What statistical test is appropriate in a corpus analysis of pronoun resolution preferences?

I have extracted data from a corpus that collected sentences with an antecedent and a next sentence with a pronoun or a stressed pronoun in subject position. Now, the anaphora can either refer back to a subject or an object antecedent.

For instance:

"Mary passes the book to Sally. She starts reading it." (She can corefer both to Mary (suj) or Sally (obj)

How can I measure if the cooccurrence of, say, pronoun - subject antecedent is significant?

Do you have any suggestions?

I also thought of making count, normalizing the result and compare the means (subject mean vs object mean). Comparing z-scores would be the adequat thing here, correct? But, how do I get the standard deviation?

• Since you have a binary choice, you could do logistic regression. Jun 7, 2018 at 12:11
• Do you mean that whether the pronoun is stressed or not is your independent variable, with subject/object antecedent as your dependent variable? In that case logistic regression is appropriate, but there are a large number of potential confounds. I would suggest adding person, number, etc. to your model and using variable selection methods to determine whether the data really supports an effect of stress after factoring out other relevant properties. Jun 7, 2018 at 16:17
• @robert Are you sure logistic regression works here? you normally need continuous predictors for your categorical dependent measure Jun 10, 2018 at 9:18
• @WavesWashSands Yes, exactly. You'd want to include all IVs that might plausibly influence the choice. Jun 10, 2018 at 13:09
• @PaulKremershof Yes, I am, and no, you can use categorical variables as predictors. I've done it myself (not this paprticular case, but log. reg. for linguistic, binary alternations. Jun 10, 2018 at 13:11

Sentences with ambiguous antecedents are called 'Winograd schemas', and are useful as tests of strong A.I. In this case, you must understand that possession of the book passed to Sally, so 'she' most likely refers to Sally when the verb is 'to read'. The pronoun refers to either Mary or Sally -- there is no 'co-reference'.

• Well, there is always coreference from interpreter point of view who must resolve anaphora in order to understand the phrase. In this case, we have a transfer-of-possession verb. With this kind of verb, the interpreter will more likely take "she" as pointing to the sentence object. Jun 9, 2018 at 9:30
• Thanks for the hint of Winograd Schema, this is indeed very interesting! Jun 9, 2018 at 9:33