1 of 60

Building the Linguistic Telescope

DANIEL PLESNIAK FEB 19^TH, 2022

THE SECOND ANNUAL WORKSHOP ON LANGUAGE FACULTY SCIENCE

PLESNIAK@USC.EDU

2 of 60

Introduction

3 of 60

Overview

The goal of this talk is to shed some light on recent attempts to develop a new experimental tool for research on the language faculty.

I liken this tool to the telescope, which, as we know, played a key role in the development of astronomy and physics.

Our “linguistic telescope” is not primarily a physical device, but rather a mental one, but this is simply a consequence of what each telescope is designed to “look at”.

By the end of this talk, I hope you will feel as I do that such a device is not only possible and hypothetically useful, but in fact well on its way to being built.

4 of 60

Roadmap (1)

In Section 1, I will explain what I believe to be the relevant features of the astronomical telescope for our purposes.

In Section 2, I will discuss theoretical linguistics’ historical lack of a telescope-equivalent (and why we might want one).

In Section 3, I will overview the Language Faculty Science (LFS) program initiated by Hajime Hoji and how it has led to advances in our “telescope-development” program.

5 of 60

Roadmap (2)

In Section 4, I will provide a summary of some of the more technical aspects of the ongoing project.

Finally, in Section 5, I will briefly go over some promising results, which are drawn from my recently completed dissertation.

This should hopefully highlight a variety of ways in which we have been making progress towards developing a functional linguistic telescope.

The linguistic telescope is by no means complete; I hope this talk can inspire some listeners to be curious about its development and to perhaps even want to take part in it (if they are not already doing so).

Feel free to add your name to our mailing list: https://forms.gle/dva5f9sd27hnMmqE9

6 of 60

Section 1: �The Telescope

7 of 60

The Power of Sight

The telescope is one of the most ubiquitous scientific tools in the modern world.

Not many kids ask their parents for a powerful thermometer or an oscilloscope, for example, even though those might be both cheaper and more practical!

Why? Perhaps because the most basic way of observing the world around us is using our senses.

Most scientific devices yield measurements that take a degree of abstract thinking to understand.

A telescope simply appears to enhance our vision, letting us directly see things we weren’t previously capable of.

Infrared Thermometer

Oscilloscope

8 of 60

Jupiter: Naked Eye vs. Telescope

Jupiter

Moons!

9 of 60

A Historical Question

Telescope-facilitated discoveries like the moons of Jupiter revolutionized our understanding of the universe.

This began essentially as soon as the telescope was invented, around the early 1600’s. (And the telescopes then were pretty “bad” compared to even mass-market ones today.)

This raises an interesting question though: people have intensely studied the stars since before recorded history and have known how to make the ingredients of a telescope (e.g., glass) for thousands of years.

The ancient Babylonians surely would have loved a telescope, so why didn’t they make one?

The answer turns out to be that they most likely lacked both the theoretical and implementational tools to create a functional telescope.

Glass from Roman-era Pompeii

Babylonian star catalogue

10 of 60

Theory (Optics)

A telescope relies on a precise mathematical relation between the object observed, the lenses in the telescope, and the human eye.

To achieve this, one needs a rather sophisticated knowledge of how light moves, interacts with surfaces of various properties, and is ultimately converted by the eye(/visual system) into what we see.

The relevant sophistication of optical theory seems only to have been reached in late-medieval/early renaissance Europe

This was achieved by investigations of scholars like Roger Bacon and subsequent experimentation by artisans involved in the production of optical devices like spectacles.

Painting by Conrad von Soest, 1403

11 of 60

Implementation (Glass Lenses)

As it turns out, creating a piece of glass curved to precise specifications free of defects is very hard.

For example, the YouTube channel “How to Make Everything” tries to make various things ”from scratch”.

An attempt was made at a telescope; despite using advanced optical theory, modern glass-shaping tools, and not even making the glass from scratch, persistent lens-based issues resulted in an unusable device

Being able to produce lenses to precise geometric configurations, with sufficient transparency and no internal bubbles, takes a great deal of technical sophistication and experimentation.

