Naming versus referring in the selection of words

The theory of lexical selection presented by Levelt, Roelofs & Meyer addresses the mechanisms of semantic activation that lead to the selection of isolated words. The theory does not appear to extend naturally to the referential use of words (particularly pronouns) in coherent discourse. A more complete theory of lexical selection has to consider the semantics of discourse as well as lexical semantics. A successful theory of lexical selection in speech production must explain a huge number of phenomena, ranging from the semantic makeup of words to how the articulatory form of a phonetic segment accommodates neighboring segments. The theory presented by Levelt, Roelofs & Meyer tackles an impressive number of these phenomena with substantial conceptual analysis and empirical evidence, but it does not tackle all of them. In this commentary I will analyze issues associated with one particular lexical phenomenon, the use of pronouns, that Levelt et al. do not address and that provide some challenges to the framework that they present. Pronouns are very common; “he” is the tenth most frequent word in English and other pronouns (“it,” “his,” and “I”) are not far behind (Kuçera & Francis 1967). It seems likely that this class of words is frequent because it is very useful. Researchers from a number of disciplines (e.g., Brennan 1995; Fletcher 1984; Grosz et al. 1995; Marslen-Wilson et al. 1982) have focused on how pronouns (and other reduced expressions) contribute to the coherence of discourse by implicitly marking the semantic entities that are central to a discourse. Do the mechanisms described by Levelt et al. provide a way of accounting for this use of pronouns? Levelt et al.’s theory has been developed primarily to provide a detailed account of speakers’ performance in the picture-naming task. A speaker who repeatedly exclaimed “it” in response to the stimulus pictures would justifiably be regarded as uncooperative, so it is not surprising that the task provides little evidence on pronoun selection. Yet during an ordinary conversation a speaker would be expected to use “it” frequently to refer to the kinds of objects represented in the stimulus pictures. Consider the following sentence, which Levelt et al. offered as a possible way that a speaker might describe a picture. I see a chair and a ball to the right of it. Here the expressions “a chair” and “it” are coreferential; they refer to the same thing. Presumably Levelt et al. would say that the word “chair” is selected because the conceptual and semantic preconditions stimulated by the picture lead to its being the most highly activated lemma. But then why is the same entity subsequently referred to with “it”? The semantic system used by Levelt et al. was selected to avoid the hyperonym problem, that is, in a compositional representation the activation of the semantic features of a concept will also activate any superordinates of that concept. From this perspective on semantics, pronouns might be seen as the ultimate hyperonyms, those specifying only minimal semantic features such as gender, number, and animacy. Levelt et al.’s use of a semantic system designed to avoid inadvertent intrusion of hyperonyms does not immediately suggest a reason for the widespread, systematic use of a particular class of hyperonyms. The example sentence points to the powerful ways in which sentential and discourse context can influence lexical selection, a topic I have worked on with my colleague Randall Hendrick. Building on the work of others, we (Gordon & Hendrick 1997; in press) have argued that a discourse model consists of a set of semantic entities and a series of predications, of which the entities are arguments. With respect to lexical selection, we have argued that the primary purpose of names (and other unreduced referring expressions) is to introduce semantic entities into a model of discourse, whereas the primary purpose of pronouns (and other reduced expressions) is to refer directly to entities that are prominent in the discourse model. On this view, results from a task like picture naming may provide evidence about the mechanisms that a speaker uses to initially introduce an entity into the discourse. Competitive selection based on activation levels is an attractive mechanism for accomplishing this, one that builds on the substantial explanatory use of such mechanisms within cognitive psychology to account for a variety of semantic and lexical processes. However, this appealingly straightforward mechanism may have to be substantially augmented if it is to account for the selection of words that refer to entities in an evolving discourse model as well as words generated from long-term memory. For the example sentence, we would argue that the selection of the pronoun reflects the prominence of the semantic entity CHAIR in the discourse model. To the extent that we are right, a theory of lexical selection must not only address the semantic organization of lexical concepts in long-term memory (as Levelt et al.’s theory does); it must also address the semantics of models created to represent the interrelations of entities and concepts in a discourse. The principles governing lexical selection from these two domains may be quite different. Commentary/Levelt et al.: Lexical access in speech production 44 BEHAVIORAL AND BRAIN SCIENCES (1999) 22:1 categories (e.g., animates vs. artifacts) are clustered in distinct (but contiguous) cortical regions. In this view, lemmas would also have a syntactic role. It is clear that syntactic properties cannot be directly attached to concepts, because semantic features do not directly map onto syntactic features. The syntactic properties could be attached to the phonological or orthographic word forms; however, this is computationally inefficient because syntactic information is modality-independent (but see Caramazza 1997). Therefore, the intermediate lemma level is the most adequate for accessing syntactic information. Lexical concepts acquire syntactic properties relatively late in development (between the ages of 2.6 and 4 years; see Levelt et al., sect. 1). This process is termed syntactization by Levelt et al. and refers to the development of a system of lemmas. However, the explosive growth of the lexicon takes place between the ages of 1.6 and 2.6. This means that an efficient mapping between concepts and phonological word forms is already established at that onset of the syntactization process. Within the architecture of Levelt et al.’s model, such mapping would presumably involve conceptual nodes and word forms, thus bypassing the yet-to-bedeveloped lemmas. Therefore, the later development of the lemma level would mean a massive rewiring of the lexical system. We believe that such a process is truly unlikely (both from the neural and computational standpoints). By contrast, if lemmas develop as a necessary component of the mapping between meaning and phonology, syntactization is simply the process of linking syntactic features to the existing lemma representation.