The question of how language evolved has received an increasing amount of attention in recent years. Compared to seemingly more complex phenomena such as syntax, word meanings are usually seen as relatively easy to explain. Mainstream accounts in psycholinguistics and evolutionary linguistics assume that word meanings correspond to stable concepts which are prior to language and derive straightforwardly from human perception of structure in the world. Taking a cognitive linguistic approach based on psycholinguistic evidence, I argue instead that word meanings are conventions, grounded, learned and used in the context of communication. The meaning of a word is the sum of its contexts of use, with particular features of these contexts made more or less salient by mechanisms of attentional learning and communicative inference. Evolutionarily, word meanings arise as an emergent product of humans’ adapted tendency to infer each other’s intentions using contextual cues. They are then shaped over cultural evolution by the need to be learnable and useful for communication. This thesis presents a series of experiments that test the effect of these pressures on the origins and development of word meanings. Experiment 1 investigates the origins of strong tendencies for words to specify features on particular dimensions (such as the shape bias). The results show that these tendencies arise via attentional learning effects amplified by iterated learning. Dimensions which are less salient in contexts of learning and use drop out of word meanings as they are passed down a chain of learners. Experiments 2, 3 and 4 investigate the structure of word meanings produced during either paired communication games or individual labelling of images by similarity. While communication alone leads to word meanings that are unstructured and poorly aligned within pairs, communication plus iterated learning leads to word meanings that increase in structure and alignment over generations. Finally, Experiment 5 investigates the interaction of event structure and developing conventions in shaping word meanings. The structure of events in an artificial world is shown to influence lexicalisation patterns in the languages conventionalised by communicating pairs. Event features that are less predictable across communicative contexts tend to be more strongly associated with the conventions in the language. Overall, the experiments show that rather than straightforwardly reflecting pre-linguistic conceptualisation, word meanings are also dynamically shaped by learning and communication. In addition, these processes are constrained by the conventions that already exist within a language. This illuminates the mixture of convergence and diversity we see in word meanings in natural languages, and gives insight into their evolutionary origins.

Humans possess a rich semantic knowledge of words and concepts which captures the perceivable physical properties of their real-world referents and their relations. Encoding this knowledge or some of its aspects is the goal of computational models of semantic representation and has been the subject of considerable research in cognitive science, natural language processing, and related areas. Existing models have placed emphasis on different aspects of meaning, depending ultimately on the task at hand. Typically, such models have been used in tasks addressing the simulation of behavioural phenomena, e.g., lexical priming or categorisation, as well as in natural language applications, such as information retrieval, document classification, or semantic role labelling. A major strand of research popular across disciplines focuses on models which induce semantic representations from text corpora. These models are based on the hypothesis that the meaning of words is established by their distributional relation to other words (Harris, 1954). Despite their widespread use, distributional models of word meaning have been criticised as ‘disembodied’ in that they are not grounded in perception and action (Perfetti, 1998; Barsalou, 1999; Glenberg and Kaschak, 2002). This lack of grounding contrasts with many experimental studies suggesting that meaning is acquired not only from exposure to the linguistic environment but also from our interaction with the physical world (Landau et al., 1998; Bornstein et al., 2004). This criticism has led to the emergence of new models aiming at inducing perceptually grounded semantic representations. Essentially, existing approaches learn meaning representations from multiple views corresponding to different modalities, i.e. linguistic and perceptual input. To approximate the perceptual modality, previous work has relied largely on semantic attributes collected from humans (e.g., is round, is sour), or on automatically extracted image features. Semantic attributes have a long-standing tradition in cognitive science and are thought to represent salient psychological aspects of word meaning including multisensory information. However, their elicitation from human subjects limits the scope of computational models to a small number of concepts for which attributes are available. In this thesis, we present an approach which draws inspiration from the successful application of attribute classifiers in image classification, and represent images and the concepts depicted by them by automatically predicted visual attributes. To this end, we create a dataset comprising nearly 700K images and a taxonomy of 636 visual attributes and use