language-specific-vs-cross-lingual-knowledge-graphs

name	language-specific-vs-cross-lingual-knowledge-graphs
version	1.0.0
description	Comparative methodology for language-specific versus cross-lingual knowledge graphs in implicit aspect identification for lower-resource languages, with task-specific fine-tuning strategies.
author	Lujain A. Alawwad
license	MIT
arxiv_id	2607.20056v1
tags	["knowledge-graphs","multilingual","aspect-based-sentiment","arabic-nlp","fine-tuning"]

Language-Specific vs Cross-Lingual Knowledge Graphs for Implicit Aspect Identification

Overview

This methodology provides a controlled comparison framework for choosing between language-specific and cross-lingual knowledge graphs (KGs) when performing implicit aspect identification in lower-resource languages like Arabic.

Key Strategies

Strategy Comparison

1. Cross-Lingual English KG: Reuse mature English KG through multilingual embeddings
2. Native Language KG: Build smaller native language KG specific to target language

Adaptation Approaches

• Zero-shot Prompting: Use pre-trained LLM without task-specific adaptation
• Task-Specific Fine-tuning: Fine-tune LLM on target task with domain-specific data

Implementation Guidelines

Hybrid Pipeline Architecture

1. Implement generative extractor component for aspect identification
2. Integrate chosen KG strategy (language-specific or cross-lingual)
3. Apply selected adaptation approach (zero-shot or fine-tuned)
4. Evaluate performance on multiple benchmarks for comprehensive assessment

Performance Expectations

• Native language KGs consistently outperform cross-lingual KGs (+0.199 to +0.251 micro-F1)
• Task-specific fine-tuning dramatically improves performance (0.13 → 0.66-0.76 micro-F1)
• Task adaptation proves more decisive than model scale in morphologically rich languages

Use Cases

• Aspect-based sentiment analysis in lower-resource languages
• Implicit aspect identification where aspects are never explicitly mentioned
• Multilingual NLP applications requiring knowledge integration
• Resource-constrained scenarios where building native KGs is feasible

Evaluation Benchmarks

• M-ABSA (Arabic)
• SemEval-2016 Arabic
• HAAD (Arabic)

Activation Keywords

knowledge graphs, multilingual NLP, aspect-based sentiment, implicit aspects, Arabic NLP, task-specific fine-tuning

References

• arXiv: 2607.20056v1
• Author: Lujain A. Alawwad
• Published: July 22, 2026
• Benchmarks: M-ABSA, SemEval-2016 Arabic, HAAD