Augmenting web-based tourist support system with microblog analyzed data

Rarely-visited attraction spots in large parks accumulate an insufficient number of reviews on popular online review websites. On the other hand, mapping microblogs’ unstructured text data to geospatial information can be beneficial for tourism planning-related practical applications. Therefore, our research proposal focuses on acquiring additional opinions from microblogs and using them as review supplements in a developed tourist support system. Previously, geotagged tweets were commonly used for location inference. However, recent studies suggest that microbloggers usually do not post with location geotagged. Therefore, to infer the location of ungeotagged tweets, we proposed a two-stage process. A classification framework that relies on a fine-tuned transformer neural network model that learns from tweet contents and predicts the locations from which those tweets were sent—with a limited application in the detection of widely known general locations—such as tourist spots and Impact Words extraction analysis of location likelihood using location or event mentions. Unfortunately, Twitter data is typically noisy and consists of ungrammatical or informal phraseology and non-standard vocabulary, which additionally causes the feature sparsity problem, resulting in low classifier performance. To address this, we specifically evaluate a range of pre-processing techniques for text categorization to accurately obtain a proper set that collectively contributes to the improvement of prediction accuracy. We learned that the average 0.84 F1 score of a pre-trained DistilBERT language model outperformed other tested models when tested on different pre-processing datasets. Furthermore, we computed the weighting of words using TFIDF and evaluated the degree of impact of Impact words on the model performance. We discovered model accuracy improvement of as much as 6% when Impact words are involved compared to when they are removed which suggests positive influence of Impact words in geospatial estimation. We also discovered wrong weighted IW that negatively contributes to the model performance and by eliminating them, the model F1 score improved by 3%. Extracted information further contributes as touristic information in a developed web-based system.