Academic literature on the topic 'INTEGRATING TEXT'

This work systematically investigates when and how combining speech output and visual text may facilitate processing and comprehension of sentences. It is proposed that a redundant multimodal presentation of speech and text has the potential for improving sentence processing but also for severely disrupting it. The effectiveness of the presentation is assumed to depend on the linguistic complexity of the sentence, the memory demands incurred by the selected multimodal configuration and the characteristics of the user. The thesis employs both theoretical and empirical methods to examine this claim. At the theoretical front, the research makes explicit features of multimodal sentence presentation and of structures and processes involved in multimodal language processing. Two entities are presented: a multimodal design space (MMDS) and a multimodal user model (MMUM). The dimensions of the MMDS include aspects of (i) the sentence (linguistic complexity, c.f., Gibson, 1991), (ii) the presentation (configurations of media), and (iii) user cost (a function of the first two dimensions). The second entity, the MMUM, is a cognitive model of the user. The MMUM attempts to characterise the cognitive structures and processes underlying multimodal language processing, including the supervisory attentional mechanisms that coordinate the processing of language in parallel modalities. The model includes an account of individual differences in verbal working memory (WM) capacity (c.f. Just and Carpenter, 1992) and can predict the variation in the cognitive cost experienced by the user when presented with different contents in a variety of multimodal configurations. The work attempts to validate through 3 controlled studies with users the central propositions of the MMUM. The experimental findings indicate the validity of some features of the MMUM but also the need for further refinement. Overall, they suggest that a durable text may reduce the processing cost of demanding sentences delivered by speech, whereas adding speech to such sentences when presented visually increases processing cost. Speech can be added to various visual forms of text only if the linguistic complexity of the sentence imposes a low to moderate load on the user. These conclusions are translated to a set of guidelines for effective multimodal presentation of sentences. A final study then examines the validity of some of these guidelines in an applied setting. Results highlight the need for an enhanced experimental control. However, they also demonstrate that the approach used in this research can validate specific assumptions regarding the relationship between cognitive cost, sentence complexity and multimodal configuration aspects and thereby to inform the design process of effective multimodal user interfaces.

In questa tesi saranno usate tecniche di deep learning per affrontare unodei problemi più difficili dell’elaborazione automatica del linguaggio naturale:la generazione automatica di riassunti. Dato un corpus di testo, l’obiettivoè quello di generare un riassunto che sia in grado di distillare e comprimerel’informazione dall’intero testo di partenza. Con i primi approcci si é provatoa catturare il significato del testo attraverso l’uso di regole scritte dagliumani. Dopo questa era simbolica basata su regole, gli approcchi statistici hanno preso il sopravvento. Negli ultimi anni il deep learning ha impattato positivamente ogni area dell’elaborazione automatica del linguaggionaturale, incluso la generazione automatica dei riassunti. In questo lavoroi modelli pointer-generator [See et al., 2017] sono utilizzati in combinazionea pre-trained deep contextualized word embeddings [Peters et al., 2018]. Sivaluta l’approccio sui due più grossi dataset per la generazione automaticadei riassunti disponibili ora: il dataset CNN/Daily Mail e il dataset Newsroom. Il dataset CNN/Daily Mail è stato generato partendo dal dataset diQuestion Answering pubblicato da DeepMind [Hermann et al., 2015], concatenando le frasi di highlight delle news e formando cosı̀ dei riassunti multifrase. Il dataset Newsroom [Grusky et al., 2018] è, invece, il primo datasetesplicitamente costruito per la generazione automatica di riassunti. Comprende un milione di coppie articolo-riassunto con diversi gradi di estrattività/astrattività a diversi ratio di compressione.L’approccio è valutato sui test-set con l’uso della metrica Recall-Oriented Understudy for Gisting Evaluation (ROUGE). Questo approccio causa un sostanzioso aumento nelle performance per il dataset Newsroom raggiungendo lo stato dell’arte sul valore di ROUGE-1 e valori competitivi per ROUGE-2 e ROUGE-L.

In a large urban school system located in a metropolitan city in the southeastern United States, third- and fifth-grade minority students in Title I elementary schools are performing below proficiency in social studies on the statewide standardized assessments. The lack of exposure to the social studies curriculum continues to hinder minority students from successfully comprehending complex informational text, which is important to their success in the newly adopted Common Core State Standards (CCSS). The purpose of this qualitative study was to explore the problem teachers faced with an insufficient amount of time for teaching social studies content and the recent requirement to increase student exposure to informational text. The research used Lev Vygotsky's theory of social constructivism to provide a framework for the methods used in this paper. To address these problems, this study explored two third-grade and two fifth-grade language arts teachers' perceptions of integrating social studies text during their reading-language arts block. Further, the study observed teachers as they integrated social studies text to teach reading. Data for this case study were compiled from interviews, observations, and focus group discussions. The data were reviewed and coded to identify major themes and were then analyzed to generalize data findings. Teachers reported integrating social studies text afforded them the opportunity to maximize instructional time, teach the CCSS, and expose students to more informational text. Implications for positive social change include enabling teachers to identify the benefits of integrating social studies text during reading-language arts instruction and enabling minority students to increase their scores on the statewide social studies assessment.

