主题建模 LDA 和 BERTopic

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.decomposition import LatentDirichletAllocation
import numpy as np


def fit_lda(documents, n_topics=5, max_features=1000):
    cv = CountVectorizer(
        max_features=max_features,
        stop_words="english",
        min_df=2,
        max_df=0.9,
    )
    X = cv.fit_transform(documents)
    lda = LatentDirichletAllocation(
        n_components=n_topics,
        random_state=42,
        max_iter=50,
        learning_method="online",
    )
    doc_topic = lda.fit_transform(X)
    feature_names = cv.get_feature_names_out()
    return lda, cv, doc_topic, feature_names


def print_top_words(lda, feature_names, n_top=10):
    for idx, topic in enumerate(lda.components_):
        top_idx = np.argsort(-topic)[:n_top]
        words = [feature_names[i] for i in top_idx]
        print(f"topic {idx}: {' '.join(words)}")

from bertopic import BERTopic

topic_model = BERTopic(
    embedding_model="sentence-transformers/all-MiniLM-L6-v2",
    min_topic_size=15,
    verbose=True,
)

topics, probs = topic_model.fit_transform(documents)
info = topic_model.get_topic_info()
print(info.head(20))
valid_topics = info[info["Topic"] != -1]["Topic"].tolist()
for topic_id in valid_topics[:5]:
    print(f"topic {topic_id}: {topic_model.get_topic(topic_id)[:10]}")

---
name: topic-picker
description: Pick LDA or BERTopic for a corpus. Specify library, knobs, evaluation.
version: 1.0.0
phase: 5
lesson: 15
tags: [nlp, topic-modeling]
---

Given a corpus description (document count, avg length, domain, language, compute budget), output:

1. Algorithm. LDA / NMF / BERTopic / Top2Vec / FASTopic. One-sentence reason.
2. Configuration. Number of topics: `recommended = max(5, round(sqrt(n_docs)))`, clamped to 200 for corpora under 40,000 docs; permit >200 only when the corpus is genuinely large (>40k) and note the increased compute cost. `min_df` / `max_df` filters and embedding model for neural approaches also belong here.
3. Evaluation. Topic coherence (c_v) via `gensim.models.CoherenceModel`, topic diversity, and a 20-sample human read.
4. Failure mode to probe. For LDA, "junk topics" absorbing stopwords and frequent terms. For BERTopic, the -1 outlier cluster swallowing ambiguous documents.

Refuse BERTopic on documents longer than the embedding model's context window without a chunking strategy. Refuse LDA on very short text (tweets, reviews under 10 tokens) as coherence collapses. Flag any n_topics choice below 5 as likely wrong; flag >200 on corpora under 40k docs as likely over-splitting.