Annotation Manual


This annotation spectification was developed to label incremental understanding in the fruitcarts corpus. Lexical entries at the transcript level should be normalized to fit the domain specific vocabulary. This way "apple" is labelled as "tomato" since in this domain there is only tomatos, and speaker intents to refer to a tomato instead. Equally, a Move action can be realized with many different words such as "put", "move", "take", etc. Each of these words are all mapped to the Move action if that is the speaker's intentions. Labels in most cases should be alinged with some word, clause, or utterance in the transcript layer. It may be the case that some words required multiple labels, since Anvil does not allow overlapping tags, we have created a Merged checkbox which indicates that the current label should be merged with the previous unmarked label. The last simultaneous label marks the end of the set of overlapping labels.

This track has specified only the necessary attributes for each action. Utterances such as "we want a xxxx in yyyy ..." show that at the point of hearing "we want" we are unaware which action speaker wants the actor to perform (i.e. it could be a move or a paint). The disambiguation of intended action occurs later in the utterance. This means that the actual annotation of the intented action must occur when there is no doubt of what the speaker meant.

Generated by Anvil on Mon Oct 29 20:53:32 EDT 2007