Sentiment analysis on YouTube comments with Python: handling sarcasm, emoji and spam

Analyzing sentiment in YouTube comments with Python can be tricky, as the data is often noisy and difficult to interpret. Library-based sentiment scores tend to mislead due to various factors. This article outlines the challenges and provides a simple pipeline to address them. The main issues include sarcasm, emoji and slang, short text, spam, and context.

To tackle these problems, start by cleaning the data before scoring. Exclude creator comments, links, very short comments, and duplicates. Instead of reading absolute sentiment values, compare scores to see if a video is better or worse than previous ones. Weight negative comments with more likes more heavily than those with none.

Validate the tool by checking against a hand-labeled sample, labeling 50 comments yourself, and assessing how often the tool agrees. If the results are poor, try a different model or an LLM with a fixed three-label prompt and repeat the validation.

Always examine the most-liked comments in each sentiment category. Use the official YouTube Data API v3 to collect comments, handling paging, recent videos, replies, and quota yourself. The Best Damn YouTube Comments Scraper is an Apify Actor that wraps the API and returns one flat record per comment. Input your desired video or channel, set the maximum number of videos per channel, the maximum number of comments, and use a hash to replace author data for privacy.

The resulting JSON contains all necessary fields, such as text, like count, reply count, author information, publication date, video ID, video title, and comment URL. After obtaining the comments, classify them using VADER, a lexicon tool that can handle some slang and emoji. Remove unwanted comments, create a sentiment analyzer, and assign labels based on the compound score. Finally, summarize the data per video and analyze the results.