It was in the early 1960s, on a wave of progress and optimistic faith in technological solutions, that everything seemed to come together. The first automated computing machines were not much over a decade old when researchers seriously considered tools to teach these machines to comprehend human language. Thus, engineers in the computing sciences started working together with linguists. Ideas, even good ideas, need incubation time, however. Other people have to work on these ideas, coming up with new techniques; new people might think of new applications for the same idea but in very different fields. It is a bit like the first steps of human bipeds to leave the ground they were standing on. Apart from special geographical features (like hills and mountains), all man could see was their surrounds. The moment a balloon took people up into the air, however, a completely new view of the familiar surroundings was possible. Similarly, modern technology allows us to delve into subatomic geographies. The same experience is true when going beyond the use of single books as a basis to understand and manipulate language. With the second decade of the second millennium approaching, trillions of words from different sources can be collated and used for computer-based calculationsāwith tools that are available to most people who have nothing more than a simple PC or mobile device with the appropriate application. This then allows a new, different, better-informed (because fully empirical) vision of language. Yet it is much more than thatāthis knowledge can now be harvested to have machine-mediated understanding of spoken utterances; algorithms can now be designed to mimic natural human speech. As a result, we are witnessing a whole new dimension in communication. 1
Consequently, a parallel set of conclusions can be drawn: it can be seen that linguistic knowledge underpins the ability of a computational device to process human language in written or spoken form. Conversely, such electronic devices are getting closer and closer in creating a mirror image to how language is produced, processed and understood thus providing support for the theories of the underlying structure of language, while undermining rival claims: if a form of AI works, this can be seen as a result of successfully turning one theory into practise.
The genesis of the book is a story of coincidental discoveries which, over time, have built up to draw connections that changed the intended outcome several times. While I was preparing my first book, Lexical Priming in Spoken English (Pace-Sigge 2013), I happened to read Steven Levyās 2011 book about the search engine company, Google. As an aside, Levy encourages the readers to have a look at the personal page of Googleās first head of research, Peter Norvig (2017)āto see something devoid of gimmicks: a proper engineerās web page. Curious, I went and had a look, only to find that Norvig himself had made, in a number of his published articles, reference to Ross Michael Quillianāthe very man that I had identified in my book as a key figure in the development of the concept of priming . Looking at the processes to retrieve the best possible search results for any given Google search, as described by Levy, the connection to the concept of lexical priming became quite obvious. This connection has subsequently been described, albeit not in too much detail, in my earlier book.
A few years down the line, my partner suggested I could write a primer on the concept of lexical priming as part of the Palgrave Pivot series. It took, however, another year or two before I had time to think about that project. Yet, as I started to investigate the matter, it became clear that a far more interesting project offered itself: the link between the psycholinguistic theory of lexical priming developed by Michael Hoey (2005) and the current developments in speech recognition and speech production technology which are born out of current advances in AI. For both appear to have a common root in the concepts of the semantic web and spreading activation , first developed by Quillian. The task for Quillian (1969) was to create a theoretical framework explaining how to programme a machine to understand natural human speechāthe Teachable Language Comprehender ( TLC ) as he called it. The core to this task was, for Quillian, to create a form of Word-Sense Disambiguation (WSD). Tellingly, research into what is now referred to as WSD is at the heart of many AI and computational linguistics projects in the twenty-first century. 2 As a consequence, it seems to make sense to write a book that shows the development of the theory, then outline the two strands of research which developed out of it and finally see what these two communities of researchers can learn from each other.
There are, of course, a number of hugely important books available that cover the concepts in this book in far greater detail. First and foremost, the magisterial bible on AI, Stuart Russellās and Peter Norvigās Artificial Intelligence. A modern approach. Originally published in late 1995, the latest updated edition came out in 2016. The book more focussed on the area discussed hereālanguageāis the equally impressive Speech and Language Processing : An Introduction to Natural Language Processing, Computational Linguistics and Speech Recognition by Daniel Jurafsky and James H. Martin. Its first edition was published in 2000, with the second edition announced for 2018. One might want to take a shortcutāas the publisher, Prentice Hall (now Pearson) must have thoughtāand go for the 2014 book Speech and Language Processing by Jurafsky, Martin, Norvig and Russell. All three bo...