Querying a database to retrieve an answer, telling a robot to perform an action, or teaching a computer to play a game are tasks requiring communication with machines in a language interpretable by them. Semantic parsing is the task of converting human language to a machine interpretable language. While human languages are sequential in nature with latent structures, machine interpretable languages are formal with explicit structures. The computational linguistics community have created several treebanks to understand the formal syntactic structures of human languages. In this thesis, we use these to obtain formal meaning representations of languages, and learn computational models to convert these meaning representations to the target machine representation. Our goal is to evaluate if existing treebank syntactic representations are useful for semantic parsing. Existing semantic parsing methods mainly learn domain-specific grammars which can parse human languages to machine representation directly. We deviate from this trend and make use of general-purpose syntactic grammar to help in semantic parsing. We use two syntactic representations: Combinatory Categorial Grammar (CCG) and dependency syntax. CCG has a well established theory on deriving meaning representations from its syntactic derivations. But there are no CCG treebanks for many languages since these are difficult to annotate. In contrast, dependencies are easy to annotate and have many treebanks. However, dependencies do not have a well established theory for deriving meaning representations. In this thesis, we propose novel theories for deriving meaning representations from dependencies. Our evaluation task is question answering on a knowledge base. Given a question, our goal is to answer it on the knowledge base by converting the question to an executable query. We use Freebase, the knowledge source behind Google’s search engine, as our knowledge base. Freebase contains millions of real world facts represented in a graphical format. Inspired from the Freebase structure, we formulate semantic parsing as a graph matching problem, i.e., given a natural language sentence, we convert it into a graph structure from the meaning representation obtained from syntax, and find the subgraph of Freebase that best matches the natural language graph. Our experiments on Free917, WebQuestions and GraphQuestions semantic parsing datasets conclude that general-purpose syntax is more useful for semantic parsing than induced task-specific syntax and syntax-agnostic representations.

Traditional supervised visual recognition methods require a great number of annotated examples for each concerned class. The collection and annotation of visual data (e.g., images and videos) could be laborious, tedious and time-consuming when the number of classes involved is very large. In addition, there are such situations where the test instances are from novel classes for which training examples are unavailable in the training stage. These issues can be addressed by zero-shot learning (ZSL), an emerging machine learning technique enabling the recognition of novel classes. The key issue in zero-shot visual recognition is the semantic gap between visual and semantic representations. We address this issue in this thesis from three different perspectives: visual representations, semantic representations and the learning models. We first propose a novel bidirectional latent embedding framework for zero-shot visual recognition. By learning a latent space from visual representations and labelling information of the training examples, instances of different classes can be mapped into the latent space with the preserving of both visual and semantic relatedness, hence the semantic gap can be bridged. We conduct experiments on both object and human action recognition benchmarks to validate the effectiveness of the proposed ZSL framework. Then we extend the ZSL to the multi-label scenarios for multi-label zero-shot human action recognition based on weakly annotated video data. We employ a long short term memory (LSTM) neural network to explore the multiple actions underlying the video data. A joint latent space is learned by two component models (i.e. the visual model and the semantic model) to bridge the semantic gap. The two component embedding models are trained alternately to optimize the ranking based objectives. Extensive experiments are carried out on two multi-label human action datasets to evaluate the proposed framework. Finally, we propose alternative semantic representations for human actions towards narrowing the semantic gap from the perspective of semantic representation. A simple yet effective solution based on the exploration of web data has been investigated to enhance the semantic representations for human actions. The novel semantic representations are proved to benefit the zero-shot human action recognition significantly compared to the traditional attributes and word vectors. In summary, we propose novel frameworks for zero-shot visual recognition towards narrowing and bridging the semantic gap, and achieve state-of-the-art performance in different settings on multiple benchmarks.

