Unsupervised Keyphrase Extraction with Multipartite Graphs

Motivation

• Graph-based approaches build document graphs and rank their nodes, offering strong performance while remaining completely unsupervised.Typical graphs use lexical units as nodes and semantic relations as edges.Source: §1 · p. 1
• Conventional graph representations do not naturally ensure topical coverage and diversity, a central challenge in keyphrase extraction.Prior work largely focused on semantic relatedness or node-ranking functions, with limited attention to graph types.Source: §1 · p. 1
• Topic-biased ranking models can be computationally prohibitive for long documents, require large datasets for word-topic distributions, and assume independent topics.These limitations make topic diversity difficult to ensure without additional supervised approaches.Source: §1 · p. 1

Prior approaches

• TopicRank improves ranking and topic coverage but treats candidates within one topic as equally important, requiring post-ranking heuristics for representative selection.Its topic-level representation can also propagate errors from topic formation.Source: §1 · p. 1

Proposed approach

• The proposed multipartite graph encodes candidates in topic-based independent sets and ranks them in one operation, using mutually reinforcing topic-candidate relations.The representation is illustrated against TopicRank for a Hulth-2003 document; the supplied figure caption identifies nodes and co-occurrence edges.Source: §1 · p. 2, Figure 1 · p. 2
• A selection-preference mechanism adjusts edge weights to promote candidates with properties such as thesaurus matches or early document occurrence.The introduction specifically demonstrates a bias toward candidates occurring first in the document.Source: §1 · p. 2

Candidate and topic construction

• Candidates are extracted from adjacent adjective–noun sequences and grouped into topics by hierarchical agglomerative clustering over shared word stems.The setup follows Bougouin et al. (2013) for direct comparability.Source: §2 · p. 2

Graph construction

• Candidate edge weights sum inverse distances between their document occurrences, matching window-based co-occurrence performance without parameter tuning.P(ci) contains the word-offset positions of candidate ci.Source: §2.1 · pp. 2–3

Multipartite graph representation

• The complete k-partite graph connects candidates only across different topics, while each topic forms an independent set of unconnected nodes.Here, k denotes the number of topics.Source: §2.1 · pp. 2–3
• Removing intra-topic edges distributes recommendation across topics and implicitly promotes topical diversity through mutually reinforcing topic–candidate relations.The representation does not assume a particular topic-decomposition method.Source: §2.1 · p. 3

Graph weight adjustment mechanism

• The model adjusts incoming edges of each topic’s first-occurring candidate, promoting candidates according to other candidates in the same topic.The selection heuristic can instead prioritize properties such as thesaurus membership.Source: §2.2 · p. 3, Figure 2 · p. 3

Ranking and extraction

• Weighted TextRank orders candidates by relevance, after which the top N candidates are selected as keyphrases.The damping factor λ is set to 0.85, although other ranking algorithms can be applied.Source: §2.3 · p. 3

3.1 Datasets and evaluation measures

• Performance is measured by F1 at the top 5 and 10 keyphrases and by Mean Average Precision (MAP), with stemming applied before scoring.Significance markers in Table 1 indicate the 0.05 level using Student’s t-test.Source: §3.1 · p. 4, Table 1 · p. 4

3.2 Baselines and parameter settings

• The model is compared with TopicRank, Single Topical PageRank, and PositionRank, covering topic-based and feature-enhanced graph-ranking approaches.LDA topic distributions are estimated on each dataset’s training set, and the proposed model tunes graph-adjustment strength using α.Source: §3.2 · p. 4

3.3 Results

• Overall, the proposed model achieves the best results and significantly outperforms the baselines on most metrics, with smaller relative gains on the shorter Hulth-2003 and Marujo-2012 documents.TopicRank has the highest baseline precision, while PositionRank is the strongest baseline by MAP.Source: §3.3 · p. 4
• Without weight adjustment, results remain higher or on par with topic-informed baselines, supporting the model’s use of reinforcing topic–candidate relations.The performance drop is especially severe for F1@5 on SemEval-2010, where the model moves from best to worst performance.Source: §3.3 · pp. 4–5
• Over 92% of the top-10 keyphrases are assigned to different topics, indicating topical diversity without hard constraints.Some apparent redundancy reflects clustering errors or hypernym–hyponym pairs that both occur in the gold references.Source: §3.3 · p. 5

• The unsupervised multipartite-graph model is demonstrated on three public datasets.Source: §4 · p. 5
• The authors make their code and data publicly available through the project repository.Source: §4 · p. 5
• Future work will apply ranking algorithms that leverage the specific structure of the multipartite graph.Source: §4 · p. 5