Spaces:
Sleeping
Sleeping
File size: 16,939 Bytes
acb0568 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 | """
absa.py
Aspect-Based Sentiment Analysis: offline pre-computation, query intent parsing,
and validation utilities.
Precompute output writes four per-aspect columns
(aspect_food, aspect_service, aspect_price, aspect_wait_time)
directly into data/processed/meta-NYC-restaurant.parquet.
Each column is globally min-max normalized to [0,1] across the restaurants
that pass the β₯15-review filter. Restaurants without reviews get NaN.
Price is stored *unblended* here β blending with the Google Maps tier ($/$$/β¦)
happens at query-time inside the ranker so it stays transparent.
"""
import re
import numpy as np
import pandas as pd
import nltk
from nltk.stem import WordNetLemmatizer
from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
from collections import Counter, defaultdict
from tqdm import tqdm
nltk.download('punkt', quiet=True)
nltk.download('punkt_tab', quiet=True)
nltk.download('wordnet', quiet=True)
lemmatizer = WordNetLemmatizer()
analyzer = SentimentIntensityAnalyzer()
# ββ Aspect keyword table (update after running frequency_analysis in Step 0) ββ
ASPECT_KEYWORDS = {
"food": ["food", "dish", "meal", "taste", "flavor",
"cuisine", "menu", "portion", "ingredient", "quality",
"pizza", "chicken", "coffee", "cheese", "burger", "sauce",
"fries", "rice", "salad", "sandwich", "meat", "soup",
"sushi", "wine", "steak", "fish", "bread", "pork",
"ramen", "tacos", "beef", "wings", "shrimp", "pasta",
"dessert", "noodles", "seafood", "dumplings", "bagels", "beer",
"drinks", "cocktails"],
"service": ["service", "staff", "waiter", "waitress",
"server", "host", "bartender", "manager", "employees",
"friendly", "attentive", "rude", "helpful"],
"price": ["price", "cost", "cheap", "expensive",
"affordable", "value", "worth", "overpriced", "reasonable",
"pricey", "priced", "money"],
"wait_time": ["wait", "waiting", "line", "queue", "reservation",
"delay", "busy", "packed", "crowded",
"hour", "slow", "minutes", "quick", "fast"],
}
LEMMATIZED_KEYWORDS = {
aspect: set(lemmatizer.lemmatize(kw) for kw in kws)
for aspect, kws in ASPECT_KEYWORDS.items()
}
NEGATION_WORDS = {"not", "no", "never", "isn't", "wasn't", "don't", "doesn't"}
CLAUSE_SPLITTERS = ["but", "however", "although", "though", "yet", "while"]
DEFAULT_PREFS = {
"food": 0.40, "service": 0.30, "price": 0.20, "wait_time": 0.10,
}
POSITIVE_BOOST = 0.20
PRIOR_STRENGTH = 5 # Bayesian smoothing hyper-parameter k
ASPECT_COLS = [f"aspect_{a}" for a in ASPECT_KEYWORDS] # 4 cols written into meta
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Step 0 β Frequency Analysis
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def frequency_analysis(reviews_df, top_n=30):
"""Word-frequency audit to validate ASPECT_KEYWORDS."""
all_words = []
for text in reviews_df["text"].dropna():
words = re.findall(r'\b[a-z]+\b', text.lower())
all_words.extend(words)
freq = Counter(all_words)
print(f"Top {top_n} most common words in reviews:")
for word, count in freq.most_common(top_n):
print(f" {word:<20} {count:>8,}")
print("\nFrequency check for ASPECT_KEYWORDS:")
for aspect, keywords in ASPECT_KEYWORDS.items():
print(f"\n[{aspect}]")
for kw in keywords:
count = freq.get(kw, 0)
flag = " β low (<500)" if count < 500 else ""
print(f" {kw:<20} {count:>8,}{flag}")
return freq
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Step 1 β ABSA Helpers
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def split_into_sentences(text):
if not isinstance(text, str) or not text.strip():
return []
return nltk.sent_tokenize(text)
def split_clauses(sentence):
result = [sentence]
for splitter in CLAUSE_SPLITTERS:
new_result = []
for clause in result:
parts = clause.split(f" {splitter} ")
new_result.extend(parts)
result = new_result
return [c.strip() for c in result if c.strip()]
def lemmatize_words(words):
return [lemmatizer.lemmatize(w) for w in words]
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Step 1 β Core ABSA (with Bayesian Smoothing)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def compute_aspect_scores_from_texts(texts, priors, k=PRIOR_STRENGTH):
"""
Compute aspect sentiment scores from an iterable of review texts.
Logic identical to compute_aspect_scores_single but decoupled from
the review dataframe β the caller does the grouping.
