. We also used the DCGAN algorithm to balance samples. We examined the new model with ten cliques of ceramics based on their types, crafts, and artistic styles. [Results] The proposed model could more effectively extract ceramic image features and recognize ceramic cliques than the baseline models with manually designed feature engineering. Transfer learning enables the extracted features to be effectively transferred to the fine-grained downstream tasks. The accuracy of the new model reached 73.16%. The improved Stacking method integrated knowledge from the proposed models and increased the final accuracy to 81.39%. [Limitations] The data used in this paper is from Baidu pictures, which need to be expanded to improve the model’s performance. [Conclusions] The new model could effectively classify and identify ceramic images.

[Objective] This paper proposes an RB-LCM model to improve the fine-grained sentiment analysis of Weibo texts. [Methods] First, we used the RoBERTa to encode the character and sentence-level features of Weibo posts. Then, we utilized the Bi-LSTM and capsule network to capture in-depth global and local features of Weibo sentences. Third, we deployed multi-head self-attention feature fusion to fuse the relevant multi-dimensional features. Finally, we used improved Focal Loss and FGM to train the model and improve the dataset labels’ imbalance and the model’s robustness. [Results] The accuracy and F1 value of the proposed model on the SMP2020-EWECT dataset reached 80.64% and 77.41%. The model’s accuracy and F1 value on the NLPCC2013 task 2 dataset were 67.17% and 51.08%. The model’s accuracy and F1 value on the NLPCC2014 task 1 dataset reached 71.27% and 58.25%. The model’s accuracy and F1 value on the binary sentiment dataset weibo_senti_100k dataset were up to 98.45% and 98.44%, respectively. All results were better than the advanced sentiment analysis models on each dataset. [Limitations] Our model did not include relevant pictures, videos, voice, or other information for sentiment analysis. [Conclusions] The proposed model can effectively analyze the sentiment of Weibo posts.

[Objective] This paper proposes a deep learning recommendation model with item audience features. It captures collaborative information and the high-order features from users and items the interactions. [Methods] First,we used the attention mechanism to analyze the historical interaction information between items and users. Then, the system adaptively constructed personalized audience features of items. Third, we introduced these features to the model as important supplementary information for preference predictions. We also developed an explicit feature crossing and introduced residual connections to enrich the high-order features. [Results] We examined the new model with three public datasets. It improved the Precision, Recall, F1, and NDCG by up to 9.1%, 9.4%, 9.2%, and 12.1% compared with the sub-optimal method (the recommendation length = 10). [Limitations] The performance of our model relies mainly on the historical interaction data volumes. [Conclusions] The proposed model improves the recommendation quality and shows good application potential.

[Objective] This paper identifies the influencing factors of online health information sharing intention (OHISI) from the existing literature. It aims to explore the intensity of each influencing factor and the roles of different situational variables. [Methods] A total of 62 research articles were included, and 245 independent effect values corresponding to 24 antecedent variables and five moderator variables affecting OHISI were selected for meta-analysis. [Results] Five variables of perceived risk have no significant impact on OHISI, and the other 19 variables positively impact OHISI. The influence of sharing attitude is the strongest, while the impact of disease severity is the weakest. Education level, identity, social culture, sharing channels, and information types regulate some of the above relationships. [Limitations] Due to the small number of papers, we could not thoroughly examine some moderator variables. [Conclusions] Based on the meta-analysis, we constructed the overall effect model by integrating relevant theories. We also obtained the general knowledge affecting the willingness of online health information sharing. This study also helps the operation optimization of online health platforms and future research in related fields.

[Objective] Taking the crime of intentional injury as an example, this paper proposes a reasoning and visualisation method for the cause of injury crimes based on knowledge graph and D-S evidence theory. [Methods] First, we constructed the crime knowledge ontology of intentional injury crimes and supplementing relevant knowledge. Then, we used the case trial records as a data source. Third, we used text mining technology to extract knowledge and instantiate it to form a case knowledge map. Forth, we used D-S theory to resolve evidence conflicts and complete knowledge fusion. Finally, we utilized custom inference rules to achieve and visualize the results of reasoning. [Results] The accuracy of the truth value determination by D-S evidence theory reached 95.45%, which showed the effectiveness of the proposed method. [Limitations] The method of this paper is influenced by the degree of linguistic standardisation of the interrogation records. [Conclusions] The proposed method does not limit the interrogation process and frequency, which can improve the accuracy of knowledge fusion of multiple interrogation records, form the results of case cause analysis based on objective facts, and improve the efficiency of law enforcement case handling and the level of power supervision.

[Objective] This paper proposes a text retrieval model for engineering consulting reports that combines graph convolution topological and keyword features. It addresses the insufficient semantic feature extraction issues in existing retrieval methods. [Methods] First, we built a text retrieval corpus of engineering consulting reports. Then, we fed the corpus into a BERT model to obtain contextual vectors. Third, we obtained the first matching score through a graph convolutional network and a deep interactive matching model. We also mapped the paragraph keywords to vectors using a Word2Vec model and calculated their similarity scores with the titles to obtain the second matching score. Finally, we got their final matching score by averaging the two matching scores. [Results] Compared with the joint ranking model CEDR, our new model was up to 3.06% higher in the P@20 metric. [Limitations] The data was mainly from engineering consulting reports of a large state-owned company, which needs to be expanded. [Conclusions] The GKTR model could effectively improve text retrieval for engineering consulting reports.

[Objective] This paper combines the knowledge graph and reader profiling technology to address the data sparseness and cold start issues of book recommendation. [Context] We examined the proposed model with the library management system of Zhejiang University of Technology, including 220,636 circulation records from May 2020 to May 2022. A total of 60,162 books and 15,916 readers were included in this study. [Methods] First, we constructed a reader-book knowledge graph. Then, we modeled semantic associations between books and reader preferences utilizing book theme modeling and reader profiling. Finally, we explored semantic connections among reader-reader, reader-book, and book-book relationships, strategically addressing data sparsity and cold start challenges. [Results] The proposed method based on GraphSAGE improved the precision by 0.151 compared to the existing collaborative filtering algorithm. Its recall rate reached 51.44% in the cold start environment. [Conclusions] The book recommendation method based on knowledge graph and reader portraits can effectively improve the data sparseness and cold start problem.