1. Emotion Analysis with Twitter More details
We applied a deep learning model (BERT) as well as a new model for sentiment analysis, categorizing each tweet into different sentiments; trust, fear, joy, sadness, anticipation, anger, disgust, and surprise.
Visit the pre-print version of initial results and analysis at https://arxiv.org/pdf/2004.10899.pdf.
2. Emotion with Instagram More details
We performed statistical analysis on posts related to mental health topics. More than 1.7 million posts with #mentalhealth have been posted on Instagram since January 2020. The number of posts increased in mid-March, when countries in Europe and North America issued stay-at-home orders, although the number of posts with #depression did not change.
Topic 1. Emotion Change Analysis with Twitter
Contributors: Irene Li, Tianxiao Li, Yixin Li
We applied deep learning methods for sentiment analysis, categorizing each tweet into different sentiments; trust, fear, joy, sadness, anticipation, anger, disgust, and surprise, etc.
Section 1: Sentimental distribution on 8 categories
We applied a deep learning model (BERT) trained on 750 manually labeled cases to 1 million English tweets. Fear, Anger and Sadness ranked first. (A multilingual version will be available soon!)
Section 2: Sentiment trend among topics
We now look at the emotion trends on different topics. Using the same model from Section1, we analyzed two topics: “mask” and “lockdown”.
Section 3: The reasons why people feel sad or fear.
To understand why people feel fear and sadness, we calculated correlation on the tweets categorized by fear and sadness, and then kept nouns and noun phrases with the help of Stanford Stanza tool. The first graph shows some founded keywords from English tweets. The second and third graph show multilingual keywords which contain English, Spanish, German, Portuguese, Japanese and Chinese based on different models (2nd is a lexicon-based method; 3rd is an attentional method).
Image tool credit: https://worditout.com/word-cloud/create
Section 4: The topics on tweets.
We utilized the LDA(Latent Dirichlet allocation) topic modeling to analyze the topics on people’s tweets. Each “topic” learned by the model is a bunch of key words, then we manually labeled these topics as meaningful concepts. The following are five topics we learned.
Topic 0: Covid19 testing, deaths cases, positive caes
Topic 1: President Trump, government, federal affairs
Topic 2: lockdown, stay at home, physical distancing
Topic 3: (Spanish) pandemic, health conditions
Topic 4: the peak, serious treatment, Boris Johnson
We choose the data on April 7th, and first we do inference on all the data, and show the ratio for each topic learned above (All). Then we do inference on the tweets that are only labeled as sadness or fear (Sad and Fear). And the following is the ratio of each topic learned. In general, the public may be worried about Topic 3 and 2, mainly, the pandemic and lockdown, which are making people stressed.
Topic 2. Emotion Change Analysis with Instagram
Contributors: Keita Suzuki, Hiroki Kanezashi, and Naomi Nakagawa
We collected 71,737 posts with #depression from March 31 to April 5, 2020. During this period, the number of the posts are steadily increasing as below. Among the posts with location information, #depression was mostly posted in the U.S., the U.K. and India. In those countries, users posted the hashtag during the local daytime. During the long quarantine, people might struggle to keep their mental conditions healthy and the trend of #depression on Instagram might reflects their attitudes.