The amount of biomedical literature available is increasing at an exponential rate and is becoming increasingly difficult to navigate. Text-mining methods can potentially mitigate this problem, through the systematic and large-scale extraction of structured information from inherently unstructured biomedical text. This thesis reports the development of four text-mining systems that, by building on each other, has enabled the extraction of information about a large number of published statements in the biomedical literature. The first system, LINNAEUS, enables highly accurate detection ('recognition') and identification ('normalization') of species names in biomedical articles. Building on LINNAEUS, we implemented a range of improvements in the GNAT system, enabling high-throughput gene/protein detection and identification. Using gene/protein identification from GNAT, we developed the Gene Expression Text Miner (GETM), which extracts information about gene expression statements. Finally, building on GETM as a pilot project, we constructed the BioContext integrated event extraction system, which was used to extract information about over 11 million distinct biomolecular processes in 10.9 million abstracts and 230,000 full-text articles. The ability to detect negated statements in the BioContext system enables the preliminary analysis of potential contradictions in the biomedical literature. All tools (LINNAEUS, GNAT, GETM, and BioContext) are available under open-source software licenses, and LINNAEUS and GNAT are available as online web-services. All extracted data (36 million BioContext statements, 720,000 GETM statements, 72,000 contradictions, 37 million mentions of species names, 80 million mentions of gene names, and 57 million mentions of anatomical location names) is available for bulk download. In addition, the data extracted by GETM and BioContext is also available to biologists through easy-to-use search interfaces.

Identifying the boundary between relevant and irrelevant objects in text classification is a significant challenge due to the numerous uncertainties in text documents. Most existing binary text classifiers cannot deal effectively with this problem due to the issue of over-fitting. This thesis proposes a three-way decision model for dealing with the uncertain boundary to improve the binary text classification performance by integrating the distinct aspects of three-way decisions theory and the capacities of the Support Vector Machine. The experimental results show that the proposed models outperform baseline models on the RCV1, Reuters-21578, and R65CO datasets.

Suffix Tree Clustering (STC) is a search result clustering (SRC) algorithm focused on generating overlapping clusters with meaningful labels in linear time. It showed the feasibility of SRC but in time, subsequent studies introduced description-first algorithms that generate better labels and achieve higher precision. Still, STC remained as the fastest SRC algorithm and there appeared studies concerned with different problems of STC. In this thesis, semantic relations between cluster labels and documents are exploited to filter out noisy labels and improve merging phase of STC. Wikipedia is used to identify these relations and methods for integrating semantic information to STC are suggested. Semantic features are shown to be effective for SRC task when used together with term frequency vectors. Furthermore, there were no SRC studies on Turkish up to now. In this thesis, a dataset for Turkish is introduced and a number of methods are tested on Turkish.

This dissertation aims to discover synergistic combinations of top-down (ontologies), interactive (relevance feedback), and bottom-up (machine learning) knowledge encoding techniques for text mining. The strength of machine learning techniques lies in their coverage and efficiency because they can discover new knowledge without human intervention. The output, however, is often imprecise and irrelevant. Human knowledge, top-down or interactively encoded, may remedy this. The research question addressed is if knowledge discovery can become more precise and relevant with hybrid systems. Three different combinations are evaluated. The first study investigates an ontology, the Unified Medical Language System (UMLS), combined with an automatically created thesaurus to dynamically adjust the thesaurus' output. The augmented thesaurus was added to a medical, meta-search portal as a keyword suggester and compared with the unmodified thesaurus and UMLS. Users preferred the hybrid approach. Thus, the combination of the ontology with the thesaurus was better than the components separately. The second study investigates implicit relevance feedback combined with genetic algorithms designed to adjust user queries for online searching. These were compared with pure relevance feedback algorithms. Users were divided into groups based on their overall performance. The genetic algorithm significantly helped low achievers, but hindered high achievers. Thus, the interactively elicited knowledge from relevance feedback was judged insufficient to guide machine learning for all users. The final study investigates ontologies combined with two natural language processing techniques: a shallow parser and an automatically created thesaurus. Both capture relations between phrases in biomedical text. Qualified researchers found all terms to be precise; however, terms that belonged to ontologies were more relevant. Parser relations were all precise. Thesaurus relations were less precise, but precision improved for relations that had their terms represented in ontologies. Thus, this integration of ontologies with natural language processing provided good results. In general, it was concluded that top-down encoded knowledge could be effectively integrated with bottom-up encoded knowledge for knowledge discovery in text. This is particularly relevant to business fields, which are text and knowledge intensive. In the future, it will be worthwhile to extend the parser and also to test similar hybrid approaches for data mining.