File size: 11,570 Bytes
a882be6 | 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 | """
Topic modeling inference: BERTopic, LSI, HDP, LDA, NMF.
Models are lazy-loaded and cached on first use.
"""
from typing import Dict, List, Tuple
from .models import ModelType, SUPPORTED_MODELS, TopicInfo, DocumentResult, TopicResult
from .preprocessor import preprocess_batch, preprocess_batch_classical, full_preprocess
# ── Model caches ──────────────────────────────────────────────────────────────
_bertopic_models: Dict[str, object] = {} # key → BERTopic instance
_vectorizer_cache: Dict[str, object] = {} # key → (vectorizer, model)
# ── BERTopic ──────────────────────────────────────────────────────────────────
def _run_bertopic(texts: List[str], model_type: str) -> Tuple[TopicResult, Dict]:
from bertopic import BERTopic
from sentence_transformers import SentenceTransformer
from umap import UMAP
cfg = SUPPORTED_MODELS[model_type]
embedding_model_id = cfg["embedding_model"]
cleaned = preprocess_batch_classical(texts) # lowercase + no punctuation + no stop words
n = len(cleaned)
if n < 3:
raise ValueError("BERTopic requires at least 3 documents.")
# UMAP constraints for small datasets:
# n_neighbors must be in [2, n-1]
# n_components must be < n
# init="random" skips spectral eigsh decomposition (which requires n_components+1 < n)
umap_model = UMAP(
n_neighbors=max(2, min(n - 1, 15)),
n_components=max(2, min(n - 1, 5)),
min_dist=0.0,
metric="cosine",
init="random",
random_state=42,
)
embedding_model = SentenceTransformer(embedding_model_id)
topic_model = BERTopic(
embedding_model=embedding_model,
umap_model=umap_model,
language="english",
calculate_probabilities=True,
verbose=False,
min_topic_size=2,
)
topics, probs = topic_model.fit_transform(cleaned)
topic_info_df = topic_model.get_topic_info()
topic_infos: List[TopicInfo] = []
for _, row in topic_info_df.iterrows():
tid = row["Topic"]
if tid == -1:
continue
words_scores = topic_model.get_topic(tid) or []
keywords = [w for w, _ in words_scores[:10]]
scores = [s for _, s in words_scores[:10]]
topic_infos.append(TopicInfo(
topic_id=tid,
keywords=keywords,
scores=scores,
doc_count=int(row["Count"]),
))
doc_results: List[DocumentResult] = []
for i, (text, tid) in enumerate(zip(texts, topics)):
prob = float(probs[i][tid]) if tid != -1 and probs is not None else 0.0
kws = []
if tid != -1:
words_scores = topic_model.get_topic(tid) or []
kws = [w for w, _ in words_scores[:5]]
doc_results.append(DocumentResult(
doc_id=i, text=text, topic_id=int(tid),
topic_keywords=kws, probability=prob,
))
outliers = sum(1 for t in topics if t == -1)
result = TopicResult(
model_type=model_type,
num_topics=len(topic_infos),
topics=topic_infos,
documents=doc_results,
outlier_count=outliers,
)
viz_data = {"type": "bertopic", "model": topic_model, "docs": list(texts)}
return result, viz_data
# ── NMF (scikit-learn / TF-IDF) ──────────────────────────────────────────────
def _run_sklearn(texts: List[str], model_type: str, n_topics: int = 5) -> Tuple[TopicResult, Dict]:
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.decomposition import NMF
prep = full_preprocess(texts)
cleaned = [" ".join(tokens) for tokens in prep["bc_texts"]]
n_topics = min(n_topics, len(texts))
vectorizer = TfidfVectorizer(max_df=0.95, min_df=1, max_features=1000)
dtm = vectorizer.fit_transform(cleaned)
model = NMF(n_components=n_topics, random_state=42, max_iter=400)
doc_topic_matrix = model.fit_transform(dtm)
feature_names = vectorizer.get_feature_names_out()
topic_infos: List[TopicInfo] = []
for tid, component in enumerate(model.components_):
top_idx = component.argsort()[-10:][::-1]
keywords = [feature_names[i] for i in top_idx]
scores = [float(component[i]) for i in top_idx]
doc_count = int((doc_topic_matrix.argmax(axis=1) == tid).sum())
topic_infos.append(TopicInfo(topic_id=tid, keywords=keywords,
scores=scores, doc_count=doc_count))
doc_results: List[DocumentResult] = []
for i, (text, row) in enumerate(zip(texts, doc_topic_matrix)):
tid = int(row.argmax())
prob = float(row[tid] / row.sum()) if row.sum() > 0 else 0.0
doc_results.append(DocumentResult(
doc_id=i, text=text, topic_id=tid,
topic_keywords=topic_infos[tid].keywords[:5], probability=prob,
))
result = TopicResult(model_type=model_type, num_topics=n_topics,
topics=topic_infos, documents=doc_results, outlier_count=0)
viz_data = {
"type": "nmf", "model": model, "dtm": dtm,
"vectorizer": vectorizer, "doc_topic_matrix": doc_topic_matrix,
}
return result, viz_data
