Probabilistic Tree Transducers for Syntax-to-Semantics Conversion

This work investigates the use of Probabilistic Tree Transducers (PTTs) for syntax-to-semantics conversion in natural language processing. The study explores semantic parsing as a structured transformation problem, where constituency parse trees are converted into machine-interpretable logical representations through probabilistic rewrite rules grounded in formal language theory and tree automata. A PTT-based semantic parsing framework was implemented using Python, OpenFST, Stanford CoreNLP, spaCy, and NLTK. Experiments were conducted on the GeoQuery and ATIS benchmark datasets to evaluate exact-match accuracy, derivation coverage, and interpretability. The proposed system achieved 78.4% exact-match accuracy on GeoQuery and 74.1% on ATIS while maintaining complete derivation traceability. The study compares symbolic, rule-based approaches with neural seq2seq semantic parsers, highlighting the trade-off between predictive performance and interpretability. Detailed error analysis is presented, focusing on prepositional phrase attachment ambiguities and coverage limitations in rule-based systems. Results demonstrate that probabilistic tree transducers provide a transparent and auditable alternative for semantic parsing, making them particularly suitable for applications where explainability and diagnostic traceability are critical.