Project (National)

Research on Key Techniques of User-centered Emotion Elicitation and Regulation Dialogue Generation. Natural Science Foundation of China, 62272092, 2023.1~2026.12

The emotional interaction between humans and computers via dialogue system is a research hotspot in artificial intelligence. Most of the existing emotional dialogue generation techniques are system-centered, and fail to leverage the user's personality and dialogue situation to actively affect the user's emotional state. This proposal introduces a new research problem called User-centered Emotion Elicitation and Regulation Dialogue Generation and conducts the following research: (1) We study on dataset construction and external knowledge reconstruction techniques to prepare the task-specific corpus and knowledge base; (2) We study on user persona and dialogue situation modeling techniques to detect persona and situation in human-computer interaction; (3) We study on human-computer emotional interaction reasoning technique to predict user' elicited emotional state; (4) We study on emotion elicitation and regulation dialogue generation technique to generate the responses in single-turn, multi-turn or multi-modal scenarios; (5) We study on content safety detection and dialogue quality evaluation techniques to ensure the safety and quality of training data; (6) We study on the dialogue generation techniques after model development for continuous learning and updating of the model. This proposal creatively proposes to endow the dialogue system with the ability to elicit and regulate users' emotional states. The research results have broad application prospects in the emotional companionship and psychological counseling in chatbots.

Research on Interpretable Emotion-Aware Dialogue Generation. Natural Science Foundation of China, 62172086, 2022.1~2025.12

The "sentiment interaction" between human and computer in the dialogue system plays an important role in the emotion companionship and psychological counseling of robots. Due to the lack of explainability in the process of dialogue generation, the computer is difficult to convince itself and the user, the user is difficult to know, and know why, so we study the interpretable dialogue generation technology of emotion awareness. It includes: (1) Interpretable model-oriented conversational knowledge construction technology to solve the problem of training corpus required for modeling; (2) Interpretable user utterance emotion awareness technology, to solve the problem of computer emotion awareness for users; (3) Interpretable dialogue generated emotion prediction technology to solve the problem of empathy between computer and users; (4) Interpretable dialogue generation technology based on emotion prediction to solve the problem of computer response to users; (5) Criterion for the relationship between emotion awareness, dialogue generation results and their interpretation and model optimization technology, to solve the evaluation and optimization problems of models and processes. Finally, a prototype system of interpretable dialogue generation for human-computer sentiment interaction with the functions of knowledge construction, dialogue understanding, dialogue prediction, dialogue generation and dialogue management will be implemented. Achieve: (1) users get better experience through explanations. (2) The results generated by models and their corresponding explanations are mutually promoted and co-enhanced.

Study on Human-Machine Conversation Text Generation Technique for Emotional Interaction. Natural Science Foundation of China, 61872074, 2019.1~2022.12

Implementing human-machine dialogue using natural language is the hot research topic in artificial intelligence. At the early stage, the dialogue system has no emotion and no personality. Therefore, embedding the machine with human emotions has become one of the long-term goals for intelligent dialogue system. In this proposal, we focus on the realization of communications between human beings and machines using text with emotions and opinions. We attempt to improve the human-machine dialogue system with a series of new features, based on which the new text generation techniques are studied for emotional interaction. 1) Considering different personalities of users and different traits of machines, we study on modeling techniques for user personality and machine trait in conversation context. 2) We study on emotion and stance perception techniques for tracking users' dynamic emotion states and detecting users' opinion stances. 3) We study on controllable emotion and opinion text generation techniques for generating the text with specific emotions or opinions. 4) Based on above conditions, we study on personalized emotional response generation as well as ingratiating and refutation opinionated response generation. This proposal innovatively injects user personalities and machine traits into emotional conversations. Moreover, we propose a novel research problem called opinionated conversation generation. The research achievement of this proposal has a broad application prospect for emotional companionship and psychological counseling using chatbot system.

Real-Time and High Precision Dynamic Scholars Profile Construction. National Key Research and Development Projects (Sub-Projects), 2018YFB1004702, 2018.7~2021.6

The subject is oriented to the high-precision evolutionary scholar profile of the knowledge graph. Aiming at the characteristics of scholars with diverse fields and great differences, the modeling method of accurately describing the scholar profile is studied. Through the real-time collection of multidimensional research behavior data, the NSF classification, Chinese library method and knowledge graph analysis of academic data are used to accurately construct the academic profile. In view of the complexity of expert evaluation and ranking in the profile of scholars, an evaluation system based on distributed spectrum analysis algorithm is studied, so as to efficiently extract high-level experts in the field and complete comprehensive evaluation and multidimensional ranking of experts. Aiming at the real-time evolution of scholar's profile, the dynamic updating rules based on the profile activation model are studied, including the self-updating of the portrait library and the dynamic evolution of the relational graph. Aiming at the complexity of scholars' profile modeling, this paper studies the three-dimensional, accurate and credible methods of academic portrait, including the methods of peer evaluation and expert recommendation.