A biconvex lens

The moon, apparently

12 of 60

Just “Seeing”?

As it turns out, a telescope is not really just a more powerful version of the naked eye.

The image we see from a telescope is highly “processed” according to the theory and implementational techniques that go into its design.

It is perhaps best to think of the telescope as a measurement device whose output is a pattern of light; unless we are confident in the validity of the steps that produce that pattern, the light may be essentially meaningless.

None of this is to say that telescopes are not great tools; they simply have the baggage that comes with all scientific measuring devices.

A disassembled telescope

13 of 60

(Astronomical) Telescope Summary

Telescopes allow for powerful and easy to understand investigation of the physical world, because they let us clearly “see” things otherwise hidden from our vision.

Nevertheless, making a telescope that yields reliable data requires both optical theory and precise construction techniques.

Simple results ≠ Simple way of obtaining those results

Telescope

14 of 60

Section 2: �Linguistic Intuitions

15 of 60

Chomsky’s Insight

Chomsky’s perhaps most important contribution to science was his observation that languages are not the abstract properties of speech communities, but are instantiated in the mind of each individual.

Humans are able to map between linguistic sounds(/signs/symbols) and meanings, and there must be something(s) in our minds that let us do that.

We call that part of each person their “Language Faculty”, and we want to know what its properties are.

This is especially true in the universal sense; what properties define the human language faculty and are thus common to the language faculties of all humans?

16 of 60

Internal “Sight”

Following Chomsky, generative linguistics has taken linguistic acceptability (sometimes called grammaticality) judgements as its primary data.

Each individual can ask themself whether a given combination of sounds, pairing of sentence and interpretation, etc., is possible in their language, according to the intuitions coming from their language faculty.

These judgements are direct sensations, letting us “see” the output of our language faculties with our own internal “naked eye”.

17 of 60

The Naked Eye Search

By using our “internal eye” of intuitions, we can investigate our language faculties by examining what they tell us is/is not possible to accept.

In particular, to the extent that a given hypothesized property of the language faculty makes predictions about what should/should not be acceptable, we can test that theory against our own intuitions.

Ideally, such investigations would result in patterns of (un)acceptability that are replicable across all (relevant) individuals, indeed demonstrating universal language faculty properties.

18 of 60

Limitations

Unfortunately, depending on the domain of investigation, such reliable patterns often fail to emerge.

From here, I will confine my remarks to issues in the domain of sentence-interpretation pairs, though I suspect they will be applicable elsewhere.

Most problematically, different individuals, and even the same individual at different times, frequently disagree on the acceptability of a given sentence-interpretation pair.

In such cases, it seems our “naked eye” observations are insufficiently powerful to investigate the hypothesized properties of the language faculty.

19 of 60

A Crucial Example

Chomsky’s long-held contention is that all language faculties represent sentences as corresponding to an abstract syntactic structure, and that this structure is a basic input into the computation of the sentence meaning.

We can call the part of the language faculty that computes these structural representations the Computational System (CS).

Based on Chomsky 1995’s “Merge” hypothesis, we expect all sentences to be built/represented in the CS by binary-branching set-based hierarchical structures.

Basically, syntactic structure is built up by combining two units at a time into bigger units, which can then recursively combine.
For example, ‘every man praised his mother’ might be built in the CS as on the right. (I am simplifying a bit; for example, the nodes should be sets, not strings of words)

20 of 60

Meaning Relations

There is an initial hurdle, namely that Chomsky’s theory by itself doesn’t make any predictions about un(acceptability).

We can’t ”see” our structural representations (they’re unconscious), so we can’t directly sense whether ‘every man praised his mother’ really is represented as on the right.

This is actually not so different from naked-eye astronomy though: we can’t “see” stars either, we can see the light stars emit.

So, we need to find something equivalent to “light” for syntactic structures.

