Brown bag lunch (March 2013)

Implicit Learning of Verb Selectional Preferences

Albertyna Paciorek, John N. Williams, Department of Theoretical and Applied Linguistics

Keywords: Implicit learning; Second Language Acquisition; Vocabulary Acquisition

Most implicit learning research to date has focused on form-level regularities, e.g. in strings of letters or syllable strings. Some recent research has shown that regularities involving meaning can also be learned implicitly, e.g. the use of semantic information to guide visual attention, or learning arbitrary sequences of semantic categories. However, it is still not clear to what extent meaning can participate in implicit learning of natural language. On the one hand, it has been claimed that referential word meaning is essentially learned explicitly (Ellis, 1994). On the other, intuition suggests that grammatical and collocational aspects of words may be known and acquired implicitly (Paradis, 2004).

The present research examined implicit learning of selectional preferences of novel verbs using a false memory paradigm as an indirect test of learning. Selectional preferences are a word’s tendency to co-occur with words of certain semantic classes in a given role, e.g., drink is followed by words denoting LIQUID as direct object. In Experiment 1 participants first read sentences, some of which contained one of four novel verbs, and indicated whether the novel verb meant roughly ‘increase’ or ‘decrease’ (e.g. “Nightingale worked tirelessly to improve public health and POWTER the status of nurses” - POWTER means increase). Unbeknown to them, novel verbs regularly took either an abstract or a concrete noun as a complement (cf. MOUTEN the calcium). In a subsequent surprise test phase participants saw verb-noun pairs, and were asked whether these pairs had appeared in the training (e.g. POWTER the greatness). Although all of the words presented in the test had appeared in the training individually, most of them did not occur together. Crucially, half of the new combinations followed the pattern of selectional preferences from training, the other half violated it. We found that pairs that followed the pattern were erroneously judged as familiar significantly more often than those that did not. This was the case even for participants for whom no conscious knowledge of the regularity could be elicited in a post-test task. An analysis of confidence judgements confirmed that knowledge of the semantic regularity was affecting responses unconsciously. Experiment 2 replicated this laboratory experiment in an internet-based version using a large, and highly heterogeneous, participant sample. Experiment 3 showed that the false memory effect could still be obtained when the nouns in the test phase were semantically dissimilar to those that occurred with the nonwords in training, showing broad generalisation of the abstract/concrete regularity. Experiment 4 showed that explicitly instructed participants did not differ from implicit groups in their performance on the false-memory task, suggesting that even when participants have conscious knowledge they do not use it strategically when judging familiarity, confirming that false memory effects provide an indirect measure of learning.

We conclude that implicit learning of generalizable knowledge of selectional preferences is possible. Even if awareness is required for establishing initial form-meaning connections (in this case increase/decrease meanings), tuning of meaning through usage can occur unconsciously.

Click here to view the poster

Learning to read irregular and regular words: Combining artificial orthography learning with fMRI

Joanne Taylor, Kathleen Rastle, Matthew Davis, MRC-CBU

Cognitive models of reading proposed that item-specific knowledge is necessary for irregular word reading (SEW, YACHT). Whereas item-specific knowledge is lexical (whole-word forms) in the Dual Route Cascaded model (Coltheart et al., 2001), it is semantic in the triangle model (Plaut et al., 1996).  Neuroimaging contrasts are often confounded by difficulty, familiarity, and word meaning and, as such, have not discriminated between these possibilities. This fMRI experiment used an artificial language to avoid these confounds and delineate the neural systems supporting irregular and regular word learning.

Twenty-two adults learned to read 24 new words written in novel symbols, whilst in an MRI scanner. Some words were regular - all symbols had one pronunciation, some words were irregular - vowel symbols were pronounced differently in different words. Regular symbols occurred in 8 (high frequency) or 4 items (low frequency). Learning involved interleaved training (see word-hear pronunciation) and testing (read words aloud) phases, and was followed by generalization to untrained words.