Diversified Social Relation Active Learning Based on Social Media Mining. Natural Science Foundation of China, 61772122, 2018.1~2021.12

Social relation learning on social media plays significant role for many applications, such as information retrieval, personalized recommendation, and community discovery. In recent studies, there are more applications based on specialized social relations, but less work of actively discovering potential social relations. There are more applications about direct, explicit, and static social relations, but less work about indirect, implicit, and dynamic social relations. There are more efforts on detecting the existence of social relations, but less work of learning semantics of social relations. Using the resources and users in social media as data sources, this proposal defines diversified social relation from the perspective of form, semantic, and operation for social relations, and analyzes the challenges for discovering the diversified social relations. This proposal applies the social media mining, deep learning, and especially active learning techniques to address the aforementioned challenges, and the major research contents include: (1) the active learning of direct social relation based on relation definition and annotation; (2) the active learning of static social relation for recognizing the semantic of social relation; (3) the active learning of dynamic social relation for discovering evolving patterns and (4) the active learning of the correlation between relations for multi-relation fusion and refinement. As result, this proposal will construct a multigraph model for effectively demonstrating the diversified social relations in social media. For this purpose, this proposal also studies the relation storage, update and visualization based on multigraph structure. The results of the proposed research subjects can better support many applications such as social relation search.

Study on Multi-Channel Feature Learning for the Sentiment Analysis on Micro Video. Natural Science Foundation of China, 61402091, 2015.1~2017.12

Micro video, which is becoming a pervasive media type on the web, provides a rich repository of people's opinion and sentiment about a vast spectrum of topics. The analysis of such information either in the area of brand monitoring, market prediction or personalized recommending plays an important role in behavior sciences. This problem poses a set of unique challenges as micro video data present their intrinsic characteristics as shortness in content, uniqueness in scenario, diversity in transmitting modals, and relativity in data. The existed algorithms on sentiment analysis can not satisfy the requirement from micro videos. For that, study on multi-channel feature learning for sentiment analysis of micro videos is proposed in this task. aiming at the above characteristics on micro video data, we study sentiment collecting on real-time Web data, sentiment semantics learning in the micro video space, the multi-channel feature expressing model based on the video sentiment ontology, and multi-channel feature fusing for sentiment analysis. Finally, we will implement a prototype system combining the above studies. This task will provide the relative technology supports on the applications of micro videos, such as affective auto-analyzing, collecting, real-time monitoring or personalized recommending.

Modeling Multi-Level Resource Recommendation of Social Media Based on Broad-Sense Content Analysis. Natural Science Foundation of China, 61370074, 2014.1~2017.12

A recommender system concerns with two main components which are users’ interest analysis and recommended information selection. The plenty of information resources and user communities with various topics in social media provide broader space for the analysis and selection. In social media based recommendation, the analysis and selection require analyzing all contents including visual and aural features, semantic information, and meta-data. In this project, we define above contents as broad-sense content, classify the social media resources into single, compositive, integrated, and sub-resource categories, and study on social media multi-level resource recommendation based on broad-sense content analysis. For foundation, broad-sense content analysis approach is explored. For users’ interest analysis, the studies include constructing user-resource heterogeneous information network, discovering user community, and mapping user community topics into users’ interest profile. For resource selection, the studies include mining other categories of resources based on single resources, modeling multi-level resource with conditional correlation, and identifying quality of the resources. For the model application, we study on the resource match and recommendation approach based on above models. The work can both reflect the characteristics in recommending multi-level resource proposed in our project and emphasize the user social relationship in the current social media researches. As a result, we will implement prototype systems based on current main stream social media for showing above characteristics.

