Distinctive Human Dynamics of Semantic Uncertainty: Contextual Bias Accelerates Lexical Disambiguation.

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Neuroscience

Background:

• Lexical-semantic ambiguity poses a challenge for sentence comprehension.
• Understanding how contextual information resolves ambiguity in real-time is crucial.
• Existing models of semantic processing vary in their dynamic and context-sensitive aspects.

Purpose of the Study:

• To investigate the temporal dynamics of lexical-semantic ambiguity resolution during sentence comprehension.
• To quantify semantic uncertainty using a novel entropy-based measure.
• To examine the influence of contextual bias strength on disambiguation speed and trajectory.

Main Methods:

• Time-resolved eye-tracking was employed to capture fine-grained temporal data.
• A novel entropy-based measure quantified semantic uncertainty from group-level semantic choice distributions.
• Parametric manipulation of contextual semantic bias strength was used to probe model sensitivity.

Main Results:

• Semantic uncertainty decreased gradually over time and sharply after ambiguous words, indicating incremental integration and syntactic anchoring.
• Stronger contextual bias accelerated uncertainty reduction, showing a near-linear relationship.
• A Chinese BERT model (RoBERTa-wwm-ext) replicated general uncertainty reduction trends but lacked sensitivity to contextual bias strength.

Conclusions:

• Human semantic processing is dynamic, context-sensitive, and shows continuous convergence towards intended meanings.
• Current language models, while broadly capable, do not fully capture the nuanced, context-driven modulation of semantic processing observed in humans.
• The findings offer a new empirical characterization of disambiguation dynamics and highlight potential architectural differences between human and artificial semantic systems.