Comment topic and sentiment analyzer
Video comments collected by search term are grouped by topic and shown with their positive and negative distribution. Researchers set the number of topics on screen and can adjust the criteria and rerun the analysis themselves.
Comment topic analysis
Before
Skimmed by eye and judged by feel
After
Distributions by topic and sentiment
To protect client information, these screens recreate the same features with mock data.
Background
Collecting YouTube comments with a topic as the search term returns thousands of results. The research needed to know which themes run through them and how many there are, but reading every comment does not reveal those themes. Readers get through a few hundred and guess at the rest.
The problem
Counting the percentage of positive comments was not enough. Comments first had to be grouped by what they discuss to show what each sentiment was directed at. The scope covered topic grouping and sentiment analysis as separate steps, and viewing the two together.
Researchers also needed to adjust the criteria and rerun the analysis themselves. The number of topics has no single correct value. It is decided by looking at the data. If that value were fixed in the code, every change would require a request to us.
What we built
- Topic modeling — Comments are grouped into topics. The number of topics, training iterations, and top words can be changed on screen
- Frequency charts — The 10 most frequent words are shown as a bar chart
- Topic share — A pie chart shows what percentage of all comments include a given topic
- Sentiment distribution — Positive and negative ratios are shown alongside
Results
Thousands of comments collected for each search term are organized into topic groups and sentiment distributions. In the two cases shown, 21.6% and 24.4% of all comments included the given topic.
Researchers rerun the analysis with new criteria on their own.
Related work is listed under Data analysis.
Related work is also listed under AI adoption.