Foundational Issues Seminar (TT 24)
Undergraduate seminar on long-distance dependencies
Colin Phillips|
Trinity Term 2024
Tuesdays, 11:00 – 12:30, Weeks 1-3, 6-8, Clarendon Institute Lecture Theatre
This is an interactive undergraduate seminar that aims to explore links between different foundational topics in linguistics and language science, connecting them to recent research, especially in the light of the greatly expanded datasets and analyses now available to linguists.
The seminar will pay particular attention to the role of long-distance relations in sentence structure (“long-distance” or “unbounded” dependencies). Discoveries about these phenomena played an important role in early modern linguistics (1960s-1980s), especially in motivating claims about human specialisation for language. These arguments are ripe for reevaluation, in light of new data (from many languages) and a wide array of new tools.
The seminar is listed under FHS Paper A (“General Linguistics”), which is primarily covered through tutorials rather than lectures. This is NOT a series of Paper A lectures. Rather, it aligns with Paper A goals of exploring foundational questions in linguistics, and integrating different perspectives and approaches. Key themes include constraints on language diversity, innateness and critical periods, levels of analysis, big data and large language models, and connections between speaking and understanding. The seminar will encourage active student participation via targeted readings and brief writing/reflection exercises. We aim to closely evaluate the relation between specific evidence and theories/models. By highlighting recent advances in data and tools the seminar may be of interest to students who are considering future research in linguistics.
Materials from the seminar will be shared here once they are available. Meetings will not be recorded, in order to facilitate discussion.
Week 1 – Scene Setting
Theme: setting the scene; why we care; early findings, especially from 60s-80s.
Why we care: island phenomena have played a (justifiably) prominent role in linguistic theories over the past 60 years. In the 1960s, linguists started discovering a goldmine of regularities in human language(s) that they hadn’t been aware of before. Educated people had been studying the details of languages for centuries, and yet here were these complex-yet-robust generalizations about languages that hadn’t been noticed until people started to construct explicit, generative grammars. Surely this must be telling us something important about human cognition.
Follow-up: thanks for coming along in Week 1. We had a useful discussion of some core island phenomena, and addressed the appeal of linking constraints on cross-language variation to constraints on learners. We also did some calibration around student backgrounds and relevant experience.
Recommended reading for Week 1: Phillips 2013. On the nature of island constraints. I: Language processing and reductionist accounts.
The first half of this article gives a summary of basic island phenomena and variation across languages. We’ll be coming back to these and related examples repeatedly.
Slides: Foundations Week 1. Note: we did not cover all of this, and not in order. In particular, we skipped a section about the earliest attempts to classify and explain island effects, which is nicely summarised in the early sections of Haj Ross’s 1967 PhD dissertation. Worth a look!
Week 2 – Levels of analysis and reductionism
Theme: are island effects epiphenomena of human cognitive limitations?
Why we care: for as long as linguists have been studying island effects, there has been a debate about whether they should be understood as consequences of explicit grammatical rules/constraints or as consequences of more general cognitive limitations. This long-standing debate is a useful one for us to sink our teeth into, because it forces us to be as clear as possible about the link between specific cognitive hypotheses and their empirical consequences.
This debate is partly about figuring out what is the right level of analysis for explaining island effects. In this context, it’s useful to be familiar with vision scientist David Marr’s famous distinction between computational, algorithmic, and implementational levels of analysis for explaining neurocognitive phenomena (Marr 1980).
Discussion note: Take a look at my summary of arguments for/against reductionist accounts of island phenomena in the second half of Phillips 2013. Also, feel very free to take a look at something very new on this topic, such as Jiayi Lu’s new synthesis of findings on syntactic satiation effects (Lu 2024). I also plan to look at a brand new piece by Giacomo Presotto that arrived in my mailbox this week. Question: what arguments/evidence do you take to be more or less compelling for or against a reductionist account of island phenomena? You don’t have to give a full ranking, but it will be very useful if you take a position. To be clear, you should feel no obligation to be persuaded by evidence that I or my collaborators introduced. We are fair game for criticism, and we always learn useful things when we are wrong. Please send me an email with your thoughts by 1 hour before the seminar time, i.e., 10am on Tuesday April 30th. This will give me a chance to read your thoughts before the seminar, and we’ll then have a better discussion among the group. If you have questions about something that you read that is unclear to you, please feel free to mention that. We will be reading and discussing some challenging material in the coming weeks (a good thing!).
One important early contribution in this debate is Janet Fodor’s 1978 paper Parsing strategies and constraints on transformations.
Slides: Foundations Week 2. Note that we did not cover all of the material here. Not even close. This is fine. We can return to some of it later.
Week 3 – Levels of analysis and mentalistic commitments
Theme: Describing neurocognitive systems at different levels of analysis (“Marr levels”). The relationship between acceptability and representability. Active dependency formation.
