On the acquisition of either and too

Naomi Francis, Shuli Jones, Leo Rosenstein, & Martin Hackl∗

Abstract. This paper presents an experimental investigation of how English-learning
children acquire the additive discourse particles either and too. In the target gram-
mar these items exhibit near-complementary distribution conditioned on the polarity
of their host sentence. The path leading to that grammar appears to be rather intri-
cate. We present comprehension data showing that for an extended period of time
(ages 3–5) learners find both items acceptable in both polarity environments, exhibit-
ing only a weak adult-like tendency of preferring either in negative and too in positive
sentences. At age 6 their grammar appears categorical with respect to either in that
they no longer tolerate it in positive sentences while still exhibiting only a weak dis-
preference for too in negative environments. These findings are even more striking
in the context of production data. We find that child-directed speech is essentially
categorical, providing unambiguous evidence for the adult grammar. Moreover, we
find essentially categorical, adult-like use of either and too in child production from
the earliest stage of development. These observations raise a number of challenges for
theories of either and too and for approaches to learning focus particles more generally.
Perhaps most strikingly, the protracted insensitivity of the learner’s grammar to accu-
mulation of unambiguous evidence constitutes a novel argument from the abundance
of evidence for encapsulated learning.
Keywords. Additivity; focus particles; polarity sensitivity; L1 acquisition

1. Introduction. This paper presents an experimental investigation of how English-learning chil-
dren acquire the additive focus particles either and too. Just like other non-scalar, additive dis-
course particles (items such as also, additionally, as well, . . . ), their basic function is to signal that
the sentence they attach to is part of a sequence of sentences that constitutes a (potentially partial)
answer strategy for addressing a question under discussion (see e.g. Beaver & Clark 2008). What is
special about either and too is that they appear to be designated to mark, respectively, negative and
positive sequences. That is, they exhibit (near) complementary distribution, which is conditioned
on the polarity of the sentence they accompany. This is exemplified in (1).

(1) a. Sam is eating cake. Sam is eating ice cream too/*either.
b. Sam isn’t eating cake. Sam isn’t eating ice cream either/*?too.

One approach to capturing the divergent polarity sensitivity of either and too while maintain-
ing their shared discourse functionality is to assume that they trigger the same canonical additive
presupposition (while being assertorically inert), as shown in (2), and to stipulate that they are

∗We are grateful to Yadav Gowda for his help with building the scripts that extracted and coded our CHILDES
data, to Anya Keomurjian, Abena Peasah, and Mika Thakkar for constructing stimuli for the adult experiment, and
to Steve Worthington for stats guidance. This research was supported by the Social Sciences and Humanities Re-
search Council of Canada (Doctoral Fellowship to the first author). Authors: Naomi Francis, University of Oslo
(n.c.francis@iln.uio.no), Shuli Jones, Massachusetts Institute of Technology, Leo Rosenstein, Massachusetts Institute
of Technology, & Martin Hackl, Massachusetts Institute of Technology.

Proceedings of ELM 1: 159-171, 2021

c©2021 Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl
Published by the LSA with permission of the author(s) under a CC BY license. 159

https://doi.org/10.3765/elm
https://www.elm-conference.net/


(morpho-) syntactically specified for specific environments. Either would be marked to occur with
negative sentences, taking necessarily wide scope over the negative operator, while too would be
confined to positive sentences either via designated marking or as the result of competition with its
more marked counterpart either.1

(2) Jeither/tooK(p) = λw: ∃q ∈ ALT(p) s.t. q(w)=1. p(w)=1

Rullmann (2003) provides compelling arguments against such a treatment, however, and instead
proposes an analysis on which either and too are not synonymous. Concretely, while too receives a
treatment like any run-of-the-mill additive particle, either is assumed to trigger an “anti-additive”
presupposition demanding that there is an alternative to its host clause that is false, (3).2

