MUDY: Multi-Granular Dynamic Contextualization for Unsupervised Keyphrase Extraction

Forum topic · 小凯 · 2026-05-04

Summary

This post discusses MUDY (Multi-Granular Dynamic Candidate Contextualization for Unsupervised Keyphrase Extraction), a paper by Hyeongu Kang and Susik Yoon (arXiv 2605.00597). The author argues that existing keyphrase extraction methods suffer from a single-granularity blind spot: global methods capture document-level importance but miss locally significant phrases (e.g., a specific methodological innovation in one section), while local methods lack overall structure and may pick up secondary details. MUDY addresses this by evaluating candidate keyphrases at two granularities — global relevance to the whole document via pretrained language model embeddings, and local contextual importance relative to surrounding sentences — then fusing the two scores complementarily. It also applies dynamic contextualization, adjusting scores based on where a candidate appears in the document, so the same phrase can carry different weights in different positions. The post explains the approach with examples (a paper on deep learning where "new attention mechanism" is locally key) and an analogy to evaluating an employee by both overall contribution and project-specific impact, framing multi-granular understanding as "seeing both the forest and the trees."

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

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

The "Keyphrases That Aren't Key" Problem

Ask an AI to extract keyphrases from a paper structured as: (1) history of deep learning, (2) Transformer architecture, (3) a new attention mechanism, (4) NLP experiments. Typical methods return broad terms like "deep learning," "Transformer," and "NLP" — while missing the specific innovations ("new attention mechanism") and contributions ("experimental results"). Existing approaches excel at globally important phrases but ignore locally important ones.

Global vs. Local: The Blind Spot

Global methods score candidates by semantic relevance to the entire document, so phrases like "deep learning" dominate even when a section's real contribution is, say, an improved dropout technique. Locally crucial phrases get low scores.
Local methods capture position-specific importance but lack a global view, potentially extracting secondary details and fragmenting the keyphrase set.
The solution requires multi-granular evaluation: both a global view and local sensitivity.

MUDY's Approach

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

1. Dual-granularity scoring

Global: semantic similarity between the candidate and the whole document, computed with pretrained language models (document-level importance).
Local: semantic association between the candidate and its surrounding sentences (paragraph-level importance).

Dynamic contextualization

Complementary fusion

It's like evaluating an employee by both overall contribution to the company and their pivotal role in a specific project — only together do they give a complete picture.

Why Multi-Granularity Matters

Global-only: keyphrases become generic ("deep learning" matters in every AI paper) and fail to distinguish actual contributions.
Local-only: secondary details get picked up while core themes are lost.
MUDY's multi-granularity gives hierarchical understanding (core themes plus specific contributions), complete coverage without drowning in global noise, and interpretability — you know whether a phrase was chosen for global importance, local importance, or both.

Takeaways

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

1. Does my method focus on only a single 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 that text understanding is hierarchical, not flat. The best keyphrase extraction "sees both the forest and the trees" — grasping global themes while capturing local highlights. Good extraction doesn't catch the biggest fish in the ocean of information; it catches the most representative ones, regardless of size.