How Sememic Components Can Benefit Link Prediction for Lexico-Semantic Knowledge Graphs?

Hansi Wang, Yue Wang, Qiliang Liang, Yang Liu2025 · DOI

Abstract

Link Prediction (LP) aims to predict missing triple information within a Knowledge Graph (KG). Existing LP methods have sought to improve performance by integrating structural and textual information. However, for lexico-semantic KGs designed to document fine-grained sense distinctions, these types of information may not be sufficient to support effective LP. From a linguistic perspective, word senses within lexico-semantic relations usually show systematic differences in their sememic components. In light of this, we are motivated to enhance LP with sememe knowledge. We first construct a Sememe Prediction (SP) dataset, SememeDef, for learning such knowledge, and two Chinese datasets, HN7 and CWN5, for LP evaluation. Then, we propose a method, SememeLP, to leverage this knowledge for LP fully. It consistently and significantly improves the LP performance in both English and Chinese, achieving SOTA MRR of 75.1%, 80.5%, and 77.1% on WN18RR, HN7, and CWN5, respectively. Finally, an in-depth analysis is conducted, making clear how sememic components can benefit LP for lexico-semantic KGs, providing promising progress for their completion.

Introduction

Knowledge Graphs (KGs) like WordNet are essential for Natural Language Processing but are often incomplete. The task of Link Prediction (LP) aims to fill these gaps by predicting missing triples, such as (boy, antonym, ?). Current state-of-the-art methods typically rely on graph structure and textual definitions. However, these methods treat word senses as atomic "black boxes," ignoring the internal linguistic logic that defines their relationships.

The Core Idea: Sememes as Semantic DNA

From a linguistic standpoint, word senses are composed of sememes—the smallest indivisible units of meaning. For example, the sense of "boy" can be decomposed into sememes like {human, male, immature}. Understanding these components allows a model to treat relations as logical operations. Predicting that "girl" {human, female, immature} is an antonym of "boy" becomes a matter of identifying the gender component flip rather than a purely statistical inference.

Examples of word senses decomposed into sememes and main sememes.

Methods: The SememeLP Framework

The authors introduce SememeLP, a framework designed to acquire and integrate sememic knowledge into the link prediction process.

Sememe Knowledge Acquisition

Since many KGs lack manual sememe annotations, the authors fine-tuned a BERT-based encoder on a newly created dataset, SememeDef. This dataset contains tens of thousands of (definition, sememes) pairs in both English and Chinese, allowing the model to predict both the "main sememe" (core category) and the full set of attributes for any given word sense.

Three-Stage Fusion Module

To integrate these predicted sememes, SememeLP employs a novel three-stage fusion module:

  1. Independent Gated Fusion (IGF): Uses a gating mechanism to filter out noise from predicted sememes, ensuring only relevant features are merged.
  2. Weighting Fusion (WF): Dynamically balances the importance of the "main sememe" versus the full sememe set.
  3. Final Gated Fusion (FGF): Combines the refined sememic information with the original BERT-based definition embedding to create a robust final representation.
The SememeLP architecture featuring the three-stage fusion module (IGF, WF, and FGF).

Experimental Results

SememeLP was evaluated on the standard English benchmark WN18RR and two new Chinese benchmarks, HN7 and CWN5. The model achieved new state-of-the-art (SOTA) performance across all datasets:

  • WN18RR: 75.1% MRR (Mean Reciprocal Rank).
  • HN7: 80.5% MRR, a significant +5.2 point improvement over the baseline.
  • CWN5: 77.1% MRR.

The results indicate that sememic information is particularly effective at helping models distinguish between highly similar candidates (improving Hits@1 metrics) by providing a logical basis for differentiation.

Conclusion

By decomposing word senses into their sememic components, SememeLP moves beyond treating words as atomic units. This approach not only improves link prediction accuracy but also aligns model behavior with linguistic theory. Future work may extend this decomposition strategy to other specialized domains, such as biomedical knowledge graphs, where complex terms can be broken down into simpler semantic building blocks.

Summary updated

Abstract

Link Prediction (LP) aims to predict missing triple information within a Knowledge Graph (KG).Existing LP methods have sought to improve the performance by integrating structural and textual information.However, for lexico-semantic KGs designed to document fine-grained sense distinctions, these types of information may not be sufficient to support effective LP.From a linguistic perspective, word senses within lexico-semantic relations usually show systematic differences in their sememic components.In light of this, we are motivated to enhance LP with sememe knowledge.We first construct a Sememe Prediction (SP) dataset, SememeDef, for learning such knowledge, and two Chinese datasets, HN7 and CWN5, for LP evaluation; Then, we propose a method, SememeLP, to leverage this knowledge for LP fully.It consistently and significantly improves the LP performance in both English and Chinese, achieving SOTA MRR of 75.1%, 80.5%, and 77.1% on WN18RR, HN7, and CWN5, respectively; Finally, an in-depth analysis is conducted, making clear how sememic components can benefit LP for lexico-semantic KGs, which provides promising progress for the completion of them 1 .
BibTeX
@misc{wang2025how,
  author = {Hansi Wang and Yue Wang and Qiliang Liang and Yang Liu},
  title = {{How Sememic Components Can Benefit Link Prediction for Lexico-Semantic Knowledge Graphs?}},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.740},
}
Figure