Realizing and implementing vocal services using natural language
Natural language is the capacity for an automatic service to offer users a simplicity in interactions which is close to what they would experiment with a human being.
Interactive vocal services are generally limited by the keypad possibilities, in particular by the number of possible combinations leading to sometimes complex tree structures in dialogue services. They are also limited by systems that rely on speech recognition with command words, because of the need for callers to know the authorized list of words. In addition to the apparent complexity of menus composed of numerous choices or of a long list of command words, this type of interface forces users to make the effort to identify the branch that corresponds to their need within the service structure.
Interactions in natural language offer more flexibility and liberty to users to express their requests. They also offer the possibility to develop more complex functionalities within vocal services (e.g. "self-care"). Consequently, the effort users make to identify the branch of the service that corresponds to their request is transferred onto the technology, thus letting users simply express their request.
Within a service, of course, dialogue steps and branches vary in complexity and do not all require natural language. For example, responses such as yes/no do not necessarily require understanding complete sentences. It may be more natural to enter numbers using the DTMF keypad of the telephone, which in addition allows keeping the confidentiality in the case of PIN code for example.
Implementing natural dialogue within a service thus means implementing the right technological levels to maintain simplicity and effectiveness throughout the interaction.France Telecom's software
With a wealth of several years of experience behind it, France Telecom has developed a complete solution for the design and the development of vocal services, including the capability of handling natural language interactions. This solution includes a graphical service creation environment as well as a runtime environment.
This solution supplies the full realization process with tools, and includes:
For the specification and design step:
• a graphical studio for detailed specification and design
• a graphical studio for the definition and development of the interpretation of callers' requests
For the development phase:
• a code generator
• a graphical interface to facilitate the coding of the non-generated parts of the service
• the possibility to use any external development environment in Java
For the deployment and tuning phase:
• a runtime –called phase engine- executing the developed service
• a semantic analyzer
• a dialogue database including tools for transcription and semantic annotation of the corpus, as well as usage statistics production.
The usage analysis functionalities facilitate the tuning of the service during a pilot phase or during the deployment of the service. They require the same tools as for the design and development phases.
The solution proposed by France Telecom makes it possible to develop services that let users interact with the machine by using either a touch-tone keypad (DTMF), single-word speech recognition ("help", "subscribe", etc), natural dialogue ("I want to access my account", etc.) or a various mix of these possibilities.
Features of the solution
The solution is based on the VoiceXML recommendation of the VoiceXML Forum and of the W3C. The phase engine thus produces VoiceXML pages interpreted by a vocal platform related to the caller.
The code run by the application server is Java code and includes the runtime, the semantic analyzer and the application.
The platform architecture required for service execution is the architecture recommended by the VoiceXML Forum, called 3-tier architecture:
The dialogue database can be hosted on one of these servers or on a dedicated server. Traces and logs collection is done automatically on the appropriate media servers and application servers.
