patent-wireframes / scripts /eval /intent_elicitation.py
midah's picture
Add intent elicitation eval (Task 3)
09d0e17 verified
"""Task 3 — Intent Elicitation: recover what the designer intended.
Given N-1 views of a patent (as text descriptions, no images), produce a
structured schema describing the patent's design intent. Score against the
gold drawing description using text embedding cosine similarity.
Two sub-tasks:
MC — select the correct component description from a pool of candidates
Free — fill in the structured schema; score vs. gold via cosine similarity
The structured schema forces outputs into comparable format regardless of
writing style, reducing noise in the free-form comparison.
Schema:
{
"viewpoint_set": "what views the patent shows and why",
"primary_view": "which view is the reference/primary",
"component_vocab": ["list", "of", "component", "names"],
"spatial_layout": "how components relate to each other",
"design_intent": "what the designer is communicating about the object"
}
Usage:
python scripts/eval/intent_elicitation.py \
--enriched data/enriched/enriched_2022.parquet \
--n 18 --out results/intent_elicitation_results.json
"""
import argparse
import json
import re
import time
from pathlib import Path
import numpy as np
import pandas as pd
import sys
sys.path.insert(0, str(Path(__file__).parent))
from provider import chat, get_client, get_model
# ── Prompts ───────────────────────────────────────────────────────────────────
FREE_FORM_PROMPT = """You are analyzing a US design patent.
Patent title: {title}
The patent has {n_figs} figures described as follows (some viewpoints may be redacted):
{drawing_description}
Based on this, fill in the design intent schema below as specifically as possible:
{{
"viewpoint_set": "...",
"primary_view": "...",
"component_vocab": ["...", "..."],
"spatial_layout": "...",
"design_intent": "..."
}}
Rules:
- viewpoint_set: name which projection types appear and what each reveals
- primary_view: which single view best communicates the overall form
- component_vocab: list the distinct named parts of this design
- spatial_layout: describe how parts are arranged relative to each other
- design_intent: what visual/functional story is the designer communicating?
Reply with the JSON only."""
MC_PROMPT = """You are analyzing a US design patent to identify its design intent.
Patent title: {title}
Drawing description (N-1 views shown, one view's label is redacted):
{drawing_description}
Which of the following best describes the OVERALL design intent of this patent?
{choices}
Reply with just the letter (A-D)."""
# ── Scoring ───────────────────────────────────────────────────────────────────
def schema_to_text(schema: dict) -> str:
"""Flatten schema to a single text string for embedding."""
parts = []
for k, v in schema.items():
if isinstance(v, list):
parts.append(f"{k}: {', '.join(str(x) for x in v)}")
else:
parts.append(f"{k}: {v}")
return " | ".join(parts)
def cosine_sim(a: np.ndarray, b: np.ndarray) -> float:
a = a / (np.linalg.norm(a) + 1e-8)
b = b / (np.linalg.norm(b) + 1e-8)
return float(np.dot(a, b))
def embed_text(text: str, client) -> np.ndarray | None:
"""Get text embedding via the chat model's implicit representation.
Falls back to a bag-of-words overlap score if embedding API unavailable.
"""
# Use the model to score text similarity via a direct comparison prompt
# (proper embedding would use a dedicated embedding model)
return None # placeholder — see scoring below
def token_overlap_score(pred: str, gold: str) -> float:
"""Token-level F1 between predicted and gold text (no embedding needed)."""
