Classifying Movie Scripts by Genre

1
Classifying Movie Scripts by Genre

• Alex Blackstock
• Matt Spitz
• 6/9/08

2
Overview

• Motivation
• classifying movie scripts may identify box office
  flops and successes before they're even produced!
• Data
• freely-available movie scripts (DailyScripts.com,
  etc)?
• IMDB genres (several labels/movie)?
• Tools
• Lucene
• MEMM from PA3
• jBNC (naïve Bayes classifier)?
• Stanford Named Entity Recognizer
• Stanford Part-Of-Speech Tagger

3
Processing Scripts
4
Features

• Non-NLP
• dialogue shape
• character information
• NLP
• POS ratios
• Named Entity appearances
• Character-Based NLP
• analyze individual characters
• exclamations
• main vs. secondary

5
Evaluation Metrics

• Example output
• Blade II (gold labels Action, Thriller, Horror)?
• guessed labels Action, Adventure, Horror,
  Thriller, ...
• F1 Score
• per genre
• weighted-average over all genres
• of guesses allowed of gold labels
• Partial Credit Score
• allows for some error
• guesses allowed of gold labels 1.5
• penalized for guesses that are beyond gold
  labels, but still get points

6
Conclusions

• Success!
• best feature set basic NLP POS tagging
• PC Score 0.601
• F1 Score 0.551
• Classifier comparison (jBNC)?
• N-way classification problem
• 22 genres
• average of 3.02 genres/datum
• Dataset Issues
• consistency
• diversity
• size