In order to fully understand the phenomenon to be modeled, we based our work
on the analysis of a corpus of data collected from natural spoken conversations.
The corpus contains 63 instances of misunderstandings;
all the excerpts are in Italian, except two instances taken from
an English transcription.
The analysis of the data well evidenced the complexity of the phenomenon and highlighted some important features of natural occurring misunderstandings, such as the flexibility of repair mechanisms in human conversation, the need to take different levels of misunderstanding into account and the nature of resources exploited by human conversants in accomplishing the repair task. All these features served as guidelines during the development of the model, hopefully contributing to make its treatment closer to reality.
First of all, the empirical data show the impact of misunderstanding on dialogue, and confirm, apart from the theoretical expectations, the need for a treatment of this phenomenon in a computational model of dialogue. In fact, the episodes of misunderstanding found in 59 hours and 12 minutes of conversation amount to 52, with a quite high rate of occurrence of one every 68.3 minutes approximately. This datum justifies the intuition that, even if they can be considered a sporadic event, misunderstandings do still play a role in conversation and consequently need to be treated by an appropriate mechanism (see also [Perlis & Purang1996]).
By observing the misunderstood turns, for example, it is possible to attempt a distinction concerning what is misunderstood.
With reference to the process of interpretation of an utterance, five levels
can be identified as possible object of misunderstanding,
namely the phonetic level, the syntactic level, the lexical level,
the semantic level and the pragmatic level.
If we consider the origin level of misunderstandings, we have the following
distribution: the semantic level gathers the highest amount of instances,
with 54% of the total, followed by the pragmatic level (19%), the phonetic
level (14%) and the lexical level (13%), while no instances were found
belonging to the syntactic level.
More precisely, looking at the internal composition of the semantic level, divided into propositional content and reference (including wrong reference to the allocutor, as evidenced in Schegloff's data) the rate importance of the whole category can be attributed to the misunderstandings of the referent intended by the speaker. As far as the pragmatic level is concerned, it includes several types of misunderstandings that hadn't been highlighted before. At this level, misunderstandings concern not only the comprehension of the illocutionary force, but also the speaker's plan, the role of participants, the grasping of the indirect meaning, the choice of the right interpretation context and the discourse topic. Interestingly, only a few instances (3 instances) were found in which the hearer's reaction was misunderstood, generating a wrong repair, i.e. a repair that it was not necessary.
Regarding the causes of misunderstanding the analysis of the real examples underlines the substantial role played by ambiguity in allowing for wrong interpretations through all the levels (see [Zaefferer1977], [Vendler1994]). In detail, the responsibility for misunderstanding can be attributed to the presence of a relevant ambiguity at one of the identified levels in approximately two thirds of the instances, while only one third takes place in complete absence of ambiguity (e.g. in the presence of mishearings and other phenomena like misconceptions).
Finally, it is interesting to have a look at the position of repair with respect to the misunderstood turn. The evidence from the corpus suggests that, while most repairs are executed within two or three turns of distance from the misunderstood turn, a significant set of examples shows that repairs can be delayed at up to several turns of distance from the point where the misunderstanding has occurred. There are examples of repair at the 12th, 15th, 19th and even later, at 31st turn, confirming the data collected in [Hansen et al. 1996] about the possibility for the speaker to delay the repair. Moreover, most of the analyzed repairs were performed by the misunderstood speaker on the understanding shown by the other participant (i.e. third turn repair), while only three instances can be classified as repairs initiated by one participant on his own interpretation of a previous turn. So, although this datum confirms the preference for third position repair, it underlines the need for a flexible diagnosis of misunderstanding, which doesn't rely only on a pre-ordered sequence of turns, but which is able to go backward in the dialogue in search for a misunderstood element that may reside in any previous positions and at any possible level of meaning.