(3) JeitherK(p) = λw: ∃q ∈ ALT(p) s.t. q(w)=0. p(w)=1

(3) by itself does not prevent either from being attached to positive sentences, of course. To make
that impossible Rullmann (2003) needs to assume, moreover, that either is marked as an NPI, hence
confined to occur in the scope of a suitable negative operator.3 This ensures that the host clause
of either is negative. Moreover, since either presupposes that there is an alternative (antecedent)
proposition to p that is false results in either being able to serve as a signal for a sequence of
negated propositions forming an answer strategy. The fact that too can mark a positive sequence
follows directly. To account for the fact the too cannot mark a negated sequence requires, however,
a further assumption. Rullmann (2003) proposes that too is preferentially attached low in the
structure, lower, importantly, than the attachment site of sentence negation. Attaching too below
negation in (1-b) triggers a positive presupposition which the context does not satisfy, however,
giving rise to infelicity.4 By the same token, this proposal predicts that speakers might find (1-b)
with too acceptable to the extent to which they are able to access a parse with high attachment
of too. In that case the host clause happens to express a negated proposition which is licit as
long as the context provides a true focus alternative to it. Since this is the case in (1-b) wide
scope attachment of too would produce a felicitous sequence, offering a perspective on why the
unacceptability in (1-b) is typically judged less severe than the unacceptability of (1-a).

Although the proposal in Rullmann (2003) is not without problems and there are competing
theories in the literature, notably the alternatives-based analysis of Ahn (2015), we think that it
provides a useful framework within which to explore questions about how either and too might
be acquired. In particular, we are interested in finding out whether learners initially adopt the
most basic additive semantics for either as in (2). Evidence for such a stage in development

1This could be implemented via morphological suppletion which is conditioned on the syntactic position of the
additive particle, as in König (1991), or via some form of feature probing mechanism.

2How exactly alternative propositions are generated is an open question in the literature. For our purpose it will be
sufficient to follow Beaver & Clark (2008) and assume that the the focus marking inside the host clause reflects which
QuD the sequence is meant to address.

3See Rullmann (2003) for details. Since the question of which negative operators can license either is not central
to our purpose, we will organize our discussion around sentence negation and abstract away from any complications
that arise when considering other licensors.

4However, in case there is a positive alternative provided by the context as in (i) too is predicted to be licensed.

(i) Mary ate the lasagna, but she couldn’t eat the spaghetti, too. (Rullmann 2003; 389)

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 160

https://doi.org/10.3765/elm
https://www.elm-conference.net/


would come in the form of either being fully interchangeable with too across positive and negative
environments. If such a stage can be identified, we would then hope to trace the time course and the
evidence it takes to learn their near-complementary distribution. Of particular interest is whether
learners go through a stage predicted by the proposal that Rullmann (2003) ends up rejecting.
If so, these learners should expect either and too to exhibit complementary distribution in basic
sentences such as (1) but they would not yet know that either is, in fact, an NPI.

To investigate these questions we first conducted a comprehension experiment employing a
comparative felicity judgment task. This task is well-suited for probing directly whether speakers
prefer either over too in positive environments and too over either in negative environments. We
find that children do not behave in accordance with Rullmann (2003) or the simpler proposal
rejected therein. Indeed, our findings are puzzling under all theories of either and too we are aware
of. We then conducted a corpus study on child-directed speech to see whether their comprehension
could be understood in terms of the data available to the learner being noisy, uninformative or even
misleading. Interestingly, our corpus study revealed quite the opposite: child-directed speech is in
effect categorical with respect to the polarity-dependent distribution of either and too. Finally, we
conducted a corpus study on child speech to examine whether their productions of either and too
reflect their comprehension system. We find, however, that child speech is in fact adult-like even
at the earliest stage in development. This puts our comprehension data in even starker relief.

2. Child experiment.

2.1. METHOD. To assess children’s understanding of the polarity-conditioned distribution of ei-
ther and too, we employed a comparative felicity judgment task (Chierchia et al. 2001, Foppolo
et al. 2012). This task presents participants with a forced choice between either and too in a given
environment. For positive sentences, we predict that adults will always choose too over either,
since there is no way for either to be licensed there. For negative sentences, a prediction based on
Rullmann (2003) is more nuanced; speakers who cannot access a wide-scope LF for too should
have a categorical preference for either over too. For speakers who can access a wide-scope LF for
too, the rate of selecting too will depend on them judging the either competitor to be nevertheless
better. If they are roughly comparable, we might expect non-categorical selection of both items; if
they are significantly different, as Rullmann suggests, we might expect a fully categorical pattern.5

In each trial, we used a laptop to present a scene like those exemplified in Figure 1 to the
participant. One experimenter, playing the role of narrator, introduced the scene and provided an
appropriate antecedent for the additive particle. Then two puppets provided competing sentences
completing the description; the puppets’ utterances were identical except for the additive item used
(either vs. too). Children were then asked to indicate which puppet “said it better”.

