Projects | Roberto Petrosino

EEG/MEG response comparability

Fri, 30 Apr 2021 00:00:00 +0000

Over the past few decades, a bulk of psycholinguistic evidence have shown that words decompose in morphemes when they are recognized (morphological decomposition). A series of magnetoencephalographic (MEG) studies have recently suggested that the brain response to morphological properties originates from a region of the inferior temporal cortex known as `Visual Word Fom Area’ (VWFA; Cohen et al. 2000, Cohen et al. 2002) and peaks at about 150 ms post stimulus onset (the M170 component). This area of the brain has been shown to be sensitive to the morpho-orthographic statistical regularities at play during the early stages of the processing of non-affixed (heteronyms: e.g., wind; Solomyak & Marantz 2009), affixed (e.g., bi-morphemic words involving free and bound stems: e.g. taxable, tolerable; Solomyak & Marantz 2010), and pseudo-affixed (e.g., brother; Lewis et al. 2011) words, but not to semantic properties, which seem to be indexed by a response in the superior temporal and Silvyan Fissure region peaking at about 350 ms (the M350 component; Pylkännen & Marantz 2003). These studies have developed a novel approach to MEG data analysis, which is based on the correlational data analysis approach pioneered by Hauk et al. (2006) for ERP sensor-space data analysis. While traditional data analysis pipelines group together stimuli sharing some properties in different conditions and compare them to one another, this new approach correlates a given variable (e.g., orthographic transitional probability) of each stimulus with the brain response arising at a given time window. This kind of technique is particularly beneficial for investigations on morphological phenomena, which often consists of a small number of stimuli, which would be impossible to control for all the necessary variables. Unfortunately, not many morphologists that do not have access to highly expensive MEG machines, as there are, indeed, only a few available worldwide.

This project explores the possibility of applying the MEG correlational technique just mentioned to high-density EEG data as a way to provide a reliable and a cost-effective methodology for morphologists interested in conducting electrophysiological research. So far, we have collected high-density EEG data for an identical replication of Lewis et al. (2011). In this study, the M170 response to both transparent (e.g., driver) and opaque (e.g., brother) words is shown to correlate with the morpho-orthographic transitional probability between pseudo-stem (broth) and pseudo-affix (er), which supports the morpho-orthographic nature of early visual decomposition. We are now in the process of defining the EEG-adapted pipeline and comparing the two sets of results to one another. The hope is to obtain an agreed-on pipeline for correlational data analysis, which may guarantee comparability across types of signal (i.e., MEG vs. EEG) and boost electrophysiological research in morphology.

The nature of phonological, morpho-phonological and morphological alternations

Thu, 25 Mar 2021 00:00:00 +0000

This strand of my program explores the representational and the computational relationships among phonological, morpho-phonological, and morphological alternations. Morphological and phonological theories have primarily focused on two main questions about alternations.

1. What the representational status of alternations? Derivation vs. representation; computation vs. storage

The first question concerns the definition of the grammatical procedure under which a phonological or morphological representation undergoes a given set of changes, before surfacing as the corresponding phonetic representation. One path of research seeks to define the best procedural approach to account for purely phonological changes (in which both the target and the trigger are phonological in nature). Derivational theories have claimed that phonological alternations result from application of sequential rules applying at different derivational steps (e.g., Chomsky & Halle, 1968). Representational theories have claimed instead that alternations result from the selection of the optimal candidate via evaluation against a language-specific constraint ranking (e.g., Prince & Smolensky, 1993/2004). On this issue, Petrosino & Calabrese (2022) present an overview of all the accounts proposed for Romance palatalization, and ultimately show that a derivational framework can account for palatalization in a more articulatory-grounded and less stipulative way than a representational framework.