# ── Gensim (LSI / HDP / LDA) ─────────────────────────────────────────────────
def _run_gensim(texts: List[str], model_type: str, n_topics: int = 5) -> Tuple[TopicResult, Dict]:
from gensim.models import LsiModel, LdaModel, HdpModel
prep = full_preprocess(texts)
dictionary = prep["dictionary"]
corpus = prep["corpus"]
n_topics = min(n_topics, len(texts))
if model_type == ModelType.LSI:
model = LsiModel(corpus, id2word=dictionary, num_topics=n_topics)
topic_infos: List[TopicInfo] = []
for tid in range(n_topics):
word_scores = model.show_topic(tid, topn=10)
keywords = [w for w, _ in word_scores]
scores = [abs(float(s)) for _, s in word_scores]
topic_infos.append(TopicInfo(topic_id=tid, keywords=keywords, scores=scores, doc_count=0))
doc_results: List[DocumentResult] = []
for i, (text, bow) in enumerate(zip(texts, corpus)):
vec = model[bow]
if not vec:
tid, prob = 0, 0.0
else:
tid, prob = max(vec, key=lambda x: abs(x[1]))
tid, prob = int(tid), abs(float(prob))
topic_infos[tid].doc_count += 1
doc_results.append(DocumentResult(
doc_id=i, text=text, topic_id=tid,
topic_keywords=topic_infos[tid].keywords[:5], probability=prob,
))
result = TopicResult(model_type=model_type, num_topics=n_topics,
topics=topic_infos, documents=doc_results, outlier_count=0)
viz_data = {"type": "lsi"}
return result, viz_data
elif model_type == ModelType.HDP:
model = HdpModel(corpus, id2word=dictionary)
raw_topics = model.show_topics(num_topics=50, num_words=10, formatted=False)
topic_map: Dict[int, TopicInfo] = {}
for tid, word_scores in raw_topics:
keywords = [w for w, _ in word_scores]
scores = [float(s) for _, s in word_scores]
topic_map[int(tid)] = TopicInfo(topic_id=int(tid), keywords=keywords,
scores=scores, doc_count=0)
doc_results = []
for i, (text, bow) in enumerate(zip(texts, corpus)):
vec = sorted(model[bow], key=lambda x: x[1], reverse=True)
if not vec:
tid, prob = 0, 0.0
else:
tid, prob = int(vec[0][0]), float(vec[0][1])
ti = topic_map.get(tid)
if ti:
ti.doc_count += 1
doc_results.append(DocumentResult(
doc_id=i, text=text, topic_id=tid,
topic_keywords=ti.keywords[:5] if ti else [], probability=prob,
))
active = sorted([t for t in topic_map.values() if t.doc_count > 0],
key=lambda t: t.topic_id)
result = TopicResult(model_type=model_type, num_topics=len(active),
topics=active, documents=doc_results, outlier_count=0)
viz_data = {"type": "hdp", "model": model, "corpus": corpus, "dictionary": dictionary}
return result, viz_data
else: # LDA via gensim
model = LdaModel(corpus, id2word=dictionary, num_topics=n_topics,
random_state=42, passes=10, alpha="auto")
topic_infos = []
for tid in range(n_topics):
word_scores = model.show_topic(tid, topn=10)
keywords = [w for w, _ in word_scores]
scores = [float(s) for _, s in word_scores]
topic_infos.append(TopicInfo(topic_id=tid, keywords=keywords, scores=scores, doc_count=0))
doc_results = []
for i, (text, bow) in enumerate(zip(texts, corpus)):
topic_probs = model.get_document_topics(bow, minimum_probability=0.0)
if not topic_probs:
tid, prob = 0, 0.0
else:
tid, prob = max(topic_probs, key=lambda x: x[1])
tid, prob = int(tid), float(prob)
topic_infos[tid].doc_count += 1
doc_results.append(DocumentResult(
doc_id=i, text=text, topic_id=tid,
topic_keywords=topic_infos[tid].keywords[:5], probability=prob,
))
result = TopicResult(model_type=model_type, num_topics=n_topics,
topics=topic_infos, documents=doc_results, outlier_count=0)
viz_data = {"type": "lda", "model": model, "corpus": corpus, "dictionary": dictionary}
return result, viz_data
# ── Public API ────────────────────────────────────────────────────────────────
def run_topic_model(
texts: List[str],
model_type: str = ModelType.BERTOPIC_MINI,
n_topics: int = 5,
) -> TopicResult:
cfg = SUPPORTED_MODELS[model_type]
if cfg["type"] == "bertopic":
result, _ = _run_bertopic(texts, model_type)
elif cfg["type"] == "gensim":
result, _ = _run_gensim(texts, model_type, n_topics)
else:
result, _ = _run_sklearn(texts, model_type, n_topics)
return result
def run_topic_model_with_viz(
texts: List[str],
model_type: str = ModelType.BERTOPIC_MINI,
n_topics: int = 5,
) -> Tuple[TopicResult, str]:
"""Run topic modeling and generate visualization HTML. Returns (TopicResult, viz_html)."""
from .visualizer import generate_viz_html
cfg = SUPPORTED_MODELS[model_type]
if cfg["type"] == "bertopic":
result, viz_data = _run_bertopic(texts, model_type)
elif cfg["type"] == "gensim":
result, viz_data = _run_gensim(texts, model_type, n_topics)
else:
result, viz_data = _run_sklearn(texts, model_type, n_topics)
viz_html = generate_viz_html(viz_data)
return result, viz_html
|