Research on visual word recognition: From verbal learning to parallel distributed processing

Canadian Journal of Experimental Psychology, Dec 1999 by Herdman, Chris M

It is my pleasure to welcome you to this special issue on Visual Word Recognition. Over the past 30 years, research on visual word recognition has contributed greatly to the understanding of how information is processed, represented, and accessed in the cognitive system. Today, word recognition continues to serve as a rich domain for modelling and investigating core issues in cognition. In addition, research on word recognition has been used to improve assessment and instruction of reading for children and adults. The following paragraphs provide a brief overview of some trends and developments in word recognition research over the past few decades. This overview is not meant to be exhaustive (or unbiased). My intent is to give readers from outside the area a feel for why word recognition is such an, important domain for cognitive research.

The 1960s. In the 1960s, the zeitgeist for human experimental research was largely defined by verbal learning. Verbal learning research was deeply rooted in the approach that had driven behaviourism: Researchers in verbal learning were focused on discovering the principles that govern how associations between items (e.g., pairs of nonsense syllables, lists of words) are developed and reinforced. Paradigms such as list learning were used to assess how memory traces were formed and strengthened, and how old and new memory traces could create proactive or retroactive interference. In verbal learning, the perceptual uptake and coding of an individual item through the cognitive system was not of concern. The laying down of a memory trace was thoroughly investigated, but the architecture of the representational system was not.

Information processing and mental representations. Against the massive backdrop of verbal learning research, Morton (1969) proposed his logogen model. In the logogen model, the focus was placed on how words are processed through an activation-based system in which there exists a representation (logogen) for each known word. Logogens were conceptualized as information accumulators that are sensitive to experiential and contextual factors. For example, word frequency is assumed to affect a logogen's activation threshold, whereby words that are experienced frequently have lower threshold than words that are experienced infrequently. Lower thresholds translate into quicker firing of a logogen, and thus faster response times. In contrast to the approach taken in the verbal learning literature, Morton was not concerned with modelling how logogens are acquired. Instead, the logogen system was modelled as a fully developed representational system with an emphasis on describing the properties of these representations. In sum, Morton's approach is a good example of an early cognitive model because he includes assumptions about the representation of lexical knowledge as well as assumptions concerning the coding and flow of information.

Serial or parallel? In the 1970s other classes of models were soon developed as alternatives to the logogen model. Interestingly, these models often incorporated perspectives from areas outside experimental psychology. Consider search models of word recognition. Search models are based on the notion that word recognition involves a serial search of items in the mental lexicon. Why serial and not parallel search? One reason is that through the 1970s and early 1980s, cognitive modelling was strongly influenced by the computer metaphor. Computers were serial beasts that could search large databases rapidly, so why not assume that the search for lexical entries in the human system is also serial?

Many cognitive researchers do not find the assumption that lexical search is serial to be very palatable. Neuropsychological research showing that massive parallelism exists in the brain has heightened the discomfort with serial models. However, there are at least four reasons why proponents of serial-search models should stick to their guns rather than simply re-characterizing search as parallel. First, despite evidence for massive parallelism, it is equally obvious that there is temporal ordering of brain activity that corresponds to the information processing sequence. The logical approach, therefore, is not to repress the notion that serial mechanisms can play a role in lexical processing, but instead to determine which events in the process are best described as parallel and which events are best described as serial.

Second, researchers in the 1970s realized that data per se cannot allow one to unequivocally differentiate between a search that is parallel versus a search that is serial but very efficient. Third, the evolution of serial-search models has occurred in a logical, step-by-step manner (e.g., compare Paap, Newsome, McDonald, & Schvaneveldt, 1982, to Paap, McDonald, Schvaneveldt, & Noel, 1987). This gradual evolution has allowed researchers to build upon clearly articulated assumptions. Fourth, search models of word recognition have continued to "survive" many of the empirical acid tests that have been thrown in their direction: There is no compelling empirical reason to abandon these search models.