Figure 3: A dialogue fragment annotated with move labels.

For example, in the user:constraint move, the user delimits the range of possible trips he is interested in. By contrast, in the user:ask-for-info move the user asks for information about possible trips, but the queried information does not count as content to be added to the current constraints on possible trips. The query is a ``side question'' not contributing directly to the current set of mutually understood constraints (but may, depending on the answer, lead to a new constraint). In the user:ask-for-suggestion move, the user asks for an alternative suggestion without rejecting the previous suggestions from the system (the user might very well go back and accept a previous suggestion).

We distinguish between twelve different user moves and roughly the same number of system moves. The DM catgorizes an utterance as a certain move by computing a heuristic likelihood score for each move type, based on the following factors:

• the existence of suitable contexts. For example, an utterance cannot be classified as a user:accept unless the system has proposed some train(s) and/or flight(s) that the user can accept.

• the difference between the propositional contents of the utterance and that of the context. For instance, if these two are inconsistent, the utterance cannot be classified as a user:accept; if they are consistent, it is unlikely that the utterance should be classified as a user:ask-for-suggestion.

• the precence of keywords in the utterance. For example, if the utterance contains ``accept words'' like ``yes'', ``ok'', etc., the user:accept score is increased.