认知型服务中不确定性需求理解方法研究

At present, the booming artificial intelligence makes the cognitive service system which can think like human beings become the trend of future development. Cognitive service has also become one of the hotspots in the field of service computing. However, the uncertainty user requirement makes cognitive services hard to understand user requirement. The key Challenges are as follows: 1) the incompleteness of requirement expression makes the system have to model group requirement network to infer individual requirement space; 2) the randomness of personalized requirement makes the system have to capture user's personalized requirement in time; 3) the intention of user's real-time requirement is often a number of discrete functions, which cannot form a complete requirement plan. Therefore, on the basis of Bayesian learning framework, this topic will study the methods of understanding uncertain user requirement in cognitive services: 1) Deep Bayesian network based user requirement clustering model can learn the expression of user group (domain) requirement network and individual requirement space accurately; 2) Multi-class and multi-modal user intention inference algorithm can establish constraints variables, and achieve users' real-time personalized intention; 3) Bidirectional perception based guided user requirement completion method can obtains a complete and standardized set of requirements by perceiving the semantic relationship in the requirement network and the logical relationship in the service process.

当前，蓬勃发展的人工智能使得能像人一样进行思考的认知型服务系统已经成为了未来发展的趋势。认知型服务也成为了服务计算领域的研究热点之一。然而，需求主张的不确定性使得认知型服务难以精确理解用户需求。其关键难点是：1）模糊不完整的个体需求表达要求系统通过建模群体需求网络对用户个体需求空间进行扩展；2）随意变更的用户需求使得系统必须及时捕获用户个性化意图；3）个体实时意图往往是若干离散的功能点，而系统期望完备的需求方案。因此，本课题将在贝叶斯学习框架的基础上，研究认知型服务中不确定性需求理解方法：1）基于深度贝叶斯网络的用户需求聚类模型，精准地学习用户群体（领域）需求网络和个体需求空间的表达；2）多类标多模态的用户意图推断算法，建立约束变量，得到用户实时个性化意图；3）基于双向感知的引导式用户需求补全方法，通过感知需求网络中的语义关系及服务流程中的逻辑关系得到完备规范的需求集合。

精细化辨识用户个性化需求，设计使用户满意、服务增值的解决方案，是认知型服务系统急需解决的关键科学问题。在认知服务系统中，用户原始的需求具有需求表述模糊，意图结构复杂，特征多模态、目标个性化以及变更随意等特点，导致智能服务系统在复杂场景下的意图的识别能力差。针对上述问题，本项目从用户需求的层次结构模型出发，深入研究多模态用户需求意图的识别方法，并从服务-需求双向感知的角度进一步研究用户需求意图补全集成方法。针对用户个性化需求结构模糊特性，本项目设计了需求功能特征树模型对个人需求进行形式化建模。针对用户需求表述模糊的特性，本项目提出了基于双重双向长短期记忆网络的用户需求意图分析模型，通过捕捉意图槽位标记之间的依赖关系提升用户意图解析的准确性。针对需求表达的多目标场景，提出基于多重注意力机制的用户多意图分析模型。针对需求表达的多模态性，提出基于多维度时空循环神经网络的用户兴趣点预测方法，研究在特定场景中用户行为对其需求表达的影响。本项目在进展过程中发表论文12篇，申请或获得国家发明专利6项，获得软件著作权2项，培养硕士研究生6名。设计并实现了用户需求解析工具，并将其应用到智慧养老领域，取得了较好的社会效益和经济效益，获得了省部级人工智能科技进步奖一等奖，验证了本项目成果的有效性。