• No results found

A. Simple lexical decision

6. Discussion

The results support all three of the hypotheses proposed above. Nouns that appear in more syntactic contexts are recognized faster (categorical hypothesis); nouns that are distributed more uniformly across these contexts are recognized faster (probabilistic hypothesis); and more prototypical nouns were recognized faster (prototypicality hypothesis). As predicted by the neurophysiological findings of Linzen et al. (2013), syntactic diversity and

prototypicality showed independent, additive effects. However, unlike Linzen and colleagues, both types of effect were observed for RTs. This finding suggests that the

Figure 7: Significant effect of prototypical component. Left panel: Component loadings of prototypicality component 1. This component reflects general distance from

the prototype. Distances in the modifier and particularly the rightward modifier distributions are prioritized. Right panel: Effect of probabilistic component 6 on

response times.

Several novel effects were also observed. First, syntactic diversity breaks down into additive effects of categorical and probabilistic diversity. Second, the diversity and prototypicality effects depend most heavily on rightward modifier dependencies.

This pattern of findings is inconsistent with theories that posit only categorical syntactic representations in the lexicon. In these theories, words either are or are not licensed in a particular structure (e.g., Chomsky, 1995). These theories could account for the categorical diversity effect observed here. For example, words that activate more syntactic categories are processed faster, perhaps through a feedback mechanism. However, if this account were

correct, we should not have seen effects from probability and prototypicality (both of which depend on frequency distributions). But we did see such effects, indicating that these theories are incomplete and underpredictive. Looking closer, we see that many of these theories are also unable to account for the modifier-driven diversity effect. Often, categorical theories only mark words for the structures that they may head (e.g., Bresnan, 2001;

Chomsky, 1995). But the model revealed that the as-modifier diversities contributed the most to the categorical effect (Figure 5, left panel).

These findings also differ from those observed by Linzen et al. (2013) for verbs. In that study, they found no effect of either diversity or prototypicality on RTs. Other work on nouns has reported a syntactic effect; but the measures used there were actually based on lexical variation (Baayen et al., 2011). Lexical variation is known to reflect semantics (Bullinaria & Levy, 2012). Therefore, it was possible that the findings for nouns were tainted by semantics. Linzen and colleagues were the first to use fully abstract, cross-structural syntactic diversity. Therefore, it was possible that similarly abstract measures applied to nouns would likewise show no correlation with RTs. However, the opposite was true: syntactic distributions impact the processing of isolated words. This discrepancy could stem from at least three differences between the study of Linzen and colleagues and this one. First, they based their measures on phrase-structural subcategorization frames, whereas as I based mine on binary dependencies. From a construction-grammar perspective, both levels should be involved simultaneously: an argument-structure construction embodies the entire argument configuration, as well as the lower-level constructions that fill out the individual arguments (Goldberg, 1995; Langacker, 1987). The dependencies studied here correspond to those lower-level

relationships. Perhaps these are more intimately tied to word processing given that they mark the point of entry for words into constructional frames. Second, I carefully corrected the entropy estimates to avoid underestimation biases. Noise attributable to biased estimates could have interfered with the model estimates for Linzen and colleagues. Third, regarding diversity, I made sure to remove all lexical information from the entropy estimates. Linzen and colleagues did not control for lexical information. This lexical information could interfere with the estimate, again obscuring the effect.

Several theories can explain all three effects. Usage-based construction grammar

(UBCG; Diessel, 2015; see also Goldberg, 2006) proposes that language is best modeled as a complex network of interactions between linguistic units at all levels of abstraction.

Categorical specifications on word forms are replaced by arcs between word-level and syntax-level nodes. Statistical information derived from the input tunes the strength of these connections, thus accounting for the probabilistic diversity effect. When these connections and connection strengths match up with the expectations of the system, words are recognized more quickly. System expectations can be modeled in several ways. For example, patterns of resting activation may develop over time based on the average behavior of the system. When a noun deviates from this pattern, it will not benefit as much from the activation that is fed back from the syntactic system. Alternatively, prototypical nouns, which better signal their membership to the noun category, may provide more compelling evidence to the system responsible for making the two-way lexical decision judgment. This could surface as input into the drift space between two choices (e.g., Ratcliff et al., 2004) or as an influence on prior probabilities of encountering a noun with such-and-such syntactic behavior (Norris,

2006; see Linzen et al., 2013, for a similar suggestion).

The discriminative-learning model of Baayen et al. (2011) could also account for these findings. According to this model, variability of contextual cues over time helps to solidify the bond between a target form and its meaning. Syntactic dependencies constitute one form of contextual cue. These cues are also thought to be paradigmatically bound, such that the information carried by a paradigm can influence discriminability of the connection between form and meaning. For example, the unconditioned distribution of prepositions (the

paradigm) in English prepositional phrases represents the generalized potential for a

preposition to be followed shortly by a noun. Nouns that function as objects to prepositions in proportion to the overall distribution of those prepositions will be more likely to surface when a preposition has been deployed (e.g., inversely proportional nouns would load too heavily on uncommon preposition types). Therefore, they stand to benefit the most from the information carried by the 'prepositional paradigm,' where that information comes in the form of structured contextual variability. The variability produces stronger, more stable connections between form and meaning, leading to more efficient reading. This, they argue, is the source of the typicality effect they observe.