it to train attribute classifiers. We show that their predictions can act as a substitute for human-produced attributes without any critical information loss. In line with the attribute-based approximation of the visual modality, we represent the linguistic modality by textual attributes which we obtain with an off-the-shelf distributional model. Having first established this core contribution of a novel modelling framework for grounded meaning representations based on semantic attributes, we show that these can be integrated into existing approaches to perceptually grounded representations. We then introduce a model which is formulated as a stacked autoencoder (a variant of multilayer neural networks), which learns higher-level meaning representations by mapping words and images, represented by attributes, into a common embedding space. In contrast to most previous approaches to multimodal learning using different variants of deep networks and data sources, our model is defined at a finer level of granularity—it computes representations for individual words and is unique in its use of attributes as a means of representing the textual and visual modalities. We evaluate the effectiveness of the representations learnt by our model by assessing its ability to account for human behaviour on three semantic tasks, namely word similarity, concept categorisation, and typicality of category members. With respect to the word similarity task, we focus on the model’s ability to capture similarity in both the meaning and appearance of the words’ referents. Since existing benchmark datasets on word similarity do not distinguish between these two dimensions and often contain abstract words, we create a new dataset in a large-scale experiment where participants are asked to give two ratings per word pair expressing their semantic and visual similarity, respectively. Experimental results show that our model learns meaningful representations which are more accurate than models based on individual modalities or different modality integration mechanisms. The presented model is furthermore able to predict textual attributes for new concepts given their visual attribute predictions only, which we demonstrate by comparing model output with human generated attributes. Finally, we show the model’s effectiveness in an image-based task on visual category learning, in which images are used as a stand-in for real-world objects.

No description available.

L'expression « mouvement fictif »' désigne l'utilisation de verbes dits de mouvement ou de déplacement pour décrire une configuration spatiale statique. Ce sujet n'a jusqu'ici pas été traité de manière systématique à partir de données attestées, et son étude s'est généralement limitée au niveau phrastique. En nous appuyant sur les listes de verbes du Lexique-Grammaire, sur la base textuelle Frantext ainsi que sur le cadre théorique concernant la sémantique des verbes de mouvement développé par Aurnague (2011b), nous avons dégagé pratiquement 600 occurrences de mouvement fictif en discours, et ceci afin de comprendre les mécanismes et les motivations de ce type de construction.Bien loin des explications habituellement proposées en termes d'opposition entre Path et Manner, de forme de l'entité et de contraintes bloquant parfois la production d'énoncés de ce type, nos résultats pointent l'importance du Mode de discours, montrent l'utilité de l'adoption d'une sémantique de l'espace dépassant la simple géométrie et plaident pour une approche privilégiant la structure, la configuration de plusieurs entités.Par ailleurs, cette thèse propose une classification qui se veut complète des verbes de mouvement/déplacement du français mobilisant les notions de changement d'emplacement et de relation locative élémentaire, ainsi qu'un corpus d'exemples conséquent constituant une base de travail pour la communauté des linguistes.Au-delà des approfondissements que nous effectuerons, cette étude ouvre la voie à une nouvelle manière de considérer les emplois dits "figurés" des verbes de mouvement. / The term “fictive motion” refers to the uses of verbs of motion which describe a static space configuration. This topic has thus far not been addressed systematically on empirical data, and its study was restricted to the sentence level. Basing our work on the verb lists of the Lexique-Grammaire, on the Frantext database, and on the theoretical framework proposed by Aurnague (2011b) on the semantics of motion verbs, I have extracted up to 600 instances of verb use coded as fictive motion, in discourse, in order to understand the mechanisms and motivations of such constructions.This thesis attempts to go beyond proposed distinctions between Path and Manner, or explanations based on the shape of entities, or even constraints blocking fictive motion uses. Our results highlight the importance of the Mode of discourse, stress the need for a framework in space semantics which relies on functional properties, rather than plain geometry, and advocates an approach based on the structure and the configuration between several entities.Besides, this thesis provides a classification, intended to be complete, of motion verbs in French, which elaborates on the notions of change of location and basic locative relation ; it also makes available a significant corpus of « fictive motion » verb uses for further linguistic studies.This thesis introduces several in-depth studies and paves the way for a new understanding of the ''figurative'' use of motion verbs.

