Spicy Adjectives and Nominal Donkeys: Capturing Semantic Deviance Using Compositionality in Distributional Spaces

Sophisticated senator and legislative onion. Whether or not you have ever heard of these things, we all have some intuition that one of them makes much less sense than the other. In this paper, we introduce a large dataset of human judgments about novel adjective‐noun phrases. We use these data to test an approach to semantic deviance based on phrase representations derived with compositional distributional semantic methods, that is, methods that derive word meanings from contextual information, and approximate phrase meanings by combining word meanings. We present several simple measures extracted from distributional representations of words and phrases, and we show that they have a significant impact on predicting the acceptability of novel adjective‐noun phrases even when a number of alternative measures classically employed in studies of compound processing and bigram plausibility are taken into account. Our results show that the extent to which an attributive adjective alters the distributional representation of the noun is the most significant factor in modeling the distinction between acceptable and deviant phrases. Our study extends current applications of compositional distributional semantic methods to linguistically and cognitively interesting problems, and it offers a new, quantitatively precise approach to the challenge of predicting when humans will find novel linguistic expressions acceptable and when they will not.     

Visual and Affective Multimodal Models of Word Meaning in Language and Mind

One of the main limitations of natural language-based approaches to meaning is that they do not incorporate multimodal representations the way humans do. In this study, we evaluate how well different kinds of models account for people's representations of both concrete and abstract concepts. The models we compare include unimodal distributional linguistic models as well as multimodal models which combine linguistic with perceptual or affective information. There are two types of linguistic models: those based on text corpora and those derived from word association data. We present two new studies and a reanalysis of a series of previous studies. The studies demonstrate that both visual and affective multimodal models better capture behavior that reflects human representations than unimodal linguistic models. The size of the multimodal advantage depends on the nature of semantic representations involved, and it is especially pronounced for basic-level concepts that belong to the same superordinate category. Additional visual and affective features improve the accuracy of linguistic models based on text corpora more than those based on word associations; this suggests systematic qualitative differences between what information is encoded in natural language versus what information is reflected in word associations. Altogether, our work presents new evidence that multimodal information is important for capturing both abstract and concrete words and that fully representing word meaning requires more than purely linguistic information. Implications for both embodied and distributional views of semantic representation are discussed.     

Semantic Memory Search and Retrieval in a Novel Cooperative Word Game: A Comparison of Associative and Distributional Semantic Models

Considerable work during the past two decades has focused on modeling the structure of semantic memory, although the performance of these models in complex and unconstrained semantic tasks remains relatively understudied. We introduce a two-player cooperative word game, Connector (based on the boardgame Codenames), and investigate whether similarity metrics derived from two large databases of human free association norms, the University of South Florida norms and the Small World of Words norms, and two distributional semantic models based on large language corpora (word2vec and GloVe) predict performance in this game. Participant dyads were presented with 20-item word boards with word pairs of varying relatedness. The speaker received a word pair from the board (e.g., exam-algebra) and generated a one-word semantic clue (e.g., math), which was used by the guesser to identify the word pair on the board across three attempts. Response times to generate the clue, as well as accuracy and latencies for the guessed word pair, were strongly predicted by the cosine similarity between word pairs and clues in random walk-based associative models, and to a lesser degree by the distributional models, suggesting that conceptual representations activated during free association were better able to capture search and retrieval processes in the game. Further, the speaker adjusted subsequent clues based on the first attempt by the guesser, who in turn benefited from the adjustment in clues, suggesting a cooperative influence in the game that was effectively captured by both associative and distributional models. These results indicate that both associative and distributional models can capture relatively unconstrained search processes in a cooperative game setting, and Connector is particularly suited to examine communication and semantic search processes.     

