computational linguistics, machine learning, and philosophy

Most of my work at the moment involves looking at what it means to say that a neural language model represents something and how this should inform how we interpret the text generated by such models. I have several papers under review and in preparation on the interpretation of neural networks and the metasemantics of algorithmically generated text. Right now, I am working on applying teleosemantics to the simplest neural language models (e.g., Word2Vec, ELMo) and seeing how it can constrain methods of model interpretation and probing. Feel free to get in contact if this is the kind of thing you’re interested in. There’s an extremely rough draft of some of this work here. This is connected to a longer-term project developing an appropriate epistemology for stochastic measuring devices, a category which includes deep neural networks and other computational systems often labelled ‘artificial intelligences’. While the approach I take to the representational capacities of neural networks is informed by teleosemantics, I don’t take teleosemantics to provide an adequate account of the content of machine-generated text. Here are the slides from the Philosophy of Deep Learning conference at NYU.

computation, competence, and cognition

I’m also interested in understanding the ways that generative linguistics integrates (or doesn’t) with the rest of cognitive science. My PhD thesis was on the concept of computation in generative syntax. Computation as discussed in theoretical syntax is a ‘logical’ operation. It is the atemporal relation holding between a lexicon, an operation like merge (in minimalism) or unification (in HPSG, GPSG) and the unbounded set of structures generated. In minimalist syntax, the operation applies first to the most embedded constituent which may be extracted to higher and earlier positions in a sentence leaving traces along the way. It is not clear what this notion of computation has to do with ideas of computation in computational cognitive science. This is an area I still work on. I’m also thinking about applications of sub-regular grammars across cognitive science.

online speech

For a non-negligible portion of the planet, the dominant way they use language is now through text-based platforms like WhatsApp or Twitter. This is in contrast to the previous millennia of primarily face-to-face communication. I’m interested in the ways in which new technology is forcing us to reevaluate the philosophical tools developed to analyse more traditional uses of language. I’m currently writing a chapter on Online Speech for an OUP handbook and co-editing a special issue of the journal Philosophy with Eliot Michaelson dedicated to the philosophical problems raised by online communication.

(good old-fashioned) philosophy of language

I still like to keep up with more traditional philosophy of language. I currently have papers under review on the notion of semantic competence within the common ground framework and modelling discursive power with QUDs.


mallory, f. (2023) Fictionalism about Chatbots, Ergo [forthcoming, penultimate draft]

mallory, f. (2021) Why is Generative Grammar Recursive? Erkenntnis: An International Journal of Scientific Philosophy ISSN 0165-0106. p. 1–15. doi: 10.1007/s10670-021-00492-9  

mallory, f. (2021). Critical Notice: A Spirit of Trust: A Reading of Hegel’s Phenomenology by Robert Brandom (Harvard University Press, 2019). Philosophy, 96(4), 675-682.. doi:10.1017/S0031819121000206

mallory, f. (2020) In Defence of the Reciprocal Turing Test, Minds and Machines, 30:659–680

mallory, f. (2020) Linguistic types are capacity-individuated action-types, Inquiry, 63:9-10, 1123-1148 DOI: 10.1080/0020174X.2020.1772864

mallory, f. (2020) The Case Against Linguistic Palaeontology, Topoi 40, 273–284

drafts under review/revision

  • What do large language models model? In Communicating with AI (eds. Rachel Sterken & Herman Cappelen), OUP [forthcoming]
  • Paper on semantic competence [under review]
  • Paper on discursive injustice [under review]
  • Paper on Wittgenstein and language modelling [under review]
  • Paper on computation in generative linguistics [email for draft]

notes and scraps

Here’s a link to my recent contribution to the Daily Nous Philosophers On discussion of ChatGPT (it also features some cool other people whose contributions you should read)

My Stanford Encyclopedia of Philosophy map, the code used to generate it is on my githib account

Bibliography for Computation in Generative Linguistics talk, 2022

Zellig Harris: blog post on Harris’s alleged antirealism

Some ways to think about linguistic structure (for philosophers)

Some notes on different varieties of merge (made in grad school)

dissertation (2019): a pragmatist interpretation of generative grammar

Theories of generative grammar describe a device for recursively enumerating syntactic structures. Generative theorists deny that such a device operates in real-time. Instead, it is claimed that the device characterises ‘competence’, not ‘performance’. On the surface, it looks like a device that doesn’t run — a function that isn’t computed — isn’t empirically interesting. This thesis is an attempt to make sense of the idea that generative grammar describes computations that don’t actually happen in any temporal sense. I argue that, despite claims to the contrary, generative grammar does not characterise a function in the sense of Marr’s computational level of theorising. Instead, I argue that the function characterised is more like the transition function of a Turing machine rather than any function implemented by the machine. In the process, the thesis discusses the philosophical context in which generative grammar developed, provides analyses of of the roles played by concepts like recursion, computation, and function-in-intension, and discusses the impact of the trend toward the lexicalisation of syntactic information on how computation should be understood in theoretical syntax.

Supervisors: Mark Textor, David Adger, Alex Clark