Returns {aspect: smoothed_score} covering all ASPECT_KEYWORDS.
"""
aspect_score_lists = {aspect: [] for aspect in ASPECT_KEYWORDS}
for review in texts:
for sentence in split_into_sentences(review):
for clause in split_clauses(sentence):
raw_words = clause.lower().split()
lemma_words = lemmatize_words(raw_words)
aspect_trigger_strength = {}
for i, lemma in enumerate(lemma_words):
for aspect, kws in LEMMATIZED_KEYWORDS.items():
if lemma in kws:
context_start = max(0, i - 3)
context_end = min(len(raw_words), i + 4)
context_text = " ".join(raw_words[context_start:context_end])
strength = abs(analyzer.polarity_scores(context_text)["compound"])
if aspect not in aspect_trigger_strength or \
strength > aspect_trigger_strength[aspect]:
aspect_trigger_strength[aspect] = strength
if not aspect_trigger_strength:
continue
clause_sentiment = analyzer.polarity_scores(clause)["compound"]
if len(aspect_trigger_strength) == 1:
aspect = next(iter(aspect_trigger_strength))
aspect_score_lists[aspect].append(clause_sentiment)
else:
dominant = max(aspect_trigger_strength, key=aspect_trigger_strength.get)
aspect_score_lists[dominant].append(clause_sentiment)
smoothed = {}
for aspect, scores in aspect_score_lists.items():
n = len(scores)
observed_mean = float(np.mean(scores)) if n > 0 else 0.0
smoothed[aspect] = (n * observed_mean + k * priors[aspect]) / (n + k)
return smoothed
def compute_priors(all_scores_list):
"""Estimate global priors from a sample of restaurants (no smoothing applied)."""
aggregated = defaultdict(list)
for scores in all_scores_list:
for aspect, val in scores.items():
aggregated[aspect].append(val)
return {aspect: float(np.mean(vals)) for aspect, vals in aggregated.items()}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Step 1 β Precompute All Aspects + Write to Meta Parquet
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def precompute_all_aspect_scores(
reviews_df,
meta_path="data/processed/meta-NYC-restaurant.parquet",
prior_sample_frac=0.10,
random_state=42,
):
"""
Compute per-restaurant ABSA scores and write four aspect columns
(aspect_food, aspect_service, aspect_price, aspect_wait_time) directly
into the meta parquet at `meta_path`, globally min-max normalized to [0,1].
Only restaurants present in `reviews_df` (i.e. those that passed the
β₯15-review filter upstream) are scored; all other meta rows get NaN
in the aspect columns.
Returns (priors_dict, n_scored).
"""
# Pre-group by gmap_id ONCE β avoids O(nΒ²) linear scans per restaurant
print("Grouping reviews by gmap_id...")
texts_by_gid = (
reviews_df[["gmap_id", "text"]]
.dropna(subset=["text"])
.groupby("gmap_id")["text"]
.apply(list)
)
gmap_ids = texts_by_gid.index.to_numpy()
n_total = len(gmap_ids)
print(f"Will score {n_total:,} restaurants (β₯15 reviews).")
# ββ Phase 1: prior estimation (no smoothing, 10% sample) ββ
rng = np.random.default_rng(random_state)
sample_size = max(1, int(len(gmap_ids) * prior_sample_frac))
sample_ids = rng.choice(gmap_ids, size=sample_size, replace=False)
print(f"\nPhase 1/2: estimating priors on {sample_size:,} sampled restaurants...")
sample_scores = []
for gid in tqdm(sample_ids, desc="priors"):
raw = {asp: [] for asp in ASPECT_KEYWORDS}
for review in texts_by_gid.loc[gid]:
for sentence in split_into_sentences(review):
for clause in split_clauses(sentence):
lws = lemmatize_words(clause.lower().split())
for lw in lws:
for asp, kws in LEMMATIZED_KEYWORDS.items():
if lw in kws:
raw[asp].append(
analyzer.polarity_scores(clause)["compound"])
break
sample_scores.append({
asp: float(np.mean(v)) if v else 0.0
for asp, v in raw.items()
})
priors = compute_priors(sample_scores)
print(f"Estimated priors: { {k: round(v, 4) for k, v in priors.items()} }")
# ββ Phase 2: smoothed scores for every restaurant ββ
print(f"\nPhase 2/2: computing Bayesian-smoothed aspect scores for all {n_total:,}...")