The distributional structure of grammatical categories in speech to young children

We present a series of three analyses of young children's linguistic input to determine the distributional information it could plausibly offer to the process of grammatical category learning. Each analysis was conducted on four separate corpora from the CHILDES database (MacWhinney, 2000) of speech directed to children under 2;5. We showthat, in accord with other findings, a distributional analysis which categorizeswords based on their co‐occurrence patterns with surroundingwords successfully categorizes the majority of nouns and verbs. In Analyses 2 and 3, we attempt to make our analyses more closely relevant to natural language acquisition by adopting more realistic assumptions about howyoung children represent their input. In Analysis 2, we limit the distributional context by imposing phrase structure boundaries, and find that categorization improves even beyond that obtained from less limited contexts. In Analysis 3, we reduce the representation of input elements which young children might not fully process and we find that categorization is not adversely affected: Although noun categorization is worse than in Analyses 1 and 2, it is still good; and verb categorization actually improves. Overall, successful categorization of nouns and verbs is maintained across all analyses. These results provide promising support for theories of grammatical category formation involving distributional analysis, as long as these analyses are combined with appropriate assumptions about the child learner's computational biases and capabilities.     

Probing Lexical Ambiguity: Word Vectors Encode Number and Relatedness of Senses

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.     

The differential role of phonological and distributional cues in grammatical categorisation

Recognising the grammatical categories of words is a necessary skill for the acquisition of syntax and for on-line sentence processing. The syntactic and semantic context of the word contribute as cues for grammatical category assignment, but phonological cues, too, have been implicated as important sources of information. The value of phonological and distributional cues has not, with very few exceptions, been empirically assessed. This paper presents a series of analyses of phonological cues and distributional cues and their potential for distinguishing grammatical categories of words in corpus analyses. The corpus analyses indicated that phonological cues were more reliable for less frequent words, whereas distributional information was most valuable for high frequency words. We tested this prediction in an artificial language learning experiment, where the distributional and phonological cues of categories of nonsense words were varied. The results corroborated the corpus analyses. For high-frequency nonwords, distributional information was more useful, whereas for low-frequency words there was more reliance on phonological cues. The results indicate that phonological and distributional cues contribute differentially towards grammatical categorisation.     

Usage of semantic representations in recognition memory

Meanings of words facilitate false acceptance as well as correct rejection of lures in recognition memory tests, depending on the experimental context. This suggests that semantic representations are both directly and indirectly (i.e., mediated by perceptual representations) used in remembering. Studies using memory conjunction errors (MCEs) paradigms, in which the lures consist of component parts of studied words, have reported semantic facilitation of rejection of the lures. However, attending to components of the lures could potentially cause this. Therefore, we investigated whether semantic overlap of lures facilitates MCEs using Japanese Kanji words in which a whole-word image is more concerned in reading. Experiments demonstrated semantic facilitation of MCEs in a delayed recognition test (Experiment 1), and in immediate recognition tests in which participants were prevented from using phonological or orthographic representations (Experiment 2), and the salient effect on individuals with high semantic memory capacities (Experiment 3). Additionally, analysis of the receiver operating characteristic suggested that this effect is attributed to familiarity-based memory judgement and phantom recollection. These findings indicate that semantic representations can be directly used in remembering, even when perceptual representations of studied words are available.     

The relative sensitivities of same-different and identification judgment models to perceptual dependence

Probabilistic models of same-different and identification judgments are compared (within each paradigm) with regard to their sensitivity to perceptual dependence or the degree to which the underlying psychological dimensions are correlated. Three same-different judgment models are compared. One is a step function or decision bound model and the other two are probabilistic variants of a similarity model proposed by Shepard. Three types of identification models are compared: decision bound models, a probabilistic multidimensional scaling model, and probabilistic models based on the Shepard-Luce choice rule. The decision bound models were found to be most sensitive to perceptual dependence, especially when there is considerable distributional overlap. The same-different model based on the city-block metric and an exponential decay similarity function, and the corresponding identification model were found to be particularly insensitive to perceptual dependence. These results suggest that if a Shepard-type similarity function accurately describes behavior, then under typical experimental conditions it should be difficult to see the effects of perceptual dependence. This result provides strong support for a perceptualindependence assumption when using these models. These theoretical results may also play an important role in studying different decision rules employed at different stages of identification training.We thank Robert Melara, Jerome Busemeyer and three anonymous reviewers for comments on an earlier draft of this paper.     

