Project Overview
Our project is focused on the entertainment industry and audience reactions to films. We are particularly interested in the effect that movie trailers can have on audience perceptions and beliefs about a film before it comes out, and how this can shape expectations before the movie is released. We will use the Star Wars film The Rise of Skywalker, released in 2019, as a case study for this. We chose the film in part because it is relatively recent and is still fresh in cultural memory, and because critic and audience reactions were so split (and for longtime fans, mostly negative).
Research Question
What does comparing YouTube trailer comment sections with Rotten Tomatoes audience reviews of the movie The Rise of Skywalker tell us about people's expectations for the movie versus their actual reactions, and what can this tell us about how movie trailers can sometimes mislead audiences?
We hypothesize that reactions to the trailer for The Rise of Skywalker will be more positive than film reactions because the trailers may have misdirected audiences primarily about the role that Palpatine plays in the film. This demonstrates that movie trailers, especially trailers for major blockbuster films, tend to reveal too much or set expectations too high. In the case of the newer Star Wars movies, the trailers preyed on audience nostalgia for the original Star Wars trilogy and its actors, evoking positive feelings without actually revealing any of the plot details.
Data & Methods
Dataset
Data Source: We collected our data from the YouTube comments section of the final Rise of Skywalker trailer found here and from audience reviews on the Rotten Tomatoes page found here.
Collection Method: Because of the sheer amount of data we had to collect for our research, we relied on Instant Data Scraper to scrape comments and reviews. This is a highly efficient tool that allows us to scrape large amounts of data in a short amount of time and export it in a neatly organized table.
Dataset Size: The dataset consists of 468 YouTube comments found here and 561 Rotten Tomatoes audience reviews found here. The YouTube comments range from six to five years ago, before and immediately after the movie's release, while the Rotten Tomatoes reviews range from July 25, 2022 to as recent as October 20, 2025. This gap in time was essential for us to compare how people's perceptions of the film differed over time, in part because it's fairly common for people to change their opinions of a movie after a period of time has passed.
Ethical Considerations: Because these reviews are all public and easily accessible online, we had no qualms about scraping them for data collection purposes. It's important to note that neither YouTube nor Rotten Tomatoes shows the actual full name of the commentor/reviewer and just shows a username, so the privacy of the people who wrote these stays intact.
Analysis Methods
Tools: Python (pandas, VADER, Gensim). Access to our Google Collab IPYNB found here.
- Term Frequency Analysis: By gathering the most dominant terms, we can better understand the common ideas of these comments. Therefore, we can better make claims as to what people are commenting/reviewing about.
- Sentiment Analysis (VADER): We hypothesize that sentiment analysis will reveal the YouTube comments to be generally more positive than the film reviews because many of them were written before the film's release.
- Topic Modeling (Gensim LDA): We ran 5 topics in order to receive a better understanding of the samples we were working with. We discovered how the topics varied between YouTube comments and Rotten Tomatoes Reviews. YouTube comments tend to involve more sarcasm and jokes, which makes sense considering YouTube comment sections tend to be pretty unserious and full of memes. This contrasts with Rotten Tomatoes Reviews, which seemed to express frustration over Luke Skywalker's and Emperor Palpatine's character development. This signals how Rotten Tomatoes users are generally more concerned with the actual quality of the movie and focus on technical aspects like the writing or cinematography.
Group Roles & Responsibilities
Early on, we split up the responsibilities of the project between the group members in a way that highlights our academic strengths. It was decided that David would be the data programmer due to being the most experienced and most comfortable with coding, including Python; Cora would be the website programmer due to having experience with HTML and CSS; Sam would be co-data programmer and write the final essay/copy on the website due to being the least experienced with code. Later, TJ joined our group because of scheduling conflicts with his group and we decided that he would assist with data scraping, Google collab, and writing as well, keeping the work as evenly split as possible.
Results & Analysis
Term Frequency Analysis Results
Using term frequency analysis, we examined the most common words to appear in both the YouTube comments and Rotten Tomatoes audience reviews, respectively; visualizations of our results can be found below.
Sentiment Analysis Results
Using VADER sentiment analysis, we analyzed all of the comments and reviews and automatically flagged them as having a positive, negative, or neutral tone. Visualizations of our results can be found below.
Code Example
Here's how we implemented the sentiment analysis using vaderSentiment:
from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
import pandas as pd
analyzer = SentimentIntensityAnalyzer()
# Analyze sentiment for each text
df['compound'] = df['text'].apply(
lambda x: analyzer.polarity_scores(x)['compound']
)
# Classify sentiment
df['sentiment'] = df['compound'].apply(
lambda x: 'positive' if x > 0.05
else ('negative' if x < -0.05 else 'neutral')
)Topic Modeling Results
Using Gensim's LDA implementation, we identified five (0-4) major topics for both the YouTube comments and Rotten Tomatoes reviews. Visualizations of the data can be found below.
Mixed Methods Results
The following graphs were created through a mix of the methods used above to reveal unique insights. This included analyzing the sentiment distribution of each topic, and pulling the top words that appeared in comments and reviews that the model flagged as either positive or negative.
Key Findings
The computational analysis revealed three major insights:
- Finding 1: Across the board, people were most interested in Palpatine's presence in the film, as his name is the seventh most common word in YouTube comments and the third most common in Rotten Tomatoes reviews. This makes sense considering his sudden reappearance after being absent from the previous two movies. The topic modeling was a bit all over the place, but it does seem to imply that the YouTube comments in particular are fixated on this film being marketed as the supposed "end" of the saga (all of the topics have the word "end" as the most important), which is ironic considering it is the third film to be marketed as such. This could imply that people are interested in the return of a central character to the franchise but are potentially frustrated by the ploy of calling it the "end of the saga" to get people to come and see it, which could be seen as a predatory marketing practice.