A second path of research seeks to identify an agreed-upon heuristic to best account for the different kinds of alternations sketched above. In general, all alternations involve two segments: the target (i.e., the segment undergoing the change) and the trigger (i.e., the segment causing the change). The nature of the two segments may be either phonological (i.e., involving phonological information) or morphological (i.e., involving morphological information); this taxonomy logically allows for four possible target-trigger combinations, which can all be identified in human languages. A type of alternation that has been at the center of a vivid debate involves a phonological target conditioned by a morphological, or a morpho-phonological trigger. An example of this type of alternation is the masculine forms of the Standard Italian definite determiner: it realizes as _il [il] _(singular) ~ i [i] (plural) before a noun beginning with a simple consonant (e.g., il [k]ane ‘the dog’ ~ i [k]ani ‘the dogs’), lo [lo] (singular) ~ gli [λi] (plural) before a noun beginning with a complex consonant (e.g., lo [st]ato ‘the state’ ~ gli [st]ati ‘the states’). I have argued that the forms of Italo-Romance determiners (Petrosino 2018) can be better accounted for in a derivational account, which, unlike any representational account, is able to catch language-wide generalizations via application of morpho-phonologically conditioned rules.

Currently, the processing of phonological and morphological alternations is relatively understudied in word processing. I am exploring the extent to which processing time may be a function of the number of the derivational steps the parser needs to go through to retrieve to the correct (morpho-)phonological representation. In essence, this approach applies the theory of derivational complexity (DTC: Miller & Chomsky, 1963) to the (morpho-)phonological component of the grammar, under the assumption that any potential impact on processing time may be measurable in terms of reaction times (RTs; Phillips, 1996). If that is the case, derivational theories seemingly predict that processing time increases along with the number of steps required for the parser to recognize the phonological representation, thus slowing down recognition and ultimately increasing RTs.

2. What are the grammatical conditions of alternations?

The second question regarding the nature of alternations concerns the definition of the grammatical conditions under which alternations may arise. This topic has been at center of the debate within the framework of Distributed Morphology, in which word formation is seen in purely syntactic terms, as a result of cyclic movement of functional morpho-syntactic heads (Halle & Marantz, 1993). The most-accepted hypothesis maintains that alternations may arise only when the target and the trigger heads are syntactically adjacent to one another (node adjacency hypothesis: Embick, 2010). The node adjacency hypothesis has been recently challenged and replaced with a less restrictive one, whereby alternations may be triggered by a span of different morpho-syntactic heads (span adjacency hypothesis: Merchant, 2015). On this issue, Christopoulos & Petrosino (2018) provide striking evidence in support of the node adjacency hypothesis, in showing that roots alternate when the trigger node is adjacent to the root (e.g., ser-n-i ‘drag-IMPFV.ACT-3SG’ ~ sir-θ-i ‘drag-PFV.NACT-3SG’, where the Aspect/Voice head -θ- triggers ablaut on the root), but they don’t when an additional head intervenes in between (e.g., pal-ev-i ‘fight-v-IMPFV.ACT-3SG’ ~ pal-ef-θ-i ‘fight-v-PFV.NACT-3SG’, where the verbalizer -az- intervenes between the triggering Aspect/Voice head -θ- and the root).

Under similar theoretical assumptions, Calabrese & Petrosino (2023) discuss the morpho-syntactic behavior of the so-called “thematic vowel” (TV), the root-adjacent vowel that typifies Indo-European verbal morphology, across Sanskrit, Ancient Greek, and Latin. We take issue with the common trend that, in the wake of the seminal work by Aronoff (1994) on Latin verbal morphology, considers TVs cross-linguistically “ornamental” (i.e., lacking any morpho-syntactico-semantic information) by default. We show that, unlike Latin (and therefore Romance), Sanskrit and Ancient Greek TV are fully functional, and argue for a diachronic account whereby the original functionality of such pieces was gradually lost over time, and finally gave rise to the modern Romance verbal ornamental system.

The richness of phonological and morphological representations in the lexicon

Thu, 25 Mar 2021 00:00:00 +0000

This strand of my program explores the the amount of information associated with the phonological and morphological representations that are stored in the lexicon.