The experiment consisted of eight target items and four filler items, balanced for polarity.
Positive and negative versions of each target item were created. For each target item, half of the
participants saw the positive version and half saw the negative version; the fillers were seen by all.
The ordering of positive and negative trials, the order in which the puppets spoke, and the additive
particle used by the first speaker were pseudo-randomized to avoid unwanted patterns.

We recruited 57 participants aged 3–7 years in the Boston area at the Museum of Science and

5Since this is an empirical question, we ran an adult control study reported in Section 3.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 161

https://doi.org/10.3765/elm
https://www.elm-conference.net/


Figure 1: Examples of positive (L) and negative (R) target items

local daycares. Of these, 47 L1 English acquirers completed the experiment: nine 3-year-olds,
fourteen 4-year-olds, fifteen 5-year-olds, eight 6-year-olds, and one 7-year-old. The 7-year-old
was excluded from our analysis, leaving data from 46 participants. A parent or guardian provided
informed consent for each child, and each child received a sticker in thanks for their participation.

2.2. RESULTS. The mean rates at which children selected each item in each environment are
shown in Figure 2.

Figure 2: Mean rate of either/too selection, by polarity environment and age

To examine the distribution of our responses we constructed a linear mixed effects logit model
in R (R Core Team 2019) using the lme4 package (Bates et al. 2015) with fixed effects of polarity
(pos vs. neg) and age (3–5 vs. 6). This model showed main effects of polarity and age; chil-
dren were more likely to select either in negative sentences than in positive sentences (Pr(>|z|) =
0.000171 ***), and 6-year-olds were less likely to select either overall (Pr(>|z|) = 0.009378 **).
The model also revealed a significant interaction between polarity and age (Pr(>|z|) = 0.036657 *);
the contribution of this interaction was confirmed with a likelihood ratio test (p < 0.05 *).

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 162

https://doi.org/10.3765/elm
https://www.elm-conference.net/


2.3. DISCUSSION. We take the comparative felicity judgment task to reflect comprehension, as
it tests participants’ responses to utterances of which they are not the author. The results reported
above show that children are not adult-like in their treatment of either and too. From the ages of
3 to 5, children exhibit a non-categorical response pattern, accepting both items in both environ-
ments. However, their behavior is not random; they select either more often than too in negative
environments and too more often than either in positive environments. That is, they trend toward
the adult grammar. Importantly, this trending pattern does not reflect a population-level split, with
some children being categorical in one direction and others being categorical in the other; the
non-categorical trending pattern is attested within subjects.

That this behavior is non-adult-like is transparent for either, but the interpretation of the too
data is more involved. Informal fieldwork with native English speakers and an adult felicity rating
study reported in Section 3 suggest that too in negative sentences is judged to be quite infelicitous.
A small (n=4) replication of the above experiment with adults confirmed that they deliver categor-
ical judgments on this task. Moreover, some children (n=8) explicitly indicated that both items
were equally good in the sentences they were presented with, answering “both” or “neither” to the
experimenter’s question “Who said it better?”. We take this to indicate that children’s treatment of
too in our task is non-adult-like.

The non-categorical trending behavior continues virtually unchanged until the age of 6, when
children suddenly stop selecting either in positive sentences. This is presumably responsible for
the statistical effect of age, which indicated that 6-year-olds are less likely to accept either overall.

The patterns of behavior exhibited by the 3–5-year-olds on the one hand and the 6-year-olds on
the other both raise interesting questions. What knowledge state does the 3–5-year-olds’ behavior
reflect? Their different response patterns in positive and negative environments suggest they are
not ignorant of the role of polarity in conditioning these items’ distribution, but neither do they
faithfully reproduce its categorical nature. The stability of this trending behavior is also striking;
the pattern of responses displayed by the 3-year-olds in our sample is virtually identical to that
displayed by the 5-year-olds, despite the fact that the latter group has had two more years’ worth
of exposure to the distribution of these items in the input. This suggests that whatever is responsible
for the change in behavior that occurs between ages 5 and 6 is not merely the result of the steady
accumulation of evidence; this constitutes a novel argument for encapsulated learning.

