Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams
Replicating the Ambiguity Advantage in MEG Using a Novel Method for Quantifying Number of Senses
Poster B71 in Poster Session B and Reception, Thursday, October 6, 6:30 - 8:30 pm EDT, Millennium Hall
Kyra Wilson1, Alec Marantz1,2; 1New York University Abu Dhabi, 2New York University
[INTRODUCTION] Previous lexical decision studies have reported an “ambiguity advantage” whereby words with multiple meanings and/or senses are recognized faster than unambiguous words (Rubenstein, Garfield, Millikan 1970; others). Further studies have argued that this effect is not homogenous–rather, multiple senses are facilatory and multiple meanings are potentially inhibitory (Rodd, Gaskell, Marslen-Wilson 2002; Beretta, Fiorentino, Poeppel 2005; others). However, one limitation of these studies is the method of deriving the number of senses and meanings; they are generally based on human-generated resources with little empirical motivation. In this study, we quantify words’ number of senses using unsupervised methods to reduce human bias. Additionally, we verify that this measure replicates the “ambiguity advantage” found previously and is a better predictor of brain activity than traditionally used variables. [METHODS] During a visual lexical decision task, we recorded brain activity from 21 adults using magnetoencephalography (MEG). The stimuli were 631 monomorphemic, monosyllabic, noun-verb ambiguous English words having a lexical decision accuracy > 80% plus 631 matched non-words. [NUMBER OF SENSES MEASURE] We derive our number of senses using BERT, which learns context-sensitive word representations. For each of our stimuli, we sampled 1000 sentences (or all sentences if there were fewer than 1000 occurrences) from Wikipedia, and retrieved the contextualized embedding of that word in each context. Next, we applied a hierarchical clustering algorithm to the resulting embeddings to estimate the number of senses (clusters) each word has. As a comparison, we also retrieved the number of senses for a given word from WordNet and the Wordsmyth online dictionary. Correlations among all predictors were below r=0.15, except WordNet and log frequency (r=0.52). [RESULTS] We tested three models for comparison in an ROI containing left middle and superior temporal gyri. Test statistics were computed over activation levels (dSPM) at each time point, and cluster permutation tests were conducted to identify significant clusters. Test statistics were t-values resulting from one-tailed t-tests over beta coefficients resulting from regressions at each time point and source per subject (dSPM ~ log(Lexical Frequency) + Number of Senses, 300-400ms). In the Wordsmyth model, frequency neared significance in LMTG (p=0.057, 300-400ms) and there was no effect of number of senses. In the WordNet model, frequency was facilitatory in LMTG (p=0.048, 300-400ms) and there was no effect of number of senses. In the BERT model, frequency neared significance in LMTG (p=0.081, 300-400ms) and number of senses was facilitatory in LMTG (p=0.038, 300-400ms). [CONCLUSION] Our computationally derived measure of number of senses is the only significant predictor of our brain data, outperforming more traditional measures. This suggests that cognitively valid estimations of senses can be acquired in an unsupervised manner without resorting to human-annotation or curation. Additionally, we replicate the “ambiguity advantage”, finding that words with more senses elicit less brain activity. Before SNL, we will derive a continuous measure quantifying words’ sense relatedness in order to further probe the potential differing effects of meanings and senses on word recognition.
Topic Areas: Meaning: Lexical Semantics, Computational Approaches