Generation of referring expression with an individual imprint
Bernd Bohnet · 2009
A major outcome of the last Shared Tasks for Referring Expressions Generation was that each human prefers distinct properties, syntax and lexical units for building referring expressions. One of the reasons for this seems to be that entities might be identified faster since the conversation partner has already some knowledge about how his conversation partner builds referring expressions. Therefore, artificial referring expressions should provide such individual preferences as well so that they become human like. With this contribution to the shared task, we follow this idea again. For the development set, we got a very good DICE score of 0.88 for the furniture domain and of 0.79 for the people domain.