Chapter 2 Literature Review
2.4 Anticipation in ERP studies
2.4.1 N400 and anticipation
The sentence violation paradigm (i.e., using words that are either congruent or incongruent at the ending of sentences) was used in the very first study that
discovered the N400 component (Kutas & Fedemeier, 2011; Kutas & Hillyard, 1980)
18
and has since been applied to the discourse level to examine whether listeners processed larger pieces of message in the same way as single sentences. Van Berkum, Zwitserlood, Hagoort, and Brown (2003) designed two experiments that directly compared the impact of discourse-level information on an unfolding sentence to one without a preceding context. Subjects listened to sentences that ended in equally plausible words (“Jane told her brother that he was exceptionally
quick/slow.”), but the sentences were preceded by contextual discourse in the first
experiment (“As agreed upon, Jane was to wake her sister and her brother at fiveo’clock in the morning. But the sister had already washed herself, and the brother had even got dressed.”), while being presented alone in the second experiment. In
the first experiment, the discourse-anomalous words elicited a typical N400 effect that sometimes began before the words ended (e.g. “slow” in this particular example), but the same word did not elicit an N400 component in the second experiment. The experiments showed that listeners could relate the unfolding words rapidly to the preceding context, and discourse-dependent N400 effects were identical with sentence-dependent N400 effects.Both the studies on cloze probabilities and contextual discourse revealed that for expected target words, there was an inverse relationship between the amplitude of the N400 of the target words and the level of contextual constraint. This means that, usually, the higher the level of contextual constraint, the smaller the N400 amplitude for expected target words. However, unexpected words could also follow highly constraining context (e.g. “He painted the walls in the wrong HUES.”). Federmeier, Wlotko, De Ochoa-Dewald, and Kutas (2007) demonstrated that N400 amplitudes are graded by expectancy (indexed by the cloze probability) but unaffected by contextual constraint, therefore control methods other than cloze probability of the target word is needed to determine and control the level of contextual constraint. Previous studies
19
used the Latent Semantic Analysis (LSA) methodology to control the contextual constraint by looking into the semantic distance from the context sentence to the target words (Davenport & Coulson, 2011; Kuperberg, Paczynski, & Ditman, 2011).
LSA, according to its inventor Thomas Landauer, is “a fully automatic
mathematical/statistical technique for extracting and inferring relations of expected contextual usage of words in passages of discourse” (Landauer, Foltz, & Laham, 1998, p.263). It applies a form of factor analysis called singular value decomposition (SVD) to a matrix that represents the text, and through various mathematical steps of dimension reduction, a cosine between vectors in the reduced dimensional space is produced to present the similarity between words (Landauer et al., 1998). Empirical studies have demonstrated that LSA “produces measures of word-word, word-passage, and passage-passage relations that are well correlated with several human cognitive phenomena involving association or semantic similarity…[and] closely approximate[s]
human judgments of meaning similarity between words” (Landauer et al., 1998, p.
260). For example, it has successfully simulated a lexical priming study by Till, Mross, and Kintsch (Landauer & Dumais, 1997).
Conference interpreters are routinely exposed to the processing of meaningful and contextualized discourse. Therefore, the present study adopted a discourse level sentence violation (congruent/incongruent) paradigm to investigate how whether interpreters of different levels of expertise processed sentences differently. In addition, the LSA was used to control the level of constraint between the context and target word.
Despite the sensitivity of the N400 to offline semantic expectancy (as indexed by cloze probability), there was a debate about whether the N400 could actually index
“anticipation” rather than just “integration”. Some argued that listeners did not anticipate the target word; they merely found the word easier to integrate into the
20
prior context when it was congruent, and difficult when it was not (Hagoort, Baggio,
& Willems, 2009). Hagoort and colleagues posited that the N400 “reflect[s]
processes involved in the integration of the meaning of a word into the overall
semantic representation constructed for the preceding language input” (Hagoort et al., p. 820). On the other hand, Federmeier and Kutas (1999a) believed that some degree of anticipation has occurred, because they showed that in highly constraining contexts (e.g. “He caught the pass and scored another touchdown. There was
nothing he enjoyed more than a good game of…”), even though both “baseball” and
“monopoly” had low cloze probabilities when evaluated offline, subjects in an ERP experiment would generate a smaller N400 effect for “baseball” than for “monopoly”.
They argue that this was because the language processing system was strongly predicting the word “football”, thus pre-activating conceptual features of that word, which shared more features with within category anomalous words (“baseball”) than with between category violations (“monopoly”).
An ingenious design to provide more evidence that listeners are actually anticipating specific words came from studies looking at the EEG signals of words preceding the target word that was presumably being anticipated. Otten, Nieuwland, and van Berkum (2007) asked subjects to listen to short stories in Dutch that were highly constraining for a specific noun, or stories that were less predictive but contained the same prime words as the predictive stories. The gender marking (neuter or common) on prenominal adjectives was manipulated so that it was either congruent or incongruent to the anticipated noun. Results showed that when the stories were predictive, adjectives with an unexpected gender inflection would evoke a negative potential around 300 to 600-ms. When stories were not predictive, no such effect existed. De Long, Urbach, and Kutas (2005) manipulated the English indefinite article a/an in a similar design and saw the same results.
21
One way to reconcile the debate of whether the N400 indexed anticipation or integration is by taking into consideration the roles that the left and right hemispheres of the brain might play. Federmeier and Kutas (1999b) devised a visual-half-field study to explore whether the two hemispheres played different roles. Subjects read pairs of sentences that ended in either expected words, unexpected words from the expected semantic category (within-category semantic violation), or unexpected words from an unexpected category (between-category semantic violation). The critical sentence-final target words either appeared in the left or right visual field, which corresponds to the right and left hemispheres respectively. The results showed that within-category anomalous words presented to the right visual field (left hemisphere) elicited smaller N400s than between-category anomalous words, while both kinds of unexpected words elicited similar N400 effects in when presented to the left visual field (right hemisphere), suggesting that the brain is lateralized in the sense that the right hemisphere is more “integrative”, while the left is more “predictive”.
This is because the unexpected words that shared semantic features with the expected word (e.g. football vs. baseball in the previous example) elicited a smaller N400 than the unexpected words that belonged to a totally different word category (e.g. football vs. monopoly) in the left hemisphere, showing that the left brain has probably
activated the semantic features of a particular word due to anticipation. Wlotko and Federmeier (2007) conducted a similar experiment and replicated the findings.