The answer is going to be what we can call a “meaning relation” (MR), basically any type of interpretation between two elements in a sentence that is affected by how those elements are situated with regard to one another in the abstract sentence structure

21 of 60

Bound Variable Anaphora

One MR with which we will be concerned, “bound variable anaphora” (BVA), has a long history of discussion in the generative linguistics literature (and even earlier in logic).

BVA obtains when the individual “referred to” by one element in the sentence varies across the members of a quantified set.

This is what we see with ‘every man’ and ‘his’ in the most natural interpretation of ‘every man praised his mother’, i.e., that each man praised his own mother (so ‘his’ doesn’t refer to a particular man but to each man in turn).
For notational convenience, I will refer to this reading as BVA(every man, his), an instance of the more general BVA(X, Y), where X is any quantified �expression and Y is the element that varies with X.

An influential theory tracing back to Reinhart 1983 claims that BVA(X, Y) is indeed only possible under a very specific structural condition.

22 of 60

C-command

The Reinhartian claim is that BVA(X, Y) requires X to “c-command” Y in the structural representation of the sentence.

X c-commanding Y essentially means that X combines with something that contains Y.
In the structure on right, ‘every man’ combines with ‘praised his mother’, which contains ‘his’, so ’every man’ does indeed c-command ‘his’.
As predicted, BVA(every man, his) is essentially universally judged acceptable as an interpretation of that sentence.

If we reverse ‘every man’ and ‘his mother’, now ‘every man’ combines with ‘praised’, which does not contain ‘his’, leading to the prediction that BVA(every man, his) should not be possible.

The BVA reading would be something like “every man was praised by his own mother”.

23 of 60

The light of structure?

The generative syntax literature has frequently argued that this prediction is borne out.

Indeed, not only is BVA(every man, his) impossible for ’his mother praised every man’, but (it is claimed), that for all sentences, for any speaker of any language, BVA(X, Y) is never possible if X does not c-command Y (or something quite like it).

If this is the case, the BVA truly is the “light” of structure; where BVA(X, Y) is possible, we have a structure where X combines with something that contains Y, and where it is not, we have a structure where X combines elsewhere.

Under that assumption, we can indeed “see” syntactic structure through �our judgements; all we need to do is to check for the possibility of the �predicted BVA readings, and that will tell us what structure we have.

Just as what light we see tells us about what stars are where.

24 of 60

The Problem

As I have indicated though, judgements are not really that clearcut.

Many individuals accept sentences like “his mother praised very man” with a BVA reading, and as mentioned, different individuals have different judgements at different times.

This does not seem to be a particular property of this sentence type, or of English, or even of BVA.
I would wager that essentially any straightforward prediction to the effect that a given MR is possible only in a given structural configuration can be falsified with sufficient checking.

The hope of “seeing” structure with our intuitions about MR availability is not dead, but we clearly need something better than our “naked eye”; we need a linguistic telescope.

25 of 60

Section 3: �The LFS Program

26 of 60

Designing The Linguistic Telescope

The main goal: we want to be able to preserve our ability to “see” through our intuitions (that is, to sense in a direct way the truth or falsity �of a given prediction), but we need to somehow “enhance” that ability.

Looking at the “light” of MR’s like BVA was a good start, but it proved too chaotic. We did not have a good enough “resolution” to consistently see what was going on.

The reason we can’t see Jupiter’s moons is that they’re too small. Is syntactic structure “too small” somehow?

The answer is going to be “yes”, in the sense that it is being overwhelmed by other things.
(This is actually quite parallel to Jupiter’s moons, as some of them could be visible to the naked eye if Jupiter weren’t outshining them so brightly)

Jupiter’s moons are much dimmer than it is

27 of 60

MR as an Output (1)

Why would BVA(X, Y) require X to c-command Y in the first place?

Presumably, it’s because somewhere in the language faculty, there’s an interpretative function that takes as its input structures where X c-commands Y and outputs BVA(X, Y) interpretations, as below:

)🡪 BVA(X, Y)

28 of 60

MR as an Output (2)

From this formulation, we can ask a number of questions:

1. What if f() has more domain restrictions than just X c-commanding Y?
2. What if someone fails to apply f() to a structure where X c-commands Y?
3. What if there is another function, g(), whose output is also BVA(X, Y)?

