A time-critical simulation of language comprehension

Abstract

Language comprehension is usually not understood as a time-critical task. Humans, however, process language on-line, in linear time, and with a single pass over a particular instance of speech or text. This calls for a genuinely cognitive algorithmic approach to simulating language comprehension. A formal conception of language is developed, as well as a model for this conception. An algorithm is presented that generates such a model on-line and from a single pass over a text. The generated model is evaluated qualitatively, by comparing its representations to linguistic segmentations (e.g. syllables, words, sentences). Results show that the model contains synonyms and homonyms as can be found in natural language. This suggests that the algorithm is able to recognize and make consistent use of context–which is crucial to understanding in general. In addition, the underlying algorithm is evaluated against a baseline approach with similar properties. This shows that the generated model is able to capture arbitrarily extended dependencies and therefore to outperform exclusively history-based approaches.