Phonological representations

Within phonology, though there is widespread agreement that phonemic representations are composed of articulatory features, there are open questions about exactly how rich these representations should be. One class of theories, underspecification theories, has proposed that phonological representations only contain the features that cannot be predicted by the phonology of the language (a.o., Steriade, 1995; Lahiri & Reetz, 2002). Another class of theories, full-specification theories, has proposed that phonological representations contain all features, but crucially, certain features have a special status relative to others, such as markedness (a.o., Jakobson & Halle, 1956; Kiparsky, 1982; Prince & Smolensky, 1993/2004). One interesting development in this debate was the observation that an automatic electrophysiological response, the Mismatch Negativity (MMN) appears to be sensitive to differences between morphemes that has been attributed to featural underspecification or markedness (Eulitz & Lahiri, 2004, et seq.). Teasing apart underspecification and markedness approaches has been difficult within phonological theory, because they typically make similar empirical predictions. But MMN is much broader than phonological theory, as it is a low-level brain response generated by the auditory cortex (see Näätänen, 2001 for a review), and I have been exploiting the domain-generality of MMN elicitation to look for effects of markedness in auditory processing in domains that are adjacent to language such as the processing of biological sounds with potential social consequences. We found that the MMN response to a socially marked, non-linguistic sound such as [flatulence] was different from the MMN response to socially unmarked, non-linguistics sounds such as [cough] and [sniff] (Petrosino et al., submitted). These results seem to support the idea that the MMN response can be modulated by markedness alone, as there is no possibility of featural underspecification for these sounds. This may give support to the markedness approach for phonemes as well, since otherwise we need to assume a disjointed theory of the MMN modulation – i.e., that it arises for both markedness and underspecification. These results also demonstrate the values of research topics at the intersection of linguistics and other areas of cognitive science (like auditory processing and social cognition). As a final note, I ought to mention that this study was concluded prior to the pandemic, so it is possible that these values have changed in the current post-pandemic era. We are currently following up on this issue at NYUAD, by replicating the same experiment with a brand-new dry EEG system (i.e., without the need for conductive gel to active the electrodes) with a data quality comparable with that of traditional (gel-based) systems..

Morphological representations

Within the domain of morphology, theories of morphology have made no explicit arguments regarding the richness of morphological representations. A wealth of evidence from the masked priming literature on morphological decomposition (i.e., the real-time procedure of decomposing words into morphemes) have shown similar facilitation effects for (a) identical (city-CITY), (b) morphologically transparent (driver-DRIVE) and (c) morphologically opaque (brother-BROTH) pairs, but no effect for (d) orthographically related (brothel-BROTH) and (e) semantically related (cello-VIOLIN) pairs (Rastle et al., 2000; 2004). These results led to models of processing involving a level of morphological analysis, which is independent of the orthographic and semantic levels, and is accessed before meaning is available, as it only contains representations endowed with “morpho-orthographic” properties (such as orthographic transitional probability; Rastle & Davis, 2008). Building on these findings, my dissertation (Petrosino, 2020) asks whether morphological decomposition may be affected by other abstract linguistic properties such as whole-word lexicality, whole-word frequency, and syntactic categorization morphemes. The results suggest that at early stages of processing, morphological representations do contain, along with orthographic properties, at least some of the abstract linguistic properties that typify current theories of morphology—in particular, whole-word lexicality and frequency. Building on these results, I am currently investigating two main questions.

1. Are stem and affixes processed similarly?

In the literature, there is a sheer dichotomy between the masked priming response to English stems (boneless-BONE), which is usually found comparable in size with the repetition priming response (bone-BONE), and the masked priming response to English affixes (boneless-ILLNESS), which is usually null, or close to null in size. I am currently investigating, however, the possibility that the reported stem/affix asymmetry in the masked priming response may be an confound of the factual asymmetry between affixes (which by definition cannot be elicited as words on their own) and stems (which can be words on their own, but only in a limited subset of the world languages, like English). The experiment therefore aims to elicit the masked priming response to roots and affixes by using bimorphemic word pairs for both conditions (e.g., loveable-LOVELESS, jogger-FREEZER).

