A transfer grammar for mapping USR to MRS: Toward multilingual generation

Abstract

Semantic representation-based multilingual natural language generation (NLG) is still a challenging problem in computational linguistics. In this paper, we present universal semantic representation (USR), a new semantic representation that provides semantics at multiple levels, including the speaker's intention and discourse relations. However, from an NLG perspective, the problem with USR is its lack of compatibility with other NLG systems. In this research, we propose a rule-based grammar system to convert USR into minimal recursion semantics (MRSs). The MRS representation serves as an input to the answer constraint engine (ACE) and provides natural English sentences. This rule-based grammar is implemented in CLIPS and consists of 15 modules with more than 500 transformation rules and 1,43,320 dictionary items. For evaluation, we chose 100 Hindi sentences annotated by linguists who have more than three years of experience in making USRs. The 100 Hindi sentences include different types of constructions, such as active and passive, discourse connectives, negation, causatives, nonfinite verb constructions, and more. It should be noted that the evaluation was performed in semi-controlled circumstances, using a linguist-curated test suite designed to systematically cover a predefined range of grammatical constructions. The evaluation results show a 91% success rate, with 87.0% precision, 91.0% recall, and an F1 score of 89.0%. In the human evaluation, three evaluators reported 94.96% comprehensibility and 94.23% fluency. An error analysis revealed the following systematic patterns: 33% coverage gaps, 44% mismatches in argument structure, 11% causative complexity, and semantic loss in constructions lacking MRS equivalents. The research also highlights five distinct kinds of mappings, which represent clear distinctions between the two systems regarding their encoding of meaning. These findings demonstrate both the strengths and the inherent limitations of rule-based cross-framework semantic transfer. This paper presents the first systematic mapping from USR to MRS, establishes quantitative baselines for the process of rule-based semantic transformation, and shows how language-independent representation structures can be used for grammar-specific generation.