An Alternative Axiomatic Algebraic LLM (3A-LLM)

The following abstracts are the union of the three LaTeX sources (wording unchanged aside from light HTML conversion). Author lists are omitted per site policy.

From main40.tex (A-LLM, FOIS-style)

In our times, language processing AI systems, e.g., ChatGPT, are based on Large Language Models (LLMs). Such AI systems excel in orthography, morphology, and syntax but sometimes err with respect to semantics and/or pragmatics. In order to reduce semantic fumbles, we suggest the integration of a semantic resource into language processing AI systems. However, such a resource needs not only to be large by itself but also strictly formal and computational so that an AI system can exploit it. This paper presents the Axiomatic Large Language Model (A-LLM), a formally grounded, generative framework for semantic representations that, among other applications, meets the requirements of a semantic resource that can be integrated in an LLM. The A-LLM consists of concepts as semantic building blocks and anchors those concepts in a restricted set of elementary concepts derived from the Longman Defining Vocabulary (LDV) [Fox2014]. The paper describes how to build the A-LLM from this restricted set and presents how representations for semantic relations among represented concepts are generated. The paper also takes a look at some of A-LLM's possible applications among them its use as a semantic resource in modern language processing AI systems.

From main50.tex (3A-LLM search evaluation)

Neural large language models (LLMs) excel at fluent text generation but often lack a grounded semantic backbone for search and retrieval, leading to brittle keyword behaviour or opaque embedding similarity. We investigate how 3A-LLM (An alternative axiomatic algebraic Large Language Model)---a formally grounded, generative semantic framework with a typed conceptual graph and explicit derivation paths---can support LLM-based search. 3A-LLM provides deterministic query expansion and implicit concept prediction via semantic radius and proximity on a graph built from Longman Defining Vocabulary primitives and a transformation calculus. We run a large-scale evaluation on a concept taxonomy of 4,102 subject--property--object triples (1,077 unique concept types) from a concept taxonomy (see supplementary material), with 165 conceptual query expansion tasks (T1) and 165 implicit concept prediction tasks (T2), totalling n=330 test cases across categories such as vehicle, building, person, event, animal, device, and artifact. 3A-LLM substantially outperforms WordNet-synset and string-similarity baselines (e.g. T1 P@5: 0.88 vs. 0.44 vs. 0.17; T2 MRR: 0.88 vs. 0.49 vs. 0.20). We report 2--3 worked examples per category and discuss integration of 3A-LLM into hybrid KG--LLM search pipelines.

From main97.tex (extended manuscript)

In our times, language processing AI systems, e.g., ChatGPT, are based on Large Language Models (LLMs). Such AI systems excel in orthography, morphology, and syntax but sometimes err with respect to semantics and/or pragmatics. In order to reduce semantic fumbles, we suggest the integration of a semantic resource into language processing AI systems. However, such a resource needs not only to be large by itself but also strictly formal and computational so that an AI system can exploit it. This paper presents the Axiomatic Large Language Model (A-LLM), a formally grounded, generative framework for semantic representations that, among other applications, meets the requirements of a semantic resource that can be integrated in an LLM. The A-LLM consists of concepts as semantic building blocks and anchors those concepts in a restricted set of elementary concepts derived from the Longman Defining Vocabulary (LDV) [Fox2014]. The paper describes how to build the A-LLM from this restricted set and presents how representations for semantic relations among represented concepts are generated. The paper also takes a look at some of A-LLM's possible applications among them its use as a semantic resource in modern language processing AI systems.