pred_tokens = set(re.findall(r"\w+", pred.lower()))
gold_tokens = set(re.findall(r"\w+", gold.lower()))
if not gold_tokens:
return 0.0
precision = len(pred_tokens & gold_tokens) / max(len(pred_tokens), 1)
recall = len(pred_tokens & gold_tokens) / len(gold_tokens)
if precision + recall == 0:
return 0.0
return 2 * precision * recall / (precision + recall)
# ── Main eval ─────────────────────────────────────────────────────────────────
def run(enriched_path: str, n: int, out_path: str, seed: int = 42):
import random
client = get_client()
df = pd.read_parquet(enriched_path)
# Parse viewpoints and find eligible patents
def parse_vp(desc, fig_num):
m = re.compile(
rf"FIG\.\s*{fig_num+1}\s+is\s+(?:a\s+|an\s+)?(.{{5,80}}?)\s*(?:view|thereof|;|\n|$)",
re.I
).search(desc or "")
return m.group(1).strip().lower() if m else ""
df["vp"] = df.apply(lambda r: parse_vp(r.get("drawing_description", ""), r["figure_number"]), axis=1)
eligible = df.groupby("patent_id").filter(
lambda g: g["vp"].str.contains("perspective").any() and
(g["vp"].str.contains("front elev") | g["vp"].isin(["front elevation","front plan"])).any() and
len(g) >= 3
)["patent_id"].unique().tolist()
rng = random.Random(seed)
rng.shuffle(eligible)
eval_pids = eligible[:n]
print(f"Evaluating {len(eval_pids)} patents | model: {get_model()}")
results = []
free_scores, mc_correct, mc_total = [], 0, 0
for pid in eval_pids:
group = df[df["patent_id"] == pid].sort_values("figure_number")
title = str(group["patent_title"].iloc[0])
draw_desc = str(group["drawing_description"].iloc[0])
n_figs = int(group["n_figures_in_patent"].iloc[0]) if "n_figures_in_patent" in group.columns else len(group)
# ── Task 3a: Free-form intent elicitation ─────────────────────────────
# Input: drawing_description (text) + patent title
# Output: structured schema
# Score: token F1 against gold drawing_description
prompt = FREE_FORM_PROMPT.format(
title=title,
n_figs=n_figs,
drawing_description=draw_desc[:1000], # truncate for context
)
msgs = [{"role": "user", "content": prompt}]
response = chat(client, msgs, max_tokens=400)
predicted_schema = None
try:
m = re.search(r"\{.*\}", response, re.DOTALL)
if m:
predicted_schema = json.loads(m.group())
except Exception:
pass
pred_text = schema_to_text(predicted_schema) if predicted_schema else response
score = token_overlap_score(pred_text, draw_desc)
free_scores.append(score)
time.sleep(0.5)
# ── Task 3b: MC intent selection ──────────────────────────────────────
# Build 4 candidates: gold description + 3 distractors from other patents
other_pids = [p for p in eligible if p != pid]
rng.shuffle(other_pids)
distractor_descs = []
for dpid in other_pids[:3]:
dg = df[df["patent_id"] == dpid]
if not dg.empty:
distractor_descs.append(str(dg["drawing_description"].iloc[0])[:200])
correct_pos = rng.randint(0, 3)
choices_list = distractor_descs[:3]
choices_list.insert(correct_pos, draw_desc[:200])
letters = "ABCD"
choices_str = "\n".join(f"({letters[i]}) {c[:150]}" for i, c in enumerate(choices_list))
mc_prompt = MC_PROMPT.format(
title=title,
drawing_description=draw_desc[:500],
choices=choices_str,
)
mc_msgs = [{"role": "user", "content": mc_prompt}]
mc_response = chat(client, mc_msgs, max_tokens=5)
chosen_letter = re.search(r"\b([ABCD])\b", mc_response.upper())
chosen_idx = letters.index(chosen_letter.group()) if chosen_letter else -1
mc_ok = chosen_idx == correct_pos
mc_correct += int(mc_ok)
mc_total += 1
time.sleep(0.5)
result = {
"patent_id": pid,
"patent_title": title,
"free_form_score": float(score),
"predicted_schema": predicted_schema,
"mc_correct": mc_ok,
"mc_chosen": chosen_idx,
"mc_correct_pos": correct_pos,
}
results.append(result)
print(
f"[{pid}] {title[:35]} | "
f"free F1={score:.2f} | "
f"MC={'✓' if mc_ok else '✗'} | "
f"running MC: {mc_correct}/{mc_total} ({mc_correct/mc_total:.1%})"
)
avg_free = sum(free_scores) / max(len(free_scores), 1)
mc_acc = mc_correct / max(mc_total, 1)
print(f"\n=== INTENT ELICITATION RESULTS ===")
print(f"Free-form token F1: {avg_free:.3f} (higher = describes drawing better)")
print(f"MC accuracy: {mc_acc:.1%} (chance = 25.0%)")
output = {
"model": get_model(),
"summary": {
"free_form_avg_f1": avg_free,
"mc_acc": mc_acc,
"mc_chance": 0.25,
"n": len(results),
},
"results": results,
}
Path(out_path).parent.mkdir(parents=True, exist_ok=True)
with open(out_path, "w") as f:
json.dump(output, f, indent=2)
print(f"Results → {out_path}")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--enriched", default="data/enriched/enriched_2022.parquet")
parser.add_argument("--n", type=int, default=18)
parser.add_argument("--out", default="results/intent_elicitation_results.json")
args = parser.parse_args()
run(args.enriched, args.n, args.out)
if __name__ == "__main__":
main()