Run2 keyword classification

Keyword-Based Document Classification (Improved)

Key Improvements from Round 1

1. Use re.IGNORECASE instead of lowercasing text - preserves case-sensitive acronyms like DNA, RNA, LLM in patterns while matching case-insensitively.
2. Use \b word boundaries in regex patterns to avoid substring false positives (e.g., "bar" in "Raman" matching music keyword "bar", "opera" matching "operator").
3. Weight keywords by specificity: Highly specific terms (e.g., "Schwarzschild", "CRISPR-Cas9") get higher weights than generic terms (e.g., "model", "system").
4. Avoid ambiguous single-word keywords: Words like "bar", "note", "rest", "measure", "scale", "opera" are common across domains. Either use multi-word phrases or add word boundaries.

Algorithm

1. Extract text from each document (first 3 pages for PDFs, full text for PPTX/DOCX)
2. For each category, compute a weighted score by matching regex patterns case-insensitively
3. Assign document to the highest-scoring category
4. Default to music_history if no keywords matched (catch-all)

Pattern Design Guidelines

• Use \b word boundaries for short/ambiguous words: r"\bDNA\b" not "DNA"
• Use multi-word phrases where possible: "trapped ion" not "trapped" or "ion"
• Assign higher weights (10) to domain-specific terms that uniquely identify a field
• Assign lower weights (1-2) to generic terms that appear across domains
• Use re.findall(pattern, text, re.IGNORECASE) to count matches

Example Score Computation

import re

def classify(text, keywords_dict):
    scores = {}
    for category, keywords in keywords_dict.items():
        score = 0
        for pattern, weight in keywords:
            matches = re.findall(pattern, text, re.IGNORECASE)
            score += len(matches) * weight
        scores[category] = score

    best = max(scores, key=scores.get)
    if scores[best] == 0:
        return "music"  # catch-all
    return best

Category Keyword Strategy

• LLM: Focus on NLP/ML terminology, model names (GPT, BERT, Llama), training methods (RLHF, SFT, DPO)
• Trapped Ion & QC: Focus on quantum hardware terms (qubit, entanglement, gate), ion-specific terms (paul trap, Lamb-Dicke, sideband)
• Black Hole: Focus on GR/astrophysics terms (Schwarzschild, event horizon, spacetime, accretion)
• DNA: Focus on molecular biology terms (genome, nucleotide, CRISPR, polymerase, sequencing)
• Music History: Focus on music theory/history terms (composer, symphony, genre, melody, lyrics) - also serves as catch-all