Year: 1996
Type: book-chapter
Abstract: The problem with word boundaries lies in locating them.In most spoken language, few cues are available to signal reliably where one word ends and the next begins.However, understanding spoken language must be a process of understanding discrete words rather than utterances as indivisible wholes, because most complete utterances have never previously been experienced by the listeners to whom they are directed.To understand a spoken utterance, therefore, listeners must somehow, in the absence of explicit signals, locate the boundaries between the individual words (or more precisely, the lexically represented units, whatever these may be) of which the utterance is composed.Models of spoken-word recognition have addressed the word boundary problem in several ways, but the proposed solutions fall into two principal classes: Those which incorporate some explicit mechanism for the location (or at least postulation) of word boundaries, versus those which avoid the need for explicit word boundary location, by proposing that boundary information simply falls out of the normal processes of word recognition.The former class, which we can term Explicit Segmentation models, is differentiated according to the principles that various models propose for guiding word boundary location.The latter class, Serendipitous Segmentation models, essentially contains two candidate solutions: one based on word recognition in strictly sequential order, and one based on competition between word candidates for recognition.Explicit Segmentation models include the proposal that, in English, listeners apply a strategy of assuming that any strong syllable in the input is word-initial (Cutler & Norris, 1988;Cutler & Butterfield, 1992).This "Metrical Segmentation Strategy" (Cutler, 1990) is efficient in that most English lexical words do indeed begin with strong syllables, and most strong syllables in typical utterances are indeed word-initial (Cutler & Carter, 1987; see that article for more details on the proposal for English, including separation of the access of open-versus closedclass words).Similar prosodically based proposals for explicit segmentation exist for other languages, for example, the proposals that French listeners use a syllabic segmentation procedure (Mehler, Dommergues, 87 88 CUTLER Frauenfelder, & Segui, 1981) or that Japanese listeners use a mora-based procedure (Otake, Hatano, Cutler, & Mehler, 1993).Serendipitous Segmentation models see word boundary information as arising incidentally from the processing operations of recognition.Cole and Jakimik (1978) made a clear statement of such a model in the elaboration of their proposal that recognition of spoken utterances proceeds in strictly temporal order: "one word's recognition automatically directs segmentation of the immediately following word'' (1978, p. 93).A similar sequential recognition proposal is embodied in Marslen-Wilson and Welsh's (1978) Cohort Model, which focused on the fact that some words become unique (distinct from all other words in the language) prior to their acoustic offsets; the claim here is that the recognizer can in such cases predict in advance where the current word will end and, by implication, where the next will begin.Competition among candidate words, as embodied in recent connectionist models such as TRACE (McClelland & Elman, 1986) and SHORTLIST (Norris, 1991, 1994), provides another mechanism by which segmentation emerges from independent processing operations.Given an input string, candidates matching any part of the string will be activated, and competition will occur among them.The competition will in general be won by any highly activated sequence of competing candidates that successfully accounts for the entire string without any leftover portions.This process permits no role for boundary detection processes, since words beginning at any point in the input can be activated.A feature of the major classification into explicit versus serendipitous segmentation that is rarely acknowledged is that only the proposal that segmentation be explicit simultaneously addresses the word boundary problem from the point of view both of the adult language user and of the prelinguistic infant.As Mehler, Dupoux, and Segui (1990) spelled out in detail, the word boundary problem for an infant is substantially greater than that for an adult language user.For the prelinguistic infant, most speech input is continuous, just as it is for the adult listener; although caretakers will, in many communities, explicitly teach words, this caretaker behavior appears after the infant has begun to produce language (i.e., essentially after the initial segmentation problem has been solved; see Cutler, 1994, for a review).In solving the initial segmentation problem, though, the infant cannot rely on lexical knowledge at all; a lexicon must be constructed, and the construction process must be begun from the most minimal of bases.The most an infant can be born with in this respect might be, perhaps, the expectation that there will be words (i.e., that linguistic means will exist to express communicable knowledge in discrete memorizable chunks).Further than that there can be no specific expectations-for example, there can be no expectations regarding the structure of words, since this varies widely from language to language.Certainly the infant confronted with the first samples of speech input can have no preexisting stock of words or
Source: Max Planck Digital Library
Author: Anne Cutler
Language: English
Subfield: Experimental and Cognitive Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W2160958365