qanta-buzzer-logistic

Buzzer Model with Feature Engineering

This model represents the Tech+Research team submission for Technica 2024, This is a logistic regression buzzer model that uses engineered features to decide when to buzz in on quiz bowl questions. Features The model uses several carefully engineered features:

Length: Question and guess length analysis
Named Entities: Detection of locations, people, dates
Giveaway Phrases: Common patterns signaling answerable questions
Empty Guess Detection
Capital Analysis: Analysis of capitalized terms
Rate: Guessing rate analysis
Question Keywords: 5W+H detection
Pronoun Features

Model Performance

Buzz Ratio: 0.36

Files

final_submission.model.pkl: The trained logistic regression model final_submission.featurizer.pkl: Feature extraction pipeline

Training Details Used GPR guesser with the following command:

python3 buzzer.py --guesser_type=gpr --limit=500 \
    --gpr_guesser_filename=../models/buzztrain_gpr_cache \
    --questions=../data/qanta.buzztrain.json.gz --buzzer_guessers gpr \
    --logistic_buzzer_filename=models/with_legNgcbgr \
    --features Length Entities Giveaway EmptyGuess Capital Rate QKeyword Pronoun

Dependencies

spacy==3.7.2 en-core-web-sm nltk
numpy
pandas
scikit-learn
torch

Final Presentation https://docs.google.com/presentation/d/1rNrDNbHKAZmpZRR6rcFvgTUvSe1LAcIqZLUysMGMDOk/edit#slide=id.g4dfce81f19_0_45

Mentors
Jordan Boyd-Graber
Yoo Yeon Sung
Yu (Hope) Hou

Authors
dpal0 - Diksha Pal
Archita Naik
Christabel
Miray Ramirez
Grace Yang
Nishna Makala