This thesis is an investigation of Middle English insanity language. It analyses change in the Middle English lexical field of INSANITY, the semantic structure of lexemes wod and mad, and compares INSANITY conceptual metaphors in Middle English and present-day English. The INSANITY lexical field is an ideal one to study language change, due to socio-cultural changes since the Middle Ages such as advances in medical knowledge, the development of the field of psychiatry and legal changes protecting people with a mental illness from discrimination. The general theoretical aims were to examine a) change in conceptual metaphor, and b) semantic and lexical change with a particular focus on the decline in use of adjective wod. The theoretical frameworks are cognitive linguistics, prototype theory, and conceptual metaphor theory, and the data is derived from Middle English corpora and other sources. The INSANITY database I created for this study consisted of 1307 instances of mad, wod and near-synonyms in context. The main results can be divided into three groups. Firstly, the lexical field study demonstrates that various intra-linguistic and socio-cultural phenomena effect lexical change. Using case studies amongst others of the decline of wod in the Wycliffite Bible and of Caxton's translations from French, and a systematic variation across genre, I argue that the important factors are i) the arrival of new medical loanwords such as frensy, lunatic and malencolie; ii) the early re-emergence of the vernacular in medical texts starting in the twelfth century, and the development of a new medical register; iii) the so-called medieval 'inward turn'; iv) changes in the neighbouring lexical field of ANGER. Secondly, the semasiological study of wod and mad shows that the meanings of these two lexemes are structured and change in line with the central tenets of prototype theory, i.e. as described for diachronic prototype semantics by Geeraerts (1997). The path of mad's semantic development does not parallel that of wod after the thirteenth century. Mad's senses do not have the emphasis on wildness and fury that the senses of wod do. A particularly interesting finding is the semantic change from a sub-sense of adverb mad and adjective mad, 'unrestrained', leading in present-day English to a new delexicalised and grammaticalised sense of mad, where its use as an intensifier enhances scalar quantity and quality. Thirdly, the conceptual metaphor study demonstrates that predominantly the same conceptual metaphors are seen in both Middle English and present-day English, with some exceptions such as the concept of insanity being related to moral decline, as evidenced in the dearth of FALLING metaphors for insanity in present-day English. Conceptual metaphors such as INSANITY IS ANOTHER PLACE are evidenced in present-day English expressions such as out of her senses, or not in my right mind. In 1422, Thomas Hoccleve could write of a dysseveraunce between himself and his wit, or about his wyld infirmitie, which threw him owt of my selfe, illustrating the same underlying concepts. Other INSANITY conceptual metaphors which remain unchanged are GOING ASTRAY, LACK OF ORDER, LACK OF WHOLENESS, DARKNESS, FORCE, PRISON and BURDEN. Because of its unique approach in combining onomasiological and semasiological approaches with a conceptual metaphor study, this study reveals not only specific patterns of change, but differences in the rate of change on the lexical and conceptual levels. Lexical change driven by the need to be expressive, and reflecting socio-cultural changes such as changes in medical knowledge, can be seen to happen rapidly over the Middle English period. However, underlying conceptual change is barely discernible even over a much longer period of time from Middle English to present-day English. This research is significant because it provides a basis for future analysis of insanity language in other periods and contexts. It also contributes to the study of semantic change in general, highlighting the insights that can be gained by combining different types of data-driven analyses.

