Signal and Data Processing Journal A scientific journal officially licensed by the Commission for Scientific Publications of the (MSRT). Publisher: Research Ceter for Developmen of Technologies
Volume 23, Issue 1 (6-2026)
JSDP 2026, 23(1): 37-50 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Torabian N, Radaei H, Minaei-Bidgoli B, Jahanshahi M. Enhancing Knowledge graph with Selectional Preferences. JSDP 2026; 23 (1) : 3
URL: http://jsdp.rcisp.ac.ir/article-1-1479-en.html
URL: http://jsdp.rcisp.ac.ir/article-1-1479-en.html
Abstract: (8 Views)
In constructing large-scale knowledge graphs from heterogeneous data sources, many critical semantic relationships remain missing, undermining the efficacy of downstream applications such as question answering, recommendation systems, and semantic search. Link prediction has emerged as a key solution to this problem by identifying and adding plausible but absent triples. Although existing embedding-based approaches (trained on node and relation labels with models such as Word2Vec or GloVe) achieve high accuracy in resource-rich settings, their performance on low-resource languages or specialized domains often falls short of industrial requirements. This study introduces a novel adaptation of selectional preferences (a concept from lexical semantics that captures the typical co-occurrence patterns between predicates and their arguments) to the link prediction task in knowledge graphs, with a focus on the Persian benchmark dataset FarsPishBin.
Selectional preferences technique traditionally measures how compatible a verb is with a nominal argument by learning from large corpora in which word pairs co-occur more frequently than a random co-occurrence. For example, “bite” prefers animate or small‐animal subjects, whereas “read” expects objects such as “book” or “article.” Our core hypothesis is that applying selectional preferences to triples (head entity, relation, tail entity) captures higher-order semantic dependencies and thus improves the plausibility scoring of candidate links. To test this, we propose a two-phase framework: “data preparation” and “preference learning.”
Phase 1 comprises rigorous preprocessing and embedding initialization. We begin by filtering FarsPishBin to its 100 most frequent relations, retaining over 80 percent of the original (almost 600,000) triples while reducing the memory footprint. Textual labels of entities and relations undergo normalization, stop-word removal, Unicode unification, punctuation stripping, and tokenization. We then train both Word2Vec and GloVe on these labels to produce base embeddings for every node and edge. To accommodate relation-specific behavior, we generate a separate embedding for each entity–relation pair (relation-specific embedding) alongside a central embedding that encodes an entity’s general semantics across all relations.
Phase 2 implements selectional preferences technique, learning over multiplex embedding structure. For each positive triple, we generate two negative samples by randomly replacing either the head or tail entity, following best practices in knowledge graph completion. We optimize a cross-entropy objective with stochastic gradient descent, encouraging true triples to score higher than their corrupted counterparts. Our scoring function sums the head entity’s central embedding and its relation-specific embedding, and then, computes cosine similarity with the tail entity’s central embedding. This design integrates both general and relation-contextual semantics in evaluating triple plausibility.
We evaluate the proposed approach using the pseudo-disambiguation protocol in bi-directional and tri-directional modes. In the bi-directional setting, we assess whether the model can distinguish the correct head entity when presented with one negative head. In the tri-directional evaluation, the model must outperform negatives that replace either the head or tail, making the task more challenging. Metrics are reported as mean accuracy over all test triples.
Experimental results demonstrate that incorporating selectional preferences technique yields consistent, substantial gains across both embedding initializations. On Word2Vec, bi-directional accuracy rises from 71.2 percent to 78.9 percent, and tri-directional accuracy from 63.5 percent to 71.6 percent (absolute improvements of 7.7 and 8.1 points, respectively). GloVe shows similar trends: bi-directional accuracy increases from 73.0 percent to 80.4 percent (7.4% gain), and tri-directional accuracy from 65.2 percent to 72.8 percent (7.6% gain). These robust improvements confirm that selectional preferences technique captures semantic compatibility patterns, overlooked by traditional text embeddings and complements them rather than replacing them. The observed benefits stem from relation-specific embeddings that resolve polysemy by differentiating an entity’s representation per relation.
Our approach’s primary limitation is its memory complexity, which scales with the product of the number of entities and relations. Although subsampling the top 100 relations preserves most data coverage, future work must explore efficient compression strategies for relation-specific embeddings. Additionally, while we focused on Word2Vec and GloVe for proof of concept, integrating selectional preferences with graph-native embedding methods such as TransE, RotatE, or ComplEx may yield further gains. Extending evaluations to multilingual datasets and domain-specific graphs will help establish the generality of our findings.
