• Keine Ergebnisse gefunden

(109) shows.38

(109) A: What did he force you to do?

B: Sing. / To sing.

Moreover, as we have seen above when we looked at elaborations, res-via-id fragments do not always replace an element inα(where this element would be the antecedent for Ginzburg’s parallelism con-straint). The examples in (36) above, two of which are repeated here as (110), have shown that also non-realised optional arguments of elements ofαcan be elaborated with fragmental utterances.

(110) a. A: Peter was reading when I saw him.

A book about Montague semantics(, to be precise).

b. A: Die Verleihung war spektakul¨ar.

[Des Oskars]gen, meine ich.

(The handing over of the award was spectacular. Of the Oscar, I mean.)

Examples like this suggest that it is not directly the syntactic category of an element ofαthat is im-portant, but rather the syntactic constraint of an element ofαon its complements, be that a verb, as in (110-a) above, or a noun, as in (110). Note also that arguably we even need the requirements for a related verb, namely the transitive rather than the intransitive version of ‘read’, in (110-a).

As we said, we will return to the question of how this data that apparently pulls in different directions—

contra syntactic, but also contra semantic reconstruction—should best be analysed. For now, this closes our introduction of our taxonomy, and we conclude by comparing it to ones that can be found in the literature.

on a difference in the “type of inference which is based on a different kind of contextual informa-tion” (Barton 1990, p. xvii). She illustrates this difference with the following example, where the first fragment is interpreted in her model in the “submodule of linguistic context”, whereas the second is interpreted in the “submodule of conversational context”.39

(111) a. A: What stops the White House staff from visiting House Speaker Tip O’Neill in his congressional office?

B: [An] old grudge.

b. A: The White House staff doesn’t visit Tip O’Neill in his congressional office B: [An] old grudge.

This distinction is very similar to ours between res-via-id, and res-via-inf (as which we would classify (111-a) and (111-b), respectively). Moreover, she makes similar observations regarding syntactic influ-ences from the context on the fragment in the former case. What her approach doesn’t offer, however, is a classification that goes beyond this major distinction, or that would offer a basis to explain it. We will develop such an approach later in the thesis. Also, her approach is only very roughly formalised, and so it is difficult to test predictions made by it (see Section 4.4.1 for details).

Carberry (1990): The main aim of Carberry’s (1990) approach and the individuating criterion of her taxonomy is described in the following quote: “Understanding the intent behind elliptical fragments requires that the speaker’s discourse goals be recognised. [. . . ] I have identified fifteen discourse goals that occur during information-seeking dialogues and that may be accomplished by means of elliptical fragments.” (Carberry 1990, p.193). FollowingSDRTour aim is more modest: we want to describe the intended meaning of fragments, not necessarily all intentions behind uttering it. And we believe that this results in a much simpler approach, both in theoretical and in computational terms. This difference in aim allows us to generalise over some distinctions Carberry makes in her taxonomy. Moreover, by analysing certain constructions in more detail, we can even further generalise her classes.

The first effect is shown with the examples in (112), all of which we analyse as different types of Elaboration.

(112) a. A: What courses would you like to take?

B: For credit?

b. A: I want to get a degree.

CS major.

39A similar distinction of ellipsis in general can be found in (Rath 1979), who distinguishes between “Konstruktions¨ubernahme”

(construction-adoption) and “Eigenkonstruktion” (self-constructed).

c. A: What is Dr. Smith teaching this fall?

B: CS360.

A: CS360?

A0: The course in architecture?

d. A: Who’s teaching CS360? The course in architecture.

In Carberry’s taxonomy on the other hand, the fragments are all classified as belonging to completely different classes, which are in turn: seek-clarify-question (“IS [= information seeker] requests inform-ation relevant to clarifying a question posed by IP [= informinform-ation provider].”), provide-for-assimilinform-ation (“IS provides information pertinent to constructing his underlying task-related plan.”), seek-identify (“IS is unable to satisfactorily identify the referent of an item in IP’s utterance and requests help from IP in doing so.”) and identify (“IS attempts to identify an entity in his own utterance.”).

