Do null subjects (mis-)trigger pro-drop grammars?

Abstract

Native speakers of English regularly hear sentences without overt subjects. Nevertheless, they maintain a [−pro] grammar that requires sentences to have an overt subject. It is proposed that listeners of English recognize that speakers reduce predictable material and thus attribute null subjects to this process, rather than changing their grammars to a [+pro] setting. Mack et al. (2012) showed that sentences with noise covering the subject are analyzed as having null subjects more often with a first person pronoun and with a present tense – properties correlated with more predictable referents -- compared to a third person pronoun and past tense. However, those results might in principle have been due to reporting null subjects for verbs that often occur with null subjects. An experiment is reported here in which comparable results are found for sentences containing nonsense verbs. Participants preferred a null subject more often for first person present tense sentences than for third person past tense sentences. The results are as expected if participants are responding to predictability, the likelihood of reduction, rather than to lexical statistics. The results are argued to be important in removing a class of mis-triggering examples from the language acquisition problem.

1. Introduction

In this article we will assume that adult speakers of a language represent their knowledge of a language as a grammar¹, an abstract systematic characterization of possible structures and their interpretation, as well as representing lexical items and maintaining some record of actual language experience (e.g., in episodic memory). The question we will address is why input data inconsistent with a speaker’s grammar, but consistent with the grammar of other languages, doesn’t result in grammar change.

There are two common types of cases where speakers maintain their grammar despite incompatible input. One involves speech errors (see Frazier 2014 for the suggestion that natural errors that are easily fixed and lead to plausible meanings are automatically repaired and hence not taken as evidence that one’s grammar should change). The other type involves systematic variation in the input that can be explained by performance factors such as the reduction of highly predictable material. ‘Reduction’ appears to take place at all levels: not finishing a sentence when the continuation is highly predictable, not putting in optional grammatical markers due to predictability (Rohdenburg, 1996, Jaeger, 2010), phonologically reducing predictable materials (Fowler & Housum, 1987 and a ton of others). Indeed reducing predictable material permeates the language production system: for instance, adjuncts referring to easily inferrable material (e.g., instruments) are less likely to be included in re-telling a story (Brown & Dell, 1987). In sum, reduction of predictable material appears to take place at all levels of language production. Here we will focus on the issue of sentences with phonologically null subjects, and why they do not lead English speakers to a grammar containing a phonologically-null pronoun, i.e., a ‘pro-drop’ grammar as in Spanish or Italian. How can a grammar requiring subjects co-exist with input sentences lacking overt subjects?

Below we review a prior study on the perception of inputs which are ambiguous between having an overt subject or not, testing the kinds of sentences where English speakers actually encounter null subjects. The results suggest that listeners do sometimes assign a null subject analysis even though the input is open to an overt subject analysis. We then turn to the question of whether the null subject analysis comes about because listeners assume predictable constituents are reduced, or whether a null subject analysis is assigned based on actual experience of certain verbs co-occurring with a null subject. To address the issue, we present a new study exploring whether comparable effects are found with never before encountered nonsense verbs.

2. Background

Mack, Clifton, Frazier & Taylor (2012) investigated the hypothesis that null subjects are used in currently available situations. Presenting auditory discourses like those in (1) they varied whether the final target sentence had a first person or third person referent, and whether the sentence was in the present or the past. The target sentence had a subject that was distorted and covered with noise so that it was unclear from the signal itself whether there was an overt subject or not. Participants were asked to listen to the mini-dialogues and then repeat the final sentence, and then type the final sentence.

• (1)

  Present	A: I’m going to Walmart tonight with Anna.
  	B: Why?
  	A: It Looks to me/her like they might have laptops on sale.
  Past	A: I went to Walmart last night with Anna.
  	B: Why?
  	A: It Looked to me/her like they might have had laptops on sale.

As expected, the subject was restored more often with a third person than a first person referent, and restored more often in the past tense targets than present tense targets (see Table 1).

Table 1.

Speech restoration study, spoken responses

	present, 1st- person	present, 3rd-person	past, 1st-person	past, 3rd-person
# it responses	179	184	185	199
# zero responses	107	91	81	67
% it restoration	62.6%	66.9%	69.5%	74.8%

A possible analysis of the null subject sentences is that they are acceptable but not grammatical². Speakers implicitly know that they degrade predictable subjects and that listeners expect this and can imagine a situation where the utterance might be used. Imagining that the subject is likely to be predictable in a currently present situation may lead to the expectation of phonological degradation and thus to the judged higher acceptability of the present tense examples.

The hypothesis that listeners know that speakers tend to reduce predictable material predicts that effects like those seen in Mack et al. should extend to examples containing nonsense words. It is this prediction that was tested in the experiment reported below. The issue is whether participants in the Mack et al. study might have been responding based exclusively on their experience, i.e., how often specific verbs and specific verb forms had been encountered with a null subject. In most models of sentence comprehension, it is assumed that information about the frequency with which a verb occurs with particular internal arguments is stored in the lexicon. If information about the overtness of the verb’s external argument (subject) is also stored, then in the restoration study, the restoration of subjects might have been a matter of analyzing the input according to lexically-based expectations when the input itself was noisy.

In the case of language acquisition, it is not entirely clear how lexical expectations concerning particular verbs could help with the acquisition problem. What’s at issue is the analysis of a sentence with a missing subject: did it come about by general pragmatic processes of reduction or did it result from the availability of a particular phonologically null morpheme (‘pro’–a pronoun with syntactic features but no phonological features).

