Fig 1.
Overview architecture of BERT model.
Adapted from Devlin et al. [64].
Fig 2.
Overview architecture of SKEP model.
SKEP model contains two parts: (1) sentiment masking detects sentiment information from an input sequence using automatically mined sentiment knowledge, and generates a corrupted version by removing this information; (2) sentiment pre-training requires the transformer to recover the removed information from the corrupted version. The three prediction objectives on top are jointly optimized: Sentiment word prediction (on X9), word polarity prediction (on X6 and X9), and aspect-sentiment pairs prediction (on X1). Notably, on X6, only word polarity is calculated without the sentiment word, as its original word has been predicted in the aspect-sentiment pairs prediction on X1. Adapted from Tian et al. [21].
Table 1.
Performance comparison of the sentiment classification models.
Fig 3.
Topic generation process of the LDA model.
Adopted from Blei et al. [22].
Fig 4.
Sankey diagram of public comments characteristics on urban regeneration.
Fig 5.
Annual variation of sentiment classification and the ratio of positive to negative comments.
Fig 6.
Monthly variation of arithmetic mean of SI.
Fig 7.
Regional distribution of sentiment classification and arithmetic mean of SI.
Fig 8.
Geographical distribution of the arithmetic average of SI.
Shape file source: republished from http://www.gscloud.cn under a CC BY license, with permission from Geospatial Data Cloud, original copyright [2022]; Own Map output: using ArcGIS 10.2 Software analysis.
Fig 9.
WC visualization of the top 400 most frequently posted Chinese words out of all 74,481 Chinese words.
Fig 10.
Variation in perplexity of the LDA model from 0 to 15 topics.
Fig 11.
IDM of the three clusters categorized using the LDA model.
Table 2.
Topic categorization of all comments based on the LDA model.