Andrew Radford - Linguistics An Introduction [Second Edition]

listener can guess the identity of a particular word, using inference processes which do not themselves belong to the system of language perception. At the same time, however, parallel-interactive models are theoretically unconstrained, and it is therefore difficult to make testable predictions on the basis of such approaches.

Given parallel interaction, anything goes, and you can, for example, recognise a word without having any direct cues. This is impossible with the serial-autonomy approach. Moreover, as each module has a clearly described task in a serial-autonomous model, an output error or a recognition error can be traced back to the module that caused the error. This is impossible in a parallel-interactive model in which information is distributed over many different places which are all continuously interacting. In sum, parallel-interactive models of word recognition are extremely successful at the product level; in fact, they almost always produce an output, i.e. recognise a word. But they provide little insight into the actual mechanisms that are involved in understanding words. Serial-autonomous models are theoretically more interesting, and they make specific predictions as to which kinds of inputs are required for word recognition, but when the input is faulty or noisy, they are not efficient enough, and they cannot straightforwardly account for context effects.

It should be clear from the above where the lines of this particular dispute are drawn. Both types of model offer a story about how degenerate word forms may be perceived. For the adherent of a parallel-interactive account, such perception is due to the normal functioning of the perceptual system. It is a characteristic of this system that it is always seeking to identify words on the basis of any type of information available to it, and the only difference between a well-formed signal and a degraded signal is that, in the latter situation, one sort of information (the phonological form) is missing. From the serial-autonomous perspective, the lack of phonological form means that the language perception system breaks down at this point and another cognitive system (of guessing or inferencing) comes into play. Devising experiments which will distinguish clearly between these alternatives is a difficult task, and we have outlined above two paradigms which

produce conflicting conclusions. It is perhaps not accidental that most current models of lexical processing include both serial-autonomous and parallel-interactive features (exercises 1 and 2).

On the representation of words in the mental lexicon

A basic property of words is the arbitrary relationship they exhibit

between meaning and form: words have meaning, and they have phonological

Lexical processing and the mental lexicon

205

Figure 39 One view of the structure of the mental lexicon, illustrating the form of a lexical entry

or orthographic structure, and there is no way of recovering the former from the latter. Note that if this were not the case, we would not expect to find lexical differences between languages: if cow is the ‘natural’ sign for a bovine creature, we should be puzzled by the existence of vache in French. Given this arbitrariness of the linguistic sign, the lexicon (or the mental dictionary of a language) must include some sort of stored entry for the lexemes of a language. Most

psycholinguists believe that the mental lexicon must contain lexical entries which contain a number of separate but interconnected levels. The model of

a lexical entry in figure 39 is based on suggestions of the psycholinguist Pim Levelt.

According to this model, concepts must be distinguished from lexical entries, and lexical entries consist of two levels, one for the semantic form of the lexical entry, i.e. its meaning or content, and the other for the entry’s morphological make-up and its phonological properties. Hence, a lexical entry can be split into two parts, its lemma and its form information (note that in this literature the term lexical entry is used to refer to what we called lexemes and that the term lemma refers to the semantic representation of a lexeme). The lemma lexicon and the form lexicon are connected through lexical pointers: each lemma points to its corresponding form, i.e. it can address a particular entry in the form lexicon where the morpho-phonological properties of the lemmas are stored.

What is the evidence for distinguishing between these levels of representation in the mental lexicon? Switching our focus from perception to production, consider firstly the distinction between concepts and lemmas, and suppose a native speaker of English wants to formulate a message about the object/concept represented in figure 40.

According to the model in figure 39, this concept will activate the appropriate lemma in the lexicon, i.e. scissors, and subsequently the word form /sɪzəz/.

Concepts are represented on a prelinguistic level, whereas lemmas must be

part of the mental lexicon of a particular language. Thus, for native speakers of

206

words

Figure 40 A simple concept

English, the concept represented in figure 40 is probably the same as for native speakers of German, but at the lemma level there are differences: scissors

is inherently plural and is not countable (*one/*two/*three scissors), but the German equivalent of scissors is Schere, which is inherently singular and a

count noun (eine Schere, zwei/drei/vier Scheren, ‘one scissor, two/three/four scissors’). Similar examples are trousers (plural) versus Hose (singular), glasses (plural) versus Brille (singular). The fact that Schere, Hose and Brille are count nouns, whereas scissors, trousers and glasses are not, is a semantic difference on the lemma level, which is not a consequence of and does not result in different conceptualisations.

What about the distinction between the lemma and the form lexicon? Returning to perception, one important piece of evidence for distinguishing between these two levels of representation comes from our ability to process non-words,

i.e. words for which we have no proper meaning representation. Thus, native

speakers of English perceive a clear difference between the items in (195) and ( 196):

(195)

blatt

(196)

plaupf

The item in (195) is a potential English word, in terms of its phonological form, although it does not have any meaning in English. The item in (196), however, is an impossible word in terms of its form properties – specifically, English does not permit a syllable coda to comprise the consonant cluster /pf/ (see section 5). In other words, the difference between (195) and (196) in terms of their phonological form demonstrates that we can make judgements about the form of a lexical entry independently of its meaning. This in turn shows that the mental lexicon cannot be thought of as a set of entities (structured or otherwise) with direct form–meaning mappings; the form lexicon, it seems, can be accessed independently without

activating any links to meaning.

This idea has been confirmed in reaction-time experiments involving the recognition of non-words. In one set of experiments, subjects are presented with strings of four letters and their task is to decide as quickly as possible whether the stimulus letter string is or is not a word of English. Different conditions were tested as indicated in figure 41.

The experiment showed that decision times on totally illegal and unpronounceable sequences such as SJMF are considerably faster than for any other stimulus used; they are even faster than those for existing words of English. This indicates

Lexical processing and the mental lexicon

207

Figure 41 Five conditions in a word/non-word recognition experiment

that we possess a rapid process by which globally illegal words can be detected, and this process must be purely form-based.