Different things would happen in each case:

1. might yield systematic failure to get BVA(X, Y) when X c-commands Y.
2. might yield fluctuations between individuals/times as to whether BVA(X, Y) is available when X c-commands Y.
3. might yield cases where BVA(X, Y) is possible when X does not c-command Y, presuming that g()’s domain is different than f()’s.

)🡪 BVA(X, Y)

g(?)🡪BVA(X, Y)

29 of 60

Seeing Through Variation

The first two possibilities need not concern us if we simply focus on negative predictions (e.g., “BVA(X, Y) will be impossible if ___”)

This is actually what Hoji’s 2015 “Language Faculty Science” book does.

The third possibility, however, means that we cannot rely on BVA (or in theory any MR) to always give us results in accordance with structural constraints.

We therefore need some way of distinguishing which instances of a given MR are based on c-command and/or blocking the non-c-command-based sources of that MR.

30 of 60

Non-Formal Sources (1)

In our forthcoming book, “The Theory and Practice of Language Faculty Science” (edited by Hajime Hoji, Yukinori Takubo, and myself, with chapters/segments by the editors as well as Ayumi Ueyama and Emi Mukai), we lay out a comprehensive program to deal with non-c-command-based MR.

The core challenge is that the relevant g() functions seem to be conditioned on relatively “fuzzy” sources that do not have strictly formal properties, but which are more pragmatic/cognitive in nature.

For example, in her chapter, Ueyama notes that non-c-command BVA(X, Y) is affected by things like ”whether X can be understood as a topic” or “whether Y can be understood as non-individual denoting”.
These seem sufficiently dependent on the cognitive state of the speaker that they cannot be directly controlled.

31 of 60

Non-Formal Sources (2)

Hoji calls such a cognitively conditioned source of MR a “non-formal source” (NFS).

His chapters detail how such NFS’s can be circumvented.

Essentially, though we cannot control them directly, NFS’s are not random, and NFS-based MR’s leave distinct signatures in the judgement patterns of the affected individuals.

If we detect those NFS signatures, then the acceptability of something like BVA(X, Y) may not be informative about the structure of the sentence in question.
If, however, we do not detect those signatures, then we know that no NFS is at work, so BVA(X, Y) must be the result of X c-commanding Y.

In this way, we can zoom in on the “light” of c-command-based MR, revealing structure clearly and directly via our judgement-based intuitions.

32 of 60

Telescope Comparison

Property\Device	Astronomical Telescope	Linguistic Telescope
Object of Inquiry	Heavenly bodies in different parts of the sky.	Abstract representations of sentences in different human language faculties
Composition	Physical	Mental
Mode of sense	Vision	Intuition
Information transmitted	Light	Acceptability (of MR readings)
Main Challenge	Heavenly bodies are very distant; light appears too small for the eye to perceive many, if any, details. (esp. if there is another source of light)	NFS effects are strong and hard to predict, obscuring structure-based patterns
Solution	Expand the light through lenses, so the eye can properly focus on what is to be observed.	Ensure an environment where NFS effects are not present, so intuitions reflect structural representations.

33 of 60

Section 4: �Determining the Source of Judgements

34 of 60

Detecting NFS-BVA

Assuming everything is kept relatively constant, for a given individual at a given time, NFS effects seem to persist across different MR’s.

That is, if we find an NFS effect allowing c-command-less BVA with a given sentence, there is likely to be an NFS effect allowing other c-command-less MR’s with (roughly) the same sentence (and vice versa).

Further, NFS-based BVA(X, Y) seems to highly depend on the particular choice(s) of X and Y.

We can exploit these two facts to determine whether a given BVA judgement is guaranteed to be based on c-command.

35 of 60

A General Formula

If we are concerned whether NFS-MR1(X, Y) is possible, we can use two other MR’s (MR2 and MR3) to check separately whether X and Y induce NFS effects.

