A Natural Part of Natural Language

Article Featured Image

As natural language (NL) speech applications move into the mainstream, chances are your enterprise will install the technology during the next 12 to 18 months. While NL applications are very powerful and beneficial, these projects can become very complex and harmful if not managed correctly. 

The NL project’s lifecycle comprises five basic phases: initial planning, requirements gathering, data collection, utterance concept segmentation, and deployment. During each, several key issues can arise.  

Initial Planning

Key issue: Stakeholders want to define a project completion date prior to beginning the project. 

Advice: Because caller utterances are the foundation of an application, it is difficult to determine final requirements, deployment strategy, and a project end date until you have transcribed and analyzed them. Therefore, it is best to manage an NL initiative as two separate projects with separate completion dates. The first project is data collection and analysis to help better identify the final requirements and design strategy. The second is the development and deployment of the application. Managing a project in this manner will help set stakeholder expectations and create realistic project completion dates.

Requirements Gathering

Key issue: Stakeholder misconceptions regarding NL and how it will be used.

Advice: Because callers can say in their own words the reasons for their calls, an NL application will touch virtually every part of an organization. It is important in this stage to involve individuals from all groups within the organization. Having all stakeholders represented is beneficial for two reasons. First, it is important for the stakeholders to have representation in the process. Second, their involvement is a good way to begin the education process on NL. 

Data Collection

Key issue #1:  Using a different prompt in the production application than the one used during data gathering.

Advice: The NL prompt used to gather utterances should be the same prompt used in the final application. If you cannot decide on one NL prompt, narrow the choices to two and conduct separate data-gathering projects. Having data specific to the final NL prompt is important because any wording variations in the prompt will impact the way callers interact with the application. If caller behavior varies drastically, the Statistical Language Model (SLM), Statistical Semantic Model (SSM), or interpretation grammars will be invalid.

Key issue #2: It takes time and resources to transcribe caller utterances.

Advice: It is important that every utterance captured is accurately transcribed. This data will become the foundation for the SLM and SSM or interpretation grammars. Partial or inaccurate data will cause recognition rates to suffer, so leave ample time for this important process.

Utterance Concept Segmentation 

Key issue #1: Each stakeholder could  have a different opinion about the meaning of caller utterances.  

Advice: It is vital that all stakeholders participate in the utterance segmentation process, which dictates how callers will be handled in the application. Managing this part of the project is a delicate task because stakeholders from the various groups will have different interpretations of caller intent based on their own perspectives. It is important that the core project team manages these issues and makes the final decision as to which utterances fall into the various segments. Failing to properly manage this phase can cause misrouted calls, reductions in customer satisfaction, and rework efforts.

Key issue #2: The tagging process is time-consuming and labor-intensive and requires attention to detail.

Advice: It is essential that the results of this process are 100 percent accurate. Therefore, it is important to keep the tagging team small. Having too many people involved can lead to variations in understanding and interpretation of caller utterances, which will lead to tagging inaccuracies. 


Key issue:  Stakeholders want to rush the application into production to quickly realize the savings.

Advice:  Don’t rush deployment. It is important to follow a controlled rollout with tuning at the end of each phase. Following this plan will allow applications to be refined prior to full implementation. 

Aaron Fisher is director of speech services at West Interactive, overseeing the design, development, and implementation of speech applications for the company. He can be reached at asfisher@west.com.

SpeechTek Covers
for qualified subscribers
Subscribe Now Current Issue Past Issues