ELEXAI

Background

Large pre-trained language models have transformed natural language processing, but they continue to suffer from a number of well-recognised limitations: hallucination, poor reasoning and weak interpretability. For Europe – home to a rich and diverse linguistic landscape – these weaknesses pose particular risks. Without access to accurate semantic information, LLMs often struggle with under-resourced languages and domain-specific terminology, amplifying digital inequalities and reducing people’s trust in AI outputs.

Lexicographic resources offer a way forward. Carefully curated resources contain the explicit, structured semantic knowledge that LLMs lack. Injecting this knowledge into LLMs has the potential to dramatically improve their reasoning, factual reliability, and cross-lingual performance. However, the interaction between LLMs, lexicographic databases and heterogeneous user communities (i.e. researchers, industry, policy makers, general public) remains fragmented: multiple LLM architectures, dozens of language formats, and diverse user needs prevent seamless integration.

Recent advances in agentic AI frameworks provide the missing connective tissue. Agentic systems have already demonstrated their ability to manage unpredictable, complex environments. By applying them to the European lexicographic ecosystem, ELEXAI can enable smooth interactions between LLMs, lexical resources, and human users.

Aims

Existing infrastructures must be technologically and organisationally upgraded if they are to remain relevant in the era of LLMs and agentic AI. The ELEXAI infrastructure will deliver integration of lexicographic data and LLMs through:

1.     Lexicography-enhanced LLMs, specialised for European languages and lexicographic tasks.
2.     LLM-enhanced lexical resources, where machine learning and lexicographic expertise mutually strengthen each other.
3.     Accessible interfaces tailored to both expert lexicographers and general users.
4.     Lexicography-based evaluation benchmarks, ensuring that progress is measurable and trustworthy.

Method
Addressing the LLM blackbox requires three key advances:

1.     Deeper semantic description: an improved understanding of higher levels of linguistic meaning, as provided by lexicography, especially in contexts where humans must interpret machine-generated text (and vice versa).
2.     Machine-readable knowledge resources: the development of knowledge bases and knowledge graphs that can be systematically injected into LLMs.
3.     Robust evaluation frameworks: the creation of benchmarks and testing protocols that can assess machine-generated output in realistic NLU scenarios, ensuring that results are valid and scientifically grounded.

Therefore, ELEXAI will:

1. Expand existing ELEXIS services through AI-enhanced language models. 
2. Provide lexicographically informed knowledge representations. 
3. Deliver European benchmarks and evaluation datasets.
4. Build an agentic AI-powered ELEXAI platform. 
5. Ensure access, standards, and interoperability. 
6. Promote Open Science and responsible innovation in lexicography. 
7. Secure the long-term sustainability and growth of ELEXAI. 

Eurac Research
Eurac Research's role within ELEXAI is to share existing terminological and corpus data focusing on the specific South Tyrolean language variety and on domain-related data, test LLM performance on multilingual terminological data (e,g. for definitionwriting) and support the development of sound benchmarks.