How should we characterize the learning pattern displayed here? The early grammar of addi-
tivity seems to have compatibility of either and too as one of its core properties. At first glance, this
might suggest that children initially hypothesize that these items share a basic additive semantics
(i.e., (2)). However, on such a view it is difficult to explain the trending behavior, and equally dif-
ficult to explain its stability. Why it is so difficult for children to master the restricted distribution
of these items? This fact becomes even more striking once we understand the input is robustly
categorical (see Section 4). From the perspective of Rullmann’s (2003) account, we would ask
what it takes to move from a simple additive semantics for either (requiring a true antecedent, as
in (2)) to the non-standard one (requiring a false antecedent, as in (3)).6 The answer offered by this
framework appears simple enough at first blush: children must realize that either is an NPI. What
is left open, however, is why a wide-scope theory of either is not entertained at all, not even at an

6If we take our data seriously, it appears that the children we tested never entertain a simple König-style approach.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 163

https://doi.org/10.3765/elm
https://www.elm-conference.net/


intermediate state for which it is plausible to assume that learners have not yet encountered data
that decisively argue for the NPI-treatment of either.

3. Adult control.

3.1. METHOD. To precisify what the target grammar of too allows, we ran a variation of the
experiment described in the previous section on an adult population. Our goal was to determine
whether too is acceptable in negative sentences with negative antecedents on the adult grammar. A
direct replication of the child experiment may be too blunt a tool to investigate this question, as it
does not distinguish between what is dispreferred because it is infelicitous and what is dispreferred
because it is merely less felicitous than its competitor. To give participants an opportunity to
indicate a difference between what is allowed but dispreferred by their grammar vs. what is not
allowed, we conducted a felicity rating experiment.

Our adult study was conducted online. Participants were shown scenes similar to those used
in the child experiment and asked to rate the naturalness of positive and negative sentences with
either and too (each following an appropriate antecedent) on a Likert scale with values ranging
from 1 (“extremely unnatural”) to 7 (“perfectly natural”).

The experiment consisted of 32 Latin-squared target items and 40 filler items, both balanced
for polarity. In each trial the scene and antecedent sentence appeared on the screen; to ensure that
participants were actively engaged in the task and had time to read the text, they had to click their
mouse or press a key after a 2000 ms embargo period to advance to the target sentence and then to
the rating scale. Participants were rejected if they i) failed to complete their rating within 7000 ms
of the scale appearing in more than 10% of trials, or ii) rated more than two of eight ungrammatical
benchmarking fillers above a 2.

We used Ibex Farm (Drummond 2018) and PennController (Zehr & Schwarz 2018) with Ama-
zon’s Mechanical Turk to recruit participants and deploy the experiment. We collected data from
51 L1 U.S. English-speaking adults. Of these, 6 were excluded due to their performance on the
benchmarking fillers, leaving data from 45 individuals. Participants gave informed consent and
received 3.85 USD.

3.2. RESULTS. To account for possible variation in how participants used the 7-point scale, we
normalized the ratings using z-scores. The resulting mean ratings are plotted in Figure 3.

A maximally-specified convergent linear mixed effects model of z-scored ratings constructed
in R using package lme4, with fixed effects of polarity (pos vs. neg) and particle (either vs. too),
reveals a significant polarity by particle interaction (p < 0.001 ***). Either was judged to be more
acceptable in negative than in positive sentences; the opposite was true for too (t = 76.00).

3.3. DISCUSSION. We found that the ratings depended on a highly significant interaction be-
tween the polarity of a sentence and the additive item used. It appears that ratings for negative
too sentences cluster with those for positive either sentences rather than the perfectly acceptable
positive too and negative either sentences; note that this matches Rullmann’s (2003) description
of the data rather well. These results suggest that adults indeed treat too as degraded in negative
environments, albeit less so than either in positive environments. We take this to indicate that even
at age 6 children are not adult-like in their treatment of too in negative environments.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 164

https://doi.org/10.3765/elm
https://www.elm-conference.net/


Figure 3: Z-scores of ratings

4. Corpus study: Child-directed adult speech. We have observed that children’s comprehension
of either and too is non-adult-like until at least age 6. It is important to know whether this protracted
development could be attributed to the input; do children not receive sufficient clear evidence about
the distributions of either and too, or do they have difficulty using the evidence? To determine
whether children’s non-categorical trending behavior can be attributed to properties of the input,
we conducted a corpus study of child-directed speech.