Qualitative differences in the representation of abstract versus concrete words: Evidence from the visual-world paradigm

In the present visual-world experiment, participants were presented with visual displays that included a target item that was a semantic associate of an abstract or a concrete word. This manipulation allowed us to test a basic prediction derived from the qualitatively different representational framework that supports the view of different organizational principles for concrete and abstract words in semantic memory. Our results confirm the assumption of a primary organizational principle based on association for abstract words, different from the semantic similarity principle proposed for concrete words, and provide the first piece of evidence in support of this view obtained from healthy participants. The results shed light on the representational structure of abstract and concrete concepts.     

Conceptual similarity effects on working memory in sentence contexts: Testing a theory of anaphora

The degree of semantic similarity between an anaphoric noun phrase (e.g., the bird) and its antecedent (e.g., a robin) is known to affect the anaphor resolution process, but the mechanisms that underlie this effect are not known. One proposal (Almor, 1999) is that semantic similarity triggers interference effects in working memory and makes two crucial assumptions: First, semantic similarity impairs working memory just as phonological similarity does (e.g., Baddeley, 1992), and, second, this impairment interferes with processes of sentence comprehension. We tested these assumptions in two experiments that compared recall accuracy between phonologically similar, semantically similar, and control words in sentence contexts. Our results do not provide support for Almor's claims: Phonological overlap decreased recall accuracy in sentence contexts, but semantic similarity did not. These results shed doubt on the idea that semantic interference in working memory is an underlying mechanism in anaphor resolution.     

The phenonenological idealism controversy in light of possible worlds semantics

In the paper there is presented the semantic interpretation of idealism/realism controversy which is one of the most essential issues in Ingarden’s phenomenological project of ontology. The procedure of semantic paraphrase which is contemporary developed by Woleński, is the main interpretative tool. In the central part of the paper, there is formulated the formal theory of the semantic framework underlying idealism/realism discourse. Finally, there are formulated some notes showing that intentional conception of negation may be used for defending various idealistic positions.

Semantic transparency in free stems: The effect of Orthography-Semantics Consistency on word recognition

A largely overlooked side effect in most studies of morphological priming is a consistent main effect of semantic transparency across priming conditions. That is, participants are faster at recognizing stems from transparent sets (e.g., farm) in comparison to stems from opaque sets (e.g., fruit), regardless of the preceding primes. This suggests that semantic transparency may also be consistently associated with some property of the stem word. We propose that this property might be traced back to the consistency, throughout the lexicon, between the orthographic form of a word and its meaning, here named Orthography-Semantics Consistency (OSC), and that an imbalance in OSC scores might explain the “stem transparency” effect. We exploited distributional semantic models to quantitatively characterize OSC, and tested its effect on visual word identification relying on large-scale data taken from the British Lexicon Project (BLP). Results indicated that (a) the “stem transparency” effect is solid and reliable, insofar as it holds in BLP lexical decision times (Experiment 1); (b) an imbalance in terms of OSC can account for it (Experiment 2); and (c) more generally, OSC explains variance in a large item sample from the BLP, proving to be an effective predictor in visual word access (Experiment 3).     

Characterizing the Dynamics of Learning in Repeated Reference Games

The language we use over the course of conversation changes as we establish common ground and learn what our partner finds meaningful. Here we draw upon recent advances in natural language processing to provide a finer-grained characterization of the dynamics of this learning process. We release an open corpus (>15,000 utterances) of extended dyadic interactions in a classic repeated reference game task where pairs of participants had to coordinate on how to refer to initially difficult-to-describe tangram stimuli. We find that different pairs discover a wide variety of idiosyncratic but efficient and stable solutions to the problem of reference. Furthermore, these conventions are shaped by the communicative context: words that are more discriminative in the initial context (i.e., that are used for one target more than others) are more likely to persist through the final repetition. Finally, we find systematic structure in how a speaker's referring expressions become more efficient over time: Syntactic units drop out in clusters following positive feedback from the listener, eventually leaving short labels containing open-class parts of speech. These findings provide a higher resolution look at the quantitative dynamics of ad hoc convention formation and support further development of computational models of learning in communication.     

