N400 to lexical ambiguity and semantic incongruity in schizophrenia

Abstract

Our previous work showed a semantic bias in interpreting ambiguous words in schizophrenia, with disproportionate misinterpretation of subordinate meanings (toast at a wedding). We proposed pre-selection in schizophrenia of dominant-meaning networks at points of lexical ambiguity, thereby misleading thought. This selection bias may be due to semantic memory hyper-priming causing strong associates to dominate cognition. Alternately, later verbal memory maintenance failure may cause weaker associates to fade more quickly than stronger associates from memory due to less initial activation. To further examine this semantic bias, patients and controls were presented short 4 word long sentences (The toast was buttered). The second word was a homograph or unambiguous noun. The last word disambiguated homographs (dominant or subordinate meaning) or was congruent or incongruent with unambiguous nouns. Previously, we showed increasingly larger N400 from unambiguous associates to dominate associates to subordinate associates to unambiguous non-associates in controls. Pre-selection of dominant meanings predicts that schizophrenia patients would show small N400 to dominant associates and as large N400 to subordinate associates as to incongruous endings. Here, controls again showed graded N400 amplitudes. Patients with schizophrenia showed small N400 to congruent and dominant endings and large N400 to subordinate and incongruous endings. These data suggest early pre-selection of dominant associates in schizophrenia. This effect is unlikely solely due to verbal memory maintenance failure, as patients were able to detect incongruity, albeit with a smaller N400 effect, and displayed generally larger N400 to all stimuli. These results suggest alterations in semantic memory associative networks coupled with verbal working memory maintenance decay in schizophrenia.

Introduction

Thought disorder is a prominent symptom of schizophrenia. It is inferred through bizarre and disordered speech, which at the most extreme instances may result in word salad, an incomprehensible, inscrutable jumble of words, and neologisms, the idiosyncratic coinage of words with delusional significance. More commonly, though, thought disorder is subtle, and manifests as a certain looseness in conceptual focus by which the train of thought is normally maintained. Tangential and obliquely related concepts may sidetrack narrative, such that the overarching purpose of the discourse may be ultimately lost. Cohen and Servan-Schreiber (1992) have described this effect as the “local” sentence context having more influence than the “global” discourse context in schizophrenia.

Precisely what mechanisms underlie thought disorder remain unknown. Because it is inferred from language, memory systems related to language are likely involved. Although some have argued that the problem is one of executive, attention mechanisms (e.g., Schwartz, 1982), compelling evidence exists that verbal content influences thought disorder (described below), inconsistent with a purely executive abnormality, wherein phasic lapses in executive functions should be independent from the specific words in memory. Such evidence for processing deficits for specific word-types strongly suggests the cognitive abnormality must, to a great extent, involve language-specific processes. Of course, a general physiological abnormality may impact multiple systems beyond those language-related. However, the crucial point is that basic systems related to semantic memory must be implicated in thought disorder.

Semantic memory represents the lexicon of stored knowledge, and has been referred to as “crystallized” memory, essentially static. Collins and Loftus (1975) proposed that words were stored in a conceptual space and linked with related items (associates) in a semantic network, separated by the “semantic distance”, or strength of association in this conceptual space. This model provided the major framework for understanding the automatic spread of activation among related concepts. For example, upon seeing the word red, proximal associates such as fire and orange receive more initial automatic activation than distal associates such as sunsets or roses. This is reflected in speeded responses to close associates, termed semantic priming. Subjects are able to make judgements about semantically-related words more quickly than unrelated words (Schvaneveldt and Meyer, 1973).

