Why parsing is a part of Language Faculty Science (by Daisuke Bekki)

Why parsing is a part of Language Faculty
Science∗
Daisuke Bekki
Ochanomizu University
Faculty of Core Research, Natural Science Division
LFS workshop (online)
December 19, 2020
∗
This work is supported by JST CREST, and JSPS KAKENHI Grant Number
JP18H03284, Japan.

1 Language Faculty Science: Guess, Compute and Compare paradigm
1 Language Faculty Science: Guess, Compute
and Compare paradigm
Language Faculty Science (LFS) is a research program, advocated in
(Hoji, 2015), that aims to discover the properties of the language faculty
(Chomsky, 1965) by adopting the methodology of exact science, which
is also stated as the Guess-Compute-Compare method (Feynman, 1965).
Definition 1.1 Language faculty is that part of the human
mind/brain that is hypothesized to be responsible for our ability
to relate meaning to linguistic sounds/signs. (Hoji, 2015, p.332)
Definition 1.2 Guess-Compute-Compare method emphasizes
the deduction of definite predictions and the pursuit of rigorous testa-
bility of the definite predictions. (Hoji, 2015, p.331)
LFS workshop, December 20, 2020. 2

1.1 Guess: Weak Crossover (WCO)
1.1 Guess: Weak Crossover (WCO)
(1) Hoji (2015, p.73):
a. Every boy praised his father. (under BVA(every boy, his))
b. * His father praised every boy. (under BVA(every boy, his))
Definition 1.3 BVA(A,B) is the dependency interpretation de-
tectable by the informant such that the reference invoked by singular-
denoting expression B co-varies with what is invoked by non-singular-
denoting expression A. (Hoji, 2015, p.327)
The contrast between (1a) and (1b) implies the existence of some
condition on the relation between meaning and linguistic sounds/signs
regarding the availability of BVA readings.
Here are two examples of the hypotheses that we may come up with
by guessing what had caused the contrast in (1):
Hypothesis 1.4 BVA(A,B) is possible only if A c-commands B in LF.1
Hypothesis 1.5 BVA(A,B) is possible only if A precedes B in PF.
Definition 1.6 A c-commands B if and only if A is merged with
what contains B where we understand that the containment relation
is reflexive.
1
Hypothesis 1.4 is a combination of what Hoji (2015) calls Universal hypothesis
and Bridging hypothesis, which have the different status in LFS.

1.2 Compute
1.2 Compute
In order to compute the empirical predictions of Hypothesis 1.4, we put
it together with our general assumptions on language faculty, one of
which is the following Language-particular structural hypothesis about
English.
LE1 S V O in English corresponds to an LF representation where S
assymmetrically c-commands O. (Hoji, 2015, p.33)
Thus the following predictions are born out (but we will come back
to this step later).
(2) A predicted schematic asymmetry (Hoji, 2015, p.34,69)
a. okSchema
NP V [. . . B . . . ] (Under BVA(NP,B))
b. ∗Schema
[. . . B . . . ] V NP (Under BVA(NP,B))
Deﬁnition 1.7 ∗Schema is such that any Example that instantiates
it is completely unacceptable with the speciﬁed dependency inter-
pretation. (Hoji, 2015, p.336)
In words:
1. Every instantiation of okSchema (=okExample) may be acceptable
under the BVA(NP,B) reading.

1.2 Compute
2. Every instantiation of ∗Schema (=∗Example) is unacceptable un-
der the BVA(NP,B) reading.
From (2), we further predict that an instantiation of (2a) as (3a) can
be ok, and an instantiation of (2b) as (3b) is out:
(3) a. An okExample instantiating the okSchema:
Every boy praised his father. (under BVA(every boy, his))
b. An ∗Example instantiating the ∗Schema:
* His father praised every boy. (under BVA(every boy, his))

1.3 Compare
1.3 Compare
The status of okSchema and ∗Schema are asymmetrical:
1. An okJudgment on ∗Example disconﬁrms the hypothesis.
2. A ∗Judgment on okExample does not disconﬁrm, the hypothesis.
though a ∗Judgment on okExample suggests that the experiment is not
designed well.
Best Result: Next-best Result:
Judgment
∗Example *
okExample ok
Judgment
∗Example *
okExample *
Bad Result: Worst Result:
Judgment
∗Example ok
okExample ok
Judgment
∗Example ok
okExample *