The formal deduction is a little long to show here, but in essence, with the right hypotheses (and a number of assumptions coming from those hypotheses that need to be met), we can derive that:

1. NFS effects are induced by either X or Y.
2. If either NFS-based MR2(X, __) or NFS-based MR3(__, Y) is possible, NFS-based MR1(X, Y) is possible (where __ is standing for any relevant choice of argument)
3. If NFS-based MR2(X, __) and NFS-based MR3(__, Y) are impossible, then NFS-based MR1(X, Y) is impossible.

If this seems hard to follow, do not worry; I’m going to explain it again in a couple slides.

All you need to keep in mind for now is that we are using two other MR’s to test whether a given BVA judgement might be based on an NFS or not.

36 of 60

DR and Coref

To make an effective test for BVA(X, Y), we need two MR’s that are appropriate for the purpose.

Hoji finds that the MR’s DR (”distributive reading”) and Coref (“coreference”) are suitable for testing the properties of X and Y respectively.

DR(X, Y) is an interpretation where there is a different set of Y for each X.

For example, DR(every man, two women) for “Every man praised two women” would mean that each man praised a different set of two women (so if there are six men, twelve women were praised).

Coref(X, Y) is an interpretation where Y simply refers to the same individual as X.

For example Coref(John, his) for “John praised his mother” would mean that it was John’s mother (not someone else’s) who John praised.

37 of 60

NFS’s as Functions (2)

Restating everything in the functional notation I used before, what we essentially have is three functions:

f() which maps from a structure where X c-commands Y to any one of BVA(X, Y), DR(X, Y), or Coref(X, Y)
g1(), which maps from X to BVA(X, Y) or DR(X, Y), but only if X meets certain (cognitive) requirements.
g2(), which maps from Y to BVA(X, Y) or Coref(X, Y), but only if Y meets certain (cognitive) requirements.

Assuming that the “certain requirements” are always the same (for a given g()), then if X does not meet the requirement to be mapped by g1(X) to DR(X, __), then it will not meet them to be mapped to BVA(X, Y).

The same logic holds for g2 with Y and Coref.

)🡪

g1(X)🡪

g2(Y)🡪

if X meets certain requirements

if Y meets certain requirements

38 of 60

NFS’s as Functions (2)

Thus, if we have established (via DR) that X is not a valid input to g1(X) and (via Coref) that Y is not a valid input to g2(Y), then, assuming there are no other functions hanging around, any instance of BVA(X, Y) must be due to f(), which requires X to c-command Y.

In other words, if NFS-based DR is impossible with a given X, and NFS-based Coref is impossible with a given Y, then NFS-based BVA(X, Y) is impossible

And as such, any instance of BVA(X, Y) must be based on X c-commanding Y.

You may still be reasonably confused, so let me provide an example. (I will gloss over certain technical points for ease of exposition)

)🡪

g1(X)🡪

g2(Y)🡪

if X meets certain requirements

if Y meets certain requirements

39 of 60

(Slightly Simplified) Example (1)

In each of the examples on the right, the a. case is one where X of MR(X, Y) does c-command Y, while the b. case is one where X does not c-command Y.

It should be uncontroversial that an individual accepts the MR’s paired with the a. cases.

The key concern is whether ‘every man’ and ‘his’ induce NFS effects or not, (for the individual in question, at the particular time, with these types of sentences, etc.)

(1) With DR(every man, two women)

a. Every man praised two women.

b. Two women praised every man.�

(2) With Coref(John, his)

a. John praised his mother.

b. His mother praised John.�

(3) With BVA(every man, his)�a. Every man praised his mother.

b. His mother praised every man.

40 of 60

(Slightly Simplified) Example (2)

If the individual in question consistently does not accept sentences like (1b) with the DR(every man, __) interpretation, then we can conclude that ‘every man’ is not inducing any NFS effects.

Otherwise, they should have been able to accept it via NFS-based DR!
Similarly, if that individual does not accept sentences like (2b) with the Coref(__, his) interpretation, then we can conclude the same thing about ‘his’.