Lateralization of object-shape information in semantic processing

An experiment was conducted to examine whether perceptual information, specifically the shape of objects, is activated during semantic processing. Subjects judged whether a target word was related to a prime word. Prime-target pairs that were not associated, but whose referents had similar shapes (e.g. LADDER-RAILROAD) yielded longer "no" responses than unassociated prime-target pairs, suggesting that shape information had been activated. A visual-field manipulation showed that, in right-handed subjects, this effect was localized in the left hemisphere. This finding is consistent with behavioral, brain imaging, and lesion data, which suggest that object shape at the category level is represented in the left hemisphere.     

分类中相似性的理论与模型

相似性在分类的原型理论、样例理论、定义理论和理论解释观中都扮演着重要的角色。人们对相似性的研究由来已久,但是它在分类的领域中至今仍是一个相对模糊的概念,这部分地由于揭示相似性的真正机制将涉及到复杂的信息加工过程。本文以分类中的相似性为出发点介绍了近期相似性研究的一些理论与模型并在此基础上对概念和分类领域中的相似性研究进行了分析和总结。     

Modelling the effects of semantic ambiguity in word recognition

Most words in English are ambiguous between different interpretations; words can mean different things in different contexts. We investigate the implications of different types of semantic ambiguity for connectionist models of word recognition. We present a model in which there is competition to activate distributed semantic representations. The model performs well on the task of retrieving the different meanings of ambiguous words, and is able to simulate data reported by Rodd, Gaskell, and Marslen-Wilson [J. Mem. Lang. 46 (2002) 245] on how semantic ambiguity affects lexical decision performance. In particular, the network shows a disadvantage for words with multiple unrelated meanings (e.g., bark) that coexists with a benefit for words with multiple related word senses (e.g., twist). The ambiguity disadvantage arises because of interference between the different meanings, while the sense benefit arises because of differences in the structure of the attractor basins formed during learning. Words with few senses develop deep, narrow attractor basins, while words with many senses develop shallow, broad basins. We conclude that the mental representations of word meanings can be modelled as stable states within a high-dimensional semantic space, and that variations in the meanings of words shape the landscape of this space.     

Dissociations in performance on novel versus irregular items: single-route demonstrations with input gain in localist and distributed models

Four pairs of connectionist simulations are presented in which quasi-regular mappings are computed using localist and distributed representations. In each simulation, a control parameter termed input gain was modulated over the only level of representation that mapped inputs to outputs. Input gain caused both localist and distributed models to shift between regularity-based and item-based modes of processing. Performance on irregular items was selectively impaired in the regularity-based modes, whereas performance on novel items was selectively impaired in the item-based modes. Thus, the models exhibited double dissociations without separable processing components. These results are discussed in the context of analogous dissociations found in language domains such as word reading and inflectional morphology.     

Pervasive influence of semantics in letter and category fluency: a multidimensional approach

We used multidimensional statistical procedures to study semantic and lexical processes underlying word retrieval in verbal-fluency performance. Forty healthy participants were given a two-choice letter task (i.e., generate items beginning with the letter 'A' or 'F', in any order) and a two-choice category task (i.e., generate animal or fruit names, in any order). Using correspondence analysis (CoA) and hierarchical clustering (HC), we found evidence of prominent semantic organization in both letter and category fluency. For example, a striking categorical segregation between animate and inanimate entities emerged during the letter task. Analysis of inter-item times revealed strong sequential priming effects in both tasks. Taken together, these results indicate that semantic facilitation is pervasive in word retrieval processes, even in the letter-fluency task, and therefore suggest that the traditional view of letter fluency as a purely phonemically based task should be revised. Finally, our findings may help explain patterns of verbal-fluency measures obtained in focal brain lesion patients.