The aim of this blog post is to present a pilot study I conducted at the beginning of August 2013. The study explored two oppositions on extraction from French ”genitive
like” PPs that may — or may not — be considered as islands.
Let’s be honest, this study was insufficient in many ways. It was my first attempt at formulating a questionnaire, collecting data and conducting a statistical analysis of this data. But this study made me aware of the problems one is confronted with when dealing with acceptability judgements. It was a good lesson in ”learning by doing” for my further research in linguistics.
I will present the sentences chosen for the questionnaire, the questionnaire itself and the results. Afterwards, I will briefly discuss what we can learn from the results and what has to be improved for the next study.

1 The sentences

1.1 The first opposition

The first pair of sentences tested in the questionnaire, (1b) and (2b), are similar to those presented in (Grosu, 1974, 312). In French NPs, the possessor can be expressed by a ”genitive like” PP introduced by de or by à, the later being a colloquial but frequent form. Yet as is shown in the examples above, a WH-extraction of the PP introduced by à presents an issue, whereas a WH-extraction of the PP introduced by de is correct.

(1) a. Vous avez vu une photo de Thomas.
you did see a photography of Thomas
b. De qui avez-vous vu une photo ?
of who did-you see a photography
(2) a. Vous avez vu une amie à Thomas.
you did see a friend of(coll.) Thomas
b. * À qui avez-vous vu une amie ?
of(coll.) who did-you see a friend

1.2 The second opposition

The second pair of sentences tested in the questionnaire can be found in Zubizarreta (1979). Both are a WH-extraction of a ”genitive like” PP introduced by de.

(3) a. De qui a-t-elle vu le portrait d’Aristote ?
of who did-she see the portrait of-Aristotle
(4) a. * De qui a-t-elle vu le portrait de Rembrandt ?
of who did-see see the portrait of Rembrandt

According to Zubizarreta, sentence (4) causes more of an issue than sentence (3). The reason for this, she claims, is that the second sentence is ambiguous. It is not actually clear to the hearer/reader what entity the question is about, i.e. if they have been asked about the painter of a portrait representing Rembrandt or about the person represented in a portrait painted by Rembrandt. In sentence (3) however, Aristotle is unlikely to have been a painter. Our knowledge about the world enables us to disambiguate the question.

1.3 The control sentence

(5) À qui a-t-elle parlé du portrait de Louis XIV ?
With who did-se talk about.the portrait of Louis XIV

I also included a control question that does not present any problems but uses the same vocabulary and themes as the other sentences so that it would not be obvious that
the sentence does not have anything to do with the other ones. I chose an extraction out of a PP introduced by with. The sentence again refers to a portrait. It is a very famous portrait of Louis XIV displayed in the Louvre in Paris. It seemed to me that the object referred to would be disambiguated.

2 The questionnary (english translation)

Here is an English translation of the sheet distributed to the participants. The five sentences discussed in the first part of this article appear in a random order. However, every participant received the same answer sheet.

I am full age (over 18) : yes no
My mother tongue is French : yes no

PLEASE GIVE YOUR SPONTANEOUS AND PERSONAL ASSESSMENT OF THE PHRASING OF THE FOLLOWING QUESTIONS : TICK THE NUMBER CORRESPONDING TO THE APPROPRIATE ASSESSMENT.
1 : This phrasing doesn’t present a problem.
2 : This phrasing is unusual but correct.
3 : I’m not sure whereas this phrasing is correct or not.
4 : This phrasing is not correct in french, but I can understand it.
5 : This phrasing is so incorrect that I can’t understand it.

De qui avez-vous vu une photo ?  1 2 3 4 5
of who did-you see a photography
De qui a-t-elle vu le portrait d’Aristote ?  1 2 3 4 5
of who did-she see the portrait of-Aristotle
À qui a-t-elle parlé du portrait de Louis XIV ?  1 2 3 4 5
with who did-se talk about.the portrait of Louis XIV
De qui a-t-elle vu le portrait de Rembrandt ?  1 2 3 4 5
of who did-see see the portrait of Rembrandt
À qui avez-vous vu une amie ?  1 2 3 4 5
of(coll.) who did-you see a friend

3 The results of the questionnary

3.1 General remarks

The study took place on the 3rd of August 2013. The demographic was mainly inhabitants of the south of France.
42 answer sheets were filled out. 36 of them were considered valid (an answer sheet was valid if the participant was of full age and a native speaker).

3.2 The control sentence

Some remarks have to be made about the control sentence (5).
As expected, a huge majority (86.11%) of the participants gave the assessment 1 to the sentence. Still, 4 participants considered the sentence ”bad”. This shows us well that extractions present a processing problem per se. This is an interesting point for future research: control sentences with non-problematic extractions can allow us to identify outliers.
In this study I decided to consider 3 of the participants as outliers. Two of them gave an assessment of 5 (”not comprehensible”) to the control sentence. The third participant gave it an assessment of 4 and, furthermore, gave an assessment of 1 to sentence number(2), which became the worst average value of all.
The impact of outliers on the statistics of the study did not turn out to be very important. In the following I give statistics of the data, including outliers, and I mention the values without outliers when they add something to the discussion. Graphic representations always include outliers.
As an example, the control sentence (5) got an average assessment of 1.44 (1.15 without outliers). The standard deviation is 1.19 (0.61 without outliers).

3.3 The first opposition
3.3.1 Results

Sentence (1) received an average assessment of 2.22 (with a standard deviation of 1.34) whereas sentence (2) received an average assessment of 4.61 (with a standard deviation of 0.81). As expected, the standard deviations are even smaller without outliers (1.29 and 0.53 respectively).