In such a case, we know for certain that (3b) will not be accepted with a BVA(X, Y) reading, as neither X nor Y can induce an NFS effect, so BVA(X, Y) will require X to c-command Y.

(1) With DR(every man, two women)

a. Every man praised two women.

b. Two women praised every man.�

(2) With Coref(John, his)

a. John praised his mother.

b. His mother praised John.�

(3) With BVA(every man, his)�a. Every man praised his mother.

b. His mother praised every man.

41 of 60

(Slightly Simplified) Example (3)

We have consulted hundreds, if not thousands, of individuals at this point, and we find that this pattern always holds.

Again, simplifying a bit, if someone rejects (1b) with DR and (2b) with Coref, that person always rejects (3b) with BVA.
Likewise, if someone accepts (3b) with BVA, then they accept either (1b) with DR and/or (2b) with Coref

I will show you some evidence for this claim in the next section, but I recommend you check your own judgements when you get a chance (in your own native language, of course).

Feel free to vary choices of X and Y (as well as anything else about the sentences) and see what happens.
What we have seen is that your individual judgements may change, but the pattern of correlations between them will remain constant.

(1) With DR(every man, two women)

a. Every man praised two women.

b. Two women praised every man.�

(2) With Coref(John, his)

a. John praised his mother.

b. His mother praised John.�

(3) With BVA(every man, his)�a. Every man praised his mother.

b. His mother praised every man.

42 of 60

An “Optics” for BVA

What the preceding slides have discussed is essentially (the beginnings of) a theory of optics for judgements on sentence-MR pairs.

Specifically, it relates judgements on two other MR’s, DR and Coref, to judgements on BVA.
DR and Coref let us detect (and therefore avoid) NFS effects, allowing us to zoom in on structure-based BVA.

We may see these secondary MR’s as the lenses in our telescope, helping us to control which pieces of light eventually make it to our ”eye”, a.k.a. our judgements on BVA.

This process allows us to “see” syntactic structure in a simple and reliable way through our (telescope-assisted) intuitions, without the obscuring effects of NFS’s that would normally introduce the kind of noise that leads to seemingly chaotic judgement variation.

43 of 60

Section 5�Some Recent Progress

44 of 60

Progress on the Telescope

The basic summary of the LFS telescope design given in the previous sections gives only a brief outline of some of the topics discussed in the forthcoming edited volume.

However, at this point, I would like to pivot to talk about my dissertation, as it builds on some of the concepts discussed in ways I think are illustrative of current areas of focus in the ongoing telescope construction program.

This discussion will not be as technical as in the previous section!

The dissertation is titled “Towards a Correlational Law of Language: Three Factors Constraining Judgement Variation”

I will explain the significance of that title shortly.

45 of 60

Implementation

In Section 1, I described a telescope as requiring both significant theoretical and implementational development in order to be made functional.

I have not talked much about the actual implementation of the linguistic telescope though!

The implementational challenges are most severe when dealing with individuals who are not the experimenter themself, e.g., participants in a traditional “experiment” setting:

These participants may not understand the questions they are being asked, unlike in a self-experiment, where the experimenter understands perfectly (at least in theory).
These participants can only judge a small number of preset sentences, whereas in self-experiment, we can keep going for as long as necessary.
These participants may have language faculties with significantly different properties than the experimenter’s (e.g., they speak Japanese instead of English), making it hard to anticipate all relevant possible issues.

Such non-self replication is necessary, however, if we want to demonstrate a universal hypothesis.

46 of 60

The Previous Approach (1)

Earlier experiments (of both mine and Hoji’s) used two basic strategies to implement the relevant fragment of the telescope in an experimental form.

First, following Hoji 2015, we relied on a number of “sub-experiments” which checked whether a given participant had the relevant properties for their judgements to be considered reliable for our purposes.

That included such things as whether they were attentive, understood the directions as intended, and other more specific issues.
If a given participant did not pass all the (relevant) tests, their answers were deemed unreliable and were excluded from the final analysis.

