基于深度学习的城市热点空间情绪感知评价——以上海市为例

基于新浪微博数据，借助深度学习方法分析上海市情绪空间分布特征，期望构建一套基于社交媒体签到数据的深度学习空间情绪感知评价方法。研究发现：①上海市热门签到地点集中分布在城市中心、交通枢纽、地铁沿线等区域。②积极情绪占比随着到市中心距离的增加呈下降趋势；人们的情绪与活动空间类型高度相关。③高等院校场所与负面情绪相关的物质空间要素多与建筑相关，办公场所的空间使用者最关注通勤问题，交通枢纽空间的管理流程和服务水平诱发了较多负面情绪。研究发现对于城市公共安全、公共卫生和设计管理的决策者有着重要启示。

EVALUATING EMOTIONAL PERCEPTION OF SPATIAL HOTSPOTS VIA DEEP LEARNING: A CASE STUDY OF SHANGHAI

The study on spatial emotional perception which arises from the intersection between Spatial Psychology and Geography is an important field of Human Geography. But its development has long been limited by data access and methods. Deep learning can provide the strong support for the quantitative analysis of spatial emotional perception. This paper attempts to build a deep learning framework which based on social media sign-in data to evaluate and analyze spatial emotional perception. In this work, emotional perceptions were classified into six categories, such as joy, affection, distress, angry, disgust and null. The pre-trained language model named Bidirectional Encoder Representations from Transformers was used to analyze the emotional expression of active users and generate the emotional perception map. Compared with the traditional sentiment analysis model, the Bidirectional Encoder Representations from Transformers has a greater degree of improvement in correctness. Then part-of-speech analysis was applied to the comments extracted to investigate the cause of the emotional perception. In the last, this paper used Lexical Analysis of Chinese, a lexical analysis tool developed by Baidu, for lexical analysis and named entity recognition. In total, 813,633 geotagged social media data and 1619 POI were collected from Shanghai. The main findings were as follows:1) The most popular sign-in locations in Shanghai are concentrated in the inner-city, transportation hubs and important public facilities. 2) The proportion of positive emotions shows an overall decrease with the increase of the distance to the city center. 3) Producers' emotional perception of various activity spaces were mostly positive opinions, so the results of commentary viewpoint extraction were similar.