Skip to content

TopicalPageRank

TopicalPageRank (Sterckx et al., 2015) extends SingleRank by biasing the random walk toward topically important words. Instead of uniform teleportation, PageRank uses a personalization vector derived from per-word topic weights.

Users supply pre-computed topic weights as a {lemma: weight} dictionary. These typically come from a topic model (e.g., LDA via gensim or sklearn), but any source of word importance scores works -- TF-IDF weights, embedding similarities, domain relevance scores, or hand-picked values. Words absent from the dictionary receive a configurable minimum weight (min_weight, default 0.0).

Usage

Native Python class

from rapid_textrank import TopicalPageRank

# Topic weights from an external topic model or manual assignment
topic_weights = {
    "neural": 0.9,
    "network": 0.8,
    "learning": 0.7,
    "deep": 0.6,
}

extractor = TopicalPageRank(
    topic_weights=topic_weights,
    min_weight=0.01,  # Floor for out-of-vocabulary words
    top_n=10
)

result = extractor.extract_keywords("""
Deep learning is a subset of machine learning that uses artificial neural
networks. Neural networks with many layers can learn complex patterns.
Convolutional neural networks excel at image recognition tasks.
""")

for phrase in result.phrases:
    print(f"{phrase.text}: {phrase.score:.4f}")

# Update topic weights for a different document/topic
result = extractor.extract_keywords(
    "Machine learning enables data-driven decisions...",
    topic_weights={"machine": 0.9, "data": 0.8}
)

JSON interface

TopicalPageRank is also available via the JSON interface with variant="topical_pagerank" (aliases: "tpr", "single_tpr"). Set topic_weights and optionally topic_min_weight in the JSON config:

import json
from rapid_textrank import extract_from_json

payload = {
    "tokens": tokens,  # Pre-tokenized (e.g., from spaCy)
    "variant": "topical_pagerank",
    "config": {
        "top_n": 10,
        "topic_weights": {"neural": 0.9, "network": 0.8, "learning": 0.7},
        "topic_min_weight": 0.01,
    },
}

result = json.loads(extract_from_json(json.dumps(payload)))

TopicalPageRank vs BiasedTextRank

Both variants bias extraction toward specific terms, but they differ in how the bias is specified:

  • BiasedTextRank takes a list of focus terms and a single bias_weight. It is manual and direct -- good when you know exactly which terms matter.
  • TopicalPageRank takes per-word weights, typically from a topic model. It is data-driven -- good when you want the topic distribution to guide extraction automatically.

Use BiasedTextRank when you have a short, hand-curated focus vocabulary. Use TopicalPageRank when you have a numeric importance score for each word (e.g., from LDA, TF-IDF, or embedding similarity).

Topic Modeling is Optional

Despite the name, TopicalPageRank does not require a topic model. You can supply any word-importance dictionary:

  • TF-IDF weights from a corpus-level vocabulary.
  • Embedding similarities between candidate words and a target concept.
  • Domain relevance scores from a domain-specific lexicon.
  • Hand-picked values when you know the relative importance of key terms.

For details on computing topic weights from a gensim LDA model using the built-in topic_weights_from_lda helper, see Topic Utilities.

Reference