TRADR Nuance Mix.NLU models

Speech analysis is an essential component to keep track of the tasks performed by the team in a rescue mission with the TRADR system.

To improve the accuracy of the speech recognition, we developed and trained a NLU model using the cloud-based speech engine Nuance Mix.nlu.

In Mix.nlu, developers can model intents annotating a set of representative sample sentences, and training the resulting model to cover as far as possible the utterances users might want to use to express that intent.

The trained NLU model is powerful enough to classify unseen user utterances with similar intents.

The intents we focused on in TRADR are used to interpret utterances concerning the task management (task assignment, status, progress and completion as well as problems, etc.) in the human-human communication among the members of the TRADR team.

Data

Description

The present source files (TRSX = TRaining Set XML) include all elements required for the Nuance mix.NLU model developement and training:

  1. An ontology including the intents used in the model as well as the supported features, e.g. “Request_TakePhoto”;
  2. A dictionary covering the relevant keywords (operator, task objects, etc) and their labels (e.g. “Victim”);
  3. A set of annotated sample sentences.

Please note that you have to register as Nuance Developer to use the mix.NLU development platform: https://developer.nuance.com

Once your membership is confirmed, you can access the Mix.nlu toolkit under https://developer.nuance.com/mix. Adding a new project, you can choose either to develop the required elements from scratch, or to import an existing model source (a TRSX or XML source file).

Please refer to the Mix.nlu Documentation for further details: https://developer.nuance.com/mix/documentation/