Spaces:
Sleeping
Sleeping
File size: 10,016 Bytes
fb3e070 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | """
utils/sentiment.py
==================
Lightweight keyword-based positive-sentiment detector.
Used as a post-processing correction layer during inference to prevent
clearly positive / happy text from being mislabelled as stressed.
The detector counts positive and negative sentiment indicators in the
input text and returns a dampening factor in [0, 1]. When multiplied
by the raw model stress probability, the result ensures that sentences
like *"I am happy"* or *"I love her and she loves me back"* receive
stress scores well below 10 %.
It also detects *negated* stress phrases such as *"not stressed"* or
*"don't feel anxious"* and applies a moderate dampening factor so that
explicit denials of stress are not treated as stress indicators.
Design
------
- Only *dampens* β never *inflates* β the model score.
- Requires a strong positive signal **and** the absence of negative /
stress indicators before applying any correction.
- Keeps a generous margin so that ambiguous or mixed-sentiment text is
left to the model.
- Negated-stress detection replaces matched phrases with a placeholder
before the negative-hit scan so that "not stressed" does not block
dampening the way genuine stress words would.
"""
from __future__ import annotations
import re
# ---------------------------------------------------------------------------
# Positive-sentiment indicators (case-insensitive, word-boundary)
# ---------------------------------------------------------------------------
_POSITIVE_PATTERNS: list[re.Pattern[str]] = [
re.compile(
r"\b("
r"happy|happiness|joy|joyful|joyous|elated|bliss|blissful"
r"|cheerful|delighted|delightful|ecstatic|euphoric"
r"|glad|pleased|wonderful|amazing|awesome|fantastic|fabulous"
r"|terrific|magnificent|marvelous|superb|splendid"
r"|great|excellent|incredible|brilliant|outstanding|perfect"
r"|phenomenal|glorious|enjoyable|pleasant|overjoyed|jubilant"
r"|love|loved|loving|adore|adored|cherish|cherished"
r"|grateful|thankful|blessed|fortunate|lucky"
r"|optimistic|hopeful|enthusiastic|excited|thrilled"
r"|content|contented|satisfied|fulfilled|peaceful"
r"|proud|confident|empowered|inspired|motivated"
r"|relaxed|calm|serene|tranquil|comfortable"
r"|smile|smiling|smiled|laugh|laughing|laughed|grin|grinning|grinned"
r"|celebrate|celebrating|celebrated|celebration"
r"|beautiful|gorgeous"
r")\b",
re.IGNORECASE,
),
]
_POSITIVE_PHRASES: list[re.Pattern[str]] = [
re.compile(
r"\b("
r"feel(?:s|ing)?\s+(?:great|good|amazing|wonderful|fantastic|awesome|fine|nice|blessed|happy|excellent|incredible|brilliant)"
r"|love\s+(?:my|this|her|him|them|it|life)"
r"|loves?\s+me"
r"|(?:good|great|wonderful|amazing|fantastic|awesome|excellent)\s+(?:day|time|mood|news|life|morning|evening|night)"
r"|having\s+(?:a\s+)?(?:great|good|wonderful|amazing|fantastic|blast|ball)"
r"|(?:in|into)\s+(?:a\s+)?(?:great|good|wonderful|amazing|fantastic|awesome|excellent)\s+mood"
r"|so\s+(?:happy|glad|grateful|thankful|excited|thrilled|proud|pleased)"
r"|life\s+is\s+(?:good|great|beautiful|wonderful|amazing)"
r"|couldn['\u2019]?t\s+be\s+(?:happier|better)"
r"|on\s+top\s+of\s+the\s+world"
r"|over\s+the\s+moon"
r"|best\s+(?:day|time|thing)"
r")\b",
re.IGNORECASE,
),
]
# ---------------------------------------------------------------------------
# Negated-stress patterns β stress keywords explicitly preceded by a negator.
# These are detected BEFORE the negative-hit scan and replaced with a
# placeholder so that "not stressed" is treated as a positive/neutral
# signal rather than a stress indicator.
# ---------------------------------------------------------------------------
_NEGATED_STRESS_PATTERNS: list[re.Pattern[str]] = [
re.compile(
r"\b(?:not|no\s+longer|don['\u2019]?t|doesn['\u2019]?t|isn['\u2019]?t"
r"|aren['\u2019]?t|wasn['\u2019]?t|weren['\u2019]?t|haven['\u2019]?t"
r"|hasn['\u2019]?t|hadn['\u2019]?t|won['\u2019]?t|wouldn['\u2019]?t"
r"|can['\u2019]?t|less|barely|hardly|never)\s+"
r"(?:\w+\s+){0,3}"
r"(?:stress(?:ed|ful|ing)?|anxious|anxiety|worried|worrying|worry"
r"|depress(?:ed|ion|ing)?|overwhelm(?:ed|ing)?|panic(?:king)?"
r"|scared|afraid|fear(?:ful)?|exhausted|nervous|tense)\b",
re.IGNORECASE,
),
]
# ---------------------------------------------------------------------------
# Negative / stress indicators β if ANY of these are present we leave
# the model score alone (even when positive words also appear).
# ---------------------------------------------------------------------------
_NEGATIVE_PATTERNS: list[re.Pattern[str]] = [
re.compile(
r"\b("
r"stress(?:ed|ful|ing)?|anxious|anxiety|worried|worrying|worry"
r"|depress(?:ed|ion|ing)?|sad|sadness|miserable|unhappy"
r"|angry|anger|furious|frustrated|frustration|irritated"
r"|scared|afraid|fear(?:ful)?|terrified|panic(?:king)?"