Using custom R scripts, we extracted every instance of additive too and either from CHILDES
corpora of typically-developing, North American English-acquiring children spoken by Mother,
Father, Sister, Brother, Aunt, Uncle, Grandmother, Grandfather, Family Friend, Teacher, or Adult
(MacWhinney 2000). Each token was coded for the polarity of the environment in which it ap-
peared. The results are given in the following Table 1. Children hear approximately ten times

Item Total Positive Negative Unclear
Either 701 28 (3.99%) 670 (95.58%) 3 (0.43%)

Too 7896 7782 (98.56%) 103 (1.30%) 11 (0.14%)

Table 1: Child-directed utterances of additive either/too in CHILDES

more occurrences of additive too than either. The input is overwhelmingly categorical, with very
few instances of unlicensed either and too.7 It should be noted, however, that if children ask them-
selves not where to insert either or too but rather which item to insert in a positive or negative
environment, the data available to them in negative environments is somewhat noisier than that

7The picture is in fact even clearer than these numbers suggest; 23 of the 28 positive either tokens were found
in sentences like Me either!, uttered in response to a negative antecedent, while 42 of the negative too tokens were
sentences like Don’t I get some too?, following a positive antecedent.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 165

https://doi.org/10.3765/elm
https://www.elm-conference.net/


available in positive ones.8 Even so, the input should provide very few opportunities for confusion
about the distribution of either and too.

5. Corpus study: Child speech. In the previous section, we observed clear categorical behavior
in child-directed speech that contrasts with the children’s non-categorical behavior profile in the
comparative felicity judgment task. This raises a question: What does the production system look
like? Does it mimic the input or does it reflect the comprehension system?

We repeated the procedure described above for the target children in each of the corpora
investigated. The results are given in Table 2. Children’s productions, like adults’ productions,

Item Total Positive Negative Unclear
Either 242 22 (9.09%) 210 (86.78%) 10 (4.13%)

Too 5059 5007 (89.95%) 42 (0.83%) 10 (0.20%)

Table 2: Child utterances of additive either/too in CHILDES

are categorical; although children were willing to accept the puppets’ productions of too in negative
environments and (for the younger group) either in positive environments, they themselves very
rarely use these items in environments where they are not licensed.9 Children’s productions thus
broadly mirror the adult productions that they are exposed to, although the former are somewhat
noisier.

It is worth noting that, although children (like adults) produce more utterances containing
additive too than either, both particles appear to follow the same growth curve, Figure 4.

Figure 4: Children’s cumulative productions of either and too

6. Discussion. Our investigation has revealed a number of unexpected and indeed rather notewor-
thy features of children’s developing understanding of either and too. The most significant ones
are: i) from ages 3–5, children display a stable, non-categorical preference in the direction of the
adult grammar, ii) between ages 5 and 6, there is an abrupt transition to adult-like categorical be-
havior in one half of the paradigm (either’s restriction to negative environments) but not the other,
and iii) there is an asymmetry between children’s comprehension and production, with the former

8The percentages of adult-grammar consistent data would be 86.68% for negative environments and 99.74% for
positive ones. We are grateful to Jesse Snedeker for pointing this out to us.

9As with the adult production data, most of the positive either tokens (18 out of 22) were sentences like Me either!,
uttered in response to a negative antecedent.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 166

https://doi.org/10.3765/elm
https://www.elm-conference.net/


lagging behind the latter. We believe that these findings present serious challenges for all existing
theories of (adult) either and too. In keeping with our previous discussion, we will explicate the
nature of these challenges against the backdrop of Rullmann (2003), attempting along the way to
distill more general desiderata for any theory which hopes to account not only for the adult pattern
but aims to provide a principled basis for explaining how these items are acquired.

From the perspective of the target grammar advocated in Rullmann (2003), we expect the
acquisition path for either and too to unfold roughly along the following lines. Upon first encounter
of either and too learners should entertain a run-of-the-mill additive semantics as in (2) for both
items. This is because the treatment in (2) is arguably the simplest way of specifying an additive
particle and it is independently needed to account for particles like also.10 The resulting grammar
predicts either and too to be completely interchangeable across all environments and so cannot
account for their much more restricted distribution.

Next, as learners begin to record the distributional peculiarity of either and too in their input
data, which, recall, we found to be essentially categorical, they should amend the respective entries
in such a way that either is confined to negative environments and too to positive. The adjustment to
the grammar might take the form of specifying each item morpho-syntactically for a suppletion rule
along the lines of König (1991), where either is simply the spell-out of the particle when it takes
scope over a relevant negative operator. Alternatively, either might be equipped with a syntactic
selection feature that can only be discharged if the host clause contains a negative operator in the
scope of either. Specifying a lexical item morpho-syntactically for a restricted set of environments
is by no means unprecedented in the grammar and so should be a readily available option for
the learner. Moreover, since the required morpho-syntactic amendment can take place without
any change to the basic additive semantics of our particles we expect a stage in development that
is well-characterized by a wide-scope either grammar. It would exhibit a fully complementary
distribution of our particles for simple sentences like those in (1) and remain in place until the
learner encounters more complex sentences that cannot be captured by such a grammar.

One type of data that would provide compelling evidence against the wide-scope either gram-
mar is exemplified in (4), where the negative operator is in a higher clause but the presupposition
triggered by either does not contain the matrix verb.

(4) John refused to go to church. His parents were so angry that they did not permit him to go to
the soccer game either. (Rullmann 2003; 353)

The additive inference in (4) is quite transparently satisfied by the proposition that John didn’t
go to church. On the wide-scope either theory, however, (4) is incorrectly predicted to trigger
the presupposition that John’s parents did not permit him to go to church. This is because on the
wide-scope theory, either necessarily outscopes the negative operator of its host sentence. Only
in that position is it possible for either to generate the correct additive inference in simple cases
such as (1). For more complex cases like (4) this assumption fails. To account for the correct
presupposition the learner needs to assume that either, in fact, takes narrow scope with respect to
negation, which in turn requires abandoning the basic additive semantics for either in favor of its

10Note furthermore that while the proposal in Rullmann (2003) implies a hypothesis space that allows for an “anti-
additive” semantics for additive discourse particles, one where the presupposition requires a false antecedent, there are
no negative-positive sequences in the data which would unambiguously call for such an entry.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 167

https://doi.org/10.3765/elm
https://www.elm-conference.net/


anti-additive counterpart in (3). It also requires marking either morpho-syntactically for NPI-status
to ensure that it cannot occur with positive hosts. This adjustment to the grammar is clearly more
involved than the previous one. Moreover, since the type of evidence that would trigger the change
involves more complex constructions which are presumably rarer in the input, it is expected under
such an account that the switch to the NPI-grammar of either could occur quite late in development.

Our findings are only partially consistent with these expectations. To begin with, we do not
see a stage in development where both items are completely interchangeable across positive and
negative sentences. Of course, it is possible that such a stage occurs earlier, before the youngest
age at which we were able to gather production and comprehension data. Looking at the data we
do have, we see that even our youngest participants exhibit behavior that is not fully compatible
with this prediction. This is most clearly documented by our production data where we find that
even the earliest productions of either and too are adult-like in their polarity sensitivity. Turning
to our comprehension data, the picture gets more nuanced and more interesting. Recall that we
find that 3–5-year-olds judged both particles to be contextually appropriate with both sentence
types. By itself, this would be expected if they hypothesized the basic additive semantics in (2) for
both particles. However, what is not expected is that these children also exhibit a weak adult-like
trend towards preferring either with negative hosts and too with positive hosts, and that they do so
essentially unchanged for an extended period of time (three years of development). Both aspects
of this development are quite puzzling from the perspective of Rullmann (2003) and even more so
from the perspective of the simpler grammar in König (1991). The fact that these children exhibit
some sensitivity as evidenced by their trending behavior shows that they have caught on to the
fact that polarity matters somehow for the distribution of either and too. Even more striking, we
know from their productions that they view the effect of polarity on the choice of the particle for the
purpose of production as decisive. Why then are they not able to execute a rather simple adjustment
to the lexical entries of these items that would result in a grammar that matches their input as
well as their own productions? Put differently, we have strong indications form our learner’s
production and comprehension that the polarity of the host sentence matters for the status of either
and too. Nevertheless, our learners seem to view them as compatible with both environments in
comprehension, attributing to them the ability to trigger a contextually suitable additive inference
irrespective of the polarity of the host sentence. This is of course starkly non-adult-like. And it
is entirely mysterious from the perspective of the wide-scope grammar of either advocated e.g.
in König (1991) but also from the perspective of the NPI-grammar proposed in Rullmann (2003)
since that theory offers no reasons why learners shouldn’t entertain a wide-scope grammar at an
intermediate stage of acquisition when they may have taken note of the polarity effect on the
distribution of either and too in basic sentences but have not yet encountered data such as (4) that
are transparently incompatible with a wide-scope grammar of either.