2. What are the mechanisms of (masked) priming?

The masked priming design has been used as the primary tool for investigations on sublimninal word processing. However, its mechanisms are still unclear. I am currently undertaking a reconsideration of the masked priming response in the light of the fundamental, often glossed-over problem of statistical power, which may lead to unreliable results. To tackle this issue, large sample sizes (N > 200 participants per experiment) may be collected only thanks to the capabilities of modern online programs. In Petrosino & Almeida (submitted) we show that, with the right sample size (quantified through a full-fledged power analysis), word frequency may interact with the masked priming response, in contrast with what has been believed for the last 40 years (Forster & Davis, 1984). These results pose an overarching question about the mechanics of masked priming, and, more broadly, challenges the actual suitability of the design for investigations on early stages of word recognition, as it suggests that prime masking may be actually unable to screen from episodic strategies. Further confirmation for an urgent reconsideration of the mechanisms underpinning morphological processing (and word recognition in general) can also be found in Petrosino, Sprouse & Almeida (in prep.), who report an extensive meta-analysis of five different replications (conducted both in-lab and online) of the traditional morphological masked priming experiment first reported by the seminal works by Rastle et al. (2000; 2004; see also above). The results show that, with a large sample size, the effects for (c) are rather statistically closer to (d), than to (a, b), which suggests that the purported dissociation of the morphological level from the orthographic level may not be as clear-cut as previously thought.

Morphological decomposition

Mon, 01 Jan 0001 00:00:00 +0000

A wealth of psycholinguistic evidence has shown that words, before being visually recognized, decompose into smaller units which seem to correspond to morphemes (Rastle et al., 2000, 2004). Such a procedure of morphological decomposition seems to occur in words that are made of more than one morpheme, independently of whether they are semantically transparent (e.g., driver, which means “someone who drives”) or not (e.g., brother, which does not mean “someone who broths”). Decomposition, however, does not seem to occur in words that contain a root plus an additional, non-morphemic string (e.g., brothel, where el is not an English suffix). Current models of visual word processing assume the morpho-orthographic decomposition hypothesis, whereby decomposition obligatorily relies on islands of regularity – namely, statistical orthographic regularities across letter strings, so that a string such as er is detected as a single morpho-orthographic unit, but el is not (Rastle & Davis 2008). As such, decomposition is also argued to occur at early stages of word processing and before contact with the lexicon — namely, before semantic information is accessed, so that decomposition occurs regardless of semantic transparency. In my doctoral dissertation (Petrosino 2020), I built on these findings and used the visual masked priming response to test the sensitivity of decomposition to a few higher-level linguistic properties (in addition to islands of regularities) such as syllabification, whole-word frequency, phonological conditioning, and syntactic well-formedness. The results suggested that while not being affected by syllabification and syntactic well-formedness, decomposition is indeed sensitive to whole-word lexicality and whole-word frequency. This clear-cut divide between properties that may and may not impinge on decomposition might be due to the fact that both syntactic processing (i.e., the process whereby syntactic categories and affixal restrictions are retrieved and checked) and semantic processing (i.e., the process whereby whole-word meaning is accessed) require at least two mechanisms to occur: (i) recombination of decomposed morphemes (along the lines of Taft 2004), and, eventually, (ii) access to compatible syntactic and semantic representations. Given that these are fairly complex, these representations indeed take more time to access fully, which, therefore, may not occur until after decomposition. At the face value, the fact that whole-word lexicality and wor frequency may impinge on early decomposition procedures is surprising. These two properties are generally considered to be accessed at late stages of processing and their potential implications onto decomposition seems logically contradictory, under the current theories of early decomposition: how can an early process be affected by properties that concern the whole stimulus and, as such, are accessed only at later stages? A model of early visual decomposition is proposed that is able to account for these findings without necessarily challenging the morpho-orthographic decomposition hypothesis. In this model, decomposition accesses whole-word lexicality and frequency through a multi-step mechanism that first generates multiple possible morpho-orthographic decomposition patterns of the visual stimulus and then evaluates them in parallel in order to choose the optimal candidate for activation.

The investigation above has inspired a number of follow-up studies on morphological decomposition. Click on the links below for more information!