"""Text-based UI: highlight propaganda spans in the terminal, with technique and confidence.
Highlights are suggestions, not verdicts: SI tops out near 0.5 F1 even for the best systems.
"""
from __future__ import annotations
import html
import pandas as pd
from .techniques import DISPLAY_NAMES, TECHNIQUES
_PALETTE = [196, 208, 214, 220, 190, 118, 49, 45, 39, 33, 99, 135, 171, 205]
COLOURS = {t: _PALETTE[i % len(_PALETTE)] for i, t in enumerate(TECHNIQUES)}
RESET = "\033[0m"
def _style(tech: str) -> str:
return f"\033[1;38;5;{COLOURS.get(tech, 250)}m"
def render_terminal(text: str, preds: pd.DataFrame, min_confidence: float = 0.0) -> str:
"""Inline highlights with numbered tags, then a legend of every span."""
preds = preds[preds["confidence"] * preds["technique_confidence"] >= min_confidence]
preds = preds.sort_values(["start", "end"]).reset_index(drop=True)
out, pos, legend = [], 0, []
for n, r in enumerate(preds.itertuples(index=False), 1):
if r.start < pos: # overlapping span: list it in the legend only
legend.append((n, r))
continue
out.append(text[pos:r.start])
out.append(f"{_style(r.technique)}{text[r.start:r.end]}{RESET}\033[2m[{n}]{RESET}")
pos = r.end
legend.append((n, r))
out.append(text[pos:])
lines = ["".join(out), "", "-" * 72]
for n, r in legend:
score = r.confidence * r.technique_confidence
lines.append(f"[{n:>2}] {_style(r.technique)}{DISPLAY_NAMES.get(r.technique, r.technique):<34}{RESET}"
f" span {r.confidence:.2f} label {r.technique_confidence:.2f} -> {score:.2f}")
return "\n".join(lines)
def render_sentences(rows: list[dict]) -> str:
lines = ["Critical sentences:"]
for r in sorted(rows, key=lambda r: -r["confidence"]):
techs = ", ".join(DISPLAY_NAMES.get(t, t) for t in r["techniques"])
lines.append(f" ({r['confidence']:.2f}) {r['sentence']}\n -> {techs}")
return "\n".join(lines)
# HTML hues (degrees on the colour wheel), grouped by technique family.
# Separate from COLOURS above, which holds xterm-256 codes for the terminal.
HTML_HUES = {
# reds/oranges: attacks on people
"Name_Calling,Labeling": 0,
"Loaded_Language": 20,
"Appeal_to_fear-prejudice": 38,
# yellows/greens: distortion of facts
"Exaggeration,Minimisation": 55,
"Causal_Oversimplification": 80,
"Black-and-White_Fallacy": 105,
"Doubt": 130,
# cyans/blues: appeals to authority and the crowd
"Appeal_to_Authority": 175,
"Bandwagon,Reductio_ad_hitlerum": 200,
"Flag-Waving": 225,
# purples/pinks: slogans and thought-stoppers
"Slogans": 260,
"Thought-terminating_Cliches": 285,
"Repetition": 310,
"Whataboutism,Straw_Men,Red_Herring": 335,
"NONE": 0, # unused in HTML output; NONE spans are dropped before rendering
}
def html_colour(technique: str, lightness: int = 85) -> str:
"""Background colour for a technique in HTML output."""
hue = HTML_HUES.get(technique, 0)
saturation = 0 if technique not in HTML_HUES else 75
return f"hsl({hue}, {saturation}%, {lightness}%)"
def render_html(text: str, preds: pd.DataFrame, title: str = "") -> str:
preds = preds.sort_values(["start", "end"])
legend = "".join(
f''
f"{html.escape(DISPLAY_NAMES.get(t, t))}"
for t in sorted(set(preds["technique"]))
)
parts, pos = [], 0
for r in preds.itertuples(index=False):
if r.start < pos:
continue
parts.append(html.escape(text[pos:r.start]))
tip = f"{DISPLAY_NAMES.get(r.technique, r.technique)} (span {r.confidence:.2f}, label {r.technique_confidence:.2f})"
parts.append(f''
f"{html.escape(text[r.start:r.end])}")
pos = r.end
parts.append(html.escape(text[pos:]))
return (f"{html.escape(title)}"
f""
f""
f"{legend}"
f"
"
f"{''.join(parts)}
"
f"")