A Cognitive Approach to Resolving Semantic Ambiguity in Quantifier Interpretation for Machine Translation

This paper presents the first cognitive modelling approach to addressing scope ambiguity in machine translation. Quantifier scope ambiguity arises when multiple quantifiers appear in a single sentence, and their intended interpretation depends on their relative scope rather than merely their surface order or linear structure. This presents a challenge when translating between languages that differ in how scope relations are grammatically expressed. We study this problem in English-Japanese translation, with English allowing scope ambiguity and Japanese typically requiring scope to be made explicit, meaning that a single English sentence may correspond to multiple candidate Japanese translations, determined by contextual cues. We train a cognitive architecture to perform scope interpretation and translation through procedural neural operations that manipulate contextual representations, rather than relying on surface-level statistical associations. The architecture learns the task from a small set of examples, using substantially less parallel data than conventional neural translation approaches. We evaluate the architecture on a controlled dataset of 259 test cases against Large Language Models, which represent the current state-of-the-art in related ambiguity tasks. In scope interpretation, the cognitive architecture achieves 85% accuracy, outperforming all tested LLMs, including deepseek-r1:8b (71%), gpt-oss:20b (69%), llama3.1:8b (53%), and gemma3:4b (14%). In translation, the architecture reaches 63% accuracy, whereas LLM models range between 23% and 45%. Moreover, the architecture achieves a near-maximum translation entropy, indicating that it successfully distributes translations across alternative interpretations. In contrast, LLMs produce substantially lower entropy scores, revealing a strong bias towards a fixed interpretation. The results demonstrate that cognitive modelling may provide a data-efficient and robust approach to handling scope ambiguity in machine translation.