Reinterpreting 'the Company a Word Keeps': Towards Explainable and Ontologically Grounded Language Models

The paper by Walid S. Saba critiques the limitations of large language models in NLP and proposes a shift towards a symbolic, ontologically grounded approach to create explainable, language-agnostic models capable of deep language understanding and commonsense reasoning.

Abstract

We argue that the relative success of large language models (LLMs) is not a reflection on the symbolic vs. subsymbolic debate but a reflection on employing a successful bottom-up strategy of a reverse engineering of language at scale. However, and due to their subsymbolic nature whatever knowledge these systems acquire about language will always be buried in millions of weights none of which is meaningful on its own, rendering such systems utterly unexplainable. Furthermore, and due to their stochastic nature, LLMs will often fail in making the correct inferences in various linguistic contexts that require reasoning in intensional, temporal, or modal contexts. To remedy these shortcomings we suggest employing the same successful bottom-up strategy employed in LLMs but in a symbolic setting, resulting in explainable, language-agnostic, and ontologically grounded language models.

Method

The authors used a bottom-up reverse engineering strategy combined with symbolic and ontological methods to create explainable and ontologically grounded language models. This involved constructing symbolic embeddings based on the "company a word keeps" principle, analyzing the contexts in which words are used to extract their meaning, and discovering the ontological structure implicit in ordinary language. The symbolic embeddings were created across various dimensions of meaning, and the relationships between these embeddings were analyzed to reveal subtyping and ontological relationships.

Main Finding

The authors discovered that by applying a symbolic reverse engineering strategy to the analysis of word usage in language, they could extract meaningful relationships and ontological structures that are implicit in ordinary language. They found that this approach could lead to the creation of symbolic embeddings that capture the multidimensional meaning of words, which in turn can be used to support commonsense reasoning and deep language understanding in AI systems. This method contrasts with the statistical and subsymbolic approaches of current large language models, which the authors argue are inherently unexplainable and lack the ability to understand language in a way that aligns with human cognition.

Conclusion

The conclusion of the paper is that while large language models have shown impressive capabilities in handling language tasks, they are not the solution to the language understanding problem or to reasoning, particularly commonsense reasoning. Due to their inherent unexplainability and subsymbolic nature, LLMs do not shed light on how language works or how humans externalize their thoughts in language. The authors suggest that the relative success of LLMs is not due to their subsymbolic nature but to the successful application of a bottom-up reverse engineering strategy. They advocate for the continuation of this strategy in a symbolic setting, combined with ontological methods, to develop explainable and ontologically grounded language models that can be used in problems requiring commonsense reasoning. The authors believe that this approach can realize the vision of Frege and Sommers and potentially illuminate the "language of thought."

Keywords

symbolic embeddings, ontology, compositional semantics, explainability