prathamchopra001/skills/finding_extraction

Scientific Finding Extraction

Task

Extract scientific findings from academic paper chunks. A finding is a factual claim supported by evidence in the text.

Output Format

Always respond with valid JSON:

{
  "findings": [
    {
      "claim": "One clear factual statement",
      "confidence": 0.85,
      "evidence": "Direct quote or reference from text",
      "tags": ["keyword1", "keyword2"]
    }
  ]
}

Return {"findings": []} if no extractable findings exist.

Decision Framework

Is It a Finding?

✅ YES - Extract these:

• Experimental results: "X increased by 15% when..."
• Quantitative observations: "The model achieved 92% accuracy..."
• Causal relationships: "Higher learning rates led to..."
• Comparative results: "Method A outperformed B by..."

❌ NO - Skip these:

• Methodology: "We implemented using Python..."
• Background: "Reinforcement learning is a paradigm..."
• Hypotheses: "We hypothesize that..."
• Future work: "Further research could..."
• Definitions: "Q-learning is defined as..."

Confidence Scoring

| Score | Criteria | |-------|----------| | 0.90+ | Direct experimental result with statistics (p-values, CIs) | | 0.75-0.89 | Clear conclusion with quantitative support | | 0.60-0.74 | Qualitative observation with some evidence | | 0.45-0.59 | Inference or interpretation | | <0.45 | Weak claim, speculation |

Evidence Guidelines

• Quote directly when possible (use "...")
• Reference tables/figures: "As shown in Table 3..."
• Keep evidence concise (1-2 sentences max)

Examples

Input

"Results show that the Q-learning agent with ε=0.1 achieved convergence in 450 episodes, compared to 780 episodes for ε=0.3 (p<0.01, n=50 runs)."

Output

{
  "findings": [
    {
      "claim": "Lower exploration rate (ε=0.1) leads to faster convergence than higher rates (ε=0.3) in Q-learning",
      "confidence": 0.92,
      "evidence": "450 vs 780 episodes, p<0.01, n=50 runs",
      "tags": ["q-learning", "exploration", "convergence", "epsilon"]
    }
  ]
}

Input

"We used the OpenAI Gym environment for our experiments. The agent was trained using a neural network with two hidden layers."

Output

{
  "findings": []
}

(This is methodology, not a finding)

Anti-Patterns

• ❌ Extracting methodology as findings
• ❌ Including paper metadata (authors, publication date)
• ❌ Confidence scores without reasoning
• ❌ Vague claims: "The method works well"
• ❌ Multiple claims in one finding (split them)

Quality Checklist

1. Is output valid JSON?
2. Is each claim a single, specific factual statement?
3. Does evidence support the confidence score?
4. Are tags relevant keywords (not sentences)?
5. Did I skip methodology/background/hypotheses?