Participants learned the trained words (regular high/low frequency-83%, irregular-73% correct) and generalized their knowledge to untrained words (70% correct). Activity in left occipitotemporal (visual-form processing) and parietal (spelling-sound processing) cortices was greater during irregular than regular word learning. However, this was also true for regular words containing low relative to high frequency symbols. In contrast, activity in left inferior frontal gyrus was greater for irregular words than either type of regular word.

When confounds of difficulty, familiarity, and meaning are removed, irregular words do not engage brain regions representing visual form more than regular words. Parietal regions involved in mapping from spelling-to-sound are engaged when these mappings are either infrequent or irregular. Inferior frontal cortices are involved in resolving the phonological conflict associated with the multiple competing pronunciations available for irregular words, as embodied in both the DRC and triangle model.

 Click here to view the poster

Neural and behavioural consolidation of spoken words and meanings in L1 and L2

Viktória Havas, MRC-CBU

Two experiments investigated the neural and cognitive processes underlying word learning and consolidation. In Experiment 1, 65 adults learned: i) novel words associated with pictures of everyday objects, ii) novel words associated with novel objects, iii) novel words without picture associations. Half the participants learned novel words that were phonologically similar to their native language (L1), the other half learned novel words from Hungarian, an unfamiliar language with very different phonology (L2). Within each group half the participants were trained in the morning and tested in the evening of the same day, and half trained in the evening and tested in the morning the next day, allowing an opportunity for overnight consolidation. Training involved 5 presentations of each novel word paired with associated pictures (sets ii and iii), or a blank screen (set i). Tests included old-new decision, to evaluate learning of the novel words’ phonological form and a 4-alternative forced choice task, to evaluate picture-word associative learning. In old-new decision, phonological forms were more accurately recognized when they had been paired with everyday objects, relative to novel objects or unpaired items. Overnight sleep increased old-new decision accuracy only for novel words with unfamiliar phonology (L2 group). Results show associated meanings benefit phonological learning, and that consolidation enhances episodic knowledge of phonologically dissimilar novel words. In Experiment 2 we examined the influence of these manipulations on fMRI responses in 22 participants. Implications for theories of medial temporal lobe contributions to associative memory and word learning will be discussed.

An fMRI study of semantic similarity between individual words

Francesca Carota1, Hamed Nili2, Nikolaus Kriegeskorte2, Friedemann Pulvermuller3; 1 Neurolex, University of Cambridge 2 MRC - Cognition and Brain Science Unit, 2Freie Universität Berlin

Previous neuroimaging studies on semantic processing have provided convergent evidence that word meaning processing engages a widely distributed “semantic” network of fronto-temporal, motor and parietal regions (Binder et al., 2009; Bookheimer, 2002; Pulvermueller & Fadiga, 2010), which support general lexical semantic and category-specific processes. However, the conventional, univariate approaches to fMRI data analysis followed by this work revealed regional average activations to sets of words and do not elucidate whether these areas are sensitive to the meaning of individual words and their similarity.

Here we used Representational Similarity Analysis (Kriegeskorte et al. 2008) to compare semantic similarity between individual words referring to actions and onjects, as measured by Latent Semantic Analysis and semantic ratings, and single-word-elicited fMRI patterns (96 words, 6 repetitions each) recorded from 23 subjects.

Results showed that fine-grained patterns of cortical activity were associated to the meanings of individual words within the above mentioned “semantic” brain regions.

Semantically similar action-related verbs elicited similar brain activity patterns in left precentral gyrus, supplementary motor area, and inferior frontal gyrus (LIFG). These brain regions also exhibited similar brain activity patterns for action verbs and nouns denoting tools. In temporal cortex, semantic similarity was mapped for both action- and object-related words, especially tool and food nouns.

Interestingly, single-word-meaning similarity, computed by Latent Semantic Analysis (LSA), correlated significantly with the neural similarity in left inferior frontal and precentral areas.<

These findings suggest that the informational content within left inferior frontal and motor regions reflects individual word meanings and their semantic similarity as predicted by latent semantic distances, but previously not captured by activation-based fMRI approaches.