LUNA - SCIENTIFIC AND TECHNICAL APPROACH
The LUNA SLU can be broken down into three core modules:
- the Word Lattice Annotation (WLA) module
- the Semantic Composition (SC) module
-
the Context Sensitive Validation (CSV) module.
The following architecture illustrates the interoperability of components
that will be developed.
• The ASR output is sent to the SLU module using the HTK format for Word Lattice representation.
• Contextual information is output by the DM following an XML description of the dialog history. The structure of the dialog history will be defined in WP4 and its representation will follow the semantic representation specified in D1.3 (“Specifications of the annotation protocol for the data”).
• This contextual information is added to the word lattice file, providing an enriched word lattice that will be submitted to the SLU module. Each of the three SLU components will be able to handle this contextual information at the level that is best suited for its objectives.
• The SLU output will be an interpretation lattice, following the same HTK conventions as the word lattice, but having the property that the transitions will carry semantically structured interpretations represented through an XML string. The data structure of the SLU data streams has the following advantages:
• Based on a well known lattice format (HTK) for which binary representation libraries already exists.
• Expressive power of standard XML language for encoding the symbolic information attached to each arc of the lattice.
• Extendible to cover other modality other than speech-only input.
|