As you can see from this chart, the form of the anaphor seems also to agree in person with its antecedent. So once again we revise our hypothesis (rule):
8) An anaphor must agree in person, gender and number with its antecedent.
With this hypothesis, we have a straightforward statement of the distribution of this noun type, derived using the scientific method. In the problem sets below, and in chapter 5, you’ll have an opportunity to revise the rule in (8) with even more data.
You now have enough information to try GPS3, WBE2, and CPS2 & CPS3
3.2 Sources of Data
If we are going to apply the scientific method to syntax, it is important to consider the sources of our data. One obvious source is in collections of either spoken or written texts. Such data are called corpora (singular: corpus). There are many corpora available, including some searchable through the internet. For languages without a literary tradition or languages spoken by a small group of people, it is often necessary for the linguist to go and gather data and compile a corpus in the field. In the early part of the last century, this was the primary occupation of linguists, and it is proudly carried on today by many researchers.
The linguist Heidi Harley reports in her blog3 on an example of using search engines to do linguistic analysis on the huge corpus known as the web. Harley notes that to her ear, the expression half full of something sounds natural, but half empty of something does not. She does a comparison of half empty vs. half full and of half empty of vs. half full of. She finds that the ratio of half full to half empty without the of is roughly 1:1. The ratio of half full of to half empty of is approximately 149:1. This is a surprising difference. Harley was able to use the web to show that a fairly subtle difference in acceptability is reflected in the frequency with which the expressions are used.
But corpus searches aren’t always adequate for finding out the information syntacticians need. For the most part corpora only contain grammatical sentences. Sometimes the most illuminating information is our knowledge that a certain sentence is ungrammatical (i.e., not a sentence of normal English), or that two similar sentences have very different meanings. Consider the pair of sentences in (9) as a starting point.
9) a) Marian blew the building up.
b) Marian blew up the building.
Most native speakers of English will accept both of these sentences as acceptable sentences, with a preference for (9b). They also know that while the first sentence (9a) is unambiguous, the second one has two meanings (He destroyed the building using explosives vs. he blew really hard with his lungs up the stairwell). The second of these meanings is a bit silly, but it’s a legitimate interpretation of the sentence.
Now contrast the sentences in (9) with the similar pair in (10). In these forms I’ve replaced “the building” with the pronoun “it”:
10) a) Marian blew it up.
b) Marian blew up it.
Here we find a different pattern of interpretation. (10a) is unambiguous just the way (9a) is, it refers to an act of explosion and cannot have an interpretation where Marian was blowing hard with her lungs up something. Sentence (10b), however, is a surprise. Unlike (9b), (10b) cannot have anything to do with explosives. It can only have the interpretation where Marian is blowing air up whatever “it” is. Recall that with (9) this “puff of air reading” was the silly or strange one. With a pronoun, however, it’s the only available interpretation. This difference in interpretation would never be captured in a corpus, because the specific meanings of expressions and ambiguities are not indicated anywhere in the data source.
While corpora are unquestionably invaluable sources of data, they are only a partial representation of what goes on in the mind. More particularly, corpora often contain instances of only acceptable (or, more precisely, well-formed) sentences (sentences that sound “OK” to a native speaker). For example, the online New York Times contains very few ungrammatical sentences. Even corpora of naturalistic speech complete with the errors every speaker makes don’t necessarily contain the data we need to test the falsifiable predictions of our hypotheses. So, corpora are just not enough: there is no way of knowing whether a corpus has all possible forms of grammatical sentences. In fact, as we will see in the next few chapters, due to the productive nature of language, a corpus could never contain all the grammatical forms of a language, nor could it even contain a representative sample. It also doesn’t tell us about what sentences are ambiguous or what sentences are ungrammatical or strange. Those are really important sources of evidence for doing syntax. To really get at what we know about our languages we have to know what sentences are not well-formed. That is, in order to know the range of acceptable sentences of English, Italian or Igbo, we first have to know what are not acceptable sentences in English, Italian or Igbo. This kind of negative information is very rarely available in corpora, which mostly provide grammatical, or well-formed, sentences.
Consider the following sentence:
11) *Who do you wonder what bought?
For most speakers of English, this sentence borders on word salad – it is not a good sentence of English. How do you know that? Were you ever taught in school that you can’t say sentences like (11)? Has anyone ever uttered this sentence in your presence before? I seriously doubt it. The fact that a sentence like (11) sounds strange, but similar sentences like (12a and b) do sound OK is not reflected anywhere in a corpus:
12) a) Who do you think bought the bread machine?
b) I wonder what Fiona bought.
Instead we have to rely on our knowledge of our native language (or on the knowledge of a native speaker consultant for languages that we don’t speak natively). Notice that this is not conscious knowledge. I doubt there are many native speakers of English that could tell you why sentence (11) is terrible, but most can tell you that it is. This is subconscious knowledge. The trick is to get at and describe this subconscious knowledge. The psychological experiment used to get this subconscious kind of knowledge is called the acceptability judgment task. The judgment task involves asking a native speaker to read a sentence, and judge whether it is well-formed (i.e., grammatical), marginally well-formed, or ill-formed (ungrammatical).
There are actually several different kinds of acceptability judgments. Both of the following sentences are ill-formed, but for different reasons:
13) a) #The toothbrush is pregnant.
b) *Toothbrush the is blue.
Sentence (13a) sounds bizarre (cf. the toothbrush is blue) because we know that toothbrushes (except in the world of fantasy/science fiction or poetry or a dream) cannot be pregnant. The meaning of the sentence is strange, but the form of the sentence is okay. We call this semantic ill-formedness and mark the sentence with a #. By contrast, we can glean the meaning of sentence (13b); it seems semantically reasonable (toothbrushes can be blue), but it is ill-formed from a structural point of view. That is, the determiner the is in the wrong place in the sentence. This is a syntactically ill-formed sentence, which is marked with an *. A native speaker of English will judge both these sentences as ill- formed, but for very different
