The summary information contained in the topic frames still need some checks and extensions before it can be transferred to generation. Since the result summary will have a thematic ordering in topics we want each topic to contain as much information as possible. Hence, we implement a set of rules inferring relatively secure data from what is known in the context memory:
Consistency checks are necessary because of the known shortcomings of speech recognition and data extraction. Since human dialogue contains a lot of redundant information it is more than obvious to exploit this fact. Therefore, all suggestions have been annotated with speaker attitudes (accept, strong accept, reject etc.) which are used to compute the final agreement. If there are two conflicting dates for one meeting a score for each date is computed to select the one supported by more evidence.