import html
import json
import re
from functools import lru_cache
from pathlib import Path
import gradio as gr
import pandas as pd
from src import config as cfg
from src.aspect_dict import get_aspect_dict
from src.evaluator import aspect_to_overall_sentiment
from src.inference import AspectPredictor
ROOT = Path(__file__).resolve().parent
REPORT_DIR = ROOT / "reports"
CHECKPOINT_DIR = ROOT / "checkpoints" / "meta_acsa"
META_ENCODER_PATH = ROOT / "data" / "meta_encoder.pkl"
ASPECT_DESCRIPTIONS = {
"SIZE": "Size, fit, length, width, whether it runs large or small",
"MATERIAL": "Fabric, feel, thickness, breathability, material comfort",
"QUALITY": "Workmanship, durability, loose threads, zippers, buttons, post-wash performance",
"APPEARANCE": "Color, pattern, consistency with pictures, visual presentation",
"STYLE": "Style, cut, fashionability, whether it flatters the body shape",
"VALUE": "Price, value for money, worthiness of purchase, return/exchange tendency",
}
LABEL_COLORS = {
"Positive": "#15803d",
"Negative": "#b91c1c",
"Not_Mentioned": "#64748b",
}
POSITIVE_CUES = {
"love", "loved", "great", "good", "perfect", "excellent", "nice", "soft",
"comfortable", "beautiful", "pretty", "flattering", "worth", "recommend",
"amazing", "stylish", "quality", "durable", "cute", "fits", "fit",
}
NEGATIVE_CUES = {
"bad", "poor", "terrible", "awful", "cheap", "scratchy", "itchy", "thin",
"small", "large", "tight", "loose", "broke", "broken", "ripped", "tore",
"overpriced", "return", "returning", "waste", "disappointed", "uncomfortable",
"shrunk", "defective", "wrong",
}
def _load_json(path: Path) -> dict:
if not path.exists():
return {}
with path.open("r", encoding="utf-8") as f:
return json.load(f)
def _fmt_pct(value):
try:
return f"{float(value) * 100:.2f}%"
except Exception:
return "-"
def _metric_table(rows):
return pd.DataFrame(rows)
def baseline_tables():
data = _load_json(REPORT_DIR / "evaluation_comparison.json")
overall = data.get("overall_3class_comparison", {})
rows = []
for name, metrics in overall.items():
rows.append({
"Model": name.replace("__", " / ").replace("_", " "),
"Macro-F1": round(metrics.get("macro_f1", 0), 4),
"Accuracy": round(metrics.get("accuracy", 0), 4),
"Macro-F1 (%)": _fmt_pct(metrics.get("macro_f1", 0)),
"Accuracy (%)": _fmt_pct(metrics.get("accuracy", 0)),
})
aspect_rows = []
proposed = data.get("proposed_per_aspect", {}).get("per_aspect", {})
no_meta = data.get("acsa_no_meta_per_aspect", {}).get("per_aspect", {})
for aspect in cfg.ASPECTS:
p = proposed.get(aspect, {})
b = no_meta.get(aspect, {})
aspect_rows.append({
"Aspect": aspect,
"Proposed F1": round(p.get("macro_f1", 0), 4),
"No-Meta F1": round(b.get("macro_f1", 0), 4),
"F1 Gain": round(p.get("macro_f1", 0) - b.get("macro_f1", 0), 4),
"Proposed Acc": round(p.get("accuracy", 0), 4),
"No-Meta Acc": round(b.get("accuracy", 0), 4),
"Acc Gain": round(p.get("accuracy", 0) - b.get("accuracy", 0), 4),
})
return _metric_table(rows), _metric_table(aspect_rows), _bars_html(rows, "Model")
def ablation_tables():
data = _load_json(REPORT_DIR / "ablation_summary.json")
rows = []
for name, block in data.items():
rows.append({
"Variant": name,
"Mean Macro-F1": round(block.get("mean_macro_f1", 0), 4),
"Mean Accuracy": round(block.get("mean_accuracy", 0), 4),
"Macro-F1 (%)": _fmt_pct(block.get("mean_macro_f1", 0)),
"Accuracy (%)": _fmt_pct(block.get("mean_accuracy", 0)),
})
aspect_rows = []
proposed = data.get("Proposed", {}).get("per_aspect", {})
for aspect in cfg.ASPECTS:
row = {"Aspect": aspect}
for variant, block in data.items():
m = block.get("per_aspect", {}).get(aspect, {})
row[f"{variant} F1"] = round(m.get("macro_f1", 0), 4)
row[f"{variant} Acc"] = round(m.get("accuracy", 0), 4)
if variant != "Proposed":
row[f"{variant} F1 Delta"] = round(
m.get("macro_f1", 0) - proposed.get(aspect, {}).get("macro_f1", 0), 4
)
aspect_rows.append(row)
return _metric_table(rows), _metric_table(aspect_rows), _bars_html(rows, "Variant")
def _bars_html(rows, label_key):
if not rows:
return "No report JSON found.
"
max_f1 = max(float(r.get("Macro-F1", r.get("Mean Macro-F1", 0))) for r in rows) or 1
max_acc = max(float(r.get("Accuracy", r.get("Mean Accuracy", 0))) for r in rows) or 1
parts = [""]
for row in rows:
label = html.escape(str(row[label_key]))
f1 = float(row.get("Macro-F1", row.get("Mean Macro-F1", 0)))
acc = float(row.get("Accuracy", row.get("Mean Accuracy", 0)))
parts.append(
f"
{label}
"
f"
F1 {f1:.4f}"
f"
"
f"
Acc {acc:.4f}"
f"
"
"
"
)
parts.append("
Overall sentiment: {html.escape(overall)}
"]
parts.append("")
for aspect in cfg.ASPECTS:
label = result["aspects"].get(aspect, "Not_Mentioned")
color = LABEL_COLORS.get(label, "#64748b")
parts.append(
f"
"
f"
{aspect}
"
f"
{label}
"
f"
{ASPECT_DESCRIPTIONS[aspect]}
"
"
"
)
parts.append("
No metadata attention returned.
"
names = {
"meta_chunk_1": "Meta chunk 1",
"meta_chunk_2": "Meta chunk 2",
"meta_chunk_3": "Meta chunk 3",
"meta_chunk_4": "Meta chunk 4",
}
max_v = max(float(v) for v in attn.values()) or 1
parts = [""]
for key, value in attn.items():
value = float(value)
parts.append(
f"
{names.get(key, key)}"
f"
"
f"
{value:.3f} "
)
parts.append("
Enterprise Diagnostic Summary
",
f"
Risk level: {risk}
",
f"
Detected drivers: {', '.join(mentioned) if mentioned else 'No aspect was strongly mentioned.'}
",
]
if negatives:
parts.append("
Immediate action: prioritize review of " + ", ".join(negatives) + " related product claims, PDP copy, sizing guidance, QA, and return reasons.
")
if positives:
parts.append("
Reusable strengths: emphasize " + ", ".join(positives) + " in merchandising, ad copy, and search filters.
")
parts.append("
Interpretation note: metadata attention shows how much the model relied on product attributes in addition to the review text.
Model artifacts found. Proposed model inference is enabled.
"
items = "".join(f"{html.escape(x)}" for x in missing)
return f"Model artifacts are missing in this workspace. Upload these files to enable Tabs 3 and 4: