Techniques for scoring motive imagery in text - part B

• Main Author: Wong, Alvin Ann Ying
• Other Authors: Chen Lihui
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/77427
• Institution: Nanyang Technological University

This report summaries the width and depth contextual knowledge on the state of art technologies for natural language processing worked throughout the 40 weeks of professional final year project. Developing a Rule -Based System for Motive Imagery in Text – Part B was the objective scope of our project. Our task was to assist the Psychologist to detect motive imagery within a given list of text conversations or Thematic Apperception Test (TAT) stories. The highlights include understanding the relationship between human cognitive thinking process and writings skills with the computers’ interpretation with regards to humans’ knowledge and our capabilities in identifying motive or emotions. Also, unwinding the term ‘Black Box’ will also be deeply explained how machine learning models unwrap and replicate the ability of human to achieve high levels of ‘intelligence’. The use of recent framework like long-short term memory and convolution neural network will effectively generalises motive in a more robust manner. Application programming interface like TensorFlow and Keras/Sci-kit learn are also widely used to assist huge computational work in order for the model to operate successfully. A total of three motive imagery, Achievement, Power and Affiliation, were tested on TAT stories, a psychological data set of 17,462 written sentences. Five experimental models are constructed and validated using Precision, Recall, F-measure, accuracy and F-measures. Global Vector Embedding is the ‘twisting’ factor for achieve more promising results. Our finalized model was able to achieve a result of 67.5% F-measure with a validation loss of 0.21 for Achievement, 42.5% F-measure with a validation loss of 0.31 for Power, 49.9% F-measure with a validation loss of 0.31 for Affiliation. It successfully demonstrates the impact of using pre-trained word embeddings in binary classification task for running text. The evaluation results of difference models will demonstrate their suitability in terms of application in natural language processing.