AI Expert Discusses Conversational AI for Enterprises

The Greek philosopher Aristotle believed technology is an extension of nature: A construction crane extends man's bare hands, and a car replaces a horse-pulled wagon. Our ancestors consulted prophets and soothsayers for clues on what lay ahead in an unpredictable world.

But in a 24/7 connected world, supercomputers take the place of human oracles — logical-rational machines whose purpose is to inform options and/or solutions for man's imminent decisions.

Aristotle may be right. Artificial intelligence is becoming an extension of our brain's 100 trillion neural connections. If creating effective AI is an extremely complex, if not a near-impossible task, it reflects the complicated truth-seeking that modern man expects from AI to apprise his infinite queries and dilemmas.

The human-user increasingly wants AI to possess human-like intelligence to unmask and decode his surroundings. AI, therefore, has to decipher and interpret man's limitless mind to satisfy his labyrinthine requests and pursuits.

Interview With Cisco's AI Expert

Vijay Ramakrishnan, a fellow at Cisco’s Machine Learning/Artificial Intelligence Virtual Center of Excellence (https://www.cisco-ai.com/), talks about developing machine learning models and dialogue systems for enterprise conversational applications. Mr. Ramakrishnan is machine learning engineer at Cisco (NASDAQ: CSCO).

What does conversational AI mean?

Vijay Ramakrishnan: Conversational AI is a set of engineering disciplines centered around machine learning, natural language processing, and dialogue systems development that are used to create AI agents that can intelligently interact with humans. The input methods have traditionally been voice through automatic speech recognition (ASR) technology as well as text via smartphones and keyboards. However, we are seeing the emergence of vision (as an input method) that can improve the intelligence of AI agents. The output of such systems can be varied from natural language responses from an assistant to displays on a screen.

Image 1. Chart showing accuracy improvements in speech recognition. Photo credit: Economist.com

How are conversational AI applications for enterprises different from consumer or academic settings? 

Ramakrishnan: Enterprise applications have deep links to multiple business knowledge-bases that consumer and academic applications tend to avoid. In 2018, Amazon's Alexa Prize (an industry-leading competition) provided an objective for participants to build an agent that could chat with a human about random topics like weather or sports. (Source: https://developer.amazon.com/alexaprize/challenges/past-challenges/2018/)

In academic settings, the objective is to surpass a benchmark dataset like Stanford's Question/Answering dataset in reading comprehension. This dataset is built by posing data from Wikipedia to crowdsourced workers for labeling.

With enterprises, the AI application has to accomplish narrow domain tasks extremely well, such as booking a flight ticket. Such tasks require deep links to customer and business data that tend to be propriety and smaller in size than public datasets.

Therefore, the resolution of user speech/text to business objectives through information-retrieval techniques are key to narrow-domain task completion. Second, machine learning models that specialize in smaller datasets are important in building accurate enterprise conversational applications.

Image 2: General architecture of Enterprise Conversational AI systems. Photo credit: Vijay Ramakrishnan

What are the challenges you face in designing voice assistants for enterprises?

Ramakrishnan: A big challenge is recovering from mistranscriptions of automatic speech recognition (ASR) systems. Enterprises have company-specific pronouns like employee names, product abbreviations, and other proper nouns that have never been trained on ASR systems. An enterprise system would have to correct the often mistranscribed word for "Lufthansa," which is "Of Sansa."

One solution is to use contextual information surrounding a user’s query. For example, based on past user interactions, the enterprise application can infer that a user is trying to book a flight and isolates "Of Sansa" to be matched against airline names. By using various textual matching techniques, we can further resolve "Of Sansa" to "Lufthansa."

Another challenge is the lack of high-quality labeled data in an enterprise setting. Enterprises need to invest in data collection early to reap the benefits of incremental language understanding of users. Moreover, enterprises can augment their private datasets with larger public ones so that machine learning models can learn general patterns as well as from company-specific literature.

Finally, enterprises need to bootstrap their applications with simpler machine learning models like logistic regression on their small datasets. These models can provide enough predictive power for most use-cases. As the product matures and training data sizes increase in magnitude, deep-learning approaches can be used to gain accuracy and precision in human language understanding.

Image 3: Here are popular voice assistants in the marketplace. Photo credit: Geeksfl.com

What are the hurdles to understanding foreign languages?

Ramakrishnan: Internationalizing an enterprise AI application is key for driving