Thesis (M. A. (African Languages)) --University of Limpopo, 2001 / This study explores the lexical semantic representation of break verbs in Xitsonga. Chapter One is the introduction of this study. It describes the statement of the problem. The main aim of this study is described which m bnef is to investigate the form, struc e and interpretation of break verbs in Xitsonga. The significance of this study, which is to highlight the semantic value of break verbs in Xitsonga is discussed. The methodology, lite rature review as well as the theta - theory and its properties have been examined. Chapter Two explores the predicate argument structure. The difference between two lexical representations, which are lexical-syntactic and lexical­ semantic representations was investigated. A brief definition of break verbs as well as the six categories of the verbs of change of state have also been discussed. This chapter also analyses the lexical features of break verbs. Various sub - classes of external arguments and of internal argument are observed. Chapter Three presents the lexical - semantic representation of break verbs where focus is on argument structure, event structure, qualia structure and lexical inheritance structure. Chapter Four deals with the syntactic alternation and selection restriction of break verbs. The difference between transitive and intransitive alterna tions was also highlighted. This chapter also explores some of the different types of verbal alternations, such as ' instrument subject alternation', ' locative alternation' and the ' with/ against alternation' . Chapter Five gives the main conclusion of this study.

The semantic content (i.e., meaning of words) is the essence of retrieved autobiographical memories. In comparison to previous research, which has mainly focused on phenomenological experiences and age distribution of memory events, the present study provides a novel view on the retrieval of event information by addressing the semantic representation of memories. In the present study the semantic representation (i.e., word locations represented by vectors in a high dimensional space) of retrieved memory information were investigated, by analyzing the data with an automatic statistical algorithm. The experiment comprised a cued recall task, where participants were presented with unimodal (i.e., one sense modality) or multimodal (i.e., three sense modalities in conjunction) retrieval cues and asked to recall autobiographical memories. The memories were verbally narrated, recorded and transcribed to text. The semantic content of the memory narrations was analyzed with a semantic representation generated by latent semantic analysis (LSA). The results indicated that the semantic representation of visually evoked memories were most similar to the multimodally evoked memories, followed by auditorily and olfactorily evoked memories. By categorizing the semantic content into clusters, the present study also identified unique characteristics in the memory content across modalities.

The study of semantics in the brain explores how the brain represents, processes, and learns the meaning of language. In this thesis we show both that semantic representations can be decoded from electroencephalography data, and that we can detect the emergence of semantic representations as participants learn an artificial language mapping. We collected electroencephalography data while participants performed a reinforcement learning task that simulates learning an artificial language, and then developed a machine learning semantic representation model to predict semantics as a word-to-symbol mapping was learned. Our results show that 1) we can detect a reward positivity when participants correctly identify a symbol's meaning; 2) the reward positivity diminishes for subsequent correct trials; 3) we can detect neural correlates of the semantic mapping as it is formed; and 4) the localization of the neural representations is heavily distributed. Our work shows that language learning can be monitored using EEG, and that the semantics of even newly-learned word mappings can be detected using EEG. / Graduate

No description available.

