Prosodic information in an Integrated Lexicon

A Integrated Lexicon framework is proposed for combining insights from current separate prosodic research paradigms (language resource creation, experimental psycholinguistics, rules for speech synthesis, sentence search space restriction in speech recognition, prosody in discourse, emotional prosody) in the medium term into work in general and computational linguistics. Three principles for integrating autonomous prosodic information into the lexicon are proposed (the Compositionality Principle, the Semiotic Principle, and the Metalocutionary Principle), and used as a framework for developing hierarcical attribute-value (AV) lexical representations for nuclear contours and the prosody of discourse particles. 1. Lexical prosody for intonation The representation of the basic inventory of prosodic units and their combinatorial, semantic and phonetic properties is an open question, despite much attention over many decades. For present purposes, this inventory of prosodic units is termed a prosodic lexicon, and a lexicon of words (or idioms) a locutionary lexicon. A proposal is presented for modelling autonomous lexical prosodic items and relating them to locutionary items in an integrated attribute-value (AV) based lexicon. The proposal is related to work by Aubergé on multiparametric prototypes and Steedman on lexical functions in Combinatory Categorial Grammar. Current paradigms of prosodic description concentrate heavily on other issues such as the extensive prosodic annotation and experimental modelling of prosodic patterns in speech corpora, procedural rule formulation for speech synthesis, prosody for restricting sentence search space in speech recognition, prosody in discourse, and prosody in emotion modelling. Each approach continually uncovers new insights, but it is unclear how they relate to each other. Consequently the present contribution addresses the issue of integrating prosodic information into lexicalist approaches to language representation and computation. The approach is still fragmentary, but the medium-term goal is develop a formal representation for lexical prosody as a basis for compositional descriptions of more complex prosodic patterns, and as a reference point for combining results of different fields. To provide a non-ad hoc lexical representation involves addressing three basic principles which I identify as follows: the Compositionality Principle (CP), the Semiotic Principle (SP), the Metalocutionary Principle (MP). The domain of the present contribution is lexical prosody for intonation, rather than lexical tone or lexical accent placement. 1.1. The Compositionality Principle (CP) CP: Prosodic patterns are grounded in a prosodic lexicon and are projected compositionally from prosodic lexical items and their combinatorial properties to larger prosodic patterns. The prosodic CP is characterised as follows for a compositional prosodic item P = constructionpros(p1; ::: pn): propProsi(P ) = composition(propProsi(p1), ::: propProsi(pn)) The oldest explicit compositional approach to prosody is the stress cycle of generative phonology, in which “stress patterns” are projected from the lexicon by the Compound Stress Rule (CSR), and from syntactic structure by the Nuclear Stress Rule (NSR) as a numerical coding. The empirical basis of this coding has repeatedly been called into question. In the past decade, Steedman [16] has paid most attention to the issue of projecting compositional prosody from the lexicon. Steedman proposes a set of accent and boundary tones with specific structural meanings and combinatorial properties as functions which map boundaries into utterances, yielding (with the syntax of Combinatory Categorial Grammar) hierarchical prosodic patterns corresponding to a hierarchy of foreground and background utterance constituents: Accent tones: L+H* := Theme/Bh H* := (Utterance/Theme)/bl H* := (UtterancenTheme)/Bl Boundary tones: LH% := Bh LL% := Bl L% := bl Boundaries: bl intermediate phrase boundary Bl intonation phrase boundary Steedman’s approach, and and the tone-level lexical representation, is embedded in a long history of prosodic phonologies, relating particularly to work by Selkirk [15], Pierrehumbert & Hirshberg [13] and Bird [2]. The approach accounts not only for syntactic constraints on prosodic patterning but also for divergences between prosody and syntax, and has been successfully implemented in prosody generators for speech synthesis. The selection of prosodic lexical categories discussed by Steedman is limited, perhaps because of the formal speech style of speech synthesis. Another approach to compositionality is Finite State Prosody, represented in work by Fujisaki, ’t Hart & Cohen, Pierrehumbert, Gibbon in the 1970s and 1980s, and more recently in work by Ladd [12], Jansche [10], and Gibbon [6]. In Finite State Prosody, a basic vocabulary of tones is mapped by a Finite State Automaton (FSA) into a regular set of tonal sequences. The tonal sequences may be mapped into allotone sequences by means of a Finite State Transducer (FST), the set of toneallotone pairs and sequence pairs constituting a regular relation [11]. 1.2. The Semiotic Principle (SP) SP: Prosodic patterns have a semiotic dual interpretation consisting of a compositional mapping to a semantic/pragmatic domain and a compositional mapping to a phonetic domain. Speech Prosody 2002 Aix-en-Provence, France April 11-13, 2002 ISCA Archive http://www.isca-speech.org/archive