Computation of prosody in D2S


After the language generation module (LGM) of D2S has generated a sentence, it is sent to the Prosody component which computes the placement of accents and phrase boundaries in the sentence. The Prosody component is part of the LGM, and so it has direct access to all syntactic and discourse information available in the LGM.

Both syntactic and discourse information are used to establish which phrases should receive an accent. Phrases that express information that is considered to be `new' to the discourse are accented, while phrases expressing given information are not. By default, all major phrases (e.g., NP, PP) in a sentence are assumed to be accented. However, if checking for givenness reveals that a phrase or word expresses given information, it is deaccented. This happens if it refers to an object which was already referred to earlier in the discourse, or if it expresses a concept which was already mentioned before. Whether this is the case can be checked in the Context Model which is maintained by the LGM. An example of givenness in GoalGetter (accented words are printed in bold face):

In the sixteenth minute, Kluivert scored a goal for Feyenoord.
A few minutes later, the forward had his second goal noted.
In this example, the words minutes and goal in the second sentence are regarded as given because they express the same concepts as the words minute / goal in the preceding sentence. The phrase the forward is deaccented because it refers to Kluivert, who was mentioned in the previous sentence. However, not all phrases expressing given information are deaccented. If the information they express is contrastive, they are not deaccented. An example of contrast in GoalGetter:
In the sixteenth minute, the Ajax player Kluivert kicked the ball into the wrong goal.
Twenty minutes later, Overmars scored for Ajax.
The team name Ajax in the second sentence is given, because it was already mentioned in the first sentence. Still, it should not be deaccented, since it expresses contrastive information. The contrast stems from the fact that the goal described in the second sentence is a goal for Ajax, whereas the goal described in the previous sentence was one for the opposing team.

Once it has been established which phrases should receive an accent, syntactic information is used to determine on which words within a phrase this accent should land. The syntactic information that is required, is readily available, because the templates used by the LGM contain syntactic trees. After accent assignment has taken place, the placement of phrase boundaries is computed, distinguishing three levels of boundary strength: minor, major and sentence-final. The placement of phrase boundaries in a sentence depends on accentuation, punctuation and syntactic structure.


Back to GoalGetter homepage