2 A missing link in Compute
Let us repeat here the okSchema and ∗Schema from Hypothesis 1.4.
(4) a. okSchema
NP V [. . . B . . . ] (Under BVA(NP,B))
b. ∗Schema
[. . . B . . . ] V NP (Under BVA(NP,B))
together with the predictions:
1. Every instantiation of okSchema (=okExample) may be acceptable
under the BVA(NP,B) reading.
2. Every instantiation of ∗Schema (=∗Example) is unacceptable un-
der the BVA(NP,B) reading.
The deduction process of Compute, from the Hypothesis 1.4 to the
prediction above, factors through the following propositions.
(5) a. In every instantiation of okSchema, NP c-commands B in LF.
b. In every instantiation of ∗Schema, NP does not c-command
B in LF.
The proposition (5a) is deduced from LE1, repeated below.
LE1 S V O in English corresponds to an LF representation where S

However, to deduce the proposition (5b), we need something like the
following:
Proposition 2.1 S V O in English never corresponds to an LF repre-
sentation where some node in O assymmetrically c-commands S.
This is a missing link in the duduction process of Compute in Hoji
(2015)’s version of LFS.
Note that, it is not enough to remedy this missing link by asserting
that NP does not c-command B in the structure specified in (4b). First,
what is presented to an informant is a string without any specification
on its syntactic structure. Thus, we have to assume the following.
An instantiation of a Schema is a string of linguistic sign/sound, not
a syntactic structure.
We should also assume the following proposition:
Most sentences (as strings) are syntactically and/or lexically am-
biguous, i.e., one or more LF representations correspond to a given
sentence.
Therefore, we cannot logically deduce (5b). Actually, (5b) does
not hold in general, since an instantiation of ∗Schema may correspond
to several syntactic structures, some of which may be those specified
by ∗Schema, but the others may be different structures where NP c-
commands B.

How crutial is the missing link? Suppose that there exists an LF
where NP c-commands B in ∗Example. Consider first the case where
some informant’s judgment is okJudgment:
Judgment
∗Example ok
This is supposed to disconfirm Hypothesis 1.4 in LFS, but acutu-
ally it does not. It is completely reasonable judgment when both our
grammar and Hypothesis 1.4 are correct. On the other hand, there
also remain the possibility that either our grammar or Hypothesis 1.4 is
wrong,
Consider second the case where some informant’s judgment is ∗Judgment:
Judgment
∗Example *
This happens when the informant lacks resorcefullness enough to
find out a LF representation where A c-commans B. This is supposed to
support Hypothesis 1.4 in LFS, however, we also obtain this result when
our gramamr is wrong, or Hypothesis 1.4 is wrong.
In summary, the missing link not only is problematic for maintain-
ing the strong deducibility of predictions in LFS, but is problematic as
well from the perspective of learning from the errors (Hoji, 2015, p.61),
since the missing link obscures the factors behind the judgment for the
∗Example.

3 Parsing as a part of LFS
3 Parsing as a part of LFS
3.1 The main thesis
LFS must ensure that each instantiation of ∗Schema does not have other
syntactic structures with A c-commanding B. For this purpose, we need
a way to know every syntactic structure that corresponds to ∗Example,
or more generally, LFS needs a theoretical component that tells us all
possible syntactic structures of a given sentence (as a string). This is
exactly what parser does.
Claim 3.1 Parser is a part of LFS, which ensures that each okExample and
each ∗Example is associated only with syntactic structures that are in-
tended in okSchema and ∗Schema.
3.2 Competence versus performance
Since Chomsky (1965), parsing is a task classified as belonging to per-
formance of language faculty, thus is regarded as a target whose in-
vestigation should be postponed until the investigation of competence
gets matured. However, parsing has an aspect which purely belongs to
linguistic competence, which we call näive parsing/parser.
Definition 3.2 （Grammar） Suppose that Σ is a set of all strings
(or linearized PF-forms) and Λ is a set of all LF-(tree-)structures. A
grammar is a subset of Σ×Λ, or equally, an element of Pow(Σ×Λ).