r"|overwhelm(?:ed|ing)?|exhausted|burnt?\s*out"
r"|hopeless|helpless|desperate|despair"
r"|lonely|isolated|alone|abandoned"
r"|hate|hating|loathe|detest|resent"
r"|cry(?:ing)?|sob(?:bing)?|tears|weep(?:ing)?"
r"|suffer(?:ing)?|pain(?:ful)?|hurt(?:ing)?"
r"|fail(?:ed|ing|ure)?|ruin(?:ed)?"
r"|can['\u2019]?t\s+(?:take|handle|cope|stand|bear|sleep|breathe)"
r"|don['\u2019]?t\s+(?:know\s+what\s+to\s+do|want\s+to)"
r"|give\s+up|giving\s+up"
r"|breakdown|break\s+down|falling\s+apart"
r"|insomnia|nightmare|nightmares"
r"|debt|bankrupt|fired|layoff"
r"|suicide|suicidal|self[- ]?harm"
r"|deadline[s]?|overdue"
r"|sleep\s+depriv(?:ed|ation)|sleepless(?:ness)?"
r"|haven['\u2019]?t\s+slept|not\s+slept|no\s+sleep|no\s+rest"
r"|piling\s+up|pile\s+up|buried\s+(?:in|under)"
r"|falling\s+behind|can['\u2019]?t\s+keep\s+up"
r"|wearing\s+(?:me\s+)?down|breaking\s+down|worn\s+out"
r"|drowning\s+in|stretched\s+thin|at\s+my\s+(?:limit|breaking\s+point)"
r"|running\s+on\s+(?:empty|no\s+sleep|fumes)"
r")\b",
re.IGNORECASE,
),
]
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def compute_sentiment_dampening(text: str) -> float:
"""Return a dampening factor in ``[0.0, 1.0]`` for the stress score.
* ``1.0`` β no change (leave the model score as-is).
* Values ``< 1.0`` β reduce (dampen) the stress score.
The factor is only lowered when:
1. Negated stress phrases are found (e.g. "not stressed"), OR
2. At least one positive indicator is found **and** zero genuine
negative/stress indicators remain.
Parameters
----------
text : str
The raw user input.
Returns
-------
float
Multiplicative dampening factor for the stress probability.
"""
if not text or not text.strip():
return 1.0
text_lower = text.lower()
# ββ Negation detection ββ
# Replace negated stress phrases with a neutral placeholder so they
# are not counted as genuine stress indicators below.
processed = text_lower
negation_hits = 0
for pat in _NEGATED_STRESS_PATTERNS:
matches = pat.findall(processed)
negation_hits += len(matches)
processed = pat.sub("__negated__", processed)
# Count genuine (non-negated) negative hits on the processed text.
negative_hits = 0
for pat in _NEGATIVE_PATTERNS:
negative_hits += len(pat.findall(processed))
if negative_hits > 0:
# Genuine stress indicators remain β leave model score unchanged.
return 1.0
# Count positive hits (single words + phrases) on the original text.
positive_hits = 0
for pat in _POSITIVE_PATTERNS:
positive_hits += len(pat.findall(text_lower))
for pat in _POSITIVE_PHRASES:
positive_hits += len(pat.findall(text_lower))
# ββ Negation-based dampening ββ
# When stress keywords are explicitly negated and no genuine stress
# indicators remain, apply a moderate dampening factor.
if negation_hits > 0:
if positive_hits >= 1:
# Negated stress + positive words β strong dampening.
return 0.06
# Negated stress alone β moderate dampening.
return 0.35 if negation_hits == 1 else 0.22
if positive_hits == 0:
return 1.0
# ββ Standard positive dampening ββ
# 1 positive hit β factor 0.08 (score capped at ~8 % of raw)
# 2 positive hits β factor 0.05
# 3+ positive hits β factor 0.03
if positive_hits >= 3:
return 0.03
if positive_hits >= 2:
return 0.05
return 0.08
def get_sentiment_score(text: str) -> float:
"""Return a sentiment score in ``[0.0, 1.0]`` for a piece of text.
* ``0.0`` β strongly positive (no stress signals).
* ``1.0`` β strongly negative / stressful.
* ``0.5`` β neutral (no clear signal either way).
The score is derived from the same keyword lists used by
:func:`compute_sentiment_dampening`.
"""
if not text or not text.strip():
return 0.5
text_lower = text.lower()
negative_hits = 0
for pat in _NEGATIVE_PATTERNS:
negative_hits += len(pat.findall(text_lower))
positive_hits = 0
for pat in _POSITIVE_PATTERNS:
positive_hits += len(pat.findall(text_lower))
for pat in _POSITIVE_PHRASES:
positive_hits += len(pat.findall(text_lower))
total = positive_hits + negative_hits
if total == 0:
return 0.5
# Map to [0, 1]: more negative β closer to 1.0
return negative_hits / total
|