Search

We conducted a case study on how unreliable and/or unrepresentative stimuli in psycholinguistics research may impact the generalizability of experimental findings. Using the domain of lexical ambiguity as a foil, we analyzed 2033 unique words (6481 tokens) from 214 studies. Specifically, we examined how often studies agreed on the ambiguity types assigned to a word (i.e., homonymy, polysemy, and monosemy), and how well the words represented the populations underlying each ambiguity type. We observed far from perfect agreement in terms of how words are assigned to ambiguity types. We also observed that coverage of the populations is relatively poor and biased, leading to the use of a narrower set of words and associated properties. This raises concerns about the degree to which prior theoretical claims have strong empirical support, and offers targeted directions to improve research practices that are relevant to a broad set of domains.

Regular polysemes are sets of ambiguous words that all share the same relationship between their meanings, such as CHICKEN and LOBSTER both referring to an animal or its meat. To probe how a context embedding model, here exemplified by BERT, represents regular polysemy, we analyzed whether its embeddings support answering sense analogy questions similar to “is the mapping be- tween CHICKEN (as an animal) and CHICKEN (as a meat) the same as that which maps between LOBSTER (as an animal) to LOBSTER (as a meat)?” We found that (1) the model was sensitive to the shared structure within a regularity type; (2) the shared structure varies across regularity types, potentially reflective of a “regularity continuum;” (3) some high-order latent structure may be shared across regularity types, suggestive of a similar la- tent structure across types; and (4) there is equivocal ev- idence that the aforementioned effects are explained by meaning overlap.

Why do readers typically fixate near the center of a word, with a bias towards word onset? Alhama, Siegelman, Frost, & Armstrong (2019) proposed an account based on (1) perceptual constraints that reduce the likelihood of perceiving a letter the further it is from the fixated location, and (2) the information available from the perceived letters for identifying the word. We expand this work to predict the fixation location distributions of 7-letter words for 43 languages from 9 language families. We found that in the majority of the languages, words were most likely to be correctly recognized when fixating near the center, slightly toward onset. However, there were deviations from this trend, such as predicting more accurate recognition when fixating slightly towards offset in Slavic languages. Our results provide novel predictions for experimental work that considers a flexible language system that optimizes initial fixations based on information theory constraints.

New analyses of pseudo-homophone RTs (e.g., BRANE) from two published lexical decision studies clarify lexical involvement in pseudo-homophone processing and challenge widespread assumptions about word frequency effects. First, RTs increased along with increases in the proportion of base- word letters that appeared in the pseudo-homophone (e.g., WHELT-WELT slower than PHAWT – FOUGHT) suggesting that “No” decision-making is slowed by mutually reinforcing activation in phonological and orthographic representations of base word knowledge. Second, effects of base-word frequency were either extremely weak or nonexistent among pseudo-homophones that contained most or all the letters that make up their base word. In contrast, among pseudo-homophones that shared fewer letters with their base word (e.g., “PHAWT”), RTs for items derived from high-frequency base words were faster than RTs for items derived from low-frequency base words. These findings (i) challenge the ubiquitous assumption that lexical representations are frequency sensitive and (ii) suggest that lexical decision involves a spell-check.

How do neural network models of quasiregular domains learnto represent knowledge that varies in its consistency withthe domain, and generalize this knowledge appropriately?Recent work focusing on spelling-to-sound correspondencesin English proposes that a graded “warping” mechanismdetermines the extent to which the pronunciation of a newlylearned word should generalize to its orthographic neighbors.We explored the micro-structure of this proposal by training anetwork to pronounce new made-up words that were consistentwith the dominant pronunciation (regulars), were comprisedof a completely unfamiliar pronunciation (exceptions), orwere consistent with a subordinate pronunciation in English(ambiguous). Crucially, by training the same spelling-to-soundmapping with either one or multiple items, we tested whethervariation in adjacent, within-item context made a givenpronunciation more able to generalize. This is exactly whatwe found. Context variability, therefore, appears to act as amodulator of the warping in quasiregular domains.