The initial semantic memory automatic spread of activation may have a relatively short duration. For example, Hagoort (1993) demonstrated that aphasic patients failed to show semantic priming after approximately 1250 ms although they had shown relatively normal priming at shorter intervals. This suggests that as memory demands for semantic priming exceed around one second, a second memory process is necessary to maintain activations and observe priming. This is likely served by verbal working memory, a controlled or executive process, that is veiled, or not accessible to consciousness. Thus, unless overwritten by new information, priming effects may last for several seconds or longer (e.g., Deacon et al., 1999). An intact brain shows a seamless transition between the initial bottom–up activation and the maintenance of this activation in verbal working memory. Aphasic patients impaired at controlled processes, presumably relying on verbal working memory processes rather than initial semantic activation, appear to no longer show priming at slower presentation rates. Thus, the duration of the initial burst can be estimated to last for less than 1 s. In addition, much work in the cognitive literature indicates that, in the case of relatively weak context as provided by one priming word or a short sentence, it takes approximately 250 ms for the context to inhibit the initial activation of contextually inappropriate associates (e.g., Onifer and Swinney, 1981). Thus, a two-stage model of semantic activation and verbal working memory maintenance suggests that in the case of relatively weak context all associates of a word are activated in a bottom-up fashion, with top-down inhibitory mechanisms modulating the initial activation after a quarter of a second. Relatively pure automatic effects are thought to be evident at presentation rates under 300 ms, before controlled processes can exert their effects. The initial burst of bottom-up activation decays by at most a second, unless maintained in verbal working memory. When maintained, semantic priming effects can last for several seconds or longer due to veiled verbal working memory controlled processes, and without need for conscious awareness of the activated information.

Whether the thought disorder observed in schizophrenia involves deficits in semantic memory organization and bottom-up activation or in later controlled top-down processes such as verbal working memory maintenance or context utilization remains a central question. There is evidence for hyper-priming in schizophrenia at short stimulus onset asynchronies (SOAs). Schizophrenic subjects show an increase in the effect of priming on a speeded lexical decision task (Maher, 1983). This facilitation effect on priming has been replicated by several groups (Kwapil et al., 1990, Spitzer et al., 1994, Baving et al., 2001). Spitzer et al. further showed that schizophrenics demonstrated increased second order priming (e.g., lemon primes sour, which primes sweet; Spitzer et al., 1993). These data suggest that schizophrenics show greater initial activation in semantic networks than controls. Mathalon et al. (2002) and our work (described below) also demonstrated N400 ERP signs of hyper-priming in schizophrenia. It is important to realize that these effects appear only at very short SOAs, and that priming is generally absent or deficient at longer SOAs in schizophrenia, consistent with the two-stage model described above: initial semantic priming with subsequent verbal working memory maintenance of initial activation.

In addition to some evidence for hyper-priming, a large body of literature has argued for a lack of context utilization in schizophrenia. These data show that patients fail to use context to order thought. At SOAs over 750 ms, schizophrenics appear to perform worse than controls to all words (Barch et al., 1996). Cohen et al. (1999) argued for a pervasive late stage verbal working memory abnormality in schizophrenia, with no semantic bias. That is, schizophrenics were as likely to incorrectly select a subordinate ambiguous word meaning as a dominant ambiguous word meaning. However, the SOAs used were 1 and 5 s. At both these intervals, hyper-priming may not be apparent. As indicated earlier, Hagoort (1993) showed that aphasics showed normal priming effects at SOAs under 1 s (100 and 500 ms), but not at over 1 s (1250 ms). This suggests that time intervals approaching 1 s are not likely to accurately access semantic memory activation in subjects with defective verbal working memory processes, and thus studies that use relatively slow presentation rates cannot rule out semantic memory deficits.

Despite this confound, it is likely that schizophrenics have abnormalities using context to modulate activation in semantic networks. Several studies have demonstrated insensitivity to context in schizophrenia. Kuperberg et al. (1998) showed that thought disordered patients were less impaired in processing words preceded by syntactic or semantic violations, suggesting they were less sensitive to linguistic context. Such insensitivity could be due to a failure to maintain context or to correctly use stored context. Titone and colleagues have argued that schizophrenia is characterized by an inability to utilize context in situations that stress verbal working memory contextual inhibition. Titone et al. (2000) used weak and strong biasing contexts in schizophrenia and suggested that patients showed impaired controlled priming with weak biasing context, but normal controlled priming with strong context. Titone et al. (2002) compared priming following proverbs that could be literally interpreted (e.g., kick the bucket) or only abstractly interpreted (e.g., pay through the nose). Patients showed impaired inhibition to literal proverbs, with increased priming to literal associates, but normal priming to abstract proverbs. Titone et al. argued that context was maintained properly in schizophrenia, but inefficiently used when the controlled inhibition of initial semantic activation was necessary. Using the N400 event-related potential (ERP) Condray et al. (2003) showed brainwave reductions in unmedicated patients at slow presentation rates that relied on top-down processing and development of expectancies. Thus, the deficit in schizophrenia must involve some degree of top-down deficits. The question becomes, then, of the relative contributions of bottom-up and top-down deficits to thought disorder.