The data show a statistically significant difference between sentences (1) and (2) (results of the t-test : p < 0, 001). No participant gave (2) a better assessment than (1).
A statistical correlation can be drawn between the assessments of sentence (1) and of sentence (2) if we exclude the outliers. The correlation is significant by p < 0.05 and the strength of association calculated is 12.29%. We used the Pearson Test (r = 0.35) for the calculation.
Although they are both ”à qui” extractions, no correlation could be found between the assessments of sentence (2) and of the control sentence (5) (with or without outliers).

3.3.2 Discussion

There is a undiscutable distinction between the two structures, which is in accordance with the intuition expressed in the literature. We cannot be sure based on the data whether there is a real correlation between the assessments of both sentences as the results of the Pearson Test are quite low (furthermore, we do not have a perfect symmetry between both sentences, the object being a picture in (1) and a friend in (2)).
My remaining concern for this opposition is the reason for the strong difference made between such structures. We cannot say whether the colloquialism of [à + PP ] or the
preposition à itself causes the island.
It has been claimed that only PP complements to N introduced by de can be extracted (Godard, 1992). This point should be investigated in further research.

3.4 The second opposition
3.4.1 Results

Both sentences (3) and (4) got rather bad assessments. The average assessment was slightly better for sentence (4) (average assessment of 4.00) than for sentence (3) (average assessment of 4.14).

A t-test was conducted on the 2 sets. It actually resulted in a relatively weak statistical significance of p < 0.10.
No correlation could be found between the assessments of the 2 sentences. Indeed, there were as many participants who found sentence (3) better than (4) as participants
who found sentence (4) better than (3), namely 19.44%. The majority (61.11%) considered both sentences equally good (or equally bad).

3.4.2 Discussion

The difference between the assessments of sentence (3) and sentence (4) is not very significant. We could not claim that sentence (4) has a clear acceptability superiority on sentence (3). But the claim made by Zubizarreta (1979) that (3) is better than (4) definitely does not reflect the assessments of the native speakers.
My intuition is that ”Rembrandt” and ”portraits” are associated with the same lexical field. People are used to hearing about portraits of Rembrandt. ”Aristotle”, however, is used in association with philosophy, not with painting. Referring to a ”portrait of Aristotle” may create a clash between two semantic fields, and the clash may have an influence on the spontaneous assessment. This could explain why sentence (4) received a slightly better assessment than sentence (3).
The difference in the assessments could result from one of the inconveniences of the kind of questionnaire employed. If participants have only been asked about their assessment, they may focus more on the formal structure. It could be that the participants feel fine with sentence (4), but actually do not even try to understand what it means.
However, this interpretation relies on an intuition and is not in any way certified by the study. It is a point I want to keep in mind when working on similar sentences in the future. Reaction time by dealing with a comprehension task may lead to better results.

4 Problems

4.1 Problems with the questionnaire

The form of the questionnaire appears to be problematic.
– The order of the sentences may have played a role in the assessments.
– The sentences may have had an impact on one another. The fact that all sentences were available at the same time allowed the participant to adjust his/her assessments in relation to the other ones. Psycholinguists try to avoid such scenarios. Yet I wonder if this issue is not an advantage for the study in that it provides at least a common scale for all the sentences on the same answer sheet (though not between the participants).
– In the case of islands, the absence of context may influence the acceptability judgement. In the spontaneous use of language a context is always present and allows us to interpret the meaning of a sentence.
– I wonder if the written form is inappropriate for islands and whether the judgements would have been very different if the participants had heard the sentences.

4.2 Methodological problems

The study reveals that sentences like (3) and (4) are dispreferred in French and that a sentence like (2) is strongly dispreferred in French. Does this mean that we should exclude sentence (2) of the grammar? What about sentences (3) and (4)? It is unclear where to put the limit between ”correct” and ”incorrect”.
The definition of ”correct” is very blurred anyway. Like every questionnaire asking about acceptability, the instructions on the questionnaire left the participants free to decide what they considered ”correct”. On the one hand, everyone has a clear intuition on what he/she considers ”correct” or ”incorrect”. On the other hand, it is not easy for the scientist to distinguish in this kind of judgement what is a matter of grammaticality and what is a matter of repairability. Unfortunately, we cannot expect the participants to make the distinction.
The goal of my research is to implement a parser for French sentences (Müller, 2013). In doing so, I would like to have a grammar as opened as possible. That’s why I would probably not want to exclude sentences like (2), (3) and (4). And yet, I would like my parser to account for processing difficulty, repairability and the like. I will not be satisfied if my parser’s results do not show me a clear distinction between sentence (1) and sentence (2).

A HPSG proposal (Miller, 2013): At the last HPSG conference Miller sketched a possible way to account for usage in the semantic properties of a sign. His proposal for the verb ”do” in a certain syntax based the frequency 0.02 founded in corpus can be see in a simplified way below.

Miller introduced an attribute USAGE in the semantic of the sign that allowed him to show the weight of the preference for this structure. I found this proposal very interesting, though it is not yet clear what to do with this weight. The idea of the proposal definitely has to be developed in the future. It will be a big concern of mine in my further research.

References

Godard, D. (1992). “Extraction out of NP in French”. Natural Language and Linguistic Theory, 10:233–277.
Grosu, A. (1974). “On the nature of the left branch condition”. Linguistic Inquiry, 5:308–319.
Miller, P. (2013). “Usage preferences: The case of the English verbal anaphor do so”. In: Müller, S., editor, Proceedings of the 20th International Conference on Head-Driven Phrase Structure Grammar, Freie Universität Berlin, pages 121–139.
Müller, S. (2013). The CoreGram project: Theoretical linguistics, theory development and verification. Ms. Freie Universität Berlin.
Zubizarreta, M.-L. (1979). Extraction from NP and a Reformulation of Subjacency. MIT. unpublished paper.

---

---