records = []
for gid in tqdm(gmap_ids, desc="aspects"):
smoothed = compute_aspect_scores_from_texts(texts_by_gid.loc[gid], priors)
smoothed["gmap_id"] = gid
records.append(smoothed)
scores_df = pd.DataFrame(records)
# fix: Percentile-based normalization per aspect (robust to outliers) ββ
# Clip at [1st, 99th] percentile then min-max on the clipped range.
# Done once against the full database so rankings are stable across queries.
for aspect in ASPECT_KEYWORDS:
col = scores_df[aspect]
lo = col.quantile(0.01)
hi = col.quantile(0.99)
clipped = col.clip(lower=lo, upper=hi)
if hi > lo:
scores_df[f"aspect_{aspect}"] = (clipped - lo) / (hi - lo)
else:
scores_df[f"aspect_{aspect}"] = 0.5
# Percentile rank for price and wait_time (display only, not used in ranking)
for aspect in ("price", "wait_time"):
scores_df[f"aspect_{aspect}_pct"] = (
scores_df[f"aspect_{aspect}"].rank(pct=True) * 100
).round(0).astype("Int64")
PCT_COLS = ["aspect_price_pct", "aspect_wait_time_pct"]
out_cols = ["gmap_id"] + ASPECT_COLS + PCT_COLS
# ββ Merge into meta parquet: drop any stale aspect_* columns first ββ
print(f"\nLoading meta from {meta_path} to merge aspect columns...")
meta = pd.read_parquet(meta_path)
stale = [c for c in ASPECT_COLS if c in meta.columns]
if stale:
print(f" Dropping stale columns: {stale}")
meta = meta.drop(columns=stale)
merged = meta.merge(scores_df[out_cols], on="gmap_id", how="left")
assert len(merged) == len(meta), "merge changed row count"
merged.to_parquet(meta_path, index=False)
covered = merged[ASPECT_COLS[0]].notna().sum()
print(f"Wrote {merged.shape[0]:,} rows to {meta_path} "
f"({covered:,} with aspect scores, "
f"{merged.shape[0] - covered:,} without reviews β NaN).")
return priors, int(covered)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Step 2 β Query Intent Parser
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def get_aspect_prefs(query):
"""
Infer user aspect preferences from a natural-language query.
Returns a dict {food, service, price, wait_time} that sums to 1.0.
"""
prefs = DEFAULT_PREFS.copy()
triggered = False
for clause in split_clauses(query.lower()):
lemmas = lemmatize_words(clause.split())
for lemma in lemmas:
for aspect, kws in LEMMATIZED_KEYWORDS.items():
if lemma in kws:
prefs[aspect] += POSITIVE_BOOST
triggered = True
if not triggered:
return DEFAULT_PREFS.copy()
total = sum(prefs.values())
return {k: v / total for k, v in prefs.items()}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Validation helpers β kept as tools, not part of the precompute path
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def validate_vader_clause_length(reviews_df, sample_n=500):
lengths_before = []
lengths_after = []
split_counts = []
for text in reviews_df["text"].dropna().sample(sample_n, random_state=42):
for sentence in split_into_sentences(text):
lengths_before.append(len(sentence.split()))
clauses = split_clauses(sentence)
split_counts.append(len(clauses))
for c in clauses:
lengths_after.append(len(c.split()))
print(f"Avg words per sentence (before split): {np.mean(lengths_before):.1f}")
print(f"Avg words per clause (after split): {np.mean(lengths_after):.1f}")
print(f"Sentences that were split: {np.mean([c > 1 for c in split_counts]):.1%}")
print(f"Clauses with < 5 words: {np.mean([l < 5 for l in lengths_after]):.1%}")
def validate_absa_accuracy(sample_sentences, human_labels):
tp = defaultdict(int)
fp = defaultdict(int)
fn = defaultdict(int)
for sentence, true_aspects in zip(sample_sentences, human_labels):
predicted = set()
for lemma in lemmatize_words(sentence.lower().split()):
for aspect, kws in LEMMATIZED_KEYWORDS.items():
if lemma in kws:
predicted.add(aspect)
for aspect in ASPECT_KEYWORDS:
in_true = aspect in true_aspects
in_pred = aspect in predicted
if in_true and in_pred: tp[aspect] += 1
if not in_true and in_pred: fp[aspect] += 1
if in_true and not in_pred: fn[aspect] += 1
print(f"\n{'Aspect':<12} {'Precision':>10} {'Recall':>10}")
print("-" * 35)
for aspect in ASPECT_KEYWORDS:
p = tp[aspect] / (tp[aspect] + fp[aspect]) if tp[aspect] + fp[aspect] > 0 else 0.0
r = tp[aspect] / (tp[aspect] + fn[aspect]) if tp[aspect] + fn[aspect] > 0 else 0.0
flag = " β low recall" if r < 0.5 else ""
print(f"{aspect:<12} {p:>10.1%} {r:>10.1%}{flag}")
|