Research on Key Techniques of Opinion Mining for Real Time Public Opinion Monitoring in Microblogging Sphere. Natural Science Foundation of China, 61100026, 2012.1~2014. 12

As a new information carrier and transmission way, micro-blog is playing an increasingly important role in the initiation and dissemination of network public opinion information. Analyzing the network public opinion among them is of great significance for understanding social public opinion. However, due to the real-time, brief, community, especially multi-modal characteristics of micro-blog data, the existing view mining algorithm is not good enough to meet the needs of micro-blog public opinion analysis. Based on this, this research focuses on the key technologies of opinion mining for real-time public opinion monitoring in micro-blog space. In view of the above characteristics of micro-blog data, this work researches the real-time Web data viewpoint collection and tracking technology, the fine-grained sentiment analysis technology of short texts, the opinion leader discovery technology of micro-blog public opinion events, and the sentiment analysis technology for multi-modal micro-blog data. Combined with the above technologies, this study realizes a prototype system for monitoring public opinions in the micro-blog space. The research results will provide relevant technical support for automatic analysis, summary and tracking of real-time Web public opinion information such as microblog.

Effective Search and Service for Web Visual Media. National Basic Research 973 Program of China under Grant (Sub-Projects), 2011CB302206-G, 2011.1~2015.8

This project research the key technology of data mining: multi-modal visual media for super high dimension characteristic of multimodal information, different modal data of heterogeneous and features of the complexity of the type, the different modal contribution to describe objects of uncertainty, and the different characteristics of the same modal contribution to describe objects of uncertainty, the modal characteristics of the fusion model. Aiming at the inapplicability of traditional data mining algorithms in multi-modal data sources, the efficient data mining algorithm, algorithm fusion and mining result fusion technology based on multi-modal feature fusion model are studied.

Study on Key Techniques of Opinion Mining and Opinion-Based Community Discovery for Web Public Opinion Analysis. Natural Science Foundation of China, 60973019, 2010.1~2012.12

As the Internet become a cradleland of social opinions that reflect social hot topics, “Web public opinion analysis” has emerged as a novel approach for government to know the popular thoughts of people. “Opinion mining” is currently a new research area which concerns with the humanities and social science knowledge such as propagation and psychology as well as natural science knowledge such as text mining, information retrieval, and natural language process. Different from the traditional text mining technique which focuses on the facts of an event, opinion mining prefers to users’ opinions and views of the event. Therefore, the opinion mining technique has new characteristics and confronts new challenges, and becomes one of the main supporting techniques for the automatic Web public opinion analysis. This project studies on the opinion mining key techniques for Web public opinion analysis. By analyzing Web public opinion documents, we summarize the characteristics of public opinion including broad distribution, complex content, fuzzy emotion, time-variant orientation, and exploring requirement for the application. For these characteristics, we study on how to collect public opinion data and describe event and opinion, how to cluster and classify public opinion, how to detect opinion drift and new hot topics in opinion text stream, and how to discover opinion community and opinion leader. As the result, we build an opinion mining prototype system supporting Web public opinion analysis. The research result will be helpful for automatic Web public opinion analysis.

Research on User Motivation Deduction Model for New Generation Search Engine. Natural Science Foundation of China, 60573090, 2006.1~2008.12

New generation searchs engine have such characteristics as interactive searching, classific navigation, accurately related querying, and rapid updating. It is necessary to accurately understand the users’ motivation for implementing the functions of the new generation search engines. In this research, for deduction users’ search motivation, the techniques of information retrieval, data mining, data stream analysis and text mining are utilized, and natural language processing, behavior science and cognition science are imported. For building a user motivation deduction model, the analysis and deduction mechanism, content and structure of its assistant information are studied and implemented. Finally, a prototype for the functions of the model is built, and a front tool with the characteristics of the new generation search engine based on the model is provided in current search environment.

Study on Zero-Input Personalization Techniques for High Quality E-Services on Internet. NSFC (Natural Science Foundation of China), 60173051, 2002.1~2004.12

This project focuses on the key techniques for high quality E-Services personalization on Internet, which aims to provide the technique of information discovery and recommendation based on the personalized requirements of a user without extra input efforts besides his natural browsing operations. To support high quality E-Services on Internet, we proposed a zero-input personalization CMR approach that integrates the techniques of data mining, rule resolution, and information integrating. The research contents include personal data collecting technique, Web data base/warehousing technique, personalization-oriented Web mining technique, personalization rule resolution technique, and personalization service recommending technique. Based on the CMR technique, we have designed and implemented a personalization midware system SmartWeb. The research involves data precessing, artificial intellgence, behavior science and so on, and has important theoretical value and broad application expectation.