Before I took this course, if someone were to mention ‘Machine Learning” to me, I would automatically think of computers learning how to interpret speech and ‘talking’ with their human counterparts. For some reason, chatbots and computer generated speech has always fascinated me, and I knew that this was a route I wanted to take for the midterm project.
Intitially, I wanted to create a simple chatbot similar to ELIZA, which was an early model for natural language processing. ELIZA was created by Joseph Weizenbaum, a professor at MIT who wanted to make a chatbot to show the ‘superficiality of communication between man and machine’. Weizenbaum thought that because machines could not experience feelings and emotions attributed to language, conversation between a human and a computer is essentially meaningless. ELIZA was supposed to simulate an electronic therapist, where users would talk with the program to receive mental treatment. Ironically, Weizenbaum discovered that although users were aware that ELIZA was a robot, they projected their own human-like emotions onto the software itself, thereby creating the illusion that ELIZA actually did possess a degree of understanding and human intelligence. In fact, several patients reported very positive feedback in their interactions with ELIZA, stating that it differed very little from talking with an actual, human therapist.
To me, this is especially interesting, because being able to converse with something is a good marker that it possesses some degree of intelligence, especially if the conversation is ‘reactive’, in the sense that the replies are not scripted. However, right before I was about to start with this ELIZA styled chatbot, my psychology professor introduced an interesting psychological phenomenon called Wernicke’s Aphasia, which is suffered by patients who received damage in the area of the brain related to language processing. As a result, they are unable to produce meaningful speech, albeit retaining proper grammar and syntax usage. I thought this was the prefect opportunity to combine the two ideas into one; create a chatbot that simulated the natural language ability of a human being, but imbue it with syndromes similar to those of Wernicke’s Aphasia patients.
The ultimate goal of this project is to experiment with the meaning behind computer generated text. What if the model can ‘learn’ fluent speech, but produce individual words that lack meaning? Would users still want to talk with this bot? Would they project their own interpretations of what the bot is trying to say? Or would they get bored after a few minutes of interaction, and quit the program?
In order to tackle this midterm project, I had originally planned to use word2vec, but realized that a more efficient and accurate approach would be to utilize the spaCy library, as well as a bidirectional model instead. I would feed it a heavy dataset (very likely a novel) so that it could build up a working vocabulary, and then have it generate responses on the fly, according to the user input.
Sources:
https://www.eclecticenergies.com/ego/eliza (ELIZA chatbot)
Acharya, Aninda B. “Wernicke Aphasia.” StatPearls [Internet]., U.S. National Library of Medicine, 16 Jan. 2019, www.ncbi.nlm.nih.gov/books/NBK441951