Languages differ in what meanings their lexical items encode: The meaning covered by English 'blue' is famously split into 'sinij' (darkblue) and 'goluboj' (lightblue) in Russian. Recent years have seen novel interest in functional explanations of such variation, pointing to a correlation between greater communicative need of a lexical field and a finer-grained lexical inventory. Here, I develop the position that rather than the mere difference in “need” to mention lexical field, it is the field's discourse-pragmatic diversity that predicts whether languages “lump” or “split” more. I will demonstrate this with computational techniques and a typologically diverse corpus of spontaneous spoken data from 51 languages (DoReCo), first for the field of verbs of visual perception ('see'-'look'), then on a lexicon-wide level. There are implications: our notions of what a comparable concept is in lexical semantics, what lexical knowledge entails, and the dimensions along which languages differ require re-examining.

Complement coercion interpretation is typically operationalized in terms of verb paraphrases (she finished the book > she finished reading the book), reflecting the construal of a specific event triggered by the direct object (cf. Pustejovsky & Bouillon). However, when given the option, participants frequently, and selectively, refrain from providing verb paraphrases for naturally occurring sentences. We present a corpus study of the pragmatic factors that affect the rate of refraining from giving a verb paraphrases, followed by two experiments. First, a fill-in-the-blank experiment shows that minimally different clause types containing complement coercion yield different rates of verb paraphrase. Second, we pair complement coercion sentences with possible verb paraphrases and obtain semantic (entailment) judgements from participants. Taken together, the results present a puzzle; participants’ likelihood to produce a verbal coercion interpretation is modulated by contextual factors, while their semantic comprehension of potential verb paraphrases for the same sentences is not.

We model two patterns related to the acquisition of color termsin Russian and English: children produce overextension errorsfor some colors but not others, and language-specific distinc-tions affect color discrimination in a non-linguistic task. Botheffects, as well as a reasonable convergence with adult linguis-tic behavior, are shown by a Self-Organizing Map trained onnaturalistic input. We investigate the effect of different waysof representing colors, i.e., as perceptual features or in terms ofthe cognitive biases on categorization extracted from crosslin-guistic color naming data. We also consider the influence ofcolor term frequency. Our results suggest effects of all three ofterm frequency, cognitive biases, and perceptual features.

In semantic typology, it is desirable to have quick and easy access to crosslinguistic elicitations describing stimuli from a semantic domain. We explore the use of crowdsourcing for obtaining such data, and compare it with fieldwork data obtained through in-person elicitations. Despite potential concerns about the quality of crowdsourced data, we find no difference in the amount of between-language variation and can replicate a cognitive modeling experiment using the crowdsourced data in place of the fieldwork data. Both results suggest that crowdsourcing elicitations is a viable method for gathering data for semantic typology and cognitive modeling

Noun compounding is a productive word formation strategy in English. Low-frequency (hence: ‘novel’) noun compounds (NCs) like “animal teacher” elicit diverse responses when presented without a sentence context, suggesting that they are semantically underspecified. We explore exactly how much the role of context changes default interpretations of novel English NCs. We consider compounds with two plausible interpretations: OF-relations (teacher OF animals) and IS-relations (animal who IS a teacher). We predicted context would increase participants’ ratings of the plausibility of given compound interpretations. Corroborating our predictions, participant plausibility ratings showed OF-relations were more accessible out-of-context, and IS-relations were rated with less certainty. Mixed-effects modelling showed a significant interaction between context and relation type. The results suggest context increases interpretation certainty, particularly in OF-dominant compounds. However, several groups of cases notably defy this pattern. Finally, we use computational language models to identify aspects of context that are critical to novel NC interpretation.

We show, contrary to some recent claims in the literature, thatprototype distributional semantic models (DSMs) are capa-ble of representing multiple senses of ambiguous words, in-cluding infrequent meanings. We propose that word2vec con-tains a natural, model-internal way of operationalizing the dis-ambiguation process by leveraging the two sets of represen-tations word2vec learns, instead of just one as most workon this model does. We evaluate our approach on artifi-cial language simulations where other prototype DSMs havebeen shown to fail. We furthermore assess whether these re-sults scale to the disambiguation of naturalistic corpus exam-ples. We do so by replacing all instances of sampled pairsof words in a corpus with pseudo-homonym tokens, and test-ing whether models, after being trained on one half of the cor-pus, were able to disambiguate pseudo-homonyms on the ba-sis of their linguistic contexts in the second half of the cor-pus. We observe that word2vec well surpasses the baselineof always guessing the most frequent meaning to be the rightone. Moreover, it degrades gracefully: As words are moreunbalanced, the baseline is higher, and it is harder to surpassit; nonetheless, Word2vec succeeds at surpassing the baseline,even for pseudo-homonyms whose most frequent meaning ismuch more frequent than the other.

We present a novel statistical analysis of color categorizationusing a standard method from semantic typology. Our ap-proach shows that crosslinguistic color naming data exhibitslatent dimensions whose order of relative importance matchesthe evolutionary ordering of emergence of those distinctions.Moreover, we show that the importance ordering of these di-mensions holds even when controlling for frequency of the dis-tinctions by looking at languages within each stage of evolu-tion. Additionally, we find that the extreme points of the latentcolor dimensions correspond well to a small set of “univer-sal” focal colors. Thus we show that a simple mathematicalmethod simultaneously derives a consistent match both to theevolutionary stages and to the universal foci.

A key challenge for children in language acquisition is to learnthe mapping of words to mental categories, since this mappingvaries greatly from language to language. The errors childrenmake in this process are very informative regarding the devel-opment of lexical semantic categories; in particular, how chil-dren overextend a word to an inappropriate exemplar providesa window onto the mechanisms that underlie their categoriza-tion processes. We perform a large-scale quantitative analysisof the detailed patterns of children’s errors in the domain ofcolor, finding evidence that these error patterns are driven byan interaction between domain general principles of catego-rization, and children’s developing knowledge of the seman-tics of color. Our results suggest that, while domain generalprocesses play a role throughout development, their influencevaries across ages according to their use of domain specific(conceptual) knowledge, which gradually increases over time.

Distributional semantic models (DSMs) are substantially var-ied in the types of semantic similarity that they output. Despitethis high variance, the different types of similarity are oftenconflated as a monolithic concept in models of behaviouraldata. We apply the insight that word2vec’s representationscan be used for capturing both paradigmatic similarity (sub-stitutability) and syntagmatic similarity (co-occurrence) to twosets of experimental findings (semantic priming and the effectof semantic neighbourhood density) that have previously beenmodeled with monolithic conceptions of DSM-based seman-tic similarity. Using paradigmatic and syntagmatic similaritybased on word2vec, we show that for some tasks and typesof items the two types of similarity play complementary ex-planatory roles, whereas for others, only syntagmatic similar-ity seems to matter. These findings remind us that it is im-portant to develop more precise accounts of what we believeour DSMs represent, and provide us with novel perspectiveson established behavioural patterns.