- Finding 2: Fewer Rotten Tomatoes reviews were flagged as "neutral" which resulted in the proportion of both positive and negative being greater than YouTube comments. This can likely be explained by the sheer number of "meme" comments present in YouTube comment sections, which might have confused the model. However, the "average" sentiment score indicated on the bar graph indicates that Rotten Tomatoes reviews were very slightly more positive than YouTube comments. However, the difference is incredibly slight.
- Finding 3: Combining term frequency with sentiment analysis yields many interesting results, chief of which being "Palpatine" is listed as a top word in both positive and negative Rotten Tomatoes reviews, but is only a top word in negative YouTube comments. This indicates that people who were more interested in the technical aspects of the film were split on whether or not he should have returned, while YouTube commenters saw it as more of an explicitly negative thing. One of the most common words used in positive Rotten Tomatoes reviews is "better," which could be related to people's perceptions of this film being better than the previous one (but perhaps still not good). "Plot" is also one of the top terms in negative reviews; it is one of the top words in positive reviews, too, but to a much smaller extent.
Detailed Results
Breaking down the sentiment distribution for YouTube comments:
| Sentiment Category | Count | Percentage | Avg. Compound Score |
|---|---|---|---|
| Positive | 89 | 19.1% | 0.67 |
| Neutral | 166 | 35.5% | 0.03 |
| Negative | 212 | 45.4% | -0.84 |
Breaking down the sentiment distribution for Rotten Tomatoes audience reviews:
| Sentiment Category | Count | Percentage | Avg. Compound Score |
|---|---|---|---|
| Positive | 200 | 35.7% | 0.98 |
| Neutral | 52 | 9.3% | 0.24 |
| Negative | 308 | 55% | -0.98 |
What Surprised Us
We were greatly surprised by how small the difference was in average sentiment scores between YouTube comments and Rotten Tomatoes reviews. We had initially expected YouTube comments to be more positive because many were written before the film came out, which is when people found out the story and expressed their disappointment. However, this did not turn out to be the case.
Critical Reflection
This project demonstrates what happens when coding meets culture by revealing insights that neither computational analysis nor traditional close reading could discover alone.
Integration of Methods
What computational methods revealed: Computational methods helped reveal the more "quantifiable" aspects of our analysis such as the most-used terms, which would have taken an incredibly long time to do by ourselves. It was also immensely helpful with the "mixed methods" strategies like the most common words based on positive/negative comments/reviews.
What close reading added: We came in to our analysis with a general understanding of what the data might look like, as Star Wars fans with our own opinions of the movie who also have knowledge of the lore and terminology used. This is how we were able to determine what some patterns meant, like the word "end" being one of the most important in topic modeling. Because we know that The Rise of Skywalker was marketed as the final film in the saga, we were able to assign a meaning to this observation that goes beyond simply giving it a numerical "weight."
📐 Classification Logic
This project connects to Classification Logic by revealing how algorithmic categorization shapes our understanding of audience perceptions and reactions to the film. This is exemplified through the grouping of reviews or comments together based on whether they are "negative," "positive," or "neutral."
Critical question: What nuances are lost when we reduce complex cultural expressions to computational categories? In the context of the previous answer, categories like positive, negative, and neutral aren't always accurate because a review or comment can express more than one emotion and assign them to different aspects of the film (so some people may appreciate the music and cinematography but hate the plot, which seemed to be a common trend). Additionally, a person's tone of voice can't really be adequately conveyed over text, so a machine trying to analyze how a person feels just based on the text they wrote (which could be full of errors or obscure cultural references) can never be as accurate as hearing a person talk or interacting with them personally. As a result, it is always a good idea to be skeptical of any analysis methods that are even slightly subjective, like sentiment analysis and especially topic modeling.
🤖 AI Agency
The use of topic modeling and sentiment analysis demonstrates AI Agency concerns. While the algorithms appear to "discover" meaning, the interpretation and framing of results remains entirely human. We believe in an era where artificial intelligence is increasingly becoming more aware of human norms, it cannot fully comprehend and appreciate human sarcasm and genuine critique. The ability to think and read critically in an academic setting was necessary to better conduct our analysis. The context of the samples the algorithms discovered were presented in a "factual" light, human understanding and basic level context reveals that YouTube commentors love trolling and leaving snarky messages rather than nonsensical movie statements. The algorithms had a slightly better time "discovering" themes in the Rotten Tomatoes samples, but there were still certain examples where the algorithms would over-analyize sarcasm. Ultimately, although AI can help reduce the amount of time it takes to go through every single sample, human understanding and evaluation is desperately needed to fully comprehend data discoveries.
Limitations & Future Directions
What I would do differently: Initially, we were interested in comparing YouTube comments posted before the release date of the film with comments posted after the release date. We quickly realized that there was no easy way to sort YouTube comments by date posted and decided instead to go with our current approach (comparing comments to Rotten Tomatoes reviews). If we were given more time, comparing YouTube comments by date might have been more effective than comparing comments to RT reviews, which is a bit like comparing apples to oranges since user behavior varies between each site.This may help us get a more accurate glimps into specifically how reactions differ before and after the movie's release.
Questions that remain: As mentioned above, it would be interesting to see how some YouTube comments compare to others and if the ones before are more positive than the ones after, or if they're generally about the same and full of memes.
Confidence in conclusions: We are moderately certain about our findings. Although we did our best to draw logical conclusions from the data our analysis provided, sentiment scores and topic modeling is still quite subjective and prone to misunderstanding of sarcasm and other rhetorical devices, which we point out above in the AI section.