Second, we checked whether the X and Y of BVA(X, Y) being used were, for the participant in question, free from NFS effects, as determined by their judgements on DR(X, __) and Coref(__, Y)

If they were not, then once again, the participant’s judgements were (mostly) set aside, at least for that particular choice of X and Y, as they would not let us “zoom in” on the effects of structure.

47 of 60

The Previous Approach (2)

These strategies were highly effective; no individuals whose judgements were not set aside by the previous two strategies ever accepted BVA(X, Y) when X did not c-command Y.

However, very few individuals made it through the relevant tests.

In one of Hoji’s Japanese experiments, despite there being several different choices of X and Y examined, of nearly 200 individuals, only a dozen individuals ever produced judgements that were deemed both reliable and free of NFS effects.

Such results, while scientifically useful, are practically quite inefficient.

They may also raise concerns about the purported “universality” of the demonstrated properties, if they can only be demonstrated in <10% of the individuals surveyed.

48 of 60

Dissertation Goals

1. To improve the implementation of the telescope so as to reduce the number of participants that have only inconclusive results.

2. To demonstrate more comprehensively the empirically discovered “law of BVA” that we had been using in our telescope design.

3. To show that the telescope as it stands now can successfully detect structure-based BVA across a broad range of conditions not previously tested, including new sentence types, new varieties of X and Y of BVA(X, Y), and most crucially, new languages.

I will briefly discuss each of these three goals in turn.

49 of 60

1. Improving Efficiency (1)

There were two major steps I took to improve the efficiency of the telescope:

First, as seen on the right, I switched from a primarily “text-based” display of sentence-MR pairing to be judged to a visual one, which seemed to improve comprehension.

Second, rather than previous online surveys of hundreds of people, I instead walked participants through the experiments in one-on-one Zoom sessions

These each took about an hour, so I was able to gather data from far fewer individuals (a bit over 30 total), but these participants were able to hear me explain the instructions and ask questions.

An earlier “text-based” experiment of mine

The newer way of displaying MR “visually”

50 of 60

1. Improving Efficiency (2)

These two changes resulted in participants being much more reliably able to understand the intended tasks, meaning that eliminative sub-experiments could essentially be done away with.

As a result of all these changes, though the number of total individuals surveyed went down, the total number of individuals giving “usable” data actually went up by a large margin.

Nearly all informants gave ”usable” data by the standards of the telescope, so the rate of inclusion was almost 100% by that metric.

There are a number of issues that remain to be checked regarding this new implementation and other ways that efficiency can be improved, but we seem to be headed in the right direction.

I will show an overview of the “results” of the experiment implemented this way in a few slides.

51 of 60

2. The Key Law

Most of our thoughts on c-command and NFS can be traced back to Ueyama 1998’s theory of BVA.

This theory stated that BVA(X, Y) could essentially come about under three conditions

1. X c-commanding Y
2. NFS effects on X and/or Y (which she referred to as “quirky binding”)
3. X preceding Y, which I have been suppressing in the previous discussion.

Though LFS work on the telescope had been assuming and indirectly supporting this claim, I hoped to both formalize it as an empirical “law” of language and to more directly support it, showing clearly that :

(I) each of these three conditions can independently enable BVA(X, Y) interpretations of the relevant sentences.
(II) if none of these three conditions is present, BVA(X, Y) is categorically impossible.

52 of 60

3. Expanded Coverage

As I noted, this law had been (indirectly) demonstrated by our previous research, so I also hoped to show that it was a general law, applicable to more than just the few cases for which we had so far demonstrated it. I did this in three ways:

First, by repeating the same experiment in English, Korean, and Mandarin Chinese, the latter two of which had not been used in our previous experiments before.
Second, by considering a wider range of constructions than those we had used in the past, including passives (’his mother was praised by every man’) and “possessor binding” constructions (“every man’s wife praised his mother”)
Third, by employing a wider range of X’s and Y’s, most notably including among the Y’s elements like demonstrative phrases (“Every man praised that man’s mother”) and elements like the so-called “reflexive anaphors” jagi and ziji, both meaning something like “oneself/one’s own” in Korean and Mandarin Chinese respectively.

