TextRank

Overview

TextRank, derived from PageRank, is a graph-based ranking model for text processing. It can be used for various natural language processing tasks, including keyword extraction, keyphrase extraction, and text summarization.

• R. Mihalcea, P. Tarau, TextRank: Bringing Order Into Texts (2004)

Concepts

Text as a Graph

To apply the TextRank algorithm, the text must first be represented as a graph. The structure of the graph depends on the specific application:

• Nodes: Text units that best fit the task, such as words, collocations, or sentences, are added as nodes in the graph.
• Edges: Relationships between text units, such as semantic similarity, co-occurrence, or contextual overlap, are used to connect nodes with edges.

Sample graph build for keyphrase extraction (Source: Original paper)

TextRank Model

TextRank computes the ranks of all text units recursively using a "recommendation" mechanism, similar to the PageRank algorithm. It incorporates edge weights through a modified formula:

where,

• Out(v) is the set of nodes that node v points to;
• wvu is the edge weight between nodes v and u;
• d is the damping factor.

Compared to PageRank:

• Undirected: All edges are treated as undirected.
• Weighted: Edge weights influence rank propagation.
• Unnormalized: The base rank is (1 - d) instead of (1 - d) / n, so scores do not sum to 1.

Considerations

• The rank of isolated text units will stay the same as the value of (1 - d).
• A self-loop acts as both a successor and a predecessor, meaning a node can pass some rank to itself. If a network has many self-loops, it will take more iterations to converge.

Example Graph

GQL
INSERT (A:default {_id: "A"}), (B:default {_id: "B"}),
       (C:default {_id: "C"}), (D:default {_id: "D"}),
       (E:default {_id: "E"}), (F:default {_id: "F"}),
       (G:default {_id: "G"}),
       (A)-[:default {weight: 3}]->(E), (B)-[:default {weight: 3}]->(A),
       (B)-[:default {weight: 2}]->(E), (C)-[:default {weight: 1}]->(A),
       (C)-[:default {weight: 4}]->(D), (D)-[:default {weight: 5}]->(E),
       (E)-[:default {weight: 2}]->(G), (F)-[:default {weight: 1}]->(B),
       (F)-[:default {weight: 3}]->(G)

Parameters

Run Mode

Returns:

TextRank for all nodes:

GQL
CALL algo.textrank({
  weight: "weight",
  damping: 0.8,
  maxIterations: 50,
  order: "desc"
}) YIELD nodeId, score, rank

Result:

Stream Mode

Returns the same columns as run mode, streamed for memory efficiency.

GQL
CALL algo.textrank.stream({
  order: "desc",
  limit: 3
}) YIELD nodeId, score
RETURN nodeId, score

Result:

Stats Mode

Returns:

GQL
CALL algo.textrank.stats() YIELD nodeCount, minScore, maxScore, avgScore

Result:

Write Mode

Computes results and writes them back to node properties. The write configuration is passed as a second argument map.

Write parameters:

Writable columns:

Returns:

GQL
CALL algo.textrank.write({damping: 0.85}, {
  db: {
    property: "tr_score"
  }
}) YIELD task_id, nodesWritten, computeTimeMs, writeTimeMs