Skip to content

MultipartiteRank

MultipartiteRank (Boudin, 2018) extends TopicRank by keeping individual candidates as graph nodes instead of collapsing topics into single representatives. It removes intra-topic edges to form a k-partite graph and applies an alpha weight adjustment that boosts the first-occurring variant in each topic cluster, encoding positional preference directly into edge weights.

How It Works

  1. Candidate extraction -- Candidate phrases are identified using POS-filtered noun chunks (same as other variants).
  2. Topic clustering -- Candidates are grouped into topics based on Jaccard similarity over word sets, controlled by similarity_threshold.
  3. K-partite graph -- A co-occurrence graph is built with all candidates as nodes, but edges between candidates in the same topic cluster are removed. This prevents intra-topic competition.
  4. Position boost -- Edge weights are adjusted by an alpha factor that favors the first-occurring candidate in each topic cluster. Set alpha=0 to disable this boost.
  5. PageRank -- Standard PageRank is run on the k-partite graph.
  6. Phrase selection -- Top-scoring candidates are returned directly (no topic-level aggregation step).

Usage

Native Python class

from rapid_textrank import MultipartiteRank

extractor = MultipartiteRank(
    similarity_threshold=0.26,  # Jaccard threshold for topic clustering
    alpha=1.1,                  # Position boost strength (0 = disabled)
    top_n=10
)

result = extractor.extract_keywords("""
Machine learning is a powerful tool for data analysis. Deep learning
is a subset of machine learning. Neural networks power deep learning
systems. Convolutional neural networks excel at image recognition.
""")

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

JSON interface

MultipartiteRank is also available via the JSON interface with variant="multipartite_rank" (aliases: "multipartiterank", "multipartite", "mpr"). Set multipartite_alpha and multipartite_similarity_threshold in the JSON config:

import json
from rapid_textrank import extract_from_json

payload = {
    "tokens": tokens,  # Pre-tokenized (e.g., from spaCy)
    "variant": "multipartite_rank",
    "config": {
        "top_n": 10,
        "multipartite_alpha": 1.1,
        "multipartite_similarity_threshold": 0.26,
    },
}

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

Parameters

Parameter Type Default Description
similarity_threshold float 0.26 Jaccard similarity threshold for grouping candidates into topic clusters. Higher values produce fewer, larger clusters.
alpha float 1.1 Position boost strength. Controls how much first-occurring candidates within a cluster are favored. Set to 0 to disable the position boost entirely.

MultipartiteRank vs TopicRank

Both variants cluster candidates into topics, but they differ in graph construction and candidate handling:

  • TopicRank collapses each topic into a single node and ranks topics as a whole, then picks the best representative from each. This gives strong diversity but loses fine-grained candidate distinctions.
  • MultipartiteRank keeps every candidate as its own node but removes edges within the same topic. This preserves fine-grained candidate distinctions while still preventing intra-topic competition.

In practice, MultipartiteRank tends to produce more nuanced results because it can rank individual phrases rather than entire topic clusters. The alpha position boost further helps surface the most salient phrasing of each topic.

When to Use MultipartiteRank

MultipartiteRank is a good choice for:

  • Long, multi-topic documents where you want diverse keywords but also care about the specific phrasing of each keyword.
  • Documents with near-duplicate phrases (e.g., "machine learning" and "learning machine") where you want the most natural variant to surface.
  • Scenarios where TopicRank's topic-level collapsing is too aggressive -- MultipartiteRank offers a middle ground between per-candidate ranking (BaseTextRank) and per-topic ranking (TopicRank).

Reference