Arguably the most striking aspect of this pattern is that it persists over three years of develop-
ment in the face of categorical and readily available data to the contrary. For some reason, learners
seem to not make use of this evidence; the only reason for this seems to be that the wide-scope ei-
ther grammar is never entertained by learners to begin with. That is, it is not part of the hypothesis
space for additive discourse particles.11 If this diagnosis is correct, we have here a novel case of

11This would predict that there are no such items in the world’s languages. To our knowledge, this has not been

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 168

https://doi.org/10.3765/elm
https://www.elm-conference.net/


encapsulated learning, one that is based on the abundance rather than the poverty of the stimulus.12

Next, we turn to the change that occurs between ages 5 and 6, which affects either in that it
is now excluded from positive environments but does not seem to affect too, which is still readily
tolerated in negative environments. Rullmann (2003) offers a perspective for the timing of this
change since it is quite plausible that informative data such as (4) are rare in the input and, fur-
thermore, require a well-developed grammar to be fully appreciated by a learner.13 It is less clear,
however, whether 6-year-olds judging too in negative sentences to be fully felicitous (albeit at a
slightly lower rate than either) fits as well. On Rullmann (2003) too is disallowed in negative
sentences because the required high-scope LF is only marginally accessible due to a parsing pref-
erence to attach sentence final particles low and blocking by either which can convey the same
presupposition without violating the parsing preference. Note that our adult control data bear out
this prediction quite nicely. The behavior of our 6 years olds in the comparative felicity judgment
task seems to indicate, however, that too does not compete with either in negative sentences even
though either is already understood to be an NPI at that age. The only way this could be explained
within Rullmann (2003) is to assume that 6-year-olds have not yet adopted the low-attachment
preference for sentence-final particles.14

The non-categorical (non-adult-like) behavior observed in the 3–5-year-olds’ comprehension
performance contrasts sharply with their fully categorical (adult-like) productions. While asym-
metries between comprehension and production in language acquisition are not uncommon, such
asymmetries usually go in the opposite direction, with comprehension maturing first. One attrac-
tive way of deriving the appearance of comprehension lagging behind production is to exploit the
speaker’s advantage of knowing the message;15 however, it is not clear that our case is amenable
to such an explanation, as the stimuli in the comprehension experiment should have left no doubt
about the intended message and the context provides no invitation to infer an enriched meaning.
We suspect that our asymmetry is instead a reflection of formal complexity.16 If two structures can
express the same meaning, with one being more complex than the other, we might expect that chil-
dren would systematically choose to produce the simpler structure without necessarily insisting on

systematically investigated.
12See e.g. Babyonyshev et al. (2001) for a similar case.
13Our own corpus work on child-directed speech is not fine-grained enough to examine this prediction and we will

have to leave this issue for future research.
14Whether this is plausible has, to our knowledge, not been investigated in literature.
15See Hendriks (2014) for discussion of a variety of different relevant cases and an Optimality Theory-based per-

spective under which the advantage for production is a consequence of a non-adult ranking of grammatical constraints.
16We do not present a detailed comparison with Ahn’s (2015) account here, but we do not believe that it has a

significant advantage in capturing our data. In a nutshell, Ahn proposes that either and too assert disjunction and
conjunction, respectively, between the prejacent (p) and a covert propositional anaphor (q) that is presupposed to be a
focus alternative of the prejacent. She derives the NPI behavior of either via exhaustification of obligatorily-activated
subdomain alternatives in the style of Chierchia (2006). An alternatives-based approach along these lines might seem
to provide an explanation for the children’s behavior, as children under the age of 6 have been found to be non-adult-
like in their ability to access certain alternatives; indeed, it has been claimed that this restricted set of alternatives
allows at least some children in this age group to systematically strengthen overt disjunction to a meaning that is
fully equivalent to conjunction, though the structure that produces this meaning is more complex (Singh et al. 2016).
However, our experimental setup is known to circumvent this kind of problem, with children systematically selecting
the more informative of two options presented to them (Chierchia et al. 2001, Foppolo et al. 2012).