In conclusion, this study pioneers the application of selectional preferences to knowledge graph link prediction, demonstrating clear accuracy enhancements on a low-resource Persian dataset. For organizations reliant on manual curation of knowledge graphs, our method offers a scalable avenue to identify high-quality candidate triples, thereby accelerating ontology enrichment and boosting downstream NLP performance. Future research should aim to reduce memory overhead, validate compatibility with advanced graph embedding models, and broaden empirical validation across diverse knowledge graph domains.
Selectional preferences technique traditionally measures how compatible a verb is with a nominal argument by learning from large corpora in which word pairs co-occur more frequently than a random co-occurrence. For example, “bite” prefers animate or small‐animal subjects, whereas “read” expects objects such as “book” or “article.” Our core hypothesis is that applying selectional preferences to triples (head entity, relation, tail entity) captures higher-order semantic dependencies and thus improves the plausibility scoring of candidate links. To test this, we propose a two-phase framework: “data preparation” and “preference learning.”
Phase 1 comprises rigorous preprocessing and embedding initialization. We begin by filtering FarsPishBin to its 100 most frequent relations, retaining over 80 percent of the original (almost 600,000) triples while reducing the memory footprint. Textual labels of entities and relations undergo normalization, stop-word removal, Unicode unification, punctuation stripping, and tokenization. We then train both Word2Vec and GloVe on these labels to produce base embeddings for every node and edge. To accommodate relation-specific behavior, we generate a separate embedding for each entity–relation pair (relation-specific embedding) alongside a central embedding that encodes an entity’s general semantics across all relations.
Phase 2 implements selectional preferences technique, learning over multiplex embedding structure. For each positive triple, we generate two negative samples by randomly replacing either the head or tail entity, following best practices in knowledge graph completion. We optimize a cross-entropy objective with stochastic gradient descent, encouraging true triples to score higher than their corrupted counterparts. Our scoring function sums the head entity’s central embedding and its relation-specific embedding, and then, computes cosine similarity with the tail entity’s central embedding. This design integrates both general and relation-contextual semantics in evaluating triple plausibility.
We evaluate the proposed approach using the pseudo-disambiguation protocol in bi-directional and tri-directional modes. In the bi-directional setting, we assess whether the model can distinguish the correct head entity when presented with one negative head. In the tri-directional evaluation, the model must outperform negatives that replace either the head or tail, making the task more challenging. Metrics are reported as mean accuracy over all test triples.
Experimental results demonstrate that incorporating selectional preferences technique yields consistent, substantial gains across both embedding initializations. On Word2Vec, bi-directional accuracy rises from 71.2 percent to 78.9 percent, and tri-directional accuracy from 63.5 percent to 71.6 percent (absolute improvements of 7.7 and 8.1 points, respectively). GloVe shows similar trends: bi-directional accuracy increases from 73.0 percent to 80.4 percent (7.4% gain), and tri-directional accuracy from 65.2 percent to 72.8 percent (7.6% gain). These robust improvements confirm that selectional preferences technique captures semantic compatibility patterns, overlooked by traditional text embeddings and complements them rather than replacing them. The observed benefits stem from relation-specific embeddings that resolve polysemy by differentiating an entity’s representation per relation.
Our approach’s primary limitation is its memory complexity, which scales with the product of the number of entities and relations. Although subsampling the top 100 relations preserves most data coverage, future work must explore efficient compression strategies for relation-specific embeddings. Additionally, while we focused on Word2Vec and GloVe for proof of concept, integrating selectional preferences with graph-native embedding methods such as TransE, RotatE, or ComplEx may yield further gains. Extending evaluations to multilingual datasets and domain-specific graphs will help establish the generality of our findings.
In conclusion, this study pioneers the application of selectional preferences to knowledge graph link prediction, demonstrating clear accuracy enhancements on a low-resource Persian dataset. For organizations reliant on manual curation of knowledge graphs, our method offers a scalable avenue to identify high-quality candidate triples, thereby accelerating ontology enrichment and boosting downstream NLP performance. Future research should aim to reduce memory overhead, validate compatibility with advanced graph embedding models, and broaden empirical validation across diverse knowledge graph domains.
Article number: 3
Keywords: Link Prediction, Selectional Preferences, Word Embedding, Knowledge Graph Completion, Knowledge Graph Embedding
Type of Study: Research |
Subject:
Paper
Received: 2025/07/20 | Accepted: 2026/08/24 | Published: 2026/06/21 | ePublished: 2026/06/21
Received: 2025/07/20 | Accepted: 2026/08/24 | Published: 2026/06/21 | ePublished: 2026/06/21
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |