Lightweight Text Visualization for Close and Distant Reading

Sammanfattning

Analysis of larger corpora or longer documents can be time-consuming and given current tools for it, difficult for many users. Document and corpus analysis in this context refers to the process of providing users with both close and distant reading representations of individual documents and corpora. This thesis explores the design and development of a text analysis tool with close and distant reading representations and with the goal doing this with lightweight approach and with generalized language support. Additionally, the target user base for this thesis has been widened to encompass more than people with deep understanding of text visualization, done to address the slim scope of earlier tools. To accomplish the goal of the thesis and answer its research questions, several objectives were defined, and the work was split into implementation and evaluation phases. This split was done with an iterative process in mind, where an initial prototype was tested and evaluated, with a second prototype implemented based on the evaluation, and so on for further prototypes. To limit the thesis work, however, the objectives were defined to only encompass a first, second, and potential final prototype. The tool was developed as a web-based tool with JavaScript, relying on D3.js for the drawing of visual elements. The tool was developed with a component-based approach where each view visualized different data about singular documents within a corpus or the entire corpus itself. Within this interface was the close reading view which wrote out the text of the selected document in plain text, with possibility for word highlighting and selection from the user. In addition to this were views for word frequencies, showing the number of occurrences each word had in the document, and word trends, showing how significant specific words were in specific segments of the document. A provided corpus could also be visualized through a card-based corpus overview, and the more data-centric node view, representing how similar documents were in the corpus. For further explanation on the similarity between documents, a similarity overview was also added, accessible on clicks on the node view edges. The evaluation process, on the other hand, was done through result collection on different tests, both with and without users. The user-free tests aimed to find results on the inner workings of the tool and was conducted during the implementation phase to find results which were not part of the user tests and to find possible faults to fix during the user tests. The user tests were instead conducted to find results related to the usability of the tool, both measured and perceived by the participants. These test phases gave results on the processing times within the tool, showing which NLP methods were preferable from a performance perspective, and how the different methods represented the data. Additionally, the results highlighted strengths and weaknesses within the tool given the quantitative and qualitative data from the user tests. The reasons behind these results were evaluated and discussed to implement a second prototype addressing shortcomings of the first prototype. This second prototype, the discussion leaning up to it and the thesis results was then compared to the Objectives of the thesis, along with being used to answer the proposed Research Questions. Lastly, with the differences between the second prototype and the discussed improvements, a direction for the tool in the future was defined. This also covered uncompleted Objectives and answers to the Research Questions which were deemed to be less direct compared to other answers.