This study is about the development of a retrainable reading ability estimation system based on concepts from the Text Readability Indexing (TRI) domain. This system aims to promote self-directed language learning and to serve as an educational reinforcement tool for English language learners. Student essays were used to calibrate the system which provided realistic approximations of their actual reading levels. In this thesis, we compared the performance of two vector semantics-based algorithms, namely, Latent Semantic Indexing (LSI) and Concept Indexing (CI) for content analysis. Since these algorithms rely on the bag-of-words approach and inherently lack grammatical analysis, we augmented them using Part-of-Speech (POS) n-gram features to approximate the syntactic complexity of text documents. Results show that directly combining the content-and grammar-based feature sets yielded lower classification accuracies than utilising each feature set alone. Using a sparsification strategy, we were able to optimise the combination process and, with the integration of POS bi-grams, we achieved our overall highest mean exact agreement accuracies (MEAA) of 0.924 and 0.952 for LSI and CI, respectively. We have also conducted error analyses on our results where we examined overestimation and underestimation error types to uncover the probable causes for the systems' misclassifications.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:707621 |
| Date | January 2017 |
| Creators | Razon, Abigail R. |
| Publisher | University of Birmingham |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://etheses.bham.ac.uk//id/eprint/7260/ |
Page generated in 0.0017 seconds