Literature + Machine Learning
For a long time, the only intersection between ML and literature was as a prominent theme in science fiction writing. This has since changed, and today the commodification of books relies heavily on ML to market, promote and sell products. Further, ML is affecting the ways that books are consumed, by both humans and computers.
ML’s impact on the book industry
Artificial intelligence and machine learning have been intertwined with literature for decades as influential themes within the science fiction genre. Isaac Asimov’s 1950 collection of short stories, I, Robot has become a classic that changed readers' perception of AI at a time when the concept was very new. Recent science fiction literature includes AI as a prominent theme, to a varying degree of realism and probability.
However, books are no longer just written about machines (and on machines), but are also being written by machines. Developers have successfully trained deep networks to generate coherent text used for advertisements and reports using language models such as GPT-2 (Van Heerden & Bas, 2021). Conversely, creating a piece of literature requires creativity, nuance and coherence, which machines have not yet successfully mastered. It is likely only a matter of time until a program creates a bestselling novel (Grace et al., 2018), by which point the book industry will have to reckon with questions such as how ML literature will be credited, monetized and held accountable, how the structure of publishers may change, and possible legal considerations such as copyright and plagiarism (Van Heerden & Bas, 2021).
While machine learners may not yet be creating coherent work independently, humans are using ML to collaborate on literary works, including novels, plays and poetry collections (Van Heerden & Bas, 2021). An often cited example is 1 the Road, created by Ross Goodwin and an artificial neural network. This book was generated by an AI writing machine that used data inputs from a camera, GPS, microphone and clock attached to a car that was travelling from New York to New Orleans that fed into a system of long short-term memory neural networks trained on books and Foursquare location data (Merchant, 2018). The text was written completely by the machine, but the data inputs were being manipulated by the human driving the car.
Goodwin input literary and logistical data into machine learning algorithms to compose a novel, however Stanford University is using a similar technique to spark literary analysis (Han, 2020). Algorithmic literature, or the computational criticism of literature, has emerged as a field of study, where scholars input a corpus into programs, which then output quantitative results (Han, 2020). One benefit to outsourcing this work to ML, as opposed to a human, is the expansive volume of literary works that a program can examine to find patterns and identify outliers (Han, 2020). Computational literary criticism uses algorithms such as clustering, statistical modelling, and topic modelling to provide scholars with a set of tools that guides literary analysis. The results can offer a new perspective, or reinforce past findings (Han, 2020) with the speed and efficiency that only a machine can provide.
ML’s impact on reading
As machine learning has an increasing ability to produce books, ML is also used to support humans in reading books. ML can support learning and assessment by identifying people that have difficulties reading or helping learners overcome difficulties. For example, Huang et al. (2010) have developed a tool that monitors a learner completing activities, such as reading comprehension, and offers user-specific feedback or guidance, when needed. The reinforcement learning program uses inputs from a teacher, and then continues to learn based on interactions with the student (Huang et al., 2010). Decision-making is further improved through a Q-learning algorithm (Huang et al., 2010).
Other ML tools can help identify learning difficulties, such as dyslexia, by monitoring eye movements and reading speed (Kaisar, 2020). Some ML algorithms frequently used to identify dyslexia include support vector machine, Naïve Bayes, neural network, k-nearest neighbor, logistic regression and linear discriminant analysis (Kaisar, 2020). These tools are particularly helpful as more students engage in virtual learning, and educators become more comfortable adopting technologies in the classroom.