Este trabalho estuda a caracterizacao semântica dos argumentos dativos. Entendemos como “dativos” NPs que sao o primeiro objeto de construcões de duplo objeto (como, p. ex., John gave Mary a book “Joao deu um livro para Maria”), certos cliticos pronominais (como, p. ex., Maria me deu o livro) e NPs com marcacao morfologica de caso (como, p. ex., Hann gaf mér bókina “Ele me deu o livro”). Essencialmente, propõe-se, neste trabalho, que dativos sejam argumentos que estao envolvidos em eventos em que sao afetados. Iniciamos esse estudo a partir das caracterizacões tradicionais de argumentos dativos, de acordo com as quais essa seria uma funcao sintatica talhada para expressar o “possuidor” em um evento que envolve mudanca (potencial) de posse. Examinamos, com detalhe, as propostas feitas para dar conta da alternância dativa do inglês – principalmente, a conhecida restricao o “possuidor prospectivo” – e como essas propostas poderiam ser estendidas para cobrir os dados do português brasileiro. Essa proposta implica que, se o dativo é possuidor, o objeto direto deve ser o argumento possuido, descrevendo assim a relacao entre esses dois argumentos e os caracterizando. Levantamos, em seguida, algumas dificuldades para essas abordagens, mostrando, entre outros problemas, que muitos verbos que tomam argumentos dativos nao envolvem “posse”. Tendo mostrado que a abordagem baseada em posse falha em predizer quais verbos tomam dativos, buscamos uma caracterizacao semântica alternativa desses argumentos, procurando estabelecer uma comparacao dessa posicao sintatica com os objetos diretos. Apresentamos uma hipotese pouco explorada para o inglês, mas que parece ter respaldo translinguistico, de que dativos expressariam argumentos afetados. Essa proposta é criticada por alguns autores, pois “afetacao” parece ser um conceito difuso. Por isso, muitas vezes se diz que esta nao é uma nocao que possa ser empregada para cobrir classes coerentes de fenômenos, sendo substituida por outros conceitos, vistos como mais relevantes. Tomamos um caminho diferente desses trabalhos: a partir de uma semântica baseada na teoria de “reticulados”, buscamos articular o conceito de “afetacao” com outros geralmente empregados no estudo da semântica lexical dos verbos por meio de uma hierarquia implicacional. Assim, propomos que afetacao, mudanca, movimento, incrementalidade, “entrada” e “saida” de existência, afetacao psicologica, benefaccao e malefaccao sejam conceitos inter-relacionados, que estao em jogo na realizacao de argumentos. À luz dessa teoria, baseada na hierarquia implicacional que estabelece “graus” de afetacao, capturamos as observacões tradicionais sobre os dativos do inglês e as nossas observacões anteriormente levantadas sobre os dados do português brasileiro, bem como os dados que nao tinham uma explicacao natural em abordagens prévias. Nossa proposta, ainda, faz predicões a respeito da distribuicao de NPs com dativo marcado morfologicamente. / This work aims to discuss the semantic characterization of dative arguments. We take “dative” to be NPs that are the first direct object in double object contructions (e.g., John gave Mary a book), some clitic pronouns (for example, Maria me deu o livro “Maria gave me the book”) and morphologically case-marked NPs (for instance, Hann gaf mér bókina “He gave me the book”). Mainly, we propose datives are arguments that are involved in events in which they are affected. We start this study from the classical accounts of dative arguments, according to which this syntactic function is suitable to express the possessor in an event of “(potential) change of possession”. We examine in detail the proposal regarding English dative alternation – mainly the so-called “prospective possessor” constraint – and how it could be extended to account for data from Brazilian Portuguese. This proposal implicates that, if the dative is the “possessor”, the direct object must be the possessed argument, describing, thus, the relationship between these two arguments and characterizing them. We point out some difficulties for these approaches, showing among other problems that some verbs that can take a dative do not involve change of possession. After showing that the possession-based approach fails to correctly predict the verbs which take dative arguments, we attempt to find an alternative semantic characterization of datives, establishing a comparison between this syntactic function and direct objects. We present an underexplored hypothesis for English, which seems to be supported by cross linguistic data, according to which datives express affected arguments. This proposal is criticized by some authors since “affectedness” seems to be a diffuse notion, which would not be useful to account for coherent classes of phenomena. Therefore, “affectedness” is usually replaced by other notions considered more relevant to argument realization. We take a different way from these works: based on a “lattice” theory, we attempt to articulate the concept of “affectedness” with other notions generally used by lexical semantic researches by means of an implicational hierarchy. As a result, we propose that “affectedness”, change, motion, incrementallity, “coming” and “letting” to exist, psych “affectedness”, benefaction and malefaction are interrelated concepts, which are involved in argument realization patterns. In the light of this theory, based on an implicacional hierarchy of “degrees” of affectedness, we account for the standard observations about English double object construction and our previous observations about data from Brazilian Portuguese. Moreover, our proposal makes additional predictions regarding the distribution of morphologically casemarked datives.

