This is an NLP-centered project for capstone 2 of the Galvanize Data Science Intensive. Presented April 2, 2021.
As a life-long reader and as a librarian, I value books. Book-sellers and libraries want to help readers choose books they will enjoy and they want that process to be easy. We want the selection process to be enjoyable and the reading to be rewarding. Therefore, it makes sense to develop computational models for matching readers with books. I want to eventually build a model that uses a "taxonomy of terror" to quantify just how scary a book is in order to help readers who want a book that is "a little scary", "chilling and creepy", or "absolutely terrifying". This is sometimes accomplished through subgenre, but a scale of scariness (or romance, or humor, or any other metric of sentiment) would be very useful in helping to advise readers in a digital space.
Being able to discern one author from another inside of a given genre is a step in the right direction because authors often adhere to a given sub-genre.
I used a supervised approach to text classification among three authors.
The dataset I used came from a 2017 Kaggle competition using individual lines of text as predictors of the author.
The target would be the author and the goal is to train a model that can identify the author of an unseen piece of text, out of three authors whose work is included in the corpus.
The three authors represented are: Edgar Allan Poe, H.P. Lovecraft, and Mary Wollstonecraft Shelley.
| Author | Vocabulary Size |
|---|---|
| Edgar Allan Poe | 15,261 |
| H.P. Lovecraft | 14, 504 |
| Mary Shelley | 11,515 |
I also applied NLTK's Sentiment Intensity Analyzer on each line for each author's work. The following table shows those sentiments averaged by author:
| Author | Negative | Neutral | Positive | Compound |
|---|---|---|---|---|
| Edgar Allan Poe | .07 | .84 | .09 | .04 |
| H.P. Lovecraft | .09 | .84 | .07 | -.06 |
| Mary Shelley | .10 | .77 | .13 | .06 |
The model I used to classify the texts was Multinomial Naive Bayes. When tested, the following accuracy scores and predictive features were noted.
| Author Classifiers | Accuracy | Predictive Feature Samples |
|---|---|---|
| Poe/Lovecraft/Shelley | 81% | Aberrant, abhor, little, good, great, time, eye, think, far, word, mean, look |
| Poe/Lovecraft | 88% | Abandonment, abdication, man, old, house, say, night, know, night, dream, day, look |
| Poe/Shelley | 83% | Love, life, man, eye, heart, time, father, hope, little, friend, death, think |
I chose to keep what I saw as "victorian stop-words" because they seemed to be highly predictive for each author. Therefore, I am curious as to whether that decision may have led to an overfit model or if it's a good application in a purpose-built model for classifying authors by their syntactic signatures.
Notes: Each of these writers makes some unusual stylistic and vocabulary choices. I wondered if those are easily spotted in machine learning. In the story "X-ing a Paragrab" there is a letter wherein all of the "o"s in the text are replaced with x. So, there is a line for example that reads: "Gxxd Lxrd, Jxhn, hxw yxu dx lxxk!" and goes on to write "X-cellent" in place of excellent and "X-entric" in place of eccentric.
- Explore AI text generation and chatbots
- Explore ways models might “reward” words from the taxonomy of terror in classifying within the genre
- Incorporate more authors and time-periods in the corpus
Benzahra, Marc. 2019. “How to Evaluate Text Readability with NLP.” Medium. https://medium.com/glose-team/how-to-evaluate-text-readability-with-nlp-9c04bd3f46a2 (March 24, 2021).
Dare, Eleanor. 2021. “Rear Psycho: Hitchcock’s Horror Language Mediated by AI.” Leonardo Electronic Almanac 23(1). https://www.leoalmanac.org/rear-psycho-hitchcocks-horror-language-mediated-by-ai-eleanor-dare/ (March 29, 2021).
Pearl, Mike. “Behold the Incredible Horror Stories an AI Will Write If You Let It.” The Outline. https://theoutline.com/post/8302/ai-incredible-horror-stories (March 29, 2021).
Rahwan, Iyad. “Can AI Learn to Scare Us? With Shelley, MIT Researchers Aim for Goosebumps.” MIT Media Lab. https://www.media.mit.edu/posts/can-ai-learn-to-scare-us-with-shelley-mit-researchers-aim-for-goosebumps/ (March 29, 2021).
“Spooky Author Identification.” https://kaggle.com/c/spooky-author-identification (April 1, 2021).
Steger, Sara. 2013. “Patterns of Sentimentality in Victorian Novels.” Digital Studies/Le champ numérique 3(2). http://www.digitalstudies.org/articles/10.16995/dscn.235/print/ (April 1, 2021).
Zimmer, Ben. 2011. “The Jargon of the Novel, Computed.” The New York Times. https://www.nytimes.com/2011/07/31/books/review/the-mechanic-muse-the-jargon-of-the-novel-computed.html (April 1, 2021).