Abstract
Entity and Relation Extraction (ERE) is critical for knowledge graph construction, yet multilingual models often struggle with "language interference," where high-resource languages overshadow low-resource ones. We introduce mERE, a two-stage framework designed to mitigate this conflict. By first establishing a unified semantic space through a Language-universal Aggregator (LA) and subsequently refining language-specific features via a Language-specific Switcher (LS), mERE achieves superior performance across 14 languages. Experimental results on the SMiLER dataset demonstrate a new state-of-the-art F1 score of 77.9%, with particularly significant improvements in low-resource languages.
Introduction
Entity and Relation Extraction (ERE) involves identifying entities (e.g., "Big Ben") and the semantic relationships between them (e.g., "Located_in"). While performance in English is robust, scaling to multilingual contexts introduces the challenge of language interference. Standard joint training on multiple languages often leads to a "one-size-fits-all" model where parameters shared across languages conflict, preventing the model from capturing the unique nuances of individual languages—especially those with limited training data.
Methodology: The mERE Framework
The proposed mERE (multilingual Entity and Relation Extraction) framework addresses interference through a strategic two-stage training process.
Stage 1: Language-universal Aggregator (LA)
The first stage focuses on cross-lingual alignment. By concatenating semantically equivalent sentences from different languages during training, the model is forced to map diverse linguistic inputs into a single, unified "semantic space." This process reduces the distance between language representations, ensuring that the model recognizes core concepts regardless of the input language.
Stage 2: Language-specific Switcher (LS)
To resolve the "tug-of-war" between language-specific features, the authors introduce a "Switcher" module inspired by Mixture-of-Experts (MoE). After learning universal features, the LA parameters are frozen. The LS then employs a set of independent sub-modules and a "router" mechanism that dynamically selects the most relevant sub-modules for a given language. This allows for fine-tuning language-specific details without degrading the previously learned universal knowledge.
The framework utilizes joint training, identifying entities and relations simultaneously to minimize error propagation typically found in pipeline systems.
Experimental Setup
The model was evaluated using SMiLER, a comprehensive multilingual ERE dataset comprising:
- 1.1 million annotated sentences sourced from Wikipedia and DBpedia.
- 14 languages, including high-resource (English, French) and low-resource (Ukrainian, Farsi, Russian) examples.
- 36 relation types spanning nine distinct domains.
Key Results and Discussion
mERE established a new state-of-the-art benchmark with an average F1 score of 77.9%, outperforming the previous best baseline (MTMB) by 2.3%.
Performance Consistency
The model achieved the highest scores in 8 of the 14 languages and ranked second in the remaining 6. Notably, the most substantial gains were observed in low-resource languages:
- Ukrainian: +25% F1 improvement
- Farsi: +11.1% F1 improvement
- Russian: +11.4% F1 improvement
Insights
- Parameter Efficiency: The "top-K" selection strategy in the Switcher allows linguistically similar languages (e.g., Romance languages) to share sub-modules, enhancing efficiency.
- Subject vs. Object Bias: Analysis revealed that "head" entities (subjects) are generally easier to extract than "tail" entities (objects), likely due to their more predictable positioning in sentences.
Conclusion
The mERE framework effectively bridges the gap between universal semantic understanding and language-specific precision. While it sets a new standard for multilingual ERE, challenges remain in handling complex scenarios such as overlapping relational triples. Future work will focus on refining contextual representations to better navigate these intricate linguistic structures.
Summary updated