Extend OVMI to Conditional Language Distributions

Extend open-vocabulary mutual information (OVMI) from lexical information relative to an unconditional reference word distribution to a formulation that captures contextual information through a conditional language distribution.

Background

Open-vocabulary mutual information is developed in the paper as a measure of lexical information conveyed by a speech brain–computer interface relative to a reference distribution over intended words. The formulation evaluates whether individual words are supported and how reliably supported words are decoded, but it does not model the contextual dependencies among words in language.

The authors explicitly identify extending OVMI to a conditional language distribution as future work. Such an extension would be needed to evaluate contextual rather than purely lexical information, including the contribution of linguistic context to communication and decoding performance.