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 169

https://doi.org/10.3765/elm
https://www.elm-conference.net/


the same degree of economy from other speakers. For example, suppose that children’s develop-
mental path takes them through a stage where a sentence can have the same meaning regardless of
which additive item is inserted, but the either-sentence has a more complex LF than too-sentence
if it is positive and the reverse holds in negative environments. If children always produce the sim-
plest structure that will convey their intended message, we should expect them to systematically
produce too in positive sentences and too in negative ones. Importantly, this leaves open the possi-
bility that children will accept other speakers’ productions of both items in both environments, as
we observed in the comprehension task, without necessarily inferring an enriched meaning in case
the speaker chooses the less economical option.

References

Ahn, Dorothy. 2015. The semantics of additive either. In Eva Csipak & Hedde Zeijlstra (eds.),
Proceedings of Sinn und Bedeutung 19, 20–35.

Babyonyshev, Maria, Jennifer Ganger, David Pesetsky & Kenneth Wexler. 2001. The maturation
of grammatical principles: Evidence from Russian unaccusatives. Linguistic Inquiry 32(1).
1–44. https://doi.org/10.1162/002438901554577.

Bates, Douglas, Martin Mächler, Ben Bolker & Steve Walker. 2015. Fitting lin-
ear mixed-effects models using lme4. Journal of Statistical Software 67(1). 1–48.
https://doi.org/10.18637/jss.v067.i01.

Beaver, David I. & Brady Z. Clark. 2008. Sense and sensitivity: How focus determines meaning
Explorations in Semantics. Oxford, UK: Wiley-Blackwell.

Chierchia, Gennaro. 2006. Broaden your views: Implications of domain widening and the “logi-
cality” of language. Linguistic Inquiry 37(4). 535–590. https://doi.org/10.1162/ling.2006.37.
4.535.

Chierchia, Gennaro, Stephen Crain, Maria Teresa Guasti, Andrea Gualmini & Luisa Meroni. 2001.
The acquisition of disjunction: Evidence for a grammatical view of scalar implicatures. In
BUCLD 25 Proceedings, 157–168. Somerville, MA: Cascadilla Press.

Drummond, Alex. 2018. Ibex farm. Web. https://spellout.net/ibexfarm/.
Foppolo, Francesca, Maria Teresa Guasti & Gennaro Chierchia. 2012. Scalar implicatures in

child language: Give children a chance. Language Learning and Development 8. 365–394.
https://doi.org/10.1080/15475441.2011.626386.

Hendriks, Petra. 2014. Asymmetries between language production and comprehension. Dordrecht:
Springer.

König, Ekkehard. 1991. The meaning of focus particles: A comparative perspective. London and
New York: Routledge.

MacWhinney, Brian. 2000. The CHILDES project: Tools for analyzing talk. Mahwah, NJ:
Lawrence Erlbaum Associates 3rd edn.

R Core Team. 2019. R: A language and environment for statistical computing. R Foundation for
Statistical Computing Vienna, Austria. https://www.R-project.org/.

Rullmann, Hotze. 2003. Additive particles and polarity. Journal of Semantics 20. 329–401.
https://doi.org/10.1093/jos/20.4.329.

Singh, Raj, Ken Wexler, Andrea Astle-Rahim, Deepthi Kamawar & Danny Fox. 2016. Children

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 170

https://doi.org/10.3765/elm
https://www.elm-conference.net/


interpret disjunction as conjunction: Consequences for the theories of implicature and child
development. Natural Language Semantics 24(4). 305–352. https://doi.org/10.1007/s11050-
016-9126-3.

Zehr, Jérémy & Florian Schwarz. 2018. PennController for internet based experiments (IBEX).
Web. https://doi.org/10.17605/OSF.IO/MD832.

Proceedings of ELM 1: 159-171, 2021

Naomi Francis, Shuli Jones, Leo Rosenstein, and Martin Hackl:
On the acquisition of either and too. 171

https://doi.org/10.3765/elm
https://www.elm-conference.net/