La recherche d'information nécessite la mise en place de stratégies qui consistent à (1) cerner le besoin d'information ; (2) formuler le besoin d'information ; (3) repérer les sources pertinentes ; (4) identifier les outils à exploiter en fonction de ces sources ; (5) interroger les outils ; et (6) évaluer la qualité des résultats. Ce domaine n'a cessé d'évoluer pour présenter des techniques et des approches permettant de sélectionner à partir d'un corpus de documents l'information pertinente capable de satisfaire le besoin exprimé par l'utilisateur. De plus, dans le contexte applicatif du domaine de la RI biomédicale, les sources d'information hétérogènes sont en constante évolution, aussi bien du point de vue de la structure que du contenu. De même, les besoins en information peuvent être exprimés par des utilisateurs qui se caractérisent par différents profils, à savoir : les experts médicaux comme les praticiens, les cliniciens et les professionnels de santé, les utilisateurs néophytes (sans aucune expertise ou connaissance du domaine) comme les patients et leurs familles, etc. Plusieurs défis sont liés à la tâche de la RI biomédicale, à savoir : (1) la variation et la diversité du besoin en information, (2) différents types de connaissances médicales, (3) différences de compé- tences linguistiques entre experts et néophytes, (4) la quantité importante de la littérature médicale ; et (5) la nature de la tâche de RI médicale. Cela implique une difficulté d'accéder à l'information pertinente spécifique au contexte de la recherche, spécialement pour les experts du domaine qui les aideraient dans leur prise de décision médicale. Nos travaux de thèse s'inscrivent dans le domaine de la RI biomédicale et traitent les défis de la formulation du besoin en information experte et l'identification des sources pertinentes pour mieux répondre aux besoins cliniques. Concernant le volet de la formulation et l'analyse de requêtes expertes, nous proposons des analyses exploratoires sur des attributs de requêtes, que nous avons définis, formalisés et calculés, à savoir : (1) deux attributs de longueur en nombre de termes et en nombre de concepts, (2) deux facettes de spécificité terme-document et hiérarchique, (3) clarté de la requête basée sur la pertinence et celle basée sur le sujet de la requête. Nous avons proposé des études et analyses statistiques sur des collections issues de différentes campagnes d'évaluation médicales CLEF et TREC, afin de prendre en compte les différentes tâches de RI. Après les analyses descriptives, nous avons étudié d'une part, les corrélations par paires d'attributs de requêtes et les analyses de corrélation multidimensionnelle. Nous avons étudié l'impact de ces corrélations sur les performances de recherche d'autre part. Nous avons pu ainsi comparer et caractériser les différentes requêtes selon la tâche médicale d'une manière plus généralisable. Concernant le volet lié à l'accès à l'information, nous proposons des techniques d'appariement et d'expansion sémantiques de requêtes dans le cadre de la RI basée sur les preuves cliniques. / The research topic of this document deals with a particular setting of medical information retrieval (IR), referred to as expert based information retrieval. We were interested in information needs expressed by medical domain experts like praticians, physicians, etc. It is well known in information retrieval (IR) area that expressing queries that accurately reflect the information needs is a difficult task either in general domains or specialized ones and even for expert users. Thus, the identification of the users' intention hidden behind queries that they submit to a search engine is a challenging issue. Moreover, the increasing amount of health information available from various sources such as government agencies, non-profit and for-profit organizations, internet portals etc. presents oppor- tunities and issues to improve health care information delivery for medical professionals, patients and general public. One critical issue is the understanding of users search strategies and tactics for bridging the gap between their intention and the delivered information. In this thesis, we focus, more particularly, on two main aspects of medical information needs dealing with the expertise which consist of two parts, namely : - Understanding the users intents behind the queries is critically important to gain a better insight of how to select relevant results. While many studies investigated how users in general carry out exploratory health searches in digital environments, a few focused on how are the queries formulated, specifically by domain expert users. We address more specifically domain expert health search through the analysis of query attributes namely length, specificity and clarity using appropriate proposed measures built according to different sources of evidence. In this respect, we undertake an in-depth statistical analysis of queries issued from IR evalua- tion compaigns namely Text REtrieval Conference (TREC) and Conference and Labs of the Evaluation Forum (CLEF) devoted for different medical tasks within controlled evaluation settings. - We address the issue of answering PICO (Population, Intervention, Comparison and Outcome) clinical queries formulated within the Evidence Based Medicine framework. The contributions of this part include (1) a new algorithm for query elicitation based on the semantic mapping of each facet of the query to a reference terminology, and (2) a new document ranking model based on a prioritized aggregation operator. we tackle the issue related to the retrieval of the best evidence that fits with a PICO question, which is an underexplored research area. We propose a new document ranking algorithm that relies on semantic based query expansion leveraged by each question facet. The expansion is moreover bounded by the local search context to better discard irrelevant documents. The experimental evaluation carried out on the CLIREC dataset shows the benefit of our approaches.