3.2 Competence versus performance
Definition 3.3 （Näive parser） A näive parser for a grammar G is
a function that takes a string (or linearized PF-form) σ and returns
{λ | (σ, λ) ∈ G }. In words, a näive parser returns a set of all LF-
(tree-)structures each of which is associated with σ in G.
The same argument applies to generator/generation.
Definition 3.4 （Näive generator） A näive generator for a gram-
mar G is a function that takes an LF-(tree-)structure λ and returns
{σ | (σ, λ) ∈ G }. In words, a näive generator returns a set of all
strings (or linearized PF-forms) each of which is associated with λ
in G.
As read off from the above definitions, the notions of näive parsing
and näive generator are solely defined by a given grammar. Thus, if we
consider the notion of grammar as a part of linguistic competence, so
are the notions of näive parser and generator.
Näive parser is not computationally efficient to compute all the pos-
sible LF-structures for a given string, and it is commonly believed that
human parsing employs some language model that enables us to guess
what is the most plausible structure for a given string. Most computa-
tional parsers that have been developed in the research field of compu-
tational linguistics pose the same assumption.
These non-näive parsers, or at least the computational parsers, are
developed as a combination of the knowledge of competence grammar

3.3 Parsing versus structural hypothesis
and some mechanisms concerning linguistic performance, such as lan-
guage models and heuristic search. The adjective n¨aive is used to make
a clear distinction between the parser/generator that concerns linguistic
competence only and the parser/generator for linguistic performance.
3.3 Parsing versus structural hypothesis
Language-particular structural hypothesis about English
LE1 S V O in English correspons to an LF representation where S
LE2 O S V in English can correspond to an LF representation where S
c-commands O. (Hoji, 2015, p.35)
• Hard to pose a hypothesis for every sentence pattern.
• These hypothesis are derived by parsers.
4 LFS with a parser
4.1 Schemata in CCG
(6) okSchema
AT/(TNP) [. . . BNP/N . . . ]SNP (Under BVA(A,B))
(7) ∗Schema
[. . . BNP/N . . . ]T/(TNP) [. . . AT/(TNP) . . . ]SNP (Under BVA(A,B))

4.2 Demonostration
Lexical items for A in BVA(A,B):
every T/(TNP)/N
no T/(TNP)/N
Lexical items for B in BVA(A,B):
his NP/N
Other lexical items:
boy N
praised SNP/NP
father N
4.2 Demonostration
(8) Mary saw a boy with a telescope.
a. Mary [saw a [boy with a telescope]].
b. Mary [[saw a boy] with a telescope].
(9) a. Mary saw every astronomer with his telescope.
(under the reading BVA(every astronomer, his))
b. Mary saw his owner with every telescope.
(under the reading BVA(every telescope, his))

4.2 Demonostration
Suppose that an LFS researcher claims Hypothesis 1.4, and set up
an experiment with the following Schema.
(10) a. okSchema
A [VP [...B...]]
b. ∗Schema
NP [[TV A] [...B...]]
Then each of the following is an instantiation of (4.2).
(11) a. Every astronomer [[saw Mary] [with his telescope]].
b. Mary saw every astronomer with his telescope.
the latter of which is based on the following structure:
(12) Mary [[saw [every astronomer]] [with [his telescope]]]
But (my guess is that) most informants will judge (11b) as ok, based
on the following syntactic strucuture:
(13) Mary [saw [every [astronomer [with [his telescope]]]]]
Exercise 4.1 Conduct the Guess/Compute/Compare method on the
examples shown in (9), assume Hypothesis 1.4 again, and design a set
of experiment to test it.

REFERENCES REFERENCES
References
Chomsky, N. (1965) Aspects of the Theory of Syntax. The MIT Press.
Feynman, R. (1965) The character of physical law, Beitr¨age zur Philoso-
phie des deutschen Idealismus. New York, The Modern Library.
Hoji, H. (2015) Language Faculty Science. Cambridge.
戸次大介．(2010) 『日本語文法の形式理論—活用体系・統語構造・意味
合成—』，くろしお出版.

Why parsing is a part of Language Faculty Science (by Daisuke Bekki)

More Related Content

What's hot (18)

Similar to Why parsing is a part of Language Faculty Science (by Daisuke Bekki) (20)

More from Daisuke BEKKI (16)

Recently uploaded (20)

Why parsing is a part of Language Faculty Science (by Daisuke Bekki)