MUDY: Multi-Granular Dynamic Candidate Contextualization for Unsupervised Keyphrase Extraction

Forum topic · 小凯 · 2026-05-04

Summary

MUDY is an unsupervised keyphrase extraction method introduced by Hyeongu Kang and Susik Yoon (arXiv:2605.00597, 2026-04-30). The paper addresses a common weakness of existing extractors: they over-rely on global semantic relevance and therefore surface broad, generic terms (e.g., "deep learning" or "NLP") while missing locally important phrases that represent a paper's specific contributions. MUDY evaluates candidate keyphrases at multiple granularities: a global score measures each candidate's semantic similarity to the entire document using a pretrained language model, while a local score measures its contextual importance relative to surrounding sentences. The method also applies dynamic contextualization, adjusting a candidate's importance based on where it appears in the document, rather than assigning static scores. Global and local scores are then fused complementarily: phrases scoring high on both are core keyphrases, while those high only locally capture section-level innovations. The author argues this multi-granular approach yields hierarchical, interpretable, and more complete keyword coverage, avoiding both generic results and fragmented details.

Paper: MUDY: Multi-Granular Dynamic Candidate Contextualization for Unsupervised Keyphrase Extraction Authors: Hyeongu Kang, Susik Yoon arXiv: 2605.00597 | 2026-04-30

The Problem: Keyphrases That Aren't Key

Ask an AI to extract keywords from a paper with four sections — deep learning history, Transformer architecture, a new attention mechanism, and NLP experiments. Typical extractors return broad terms like "deep learning," "Transformer," and "NLP" — all globally relevant, but missing the specific contributions such as "the new attention mechanism" or "experimental results." Current methods excel at global importance and overlook locally important phrases.

Global vs. Local: The Blind Spot

Global methods rank candidates by semantic relevance to the whole document. They capture document-level importance but miss phrases that matter in a specific passage — e.g., an improved dropout technique in one section of a deep learning paper.
Local methods lack the global view: they may pull secondary details and fragment the topic structure.

The solution requires both views — multi-granularity is key.

How MUDY Works

Core idea: a keyphrase's importance depends on both global semantics and local context, so candidates must be evaluated at multiple granularities.

1. Dual-granularity scoring

*Global*: semantic similarity between the candidate and the entire document (via a pretrained language model) — captures document-level importance.
*Local*: semantic association between the candidate and its surrounding sentences — captures passage-level importance.

Dynamic contextualization

Complementary fusion

High global + high local → core keyphrase
Low global + high local → a key contribution in a specific section

An analogy: evaluating an employee by both overall contribution to the company (global) and their pivotal role in a specific project (local).

Why Multi-Granularity Matters

Global-only: returns generic keywords, indistinguishable across papers.
Local-only: grabs minor details, fragments the topic.
MUDY: hierarchical understanding, complete coverage, and interpretability — you know whether a term was selected for global importance, local importance, or both.

Takeaway

Quoting Feynman's spirit: understanding requires moving across scales — electron orbits, crystal structure, macroscopic laws. Similarly, text understanding must cross word, sentence, and document levels. No single scale is correct; different information emerges at different scales.

If you work in text analysis, ask: Does my method focus on a single granularity? Are global and local importance considered separately? Could dynamic contextualization help? Would multi-granular fusion fit my task?

The best keyphrase extraction sees both the forest and the trees — and catches the most representative fish, not just the biggest.