The primary evidence for discriminative learning comes from measures based on overt cues – cues that are directly available in the input, for example, inflectional endings or prepositions. These types of measure bias the results in favor of the model. This is because the model developed by Baayen and colleagues contains only two layers: an input layer for letter n-grams (usually 2- or 3-grams) and an output layer for meaning. Therefore, they are naturally capable of modeling lexical variation, but ostensibly incapable of modeling

variation for abstract categories. However, the measures considered here do not make any direct reference to lexical context. They are based solely on the syntactic dependencies that attach to the target noun. Information about the word to which the target is bound was explicitly removed. Therefore, these results constitute a challenge to the purely syntagmatic discriminative learning approach. Specifically, it appears that hierarchical, non-overt aspects of the contexts in which nouns appear also affect how well they are learned. We should therefore expect learning-through-discrimination to involve a multidimensional network of cues, including cues directly associated with the surface code (words built from graphs, phones, or signs) and higher-order, more abstract cues that emerge over time (e.g., Bybee, 2010; Diessel, 2015; Goldberg, 1995).

The contrast between rightward and leftward modifer distributions deserves further comment. Diverse and distinctively rightward modifiers were recognized more slowly, while diverse leftward modifiers were recognized faster. Why would diversity help in one context and hinder in another? Consider the nature of rightward modification. Nouns that modify words to their right participate in a head-final dependency. However, English has dominant head-initial word order, at least outside of the noun phrase (NP). Notice that NP-external relations of this sort are precisely where modifier relationships apply for nouns. Therefore, nouns with negative scores on this component fight against the typological orientation of English nouns as modifiers. Typological constraints like this should leave other traces. For example, they should affect word frequency. One would expect to find more words with low conditional entropies in the dispreferred dimension. Hence, the probability density function for rightward noun-as-modifier relations should be bunched up around 0. The density

function for leftward noun-as-modifier relations (head-initial ordering) should be spread more evenly across the range of entropy values. Figure 8 plots the probability density functions for the rightward and leftward variants of the noun-as-head and noun-as-modifier conditional entropies.

Figure 8 supports this intuition. Words are clustered clustered below H(D | L) = 0.5 for the typologically dispreferred dimension, rightward diversity for nouns-as-modifiers (shown in orange). The greatest peak in density centered on 0. By contrast, the typologically

preferred dimension –leftward diversity for nouns-as-modifiers – has a wide distribution, with higher rates of occurrence in the upper ranges of conditional entropy (up to ~1.5, a full bit higher than that observed for rightward modifier diversity).

Figure 8: Probability density functions for the conditional entropies of nouns. Different colors reflect different syntactic dimensions.

Typology should also relate to prototypicality: nouns that are diverse rightward modifiers should be atypical, hence less like the other nouns of the language. By definition, the

majority of other nouns would follow the dominant head-initial preference (a point

supported by the curves in Figure 8). Therefore, the processing disadvantage associated with rightward modifiership might actually be due to a typological prototype favoring head-initial structures. This intuition is supported by the prototypicality effect observed here. The general prototypicality effect was most strongly driven by rightward modifiership. Distance from the noun prototype was associated with longer RTs. The common thread is that distributions in the rightward modifiership space exert the strongest effect. Following the logic of

discriminative learning, nouns of this type would not have the same general opportunity to occur and hence would not receive the same discriminative benefit as words that fit the overall trend. This explanation directly links syntactic typology to the local word processing. Such a link clears the path for new predictions regarding the behavior of typologically distinct languages. For example, we should observe opposite effects in strongly head-final languages, such as Japanese: the processing disadvantage should emerge for leftward-facing noun-as-modifier diversity.

The primary take-away from this study is that reading a noun in isolation invokes the syntactic history of that word. While similar results have been observed before, this study is the first to demonstrate that purely syntactic distributions impact lexical decision RTs. These data falsify any theory that limits syntactic representation in the lexicon (e.g., Borer, 2005; Chomsky, 1995; Marantz, 1997; Ramchand, 2008). These data also provide converging

support for the notion that syntactic information obligatorily impacts lexical access,

regardless of task or whether the word is processed in isolation (e.g., Cubelli et al., 2005; de Simone & Collina, 2015; Lester & Moscoso del Prado Martín, 2016; Linzen et al., 2013). Even when the task requires no syntax, syntactic information impacts RTs. Finally, they underscore the need to decompose lexical frequency well beyond the typical type/token counts (Baayen, 2010; cf. Bybee, 2010). Each instance of a word is embedded in a multidimensional network of cues. Which cues are important to which tasks, and how frequency relates to these cues, are questions that may reveal important information about the processing mechanism.

One unexpected finding of the present study was the typological contrast in diversity effects. Words that match the typological properties of language are recognized more quickly than those that go against the grain. This question deserves further study. One possible extension would be to compare the size of these typological congruence effects across languages. I expect languages that have cross-linguistically less preferred orders in a given domain to benefit less from congruence than languages with more preferred word orders, irrespective of whether the system is consistent within those languages. Such studies would help to solidify the links between linguistic representation, processing, and typology.