Techniques for scoring motive imagery in text (part A)

• Main Author: Oh, Yicong
• 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/10356/77547
• Institution: Nanyang Technological University

This research aims to develop an automated system for scoring motive imagery in text using modern natural language processing techniques and state-of-the-art machine learning model. Motive imagery is defined as an action, wish or concern the speaker attributes to himself or others. (Winter, 1977). There are three forms of motive imagery that we are concern with: power, achievement and affiliation. Traditionally, motive scoring is done by trained personnel which requires intensive training (avg. 20 hours) to master a single motive imagery. The process for scoring motive imagery also requires a lot of time and effort. (Blankenship, 2010). Thus, there is a need to develop a system to automate this tedious process. Attempts using computers for motive scoring started in 1965 by Litwin and Williamson using dictionaries and complex decision trees which had limited success due to complexity of motive imagery and language. Recently, Marc Halusic developed a Maximum Synset-to-Sentence Relatedness (MSSR) system using deep neural network to derive a list of achievement corelated words. Despite using natural language processing and deep learning techniques, his framework is based on utilizing dictionaries for motive scoring which shared similar flaws as Litwin and Williamson system back in the 1960s. With intensive research, we have developed a BLAMCE framework for scoring motive imagery which outperformed traditional rule-based method by 1.66 times and even surpassed state-of-the-art deep learning model, BERT, the current leader in NLP applications. It uses a rule-based system for content extraction (features engineering), GloVe and ELMo embeddings for state representations and LSTM-Attention model to learn from its dependencies (words). The outstanding results highlight the importance of rule-based system and showed that its integration with modern machine learning model can greatly enhance its performance.