The N400 ERP, a negative brainwave at roughly 400 ms post-stimulus, appears to be activated by conceptual-level representations. It is larger to incongruent (contextually inappropriate) than congruent sentence endings (e.g., Kutas and Hillyard, 1980, Kutas and Hillyard, 1982) that may be words or pictures (Nigam et al., 1992); to unrelated than to related words (Kutas and Hillyard, 1989); and to weakly associated than strongly associated words (Kutas and Hillyard, 1989, Polich, 1985). The increase in N400 to less related or unrelated words is termed the N400 effect. Although the precise function underlying N400 is unknown, N400 can be utilized experimentally to examine priming effects. The N400 has also been used to examine lexical ambiguity. The first report of N400 to ambiguous words (Van Petten and Kutas, 1987) showed that associates of ambiguous words spelled the same for the different meanings (homographs) showed intermediate N400 amplitudes between unambiguous associates and non-associates.

In our previous research we have further examined the N400 to homograph associates in well subjects, and also used N400 to examine the idea that bottom-up hyper-priming in schizophrenia leads to a semantic bias towards strong associates. Although different homograph meanings necessarily have different patterns of activation in semantic space, they are identical at the feature level and have the same entry point for lexical access. Because homographs share orthography between different meanings, in the absence of context the appropriate meaning of the homograph cannot be determined. We conceive of the different semantic conceptual spaces activated by each homograph meaning to be a separate (though partially overlapping) network. By definition, the networks should not be highly interconnected at the associate level. By using tasks where the homograph precedes the disambiguating information (e.g., the toast was emotional), pre-selection biases can be identified in the responses to the disambiguating information. For example, if toast's browned bread meaning was pre-selected, then the disambiguating word emotional would be incongruent. Through the use of homographs with dominant versus subordinate meanings in the lack of context, semantic biases at points of lexical ambiguity can be examined in schizophrenia. Salisbury et al. (2000) showed larger N400 in schizophrenia to subordinate homograph associates on a passive reading task to sentence ending words than in controls that was associated with greater thought disorder. Salisbury et al. (2002) had subjects make a sensibility judgement for the sentences. Again, schizophrenia patients showed larger N400 to subordinate homograph sentence endings, even when the sentence was interpreted as making sense. Patients showed N400 to subordinate homograph endings judged sensible as large as controls did to subordinate endings they failed to comprehend. Behaviorally, patients failed to comprehend disproportionately more subordinate homograph sentences. At the same time, patients showed overall larger N400s to all sentence types, regardless of noun network type. Sitnikova et al. (2002) examined contextual utilization in schizophrenia by presenting subjects with longer sentences containing 2 clauses. The first clause biased the sentence to dominant or subordinate homograph meanings. The second clause always referred to the dominant homograph meaning. Unlike controls, schizophrenia patients did not generate larger N400 to dominant associates when the first clause biased towards subordinate meanings. However, schizophrenia patients still showed an N400 effect to non-associates, albeit smaller than in controls.

Salisbury et al. (2002) proposed a joint deficit comprising hyper-priming and maintenance failure in schizophrenia to explain the clear behavioral semantic bias and larger N400 to subordinate homographs in schizophrenia coupled with the moderate N400 effect reduction. In that experiment, the responses in schizophrenia patients to truly incongruent endings were not examined. The current study examined the interpretation of ambiguous and unambiguous words in schizophrenia contrasted with controls, by comparison with sentence ending words that were incongruent with unambiguous nouns. If patients were unable to maintain context, then they should show reductions in the N400 effect to incongruent endings and larger N400s to all sentence endings. If they had problems with hyper-priming and pre-selection of dominant networks, then the N400 to dominant homograph endings and unambiguous associates should be similar and smaller (i.e., more congruent) than the responses to subordinate and incongruent endings (i.e. less congruent), which should be similarly large.