The examples in (113) show that our analysis has the advantage of breaking down turns into smaller units, so that more general classes can be formed. Carberry classifies the fragments in (113) as a answer-question-suggest-alternative (“IS answers a yes/no-question negatively, providing a description of a desirable alternative.”) and answer-question-with-explanation (“IS answers a yes/no-question with an explanation of the answer.”) respectively. Since we analyse the negative answer “no” as an anaphor for the negation of the proposition contained in the polar question to which it is an answer, we can analyse these examples with rhetorical relations that are independently motivated, namely Correction or Contrast for (113-a) and Explanation for (113-b).

(113) a. A: Would you like to take CS360?

B: No, CS470.

b. A: Do you want to take CS865?

B: No, too late at night.

We will further discuss Carberry’s approach below in Section 4.4.2, and only note here that our tax-onomy captures all her examples, albeit with fewer classes (and hence fewer distinctions).

Fern´andez et al. (2002): As mentioned in the previous section, (Fern´andez & Ginzburg 2002) offer a taxonomy roughly based on Ginzburg’s approach to non-sentential utterances. Theirs is the only paper of those discussed here where data about coverage are given; they claim that they can classify 99.05% of the fragments in the corpora they annotated. As already alluded to above, however, we think this result is difficult to compare to ours. We mentioned above that it is not quite clear what the data-basis for their annotation was; another factor that makes a direct comparison difficult is that the definitions of the classes used in the study (at least as they are described in (Fern´andez & Ginzburg 2002)) do not seem to

be entirely consistent. Some of their classes appear to be defined according to the discourse function of the fragment (like our classes are, although they do not use relations), while others are based just on the form of the fragment; they also have classes that seem to be individuated by a mixture of these criteria, together with reasoning about the intentions of the interlocutors. In the following we will show that our taxonomy is both more consistent, being based on the sole criterion of the use made of fragments, and also more fine-grained. We first list the classes in (Fern´andez & Ginzburg 2002) for which there is a direct counterpart in our taxonomy. These are mostly fragments that are classified according to use by the authors as well.

(114) Short-Answer ≈QAP

Correction ≈Correction Acknowledgement ≈Acknowledgement

Clarification Ellipsis≈(one use of) Elaborationpq

The class they label ‘sluicing’ is an example of one defined by a mixture of form, function and inten-tions. Part of the definition is “sluices are bare question denoting wh-phrases” (all quotes here are from (Fern´andez & Ginzburg 2002, p.16)). However, as the authors make clear later in the paper, only one type of such phrases is actually supposed to belong to this class, namely that of fragments which “in-volve a request for additional information beyond what the speaker of the previous utterance thought was required.” Note that this definition is in terms of intentions and beliefs of the interlocutors about their wider discourse goals (“what the speaker thought was required”). The example they give (from the BNC), presented here as (115), suggests that the fragments in this class are instances of what we label Elabmq(where m is either p or q). Our class however is only defined in terms of speech act related goals, which is a more restricted and conventionalised kind of goal than the general discourse goal.

(115) A: Can I have some toast please?

B: Which sort?

Two more examples that show that our classes are both more general and more fine-grained: First, they have a class called “bare modifier phrase” (and the name already indicates that this class is individuated by the form of the fragment), for which they give (116) as an example. In our taxonomy, however, this is just one way of expressing elaborations, and so our taxonomy captures generalisations that theirs doesn’t capture.40

(116) A: They got men and women in the same dormitories.

B: With the same showers!

40Note that they ultimately have the same goal as we have, namely to specify how fragments are resolved, and so a classification just of different possible surface-forms of fragments is of little value to them as well.

Second, they describe a class they call “fragments introduced by connectives”. As our examples (32-c) and (87) have shown (repeated here in (117), such fragments can serve a number of functions, and so our classification is more fine-grained.

(117) a. A: I went to the cinema with Peter.

And Sandy.

b. A: He drove to Italy. And then to Spain.

Their conclusion that “with the context as given, the principles by means of which NSU content is resolved do not involve complex domain sensitive reasoning.” is not very strong, because the hard bit is to determine what exactly the context is. Computing the question under discussion (QUD), the main instrument for resolution in their approach (see below Chapter 5), will in all but the most trivial cases (QAPs) be a matter of pragmatic reasoning. The actual way of combining material once that is done is irrelevant to the question of how complex the reasoning in general is; we will expand on this point in Chapter 5.