nguyentr17 / cs685-narrative-analysis

[CS685] Examining Medical Narratives of Eating Disorder Recovery on Reddit

Code Structure (tasks are described by Section 5)

1. Narrative Detection

• /narrative_detection/explore_data.ipynb contains the code to train the classification task

2. Domain-Specific Sentiment Lexicons

• /data-collection contains both code to crawl the data nd code for sentiment polarity
  • /embeddings/word2vec_word_embeddings.ipynb and ./word_types.ipynb are used to train the embeddings and compute scores respectively

3. Instruction Prompting for Trigger and Factor Extraction

• /clinical-extractions/ contains the code for the extraction of trigger and factors
  • ./chatgpt.py is the code to extract helpful and harmful factors
  • ./chatgpt_extract_trigger.py is the code to extract triggers
  • ./utils.py contains all the prompts used and experimented with
  • ./sample_selection_trigger_extraction.ipynb uses topic modeling to generate samples for evaluation and few-shot prompting. It also contains code for evaluation.

4. Topic Modeling

• /topic_modeling/ contains the code for the topic modeling experiments.

  • /data/input/custom_stopwords.txt contains custom stopwords.

  • /data/analysis/topic_label_*.txt contains the topic labels for k=* and top 10 keywords for each topic.

  • /data/positive_topic_dist.csv contains the topic distribution for positive posts.

  • /script/lmw_result.ipynb contains code for preprocessing, training and analyzing the topic modeling results.

  • /script/lmw.py contains helper code for training the LDA model.

5. Power and Agency Analysis

• /power_frames/power-frames-ORIGINAL is the original repo from https://github.com/maria-antoniak/power-frames with slight modifications.
• /power_frames/power-frames is the modified power analysis with coreference resolution added.
• Important scripts inside /power_frames/:
  • `compute_power.ipynb': taken from the original power-frames repo, but with our own personas defined
  • 'compare_power.ipynb': make charts and figures comparing different methods
  • 'find_missing_verbs.ipynb': parse data with spaCy and extract verbs
  • 'gpt_augment_lexicon.ipynb': prompt chatgpt to label verbs