Conversing with Computers: Where We Are and Where We’re Going

AVIOS held its live “Conversational Interaction Conference” in San Jose, Calif., in April, during which we examined the progress of computers handling human language, and the diversity of the 40 talks reflects the strong momentum behind this trend. We obviously can’t get into all the presentations in our limited space here, but I’ll indicate some major trends and illustrate them with summaries of a few talks. (The conference website is still available and contains many of the presentations.)

Several talks emphasized the importance of process and platforms in delivering successful conversation solutions. Phil Cohen, chief scientist at Openstream.ai, gave a talk titled “Platform for Collaborative Multimodal Plan-Based Multiparty Dialogue Systems” in which he presented a plan-based approach to dialogue, arguing that such an approach is how we should be thinking about and building future dialogue systems. And in his lecture “Lucrative Conversational Commerce with New Chatbots,” Felipe Hlibco, developer relations engineer at Google, described Google’s Business Messages platform, which supports the objective of integrating virtual and real agents.

Another key trend is increasing the speed of machine learning (ML) by hardware enhancements. Tom Spencer, senior manager of product marketing at Achronix, said that the company’s hardware solution can achieve the result in his talk’s title: “Reduce Your Speech Transcription Costs by 90 Percent.” Achronix offers an ML-optimized architecture in its Speedster7t family of field-programmable gate arrays (FPGAs). The company’s Myrtle.ai software is an ML inferencing solution for the FPGA. Spencer showed how the company’s Myrtle.ai machine-learning compute engine, integrated into the Achronix FPGA, can lead to drastically decreased transcription costs.

Other speakers emphasized lessons learned from deployments. Lisa Falkson, a senior VUI designer at Amazon, shared her conversational AI best practices, based on having analyzed millions of chatbot and voice messages over thousands of conversational interfaces and use cases. Falkson detailed the full range of tasks required for a successful solution, including building NLP models, the role of context and personalization, handling fallbacks, integrating with back-end systems, and testing, monitoring, and measurement.

Vineet Mishra, principal user experience designer at Oracle Conversational Design, noted that despite the increase in chabot platforms over the past few years, poor conversational experiences still happened all too frequently once users went off the beaten path, or even when they asked something the systems should expect and be able to handle. The lack of contextual understanding can be surprising, so dealing well with the unexpected is critical.

Another theme was the need to analyze multiple sources to provide answers to the most frequent users of conversational technology—callers to contact centers. Marie Meteer, president of AVIOS and senior research scientist at Pryon, addressed this in her talk “Bridging the Gap between Chatbots and Question Answering,” emphasizing that chatbots need to be able to take advantage of existing structure in a company’s knowledge management and documents. And in “Creating Empathy at Scale,” Shadi Baqleh, COO at Deepgram, stressed the need to analyze the large number of voice files collected in customer service calls. He said that the company’s Deep Learning Speech Recognition can transcribe 10,000 hours of audio in fewer than 10 hours with transcription accuracies of more than 90 percent. Analyzing this rich source of data can lead to better handling of calls.

Finally, I attempted to look at the long-term implications of conversational technology in my talk, a summary of the argument in my recent book Evolution Continues: A Human-Computer Partnership. Dealing with computers using human language is the closest thing to a direct connection to the human brain we can have without connecting wires. We are in the early phases of establishing this connection, but as digital assistants get increasingly able to maintain a conversation, become increasingly personalized, and are always available through mobile devices, humans will integrate their use tightly and intuitively into their lives. Human evolution has always been driven by our tools. One thus might even call this human-computer connection the next step in evolution.

Whatever the long-term trend in conversational technology, the importance of the short-term trend, the rapid improvement of conversational applications, is evident, and the conference detailed the many exciting developments happening in this critical field. 

William Meisel, Ph.D., is executive director of AVIOS and president of TMA Associates (www.tmaa.com). His most recent book is Evolution Continues: A Human-Computer Partnership.