Before and After COVID-19 Outbreak Using Variance Representation Comparative Analysis of Newspaper Articles on the Travel Hotel Industry

Abstract

This study explores the impact of COVID-19 on Japan’s travel industry by analyzing differences before and after the pandemic through articles from the Nihon Keizai Shimbun. It employs text mining techniques like Latent Semantic Analysis (LSA) and BERT (a natural language model) to process and categorize information from newspaper titles. The analysis involves extracting nouns using MeCab, creating a frequency matrix, decomposing it with NMF for clustering, and setting topics. BERT is used for text classification, focusing on token attention weights and variance representation. The data includes articles from Nikkei Morning News pre- and post-COVID-19, specifically tagged with “Travel & Hotel,” totaling 792 articles. Analysis revealed ten topics such as vaccines, business structures, and financial results. Hierarchical clustering grouped these topics across eight clusters. Findings indicate a shift in topics post-COVID-19 towards financial impacts and business activities, highlighting tokens related to company activities and keywords associated with the pandemic. Future work aims at improving classification accuracy and leveraging data insights.

Appendices

Appendix 1: Data Code List

• Code related to title data analysis

  https://colab.research.google.com/drive/14kVhJuJSRHV0m7zCKCuPuOM0DkPcgKHQ

• 2019 Nikkei morning edition article data

  https://tus.app.box.com/file/998437276219?s=e7hx3pyypf06ekyjjignhpc8b725sdd1

• 2022 Nikkei morning edition article data

  https://tus.app.box.com/file/947471044518?s=bs8c424lfrckgm6o5k8bbh8vuf81axru

• Weighted word list - positive

  https://tus.box.com/s/53hksp6tk3o1bjnzbwfkgvxywij7ido6

• Weighted word list-negetive

  https://tus.box.com/s/stsfk4khv7sa7ihvarnwudeudtraeyjb

Appendix 2: Attention Weight Visualization Example

In English, it is easy because only the alphabet and each word is separated, but in Japanese, kanji, hiragana, katakana, emoji and other words are connected, so it is difficult to judge where to cut. So, since the data used this time is all in Japanese, the analysis results are also attached in Japanese as is. Thank you for your understanding.

Example sentences after the coronavirus pandemic and with high $P_{AC}$

Example sentences before the coronavirus pandemic and with high $P_{AC}$

Example sentences before the coronavirus pandemic and with high $P_{BC}$

Example sentences after the coronavirus pandemic and with high $P_{BC}$