Lexical ambiguity--the phenomenon of a single word having multiple, distinguishable senses--is pervasive in language. Both the degree of ambiguity of a word (roughly, its number of senses) and the relatedness of those senses have been found to have widespread effects on language acquisition and processing. Recently, distributional approaches to semantics, in which a word's meaning is determined by its contexts, have led to successful research quantifying the degree of ambiguity, but these measures have not distinguished between the ambiguity of words with multiple related senses versus multiple unrelated meanings. In this work, we present the first assessment of whether distributional meaning representations can capture the ambiguity structure of a word, including both the number and relatedness of senses. On a very large sample of English words, we find that some, but not all, distributional semantic representations that we test exhibit detectable differences between sets of monosemes (unambiguous words; N = 964), polysemes (with multiple related senses; N = 4,096), and homonyms (with multiple unrelated senses; N = 355). Our findings begin to answer open questions from earlier work regarding whether distributional semantic representations of words, which successfully capture various semantic relationships, also reflect fine-grained aspects of meaning structure that influence human behavior. Our findings emphasize the importance of measuring whether proposed lexical representations capture such distinctions: In addition to standard benchmarks that test the similarity structure of distributional semantic models, we need to also consider whether they have cognitively plausible ambiguity structure.

Most statistical learning studies focus on the learning oftransitional probabilities between adjacent elements in asequence, however, other statistical regularities may un-derpin different aspects of processing language and regu-larities in other domains. Here, we investigate how con-junctive statistical regularities (of the form A and B to-gether predict C) can be learned, and how this learningis impacted by similarity in representations analogousto that in unambiguous words, homonyms with mul-tiple unrelated meanings, and polysemes with multiplerelated meanings. We observed that provided the stimu-lus structure is relatively simple, participants are readilyable to learn conjunctive probabilities and display sen-sitivity to relatedness among representations. These re-sults open new theoretical possibilities for exploring thedomain-generality of how the learning and processingsystems merge conjunctive information in simple labo-ratory tasks and in natural language.

Proficient readers typically fixate near the center of a word,with a slight bias towards word onset. We explore a novelaccount of this phenomenon based on combining information-theory with perceptual constraints in a connectionist model ofvisual word recognition. This account posits that the amountof information-content available for word identification variesacross fixation locations and across languages. These differ-ences contribute to the overall fixation location bias in differ-ent languages, make the novel prediction that certain wordsare more readily identified when fixating at an atypical fixa-tion location, and predict specific cross-linguistic differences.We tested these predictions across several simulations in En-glish and Hebrew, and in a behavioral experiment. The resultsconfirmed that the bias to fixate closer to word onset alignswith reducing uncertainty in the visual signal, that some wordsare more readily identified at atypical fixation locations, andthat these effects vary across languages.

The diversity of a word’s contexts affects its acquisition andprocessing. Can differences between word types such asmonosemes (unambiguous words), polysemes (multiple relatedsenses), and homonyms (multiple unrelated meanings) be re-lated to distributional properties of these words? We tested fortraces of number and relatedness of meaning in vector repre-sentations by comparing the distance between words of eachtype and vector representations of various “contexts”: their dic-tionary definitions (an extreme disambiguating context), theiruse in film subtitles (a natural context), and their semanticneighbours in vector space (a vector-space-internal context).Whereas dictionary definitions reveal a three-way split betweenour word types, the other two contexts produced a two-way splitbetween ambiguous and unambiguous words. These inconsis-tencies align with some discrepancies in behavioural studiesand present a paradox regarding how models learn meaningrelatedness despite natural contexts seemingly lacking suchrelatedness. We argue that viewing ambiguity as a continuumcould resolve many of these issues.

Are different amounts of semantic processing associated withdifferent semantic ambiguity effects? Could this explain somediscrepant ambiguity effects observed between and acrosstasks? Armstrong and Plaut (2016) provided an initial set ofneural network simulations indicating this is indeed the case.However, their empirical findings using a lexical decisiontask were not clear-cut. Here, we use improved methods andfive different experimental manipulations to slow responding---and the presumed amount of semantic processing---toevaluate their account more rigorously. We also expandedthe empirical horizon to another language: Spanish. Theresults are partially consistent with the predictions of theneural network and differ in several important ways fromEnglish data. Potential causes of these discrepancies arediscussed in relation to theories of ambiguity resolution andcross-linguistic differences.