The aim of this study is to examine the semantic architecture of the motif complex the four elements, i.e. fire, air, water and earth, in the Swedish poet Gunnar Ekelöf’s poem A Mölna Elegy (1960). The poem belongs to the same polyphonic and quotative-allusive tradition as T. S. Eliot’s The Waste Land and Ezra Pound’s Cantos. The four elements may be regarded both as four separate motifs and as constituting one semantically coherent motif complex. The latter reading has to do with the fact that the phrase the four elements is not itself present in the text. Thus, the thesis includes the assumption that this motif complex, heavily suggested by items in the text but still omitted, might function as a text matrix, from which a number of themes emerge such as life–death and time. The thesis has a theoretical anchoring in ideas about semantic frames (Barsalou) when discussing semantic relations between the different element-related words in the poem, and how these words may be linked to the concepts ‘fire’, ‘air’, water’ and ‘earth’ respectively. Traditional lexical relations such as hyponymy, antonymy and meronymy only catch the more obvious relations such as fire–glow, warm–cold and tree–branch, but are unable to explain pragmatically based relations between words linked to the same conceptual domain, such as sea–jetty, water–sink, fly–air and the like. To some extent, the thesis also draws upon Riffaterre’s theories about a poem’s matrix and how meaning arises in such texts. A major finding of the study is the heavy lexical presence of the four elements in the poem, expressed and suggested by a great number of semantically heterogeneous words. This semantic pattern is analysed in detail with the aid of semantic frame theory. A further discovery is that most of the element words imply dichotomies such as motion–repose, warmth–cold, light–dark or soft–hard. The elements have most of the dichotomies in common, which strongly suggests a union of all the four elements. Such a union is also suggested by several conspicuous compounds never earlier recorded in Swedish, such as glödstänk (‘glowspray’), vindstänk (‘windspray’), eldsus (‘fire sough’) and vågsus (‘wave sough’). The meetings of element are also described at the syntactic level as an explicit amalgamation of all four elements, which suggests a theme not earlier noticed. This theme may tentatively be called the cyclical amalgamation.

This report investigates the causativization patterns of verbs of eating and drinking from a typological perspective, arguing that ingestive verbs may be grouped together with middle verbs with respect to causativization. It is argued that both ingestive verbs and middle verbs are lexically reflexive and, in some languages, their causative variants are derived from their non-causative variants by an antireflexivization operation that delinks the verbs' coidentified arguments. Evidence from English and Marathi shows that such an operation is plausible as a causativization strategy on both semantic and morphological grounds. / text

Argument marking prepositions in German are part of more complex structures referred to here as verb-preposition combinations (verb-PPs), e.g. warten auf ('to wait for') and pochen auf ('to insist on'). The preposition auf ('on') attaches to a wide range of verbs to form such combinations in which auf encodes different semantic relations that elude concrete description. Nevertheless, previous research in valency theory and related approaches could identify patterns in the distribution of verb-PP[subscript 'auf'] combinations (Eroms 1981, 1991, Lerot 1982, Bouillon 1984, Domínguez Vázquez 2005), based on perceived similarities in the meaning of the governing verbs. Cognitive linguistics provides insights into seemingly opaque senses of prepositions by analyzing them as motivated by metaphorical meaning extension (Brugman 1988, Lakoff 1987, Meex 2001, Liamkina 2007). Finally, generative approaches scrutinize the semantic relationships between verbs and their PP-arguments and systematize them under the concept of semantic roles (Fillmore 1968, Rauh 1993). However, none of these approaches can fully account for the distribution of verb-PPauf combinations in German. This dissertation proposes a novel approach towards identifying and analyzing the distributional patterns of verb-PP[subscript 'auf'] combinations by applying insights from Frame Semantics (Fillmore 1982, 1985) and Construction Grammar (Goldberg 1995, 2006). Goldberg's theory of