Slides: Foundations Week 3. Note that we had a good discussion in the session, with the consequence that we did not cover a lot of the material in the slides. This is fine, and we can return to some of it later on.
Week 4 – No meeting
CP presented at a memorial event for Janet Fodor, an important figure in the literature on island phenomena (and a graduate from PPP, the precursor to the current PPL degree).
The slides from the presentation, Janet Dean’s Path to Psycholinguistics, are available. They provide an interesting insight into the beginning steps of an academic career, using materials that have never been seen before.
Week 5 – No meeting
CP was the speaker at a Linguistics graduation ceremony in Maryland, where he spoke about the relation between research and public engagement, ahead of attending his daughter’s graduation ceremony later the same day. Sorry, this was something that just couldn’t be missed!
Week 6 – Learning models and LLMs
Theme: Learnability, poverty of the stimulus, and island constraints. What do recent computational models of learning island constraints tell us about innate learning biases?
Read: Pearl & Sprouse 2013 is a surprisingly simple computational model of learning. Wilcox et al. 2023 is a provocative new paper that uses Large Language Models (LLMs) to argue against famous arguments for innate learning biases based on Poverty of the Stimulus considerations.
Question: I failed in posting this to the website in good time (it was in the slides in Week 3, but I neglected to add it here). So there’s little time to seriously consider this. But give it some thought ahead of time if you can. … For decades, linguists have carefully analysed diverse constructions and languages in search of principles/constraints that explain consistent mastery of island effects. How do the learning arguments in this week’s readings change that project, if at all?
Slides: Foundations Week 6. In this instance, the slides here do match relatively well what we discussed.
Week 7 – Limits of variation
Topic: we will continue our discussion of learning models, this time starting with Large Language Models (LLMs), which are powerful but opaque, and then starting to look at how much scope for variation there is in island effects across languages and constructions.
Readings
Wilcox et al. 2023. This new paper argues, in effect, that LLMs undo a classic argument for innate linguistic constraints. Using measures borrowed from psycholinguistics, the authors argue that LLMs show knowledge of island effects, despite their lack of built-in biases related to islands. You probably didn’t have a chance to read this last week, so it’s worth taking a look before this week’s discussion. I found that it took me a little work to figure out what the figures were showing, but it was worthwhile.
Kush et al. 2021. This is a paper that sounds a bit obscure. It’s about the distribution of different kinds of rare constructions in Norwegian children’s books. But I find it both interesting and important. Why? Because the authors take a close look at the relation between acceptability judgments and input, in relation to an “exotic” property of Norwegian. They then use this to evaluate different possible learning algorithms.
(In our most recent class meeting, I also suggested to read McCloskey 2000, an article about quantifier float in a Northern Irish variety of English. It’s a really interesting paper, worth a read. But I would recommend to prioritize the papers above.)
Question: A long-standing argument about islands and innateness derives from their cross-linguistic robustness: if all languages show the same constraints, then surely this is not the result of learning. The current picture seems to mostly support the cross-linguistic generalisation, but not fully. What might this mean for classic arguments about learning and innateness? You could try to approach this in light of the concrete case of the Kush et al. study about Norwegian islands.
As usual, please send thoughts via email by 10am on Tuesday June 4th (7th week).
Slides from Week 7. Usual caveats apply. We also discussed the Kush et al. 2021 paper, but there are no slides on it here.
Week 8 – Comprehension-production relations
Topic: An important question about the human architecture for language is how much specialisation there is for tasks such as speaking and understanding. Are they subserved by the same cognitive machinery, or do we have distinct cognitive machinery that is dedicated to each task? Also, if so, how does the specialised machinery draw on the same basic grammar for the language, i.e., we speak and understand the same language.
Readings: A very interesting recent line of research on the production of wh-questions comes from Shota Momma (University of Massachusetts). In this work, he makes clever use of syntactic priming paradigms to argue for an account of wh-question production based on Tree Adjoining Grammar (TAG). I recommend to try to read Momma (2021) and EITHER Momma (2022) OR Momma (2023). I will try to cover all of them in our session on Tuesday. It takes a little work to get your head around the arguments, but I promise that it’s worth it!
As background to this topic, you could check out a review article on comprehension-production relations that Shota and I wrote a few years ago: Momma & Phillips (2018). It lays out an overall framework for thinking about the relation between speaking and understanding. But the newer findings about the details of producing wh-questions move enrich the discussion significantly.
Question: Summarise the key steps of Momma’s argument that wh-question production depends on a TAG-based grammatical formalism. How convincing do you find this? How much of the work is being done by the TAG formalism vs. by the nature of the task of mapping from a message to an utterance?
As usual, please send thoughts via email by 10am on Tuesday June 11th (8th week).
Slides from Week 8. To be added.
Recent Comments