My Portfolio

Logo

My GitHub
LinkedIn
Email

Back

Unsupervised Sentiment Analysis of Bridgerton YouTube Trailer Comments

The popular Netflix Series “Bridgerton” (based on the novel series by Julia Quinn) seems to cause many emotions amongst viewers whenever new season trailers are released. What do people think about the different seasons based on the trailers? Let’s find out!

Skills: webscraping, data cleaning, social media opinion mining, NLP, unsupervised sentiment analysis
Libraries: youtube-comment-downloader, pandas, spacy, emoji, transformers, wordcloud, plotly

Almack’s Assembly Rooms [1].

For this project, I first scraped the 200 most popular comments of each Netflix Bridgerton Series trailer on YouTube with the library youtube-comment-downloader [2]:

!pip install youtube-comment-downloader
!pip install emoji
import emoji
import spacy
from transformers import pipeline
import pandas as pd
from itertools import islice
from youtube_comment_downloader import *
downloader = YoutubeCommentDownloader()
nlp = spacy.load('en_core_web_sm')

#Season 1 (same for the other trailers)
s1 = []
comments_s1 = downloader.get_comments_from_url('https://www.youtube.com/watch?v=gpv7ayf_tyE', sort_by=SORT_BY_POPULAR)
for comment in islice(comments_s1,200):
    #print(comment)
    s1.append(comment)

df_s1 = pd.DataFrame(s1)
df_s1 = df_s1.drop(columns=['author', 'channel', 'cid', 'photo', 'heart', 'reply', 'time_parsed'], axis=1) #I am removing these for privacy reasons

Next, pre-processing was necessary. I remove emojis and emoticons, punctuation, usernames, tokenized, lowercased, and shaped the comments into useful dataframes that would serve as input to the model:

#Season 1 (same for the other trailers)
df_s1["text"] = df_s1["text"].apply(lambda x: emoji.replace_emoji(x, replace=""))
df_s1['text'] = df_s1['text'].str.replace('@[A-Za-z0-9]+\s?', '', regex=True)
df_s1['text_clean'] = df_s1['text'].apply(lambda x: [t.text.lower() for t in nlp.tokenizer(x) if not t.is_punct and t.is_alpha])

comment_s1_list = df_s1['text_clean'].tolist()
def edit_for_model(comment_list):
  comments = []
  for element in comment_list:
    element = " ".join(element)
    comments.append(element)
  return comments

comment_s1 = edit_for_model(comment_s1_list)

The dataset was now ready for the model to be predicted. I decided to use disilbert-base-uncased-emotion [3], as it offered a more fine-grained display of emotional categories (‘sadness’, ‘joy’, ‘love’, ‘anger’, ‘fear’, ‘surprise’) compared to the popular three-way distinction (positive, negative, neutral). The model assigned to every post a percentage of how likely each of these emotions are. I decided to keep all of them instead of choosing the most likely one (i.e. the one with the highest score), as this helps to represent a more nuanced picture of the user’s sentiment. In addition, this model is trained on Twitter data, which is fairly close to YouTube comments, as both are part of the wider social media register.

classifier = pipeline("sentiment-analysis",model='bhadresh-savani/distilbert-base-uncased-emotion', return_all_scores=True)
def sentiment(comments):
  results = []
  for c in comments:
    prediction = classifier(c, )
    [prediction] = prediction
    results.append(prediction)
  return results

results_s1 = sentiment(comment_s1)


Finally, I made some visualizations based on the results of the model:


We an immediately see that the vast majority of all trailer comments reflect the emotion “joy”. So far so good! However, we can also see that “anger” is equally present, with roughly a third of the comments reflecting this emotion. Especially season 3 seems to have stirred a lot of “anger”. This might have to do with the changes that the directors made to the original book series by Julia Quinn, which would also explain the high number of comments relating to “surprise”. The users express the emotion of “love” in some way or another in all trailers more or less equally. “Sadness” finds the greatest expression in the spin-off Queen Charlotte, which might have to do with the sadness of the story.

In order to learn more about the content of the comments, we need to look at some word-frequencies.

Season 1 Trailer Comments
In season 1, commentators were seemingly more focused on the cast and general changes to the original book series and it being presented as a show now. Comparisons are made to other series and movies such as Pride and Prejudice or 50 Shades of Grey.

Season 2 Trailer Comments
Season 2 is more obsessed with the original book, referencing and predicting parts in the series, specifically the love story between Anthony and Kate. Comparisons are drawn to the romance genre and its tropes, as well as other shows like Gossip Girl.

Season 3 Trailer Comments
In the third season, there is a greater obsession with the relationships (see friendship, friend) between the characters, rather than the actors. Great focus is put on Penelope, Cressida and Eloise, and to a degree Collin. But it seems that commentators are more concerned with how the friendship unfolds rather than the romance.

Season 4 Trailer Comments
Season four seams to be a mix of season 2 and season 3. There is again a frequent reference back to the book series (book, story), and to the two main protagonists, Benedict and Sopie. Predictions are very present again as well as markers of opinion and evidentiality (think, maybe, thought).

Queen Charlotte Trailer Comments
This series trailer is, similar to the first trailer, very interested in the cast and in how far the story was changed compared to the historical person of Queen Charlotte.

Caveats and Shortcomings

References:

[1] “Almack’s Assembly Rooms”. Wikimedia Commons. https://commons.wikimedia.org/wiki/File:Almack%27s_Assembly_Rooms_inside.jpg (last accessed Sept 3, 2024).
[2] https://pypi.org/project/youtube-comment-downloader
(last accessed Sept 3, 2024). [3] https://huggingface.co/bhadresh-savani/distilbert-base-uncased-emotion
(last accessed Sept 3, 2024).

Back