Electronic Theses and Dissertations
Measuring Semantic Textual Similarity and Automatic Answer Assessment in Dialogue Based Tutoring Systems
Date of Award
Doctor of Philosophy
Andrew McGregor Olney
This dissertation presents methods and resources proposed to improve onmeasuring semantic textual similarity and their applications in student responseunderstanding in dialogue based Intelligent Tutoring Systems. In order to predict the extent of similarity between given pair of sentences,we have proposed machine learning models using dozens of features, such as thescores calculated using optimal multi-level alignment, vector based compositionalsemantics, and machine translation evaluation methods. Furthermore, we haveproposed models towards adding an interpretation layer on top of similaritymeasurement systems. Our models on predicting and interpreting the semanticsimilarity have been the top performing systems in SemEval (a premier venue for thesemantic evaluation) for the last three years. The correlations between our models'predictions and the human judgments were above 0.80 for several datasets while ourmodels being very robust than many other top performing systems. Moreover, wehave proposed Bayesian. We have also proposed a novel Neural Network based word representationmapping approach which allows us to map the vector based representation of a wordfound in one model to the another model where the word representation is missing,effectively pooling together the vocabularies and corresponding representationsacross models. Our experiments show that the model coverage increased by few toseveral times depending on which model's vocabulary is taken as a reference. Also,the transformed representations were well correlated to the native target modelvectors showing that the mapped representations can be used with condence tosubstitute the missing word representations in the target model. models to adapt similarity models across domains. Furthermore, we have proposed methods to improve open-ended answersassessment in dialogue based tutoring systems which is very challenging because ofthe variations in student answers which often are not self contained and need thecontextual information (e.g., dialogue history) in order to better assess theircorrectness. In that, we have proposed Probabilistic Soft Logic (PSL) modelsaugmenting semantic similarity information with other knowledge. To detect intra- and inter-sentential negation scope and focus in tutorialdialogs, we have developed Conditional Random Fields (CRF) models. The resultsindicate that our approach is very effective in detecting negation scope and focus intutorial dialogue context and can be further developed to augment the naturallanguage understanding systems. Additionally, we created resources (datasets, models, and tools) for fosteringresearch in semantic similarity and student response understanding inconversational tutoring systems.
Dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.
Banjade, Rajendra, "Measuring Semantic Textual Similarity and Automatic Answer Assessment in Dialogue Based Tutoring Systems" (2017). Electronic Theses and Dissertations. 1703.
Data is provided by the student.