Automating Human's Cognitive Psychology for Opinion Mining Models

Abstract

This thesis focuses on the evaluation of automated reading comprehension of sentiment in text, which is also considered an opinion mining classification task. Previous work for opinion mining uses feature engineering machine learning (ML) or deep learning (DL) without consideration for the method’s adaptation to the human’s cognitive process. The aim of this thesis is to determine whether machines can learn better by following the human cognitive reading process or rather follow a machine-specific representation. The main difference lies in the intermediate representation of the data before classification. On the human side, and based on recent psychological studies, it has been determined that reading comprehension is not the result of one single process as was thought before 1970s. Instead, psychologists realized that a combination of several complex cognitive processes are involved. Based on Cognitive Psychology, comprehension heavily depends on inference and background knowledge to construct a Situation model, which is a mental representation of the text. In other words, humans develop an image of the context in their minds before concluding its meaning. To emulate the Human Mental Intermediate (HMI) representation, we propose a Text-to-Image-to-Task (T2I2T) model comprehension by first mapping the input text to an image which provides a semantic representation equivalent to the mental representation of the text being analyzed. From the machine’s learning perspective, we conjecture that machines may not need to learn human-specific representations. Instead, we propose to explore the machine’s ability in developing its own Machine Intermediate Internal (MII) representations through direct end-to-end (E2E) models. To emulate the machine’s cognitive process, MII, we propose the use of Transformer-based Language Models (LM) and show that pre-training acts as a suitable means for the machine to acquire background knowledge comparable to the cognitive psychological Situation model. To compare the performances of HMI and MII E2E model, we conducted experiments with applications to sentiment analysis. We developed our own data set with text-image-sentiment annotations by augmenting an exiting image captioning dataset with automated sentiment annotations. Several base models were developed for comparison including Bidirectional LSTMs with word embeddings and state of the art pre-trained LMs, such as BERT and ULMFit. The results showed that the machine’s E2E cognitive approach, MII, outperforms both LSTMs with word embeddings models and the human’s cognitive T2I2T approach, HMI, by 6\% and 26\% respectively. The thesis also explored models to represent HMI and MII for Arabic. In particular, we developed models for HMI Image Captioning in Arabic and an Arabic MII universal language model, called hULMunA. In Arabic Image Captioning (AIC), we developed the first Arabic dataset and encoder-decoder end-to-end models to show that it is necessary to build language specific datasets and end-to-end models rather that translating English captions. In hULMunA, we developed the first Arabic specific Language Model and fine-tune it to achieve state-of-the-art results on four Arabic Sentiment Analysis datasets.