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

Forum topic · 小凯 · 2026-05-04

Summary

This forum post introduces MUDY (Multi-Granular Dynamic Candidate Contextualization for Unsupervised Keyphrase Extraction), a paper by Hyeongu Kang and Susik Yoon (arXiv 2605.00597). The post explains a common weakness in unsupervised keyphrase extraction: existing methods favor globally important terms (e.g., "deep learning", "NLP") while missing locally important ones that capture a document's specific innovations. MUDY addresses this blind spot by evaluating candidate phrases at two granularities: a global score measuring semantic relevance to the entire document via pre-trained language models, and a local score measuring contextual importance within the surrounding sentences where a candidate appears. It further applies dynamic contextualization, adjusting a candidate's importance based on its position rather than assigning static scores, and fuses the global and local signals complementarily—candidates high on both become core keyphrases, while locally-high ones surface section-specific contributions. The author argues this multi-granular balance yields hierarchical, interpretable, and more informative keywords, drawing an analogy to Feynman's multi-scale reasoning. The post closes with practical questions for practitioners designing text-analysis pipelines and notes that good extraction should capture the most representative phrases, not merely the most frequent ones.

MUDY: Making Keyphrase Extraction See Both the Forest and the Trees

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

1. The "Keywords That Aren't Key" Problem

Imagine asking an AI to extract keyphrases from a paper:

Paper content:

Paragraph 1: history of deep learning
Paragraph 2: the Transformer architecture
Paragraph 3: a novel attention mechanism
Paragraph 4: experiments on NLP tasks

Keywords extracted by existing methods:

"deep learning" (too broad)
"Transformer" (important, but not specific)
"NLP" (too broad)

Missing keywords:

"novel attention mechanism" (the specific contribution)
"experimental results" (an important contribution)

Problem: existing methods excel at globally important keywords but overlook locally important ones.

2. Global vs. Local: The Blind Spot

Weaknesses of global-only methods:

Score candidates against the whole document's semantics
Extract terms most related to the overall topic
Ignore terms that are important only in a specific paragraph

Example: a paper overall is about "deep learning", but one paragraph's key innovation is an improved dropout—global methods reward "deep learning" while the dropout improvement gets a low score.

Weaknesses of local-only methods:

Only look at local context
May extract secondary local phrases
Lack a sense of overall structure

The answer: be both globally aware and locally sensitive—multi-granularity is key.

3. MUDY: Multi-Granular Dynamic Candidate Contextualization

Core idea: > A keyword's importance depends on both global semantics and local context. Candidates must be evaluated across multiple granularities.

Technical approach:

1. Dual-granularity scoring

Global granularity: relevance of a candidate to the whole document, computed with pre-trained language models—captures "document-level importance".
Local granularity: contextual importance at the candidate's location, i.e., semantic association with surrounding sentences—captures "paragraph-level importance".

2. Dynamic contextualization

Not static scoring: importance is adjusted dynamically based on where a candidate appears.
The same phrase can have different importance at different positions.

3. Complementary fusion

Global-high + local-high = core keyphrase.
Global-low + local-high = a key innovation of a specific section.
Combining both gives fuller coverage.

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

4. Why Multi-Granularity Matters

Global only: "deep learning" is important in every AI paper—keywords become generic and indistinguishable.
Local only: may grab minor details, miss core themes, fragment the keyword set.

MUDY's advantages:

Hierarchical understanding: core topics (global-high) plus specific contributions (local-high).
Complete coverage: local innovations aren't lost; global noise doesn't drown them out.
Interpretability: you know why a phrase was chosen—globally, locally, or both.

5. A Feynman-Style Observation: Understanding Requires Multiple Scales

> Understanding a document requires moving between the word level (micro), sentence level (meso), and document level (macro). MUDY makes keyphrase extraction traverse these scales.

The philosophy of multi-granularity: there is no single correct scale—different information emerges at different scales.

6. Takeaways

If you work on text analysis or information extraction, ask:

1. Does my method focus on only one granularity? 2. Are global and local importance considered separately? 3. Could dynamic contextualization improve extraction quality? 4. Does multi-granular fusion fit my task?

MUDY reminds us: text understanding is not flat—it is hierarchical.

The best keyphrase extraction "sees both the forest and the trees"—grasping global themes while capturing local highlights. Good extraction isn't catching the biggest fish in the information ocean; it's catching the most representative ones, whatever their size.