Next:Accessibility space proposalUp:Computational Approach to AnaphoraPrevious:Related work on anaphora

An annotation scheme for dialogue structure

For successful anaphora resolution in dialogues, we assume that it is essential to identify dialogue structure. Therefore, we propose an annotation scheme for Spanish dialogues that is based on work carried out by Gallardo 1996, who applies the theories put forward by Sacks et al. 1974 concerning (conversational) turn-taking.

We use an annotation scheme based on these theories for three main reasons. First, as it is a general approach to dialogue modeling, it is applicable to all types of dialogues, including both task-oriented and information-retrieval-oriented dialogues. Consequently, the use of such a model as a basis for developing our anaphor resolution procedure allows us to apply the procedure to any type of domain, thus offering an advantage over procedures based on discourse models specific to particular domains. Second, this annotation scheme can be easily applied to automatic processes without metalinguistic considerations. Although in our work the annotation task has been performed by hand, for dialogue-based applications in which our procedure might be embedded (e.g., in dialogue management systems), annotation tasks must be performed automatically. Finally, we wanted to base our own procedure on studies of the influence of dialogue structure on anaphora resolution that were carried out by Fox 1987, whose approach, in turn, is based on that of Sacks et al.

According to these theories, the basic unit of conversation is the move, which informs the listener about an action, request, question, etc. Moves are carried out by means of utterances.4 And utterances are joined together to become turns.

Since our work was done using spoken dialogues that had been transcribed, turns are annotated in the texts and utterances are delimited by the use of punctuation marks or by the ends of turns. Reading a punctuation mark (., ?, !, ...) allows us to recognize the end of an utterance. These tasks do not affect the anaphora-resolution process.

As a result, we propose the following annotation scheme for dialogue structure:

Turn (T) is identified by a change of speaker in the dialogue; each change of speaker presupposes a new turn. On this point, we make a distinction between two different kinds of turns:
Adjacency pair (AP) (also called exchange) is a sequence of turns headed by an initiation intervention turn (ITI) and ended by a reaction intervention turn (ITR). This form of anaphora, in which the reference appears within an adjacency pair, appears to be very common in dialogues [Fox 1987].
Topic (TOPIC). The topic must be a lexical item that is referred to frequently. According to Rocha 1998, four features are taken into account in the selection of the best candidate for a discourse topic: frequency, even distribution, position of first token, and semantic adequacy. A highly frequent element that occurs intensively in a passage of the dialogue but does not appear for long stretches is not likely to be a good choice for discourse topic. In the same way, neither is an element whose first appearance occurs a long way from the beginning the best choice. Moreover, semantic adequacy must be considered for the candidate, and it must be assessed by the annotator.

Based on the above-mentioned structure, then, the following tags are considered necessary for dialogue structure annotation: ITI, ITR, CT, AP, and TOPIC. The AP and TOPIC tags will be used to define the anaphoric accessibility space, and the remaining tags will be used to obtain the adjacency pairs. The ITR/I tag, representing mixed interventions, is not included since mixed interventions can be annotated as ITR plus ITI. This task is done in the annotation phase.

An example of an annotated dialogue with tags is presented in Figure 1. In the dialogue, the identifier (OP) indicates the turn of a railway company employee, and the identifier (US) indicates the client's turn.

TOPIC tren 
AP1  ITI (OP)  información de Renfe, buenos días. 
(Renfe information, good morning.)
 ITR (US)  hola, buenos días. 
(hello, good morning.)
CT (OP)  hola. 
AP2  IT (US)  me podéis decir algún tren que salga mañana por la tarde para ir 
a Monzón?
(could you tell me about any train that leaves tomorrow evening
for Monzon?)
IT (OP) sí, vamos, mira hay un talgo a las tres y media de la tarde.
(yes, let's see, there is a talgo at half-past three.)
AP3  IT (US) sí, tiene que ser más tarde. 
(yes, it has to be later.)
IT (OP)  más tarde. hay un intercity a las cinco y media, un expreso a las seis.
y media
(later. there is an intercity at half-past five, an express
at half-past six.)
AP4 IT (US)  el de las seis y media ¿llega a Monzón? 
(the one at half-past six, does it go to Monzon?) 
AP51 IT (OP)  a ver. el de las seis y media me ha preguntado ¿verdad? 
(let me see. you've asked about the one at half-past six, right?)
IT (US)  sí. 
IT (OP)  a las nueve y veinticinco. 
(twenty-five past nine.)
AP6  IT (US)  a las nueve y veinticinco está en Monzón? 
(at twenty-five past nine it is in Monzon?)
IT (OP)  sí 
CT (US)  vale, pues ya está. esto ya es suficiente. 
(ok, that's it. that's enough for now.)
AP7  IT (US)  gracias, ¿eh? 
(thank you, eh?)
IT (OP)  muy bien a usted. hasta luego. 
(very well, thanks to you. so long.)
1 This adjacency pair is included within AP4.
Figure 1: Example of an annotated dialogue from Corpus InfoTren: Person
One of the most important advantages of this annotation scheme is its compatibility with most of the dialogue-annotation schemes used in dialogue systems. Notice, for instance, that the adjacency pairs show the same structure as the conversational game applied to task-oriented dialogues defined in the dialogue structure by Carletta et al. 1997. Moreover, our scheme is also compatible with those that are based on utterance functions, such as the one defined in DAMSL by Allen and Core 1997. DAMSL indicates how utterances are related to the discourse by means of forward- and backward-looking functions. The interpretation of these functions builds the adjacency-pair structure. Finally, our topic structure exhibits the same features as the transaction structure of Carletta et al. or the task level defined by Allen and Core.

Next:Accessibility space proposalUp:Computational Approach to AnaphoraPrevious:Related work on anaphora
patricio 2001-10-17