Visual processing is the earliest core process required to support a normal reading system. However, little attention has been given to its role in any of the existing cognitive/computational models of reading. The ultimate goal of this thesis is to create a large-scale model of reading, which can generate phonology and semantics from print. Building such a model will allow for the exploration of a number of theoretically important cognitive phenomena in both normal and impaired reading including: font and size invariance; letter confusability; length effects; and pure alexic reading patterns. To achieve this goal, there are a number of important sub-goals that need to be achieved: (1) to develop a visual processing component which is capable of recognising letters in different fonts and sizes; (2) to produce a model that can develop useful intermediate (orthographic) representations as a consequence of learning; (3) to develop a set of semantic representations compact enough to allow efficient learning but that can still capture realistic semantic similarity relationships; (4) to integrate all the components together into a large-scale recurrent reading model; and (5) to extend the model to support picture naming, and to explore whether damage to the visual system can produce symptoms similar to those found in PA patients. Chapter 2 started by developing a simple feedforward network for letter recognition. The model was trained with letters in various transformations, which allowed the model to learn to deal with size and shape invariance problems as well as accounting for letter confusability effects and generalising to previously unseen letters. The model achieved this by extracting key features from visual input which could be used to support accurate letter recognition. Chapter 3 incorporated the letter recognition component developed in Chapter 2 into a word reading model. The reading model was trained on the mappings between print and phonology, with the orthographic representations which learn to emerge over training. The model could support accurate nonword naming and simulated the length by lexicality interaction observed in normal reading. A system of semantic representations was developed in Chapter 4 by using co-occurrence statistics to generate semantic codes that preserved realistic similarity relationships. Chapter 5 integrated all the components developed in the previous chapters together into a large-scale recurrent reading model. Finally, Chapter 6 extended the reading model to perform object recognition along with the reading task. When the model's visual system was damaged it was able to simulate the abnormal length effect typically seen in PA patients. The damaged model also showed impaired reaction times in object naming and preserved sensitivity to lexical/semantic variables in reading. The picture naming performance was modulated by visual complexity. In summary, the results highlight the importance of incorporating visual information into computational models of single word reading, and suggest that doing so will enable the exploration of a wide range of effects that were previously inaccessible to these types of connectionist models.

It has long been held that concrete material has a processing advantage over abstract material, as predicted by Dual Coding Theory (Paivio,1991), although this has been challenged. For example, based on evidence for behavioural and neuroscientific studies, Kousta,, Vigliocco, Vinson, & Del Campo, (2011) proposed that emotional valance had a greater influence in the processing of abstract words, and that under some circumstances there may be no concreteness effect and might even be an abstractness effect. This would not be predicted by DCT. In addition, Isen and Daubman (1984) have claimed that emotional valence, and particularly positive emotion can influence cognitive processing. Specifically, they demonstrated that positive emotion was associated with more inclusive categorization of ambiguous category members. This current study was a 2 x 2 between group design to investigate the effect of positive and negative valence on recognition memory for concrete and abstract words and on categorization. Contrary to what was predicted by Dual Coding Theory, abstract words were generally better recognized than concrete, with there being an additional interaction with valence. A significant interaction between word type and valence on categorization was also found. Results partially support Kousta et al. (2011).

Recent modeling work shows that patterns of shared perceptual features relate to the group-level order of acquisition of early-learned words (Peters & Borovsky, 2019). Here we present results for two eye-tracked word recognition studies showing patterns of shared perceptual features likewise influence processing of known and novel noun-concepts in individual 24- to 30-month-old toddlers. In the first study (Chapter 2, N=54), we explored the influence of perceptual connectivity on both initial attentional biases to known objects and subsequent label processing. In the second study (Chapter 3, N=49), we investigated whether perceptual connectivity influences patterns of attention during learning opportunities for novel object-features and object-labels, subsequent pre-labeling attentional biases, and object-label learning outcomes. Results across studies revealed four main findings. First, patterns of shared (visual-motion and visual-form and surface) perceptual features do relate to differences in early noun-concept processing at the individual level. Second, such influences are tentatively at play from the outset of novel noun-concept learning. Third, connectivity driven attentional biases to both recently learned and well-known objects follow a similar timecourse and show similar patterns of individual differences. Fourth, initial, pre-labeling attentional biases to objects relate to subsequent label processing, but do not linearly explain effects of connectivity. Finally, we consider whether these findings provide support for shared-feature-guided selective attention to object features as a mechanism underlying early lexico-semantic development.