Sequence Package Analysis and SoftComputing: Introducing a New Hybrid Method to Adjust to the Fluid andDynamic Nature of Human Speech

摘要

At Linguistic Technology Systems, we are using Sequence PackageAnalysis (SPA) to architect a new, pragmatically-based part of speech taggingprogram to better conform to the fluidity and dynamism of human speech. Thiswould allow natural language-driven voice user interfaces and audio mining pro-grams — for use in both commercial and government applications - to adapt to thein situ construction of dialog, marked by the imprecision, ambiguity and vague-ness extant in real-world communications. While conventional part of speech(PUS) tagging programs consist of parsing structures derived from syntactic (andsemantic) analysis, speech system developers (and users) are also very muchaware of the fact that speech recognition difficulties still plague such conventionalspoken dialog systems. This is because the inherent inexactitude, vagueness, anduncertainty that are inextricable to the dynamic and fluid nature of human dialogin the real world (e.g., a sudden accretion of anger/frustration may transform asimple question into a rhetorical one; or transform an otherwise simple andstraightforward assessment into a gratuitous/sardonic remark) cannot be adequate-ly addressed by conventional POS tagging programs based on syntactic and/orsemantic analysis. If we consider for a moment that the biological organism of thehuman mind does not appear (for the most part) to have much difficulty followingthe vagarious ebb and flow of dialog with remarkable accuracy and comprehen-sion, so that business transactions and social acts are consummated with a fairamount of regularity and predictability in our quotidian lives, why can't we designspoken dialog systems to emulate the human mind? To do this, we must firstuncover the special formulae that humans regularly invoke to understand human-to-human dialog which by virtue of its fluid and dynamic constitution is oftenpunctuated by ambiguities, obscurities, repetitions, ellipses, and deixes (indirectreferents) - the same stubborn and ineluctable features of natural language whichindividually and collectively impede the performance of speech systems. Using aunique set of parsing structures - consisting of context-free grammatical units,with notations for related prosodic features - to capture the fluid/dynamic natureof human speech, SPA meets the goal of soft computing to exploit the tolerancefor imprecision, uncertainty, obscurity, and approximation in order to achievetractability, robustness and low solution cost. And as a hybrid method - uniquelycombining conversation analysis with computational linguistics - SPA is com-plementary to artificial neural networks and fuzzy logic because in building aflexible and adaptable natural language speech interface, neural networks, or con-nectionist models, may be viewed as the natural choice for investigating thepatterns underlying the orderliness of talk, as they are equipped to handle the am-biguities of natural language due to their capacity, when confronted with incom-plete or somewhat conflicting information, to produce a fuzzy set.