argument structure constructions already served as a model for analyzing auf as a partially schematic argument structure construction encoding the meaning 'future orientation/future event' (Rostila 2007). Based on a large amount of corpus data, I show that such generalizing accounts are better arrived at by employing a usage-based bottom-up approach to verb-PP[subscript 'auf'] combinations. I argue that the detailed semantic and syntactic information provided by the lexical database FrameNet for each lexical unit can be used to identify distributional patterns and to describe them in detail. Furthermore, I argue that integrating the verb-PP[subscript 'auf'] combinations and the frames they evoke into a hierarchical lexical-constructional network allows us to discover substantiated generalizations about these combinations while at the same time preserving the description of their idiosyncratic features. / text

In this dissertation, we introduce a graph-based model of instance-based, usage meaning that is cast as a problem of probabilistic inference. The main aim of this model is to provide a flexible platform that can be used to explore multiple hypotheses about usage meaning computation. Our model takes up and extends the proposals of Erk and Pado [2007] and McCarthy and Navigli [2009] by representing usage meaning as a probability distribution over potential paraphrases. We use undirected graphical models to infer this probability distribution for every content word in a given sentence. Graphical models represent complex probability distributions through a graph. In the graph, nodes stand for random variables, and edges stand for direct probabilistic interactions between them. The lack of edges between any two variables reflect independence assumptions. In our model, we represent each content word of the sentence through two adjacent nodes: the observed node represents the surface form of the word itself, and the hidden node represents its usage meaning. The distribution over values that we infer for the hidden node is a paraphrase distribution for the observed word. To encode the fact that lexical semantic information is exchanged between syntactic neighbors, the graph contains edges that mirror the dependency graph for the sentence. Further knowledge sources that influence the hidden nodes are represented through additional edges that, for example, connect to document topic. The integration of adjacent knowledge sources is accomplished in a standard way by multiplying factors and marginalizing over variables. Evaluating on a paraphrasing task, we find that our model outperforms the current state-of-the-art usage vector model [Thater et al., 2010] on all parts of speech except verbs, where the previous model wins by a small margin. But our main focus is not on the numbers but on the fact that our model is flexible enough to encode different hypotheses about usage meaning computation. In particular, we concentrate on five questions (with minor variants): - Nonlocal syntactic context: Existing usage vector models only use a word's direct syntactic neighbors for disambiguation or inferring some other meaning representation. Would it help to have contextual information instead "flow" along the entire dependency graph, each word's inferred meaning relying on the paraphrase distribution of its neighbors? - Influence of collocational information: In some cases, it is intuitively plausible to use the selectional preference of a neighboring word towards the target to determine its meaning in context. How does modeling selectional preferences into the model affect performance? - Non-syntactic bag-of-words context: To what extent can non-syntactic information in the form of bag-of-words context help in inferring meaning? - Effects of parametrization: We experiment with two transformations of MLE. One interpolates various MLEs and another transforms it by exponentiating pointwise mutual information. Which performs better? - Type of hidden nodes: Our model posits a tier of hidden nodes immediately adjacent the surface tier of observed words to capture dynamic usage meaning. We examine the model based on by varying the hidden nodes such that in one the nodes have actual words as values and in the other the nodes have nameless indexes as values. The former has the benefit of interpretability while the latter allows more standard parameter estimation. Portions of this dissertation are derived from joint work between the author and Katrin Erk [submitted]. / text