Code for "Details in the Novel"

File Map:

• BertPreTrainDataMunge.R
  • Partitions annotated data set into train, dev and test sets.
• BertTrainModel.ipynb
  • Trains BERT classifiers of various sizes to determine optimal model for detail classification.
• ClauseParse.ipynb
  • Breaks a paragraph into its clauses. Used for early exploration of manually identifying a "detail" (with Zola's Germinal as a test case)
• Clustering.ipynb
  • Explores unsupervised clustering on my dataset (via k-means and hierarchical methods).
• DetailSummarize.R
  • Initial aggregation of data points following annotations.
• DetailTagging.R
  • Annotation script: iterates over a list of samples, printing out the text in the console and attaching the user's input to a data frame. Allows a reader to tag texts without knowing the source (prima facie). For now, used to apply scores to 364 800 character passages.
• Embeddings.ipynb
  • Generates and stores word embedding values from BERT. Also contains a function for "augmenting" embedding vectors with additional points of information.
• GeneralWordFrequencies.ipynb
  • Catch-all script that calculates the relative frequency of word w in a text t.
• GetTimeWords.R
  • Extracts relevant words from the Harvard Inquirer dictionary.
• KeyWordsInContext.ipynb
  • Highlights the usage of a word of interest across corpus. Associated cli version is kwic.py.
• LogOddsTables.R
  • Aggregates the results from LogOdds calculations and produces charts for each class' most distinctive words.
• LogOddsWithPriors.ipynb
  • Implements the log-odds ratio with an informative Dirichlet prior (described in Monroe et al. 2009, Fighting Words). As adapted from a (completed) INFO-256 assignment
• PrepositionChiSquared.ipynb
  • Runs the chi-squared test for independence using the frequency of prepositions to determine whether there is a significance difference in each preposition's usage in detail and not_detail samples.
• ProcessedTextsToFile.ipynb
  • Helper file for shuttling texts through a processing pipeline (see gut_tokenize.py) and writing their outputs to a /processed/ dir.
• SamplerForAnnotation.ipynb
  • Randomly draws n samples of k length from a directory of .txt files. Writes samples to a (configurable) output directory, maintaining the location (start:stop char index) in each text that a given sample was drawn from in order to easily trace the sample back to its place of context in a given text file.
• SpecificitySummarize.R
  • Investigates the outputs from WordnetSpecificity.ipynb. Tallies metrics by class "label," produces visualizations, determines normality of calculated data with stat_qq(), etc.
• TimeSummarize.R
  • Investigates the frequency of "Time" words per Harvard Inquirer
• TokensWithHighestAttention.ipynb
  • Pools the attention weights (from BERT) of each token. Outputs the 100 highest and 100 lowest attention-values. Provides an easy look up for a word of interest's attention weight.
• WordnetSpecificity.ipynb
  • Routine to calculate a "specificity" metric for an inputted block of text (described in Nelson 2020, Computational Grounded Theory: A Methodological Framework). Also applies part of speech tallies for classes of interest. Writes results to file.
• concat_txts.sh
  • Combines .txt files in a directory into one-big .txt file. (I'm trying to get better at bash...)
• get_26_gutentexts.sh
  • Invokes Wolff's Project Gutenberg wrapper for a provided series of books and their Gutenberg id's. Saves each book's output as .txt file. (Again, trying to get better at bash.)
• bert_interp.py
  • Uses the captum library to extract and visualize the attention weights of each token used by BERT during the classification of my annotated samples. As adapted from an INFO-256 activity
• gut_tokenize.py
  • Reusable/importable text-cleaning methods for books from Project Gutenberg. Attempts to advance past the opening "boilerplate," removes chapter numbers, etc. Relatively "unopinionated" -- doesn't remove numbers, punctuation, stopwords, etc. (maintains as much as possible to keep options open for downstream analysis).
• data/ contains the feature matrix, which holds:
  • sample name
  • parts of speech data
  • specificity scores
  • my detail "rating" annotation
• dh-fair/ contains the poster for my presentation at the 2022 Berkeley Digital Humanities fair
• paper/ contains the project write-up, which is 6 pages and adheres to the ACL's formatting guidelines.
• viz/ contains graphs and tables related to the log-odds and specificity work.

Corpus:

• austen pride_and_prejudice
• austen sense_and_sensibility
• carroll alice_in_wonderland
• conrad heart_of_darkness
• defoe robinson_crusoe
• dickens bleak_house
• dickens great_expectations
• dostoevsky crime_and_punishment
• eliot adam_bede
• eliot middlemarch
• fielding tom_jones
• fitzgerald great_gatsby
• flaubert madame_bovary
• gaskell north_and_south
• hawthorne scarlet_letter
• lewis babbit
• melville moby_dick
• more utopia
• shelley frankenstein
• shelley the_last_man
• sinclair the_jungle
• tolstoy anna_karenina
• verne around_the_world_in_80_days
• verne twenty_thousand_leagues_under_the_sea
• wells the_time_machine
• wilde picture_of_dorian_grey
• zola germinal
• zola the_fortune_of_the_rougons