Trainable linker improves inference between sentence parts in ai reasoning
Learning Proposition-Level Links from Sentence-Level Supervision for Neuro-Symbolic NLI
Computation and Language
Summary
Some AI systems try to understand if one sentence supports or contradicts another by breaking sentences into smaller pieces and then reasoning about how those pieces relate. The authors developed a new trainable method that better predicts how these smaller pieces connect without needing detailed example links during training. Their approach learns from overall sentence relationships and improves accuracy significantly on popular language inference tests. This helps AI systems reason more accurately by connecting parts of sentences in smarter ways.
What this means in practice
- •For natural language processing teams: Enhance AI models that determine if one sentence logically follows from another by improving links between sentence parts automatically.
- •For legal document review teams: Improve automated systems that check logical consistency between statements in contracts or laws by better connecting the components of sentences.
Authors
Xuyao Feng, Antonis Bikakis
Abstract
We study neuro-symbolic natural language inference (NLI) with independently translated propositional representations of the premise and hypothesis. Before symbolic reasoning, the system must infer semantic links between their proposition-level units (atoms). Existing systems based on Abstract Meaning Representation (AMR) and Boolean satisfiability (SAT) identify these links using fixed embedding-based source selection and pretrained sentence-level NLI models not trained for atom-level linking. We introduce a trainable atom linker that jointly predicts link existence, source selection, and relation type. Without atom-pair training annotations, we treat cross-formula link structures as latent and train the linker from sentence-level NLI labels, keeping the translator and SAT reasoner fixed. We evaluate on the full public test splits of Stanford Natural Language Inference (SNLI) and Sentences Involving Compositional Knowledge (SICK), and the matched and mismatched validation splits of Multi-Genre Natural Language Inference (MNLI). Our system improves accuracy by 11.40-27.09 percentage points over a baseline applying sentence-trained NLI models unchanged to atom pairs within the same SAT pipeline. It also improves over the prior fixed-linker AMR-to-SAT system by 8.14-26.32 percentage points on SNLI and both MNLI splits, with a smaller gain on SICK.