🤖 AI Summary
Large language models (LLMs) suffer from limited controllability and interpretability during text generation. Method: This paper introduces the *formal semantic geometry* framework, which decouples sentence semantics into semantic roles and lexical content. It proposes a Transformer-based variational autoencoder that explicitly models semantic structure in a low-dimensional Gaussian latent space. Furthermore, it designs a geometric-guided probing algorithm to enable targeted manipulation and attribution-based interpretation of latent representations. Contribution/Results: Evaluated on GPT-2, the method significantly improves fine-grained semantic editing accuracy—e.g., role substitution and attribute modulation—while enhancing human interpretability. It establishes a novel paradigm for LLM generation that jointly ensures formal rigor (via geometrically grounded semantic modeling) and neural operability (via differentiable, geometry-aware latent control), bridging symbolic semantics and deep learning.
📝 Abstract
Formal/symbolic semantics can provide canonical, rigid controllability and interpretability to sentence representations due to their extit{localisation} or extit{composition} property. How can we deliver such property to the current distributional sentence representations to control and interpret the generation of language models (LMs)? In this work, we theoretically frame the sentence semantics as the composition of extit{semantic role - word content} features and propose the formal semantic geometry. To inject such geometry into Transformer-based LMs (i.e. GPT2), we deploy Transformer-based Variational AutoEncoder with a supervision approach, where the sentence generation can be manipulated and explained over low-dimensional latent Gaussian space. In addition, we propose a new probing algorithm to guide the movement of sentence vectors over such geometry. Experimental results reveal that the formal semantic geometry can potentially deliver better control and interpretation to sentence generation.