As predicted, the law holds perfectly across all these different conditions. For the sake of time, I will just show the “overall” results, but I am happy to talk about various breakdowns (e.g., by language) in the question period.

53 of 60

The Overall Results (1)

Each circle/square represents a judgement of an individual as to whether BVA is possible in a given sentence.

The graph is a bit “inverted” for technical reasons having to do with testability.

The color of the circle/square reflects whether BVA was accepted: green for no, red for yes.
The location of the circle/square represents the conditions under which it was judged, namely whether it met one of the three conditions stated in the law to potentially allow for BVA.
Specifically, if it failed to meet a given condition, the dot is placed inside the relevant circle.

No c-command

No precedence

No NFS

No BVA

BVA

All Data

54 of 60

The Overall Results (2)

As can be seen, of the 95 (out of ~800) datapoints in the central intersection, where none of the relevant conditions was met, the answer was universally that BVA was unacceptable.

In all other regions of the graph, when at least one of the conditions was met, BVA was at least sometimes accepted

This is precisely as predicted by the law: across different individuals, languages, sentences types, etc., BVA is possible only if one of three independent conditions is met.

This strongly supports the validity of the hypotheses underlying the telescope project, as when all other factors are eliminated, c-command systematically constrains the possibility of BVA.

No c-command

No precedence

No NFS

No BVA

BVA

All Data

55 of 60

Conclusion

56 of 60

Summary

Telescopes make hidden things directly “visible”, but achieving this feat requires significant theoretical and practical innovation.

Linguistic intuitions, just like the light of distant stars, can be made to be revealing of the properties of the language faculty, but only if we can carefully disentangle the differing sources of our judgements.

We have made initial strides in both theoretical and methodological articulation, which have led to concrete results, offering never-before-seen levels of clarity as to to the workings of the language faculty.

57 of 60

The Ongoing Project

The data summarized in the graph in the previous section demonstrate that we were able to successfully disentangle BVA-judgements coming from both structural and non-structural sources, across a wide variety of individuals and conditions.

This result is encouraging, but there is so much more to be done before we have a “fully operational” telescope! Just to name a few outstanding issues:

Deepening our understanding of the different sources of various MR’s, and how we might separate out the influence of each source
How to assess the significance/strength of a given “detection” (we have made some progress on this, which I was not able to address today, but many challenges remain)
How to better improve participant comprehension and experiment efficiency, especially as we begin to address increasingly complex sentence-types and phenomena.

To say nothing of “pointing” our telescope at various things and seeing what we can discover with it!

58 of 60

Takeaway Messages

Just as the telescope, among other devices, helped usher in a scientific revolution in the physical sciences, so too may we be sitting at the cusp of a revolution in the mental sciences.

The language faculty has definite properties that can be studied through our intuitions, if we can develop rigorous ways of examining them.�
Just as the early scientists, our instruments will be far from perfect; we will learn from errors and successes as we go. That is both the joy and the challenge of building the linguistic telescope.

1 of 60

2 of 60

3 of 60

4 of 60

5 of 60

6 of 60

7 of 60

8 of 60

9 of 60

10 of 60

11 of 60

12 of 60

13 of 60

14 of 60

15 of 60

16 of 60

17 of 60

18 of 60

19 of 60

20 of 60

21 of 60

22 of 60

23 of 60

24 of 60

25 of 60

26 of 60

27 of 60

28 of 60

29 of 60

30 of 60

31 of 60

32 of 60

33 of 60

34 of 60

35 of 60

36 of 60

37 of 60

38 of 60

39 of 60

40 of 60

41 of 60

42 of 60

43 of 60

44 of 60

45 of 60

46 of 60

47 of 60

48 of 60

49 of 60

50 of 60

51 of 60

52 of 60

53 of 60

54 of 60

55 of 60

56 of 60

57 of 60

58 of 60

59 of 60

60 of 60