VROF

Work

An analysis tool that shows thousands of research comments split by topic and sentiment

Video comments gathered by search term are grouped by topic, with a positive/negative distribution alongside. How many topics to split into is decided by the researcher on screen. It was built so the researcher can change the criteria and re-run it themselves.

Understanding comment topics

Before

By skimming with the eye and guessing

After

By per-topic, per-sentiment distribution

Comment analysis tool running screen. On the left are input fields specifying topic count 5, total corpus training iterations 15, per-document training iterations 400, and top word count 10; on the right are two bar charts showing the top 10 most-mentioned words in the comments and two pie charts showing the share of comments containing each topic.
The actual delivered screen. The 21.6% and 24.4% in the pie charts are the shares of all comments that contained the specified topic.
Client
Domestic university research team
Industry
Research · Academia
Completed
2025
Duration
2 weeks
Service
AI adoption
  • Automatic comment topic classification
  • Sentiment analysis
  • Analysis screen with on-screen adjustable criteria
  • Visualization

Background

Collecting YouTube comments with a specific topic as the search term yields thousands of them. What the research needs to address is how many strands of stories are inside, but reading them from start to finish does not reveal those strands. After reading a few hundred, you start guessing at the rest.

The problem we defined

Tallying “what percentage is positive” alone was not enough. You have to first split by what is being talked about to know where that sentiment is aimed. The scope of the problem was keeping the work of splitting topics and the work of counting sentiment separate, and then viewing the two overlaid.

And the researcher had to be able to change the criteria and re-run it themselves. How many topics to split into is not a value with a correct answer but a value decided while looking at the data. If that value were hard-coded, they would have to call us every time they wanted to try a different one.

What we built

  • Topic modeling — splits comments into topic groups. Topic count, training iterations, and top word count are changed directly on screen
  • Frequency visualization — shows the top 10 most frequent words as bars
  • Topic inclusion share — shows as a pie what percentage of all comments contain the specified topic
  • Sentiment distribution — shows the positive/negative ratio alongside

Results

The thousands of comments gathered per search term come out organized into per-topic groups and sentiment distributions. The two on the screen each represent 21.6% and 24.4% of all comments that contained the specified topic.

Re-running with changed criteria is done by the researcher themselves.

Other work on the same topic is collected under data analysis.

Other work on the same topic is collected under applied AI.

Sources

Running screen of the delivered tool

We offer consultations to help you save time Free consultation · Replies within 1 business day

Contact