Best Practices to Help NL Evolve
Natural language (NL) speech applications are more common now than they were several years ago. These applications have been beneficial for the many enterprises that have already deployed them; however, speech practitioners and followers know that designing, developing, and implementing them is a complex undertaking. Prior to designing and deploying such applications, there are five best practices to consider:
This is one of the more important activities in NL deployment, mainly because data gathered during this process will serve as the foundation of your Statistical Language Model (SLM) and either your interpretation grammars—if you use robust parsing—or your Statistical Semantic Model (SSM)—if you choose a product like open call steering. You should create a preliminary application that will capture caller utterances in response to the proposed NL prompt. After callers state the purposes for their calls, they can be forwarded to a live operator or placed into the existing IVR to complete their calls.
After collecting the data, it’s time to transcribe. It is imperative that the utterance transcriptions are 100 percent accurate. This activity is very labor-intensive and time-consuming, so plan accordingly.
Utterance Concept Segmentation
This is also very complex, labor-intensive, and time-consuming, yet it is a vital part of the NL deployment process. Analyze the data for high-level concepts first and use subsequent analysis iterations to identify sublevel concepts. This process needs to be a collaborative effort between the development team and subject matter experts. It’s dangerous for any individual group or organization to attempt to develop a concept segmentation strategy without involving subject matter experts.
The development team must work with subject matter experts to ensure proper labeling of all utterances. Take the data at face value and don’t make broad assumptions about caller intent. If an utterance seems unclear, it’s better to place it in a concept segment that will prompt a reattempt rather than force the caller down a path that may not address his needs.
Deployment and Tuning
Since the data you’ve collected up to this point is not all-inclusive, follow a strict, controlled rollout plan with finite goals at each phase of the plan. Part of your goals should contemplate application performance from a tuning perspective.
Tuning NL applications is more complex than tuning a static grammars application because there are considerably more moving parts. Review your data to determine if additional utterances are needed to broaden your SLM and if new pertinent concepts are being introduced that were not captured in the data collection. Ensure that your interpretation grammars or SSM provide adequate coverage of caller utterances, review your data to determine if callers are being incorrectly routed, and tune your application extensively.
Each of the previous steps requires strong attention to detail and a clear understanding of the business and customer goals and objectives. A number of tools can be used to save time and improve quality; especially for larger NL projects, these tools could be well worth the investment.
The evolution of natural language speech recognition represents a major development for the speech industry as a whole and makes it possible for enterprises to improve customer service, reduce costs, and increase revenue. NL also gives the industry a chance to win back callers alienated by telephone-based customer service because of past confrontations with poorly designed, overly complicated telephone self-service systems.
As vendors and enterprises collaborate in designing and deploying NL speech applications, it is imperative that we focus on getting it right the first time so that we make a positive impression on callers. In doing so, we’ll be helping callers, and the companies they’re contacting, achieve their respective goals more effectively and efficiently.
Aaron Fisher is the director of speech services at West Interactive, overseeing the design, development, and implementation of speech applications for the company. He can be reached at email@example.com.