顾及地理语义的地图检索意图形式化表达与识别

Map Retrieval Intention Formalization and Recognition by Considering Geographic Semantics

主流地图检索方法多基于元数据文本匹配或图像内容相似度计算,缺乏对用户意图的主动理解,导致检索结果欠佳;而现有意图识别方法无法准确表达与识别复杂地理概念联合约束的地图资源检索需求。为此,本文提出一种顾及地理语义的地图检索意图形式化表达与识别方法,旨在利用相关反馈样本“感知”用户需求,以提升检索精度。该方法通过地理本体约束“意图-子意图-维度分量”模型的构建,实现检索需求的语义化描述;并将意图识别视为组合优化问题,基于最小描述长度准则、顾及地理概念从属关系的样本随机合并策略及贪心搜索实现最优意图识别。实验结果表明,相比基于频繁项集挖掘的RuleGO、决策树的DTHF算法,本文方法具有更高的识别准确度与噪声容忍度;随机合并策略可在不降低识别准确性的情况下有效缩短平均求解耗时;样本增强策略保证算法在样本规模仅为20时仍具有较高识别准确度。该方法可望应用于地理信息门户,提升各类地理信息资源共享与发现的服务品质。

Mainstream map retrieval methods for spatial data infrastructures are mainly based on metadata text matching or image similarity calculation,but such approaches lack active perception and understanding of user retrieval intention,and in turn fail to truly meet user requirements.While existing intention recognition methods are incapable to express and recognize map retrieval demands with joint constraints of complex geographic concepts.To address this issue,this paper proposes a map retrieval intention formalization and recognition method by considering geographic semantics,aiming to improve the accuracy of map retrieval in an intentiondriven and explainable manner by using relevance feedback samples.More specifically,a formalization model constrained by geographic ontology in the form of"intention-sub-intention-dimension component"is designed for expressing user's map retrieval intention.With the support of the formalization model,a recognition algorithm based on Minimum Description Length(MDL)principle and Random Merging(RM)strategy,named MDL-RM,is proposed by treating intention recognition as a combinational optimization problem.MDL-RM takes the description length of the sample set from relevance feedback as the optimization goal,merges samples randomly with the assistance of geographic ontologies and semantic similarities among geographic terminologies to generate sub-intention candidates,and searches the optimal intention using a greedy search approach.In order to evaluate the accuracy of recognized intention,we proposed a semantic metric,named Best Map Average Semantic Similarity(BMASS),and calculated it along with Jaccard index in five typical map retrieval scenes.Meanwhile,we analyzed the time cost and the influence of parameter settings and validated the effectiveness of random merge and sample augmentation strategy.The experimental results on the synthetic data demonstrate that the proposed method has higher accuracy and sample noise tolerance in most retrieval scenes comparing with the method based on Gene Ontology(RuleGO)and the Decision Tree learning method with Hierarchical Features(DTHF).The random merge strategy can reduce average computing time effectively without declining accuracy,and the sample augmentation strategy facilitates retrieval intention recognition even when the sample size is as low as 20.The proposed method is expected to be adapted and applied into geoportals and catalogue services to improve the service quality and user experiences upon the sharing and discovery of geographic information resources.