If this lexical frequency information was the basis for the responses in the restoration experiment, then there is no reason to expect comparable behavior with nonsense verbs. However, if listeners implicitly recognize the circumstances under which they degrade linguistic input, including dropping constituents, then experience with a particular verb is not critical. Similar behavior should result with never before encountered nonsense verbs.

3. Experiment with Nonsense Verbs

3.1. Materials

To test for ratings of acceptability of null subjects in sentences with nonsense verbs, eight phonologically possible nonsense verbs such as trif were selected from a set of the highest-rated most word-like nonsense verbs from Albright & Hayes' (2003) work on minimal generalization learning of the English past tense. Each verb appeared in a simple sentence. In one form it contained a past tense and a third person referent. In the other it contained a present tense and a first person referent, as illustrated in (2).

• (2)
  1. How does that trif to you?
     Which answer sounds more natural? 1. Trifs good to me. 2. It trifs good to me.
  2. b. How did that trif to him?,
     Which answer sounds more natural? 1. Triffed good to him. 2. It triffed good to him.

3.2. Method

The 8 sentences were counterbalanced creating two lists, and thus for each nonce verb a participant would see either its present tense (2a) or its past tense (2b) form. Each participant saw an equal number of present tense and past tense forms. The sentences were visually presented, without any fillers, in an individually-randomized order. Forty speakers of English were recruited using Amazon’s Mechanical Turk (https://www.mturk.com). Completion of the survey took approximately 5 minutes. Participation was restricted to workers with IP addresses in the United States, and all 40 participants self-reported as native speakers of English in a demographic survey.

The instructions were as follows: Welcome. In this experiment, you'll first read a question that introduces a 'new word' - not really a word of English, but please pretend that it is one that you just don't happen to know. When you have indicated that you read that sentence, you'll be asked to choose one of two possible responses to the question you read. Please choose the one that seems most natural to you by typing its number (1 or 2).

Participants were informed that they would see the question sentence by pressing the space bar on their computer when they saw an underscore in the middle of the screen. Once they had read the question, they would press the space bar again to see the possible answers. They were instructed that the answers differ in a subtle way, asked to read both options and then choose the one that seems to fit best with the question sentence they had just read.

3.3. Results

The results are presented in Table 2 in terms of the proportion of non-zero responses.

Table 2.

Proportion of non-zero subject responses: "It trifs/triffed…". (SE in parentheses)

Present, first person	Past, third person
0.64 (0.0381)	0.78 (0.0327)

Overall, there were more non-zero subject responses than zero subject responses, as expected for a language with a [−pro-drop] grammar. Crucially, more zero/null subject responses were chosen for present tense first person items than for their past tense third person counterparts³. This supports the view that speakers implicitly know that null subjects are more likely when the subject is more predictable.

Could the results be due to analogy with existing verbs? There are particular verbs in English where a null subject seems particularly tempting, e.g., seems, looks, smells, sounds. Of the eight nonce verbs tested (trif, glit, snell, shee, skell, murn, spack and stip), only two seem very close, namely, snell and skell. But these two items did not have particularly high rates of selecting null subjects. Snell showed only a 5% effect in the predicted direction, and skell was the only item that showed no difference between present first person and past third person. The results thus seem difficult to explain on an analogy account.

4. Discussion

We propose that the grammar of English really does not permit phonologically null subjects; it is [− pro-drop], and that the dropping of subjects results from performance factors, not the grammar. The puzzling inconsistency between encountering null subject sentences and maintaining a grammar requiring an overt subject can be reconciled, we claim, by assuming that speakers of the language are implicitly aware of reducing predictable material. This permits the language user to account for the absence of the grammatically obligatory subject of a sentence without changing the grammar. There is a vast literature on the dropping of subjects (sometimes called diary drop, see Weir 2012 for a recent account) and on the acquisition of null and overt subject parameter settings (see in particular, Hyams, 2012, Valian et al, 1996). We do not claim to have solved the grammar acquisition problem with respect to subjects. But we think that the present approach does help explain why null subject inputs do not trigger [+ pro-drop] grammars.

The puzzle we have addressed is an instance of what we believe to be a much more general problem. Speakers often encounter doubled quantifiers (Everyone always), doubled negation (without his shirt off), doubled or ‘harmonic’ modals (It’s advisable that he should..), and so forth. Why don’t they assume a grammar with Concord, rather than simply undoubling the doubled item when interpreting the sentence? (See Frazier & Clifton, 2011, in progress, and Frazier, in press, for evidence). Frazier (2014) suggested that children may recognize errors and other ‘deviations’ due to the performance system as being due to the performance system, and consequently not take them as evidence that their grammar must change⁴. This approach would simplify to some extent the language acquisition problem by eliminating large classes of misleading data that could trigger incorrect grammatical generalizations.

Appendix

Both forms are illustrated for item #1; only the present tense form is illustrated for the other items.

1. • How does that trif to you?
     Which answer sounds more natural?
     Trifs good to me., It trifs good to me.
   • b. How did that trif to him?},
     Which answer sounds more natural?
     Triffed good to him., It triffed good to him.

2. Glits like fun to me., It glits like fun to me.

3. Snells very cold to me., It snells very cold to me.

4. Shees humid to me., It shees humid to me.

5. Skells like a bad idea to me., It skells like a bad idea to me.

6. Murns a bit crazy to me., It murns a bit crazy to me.

7. Spacks perfect to me., It spacks perfect to me.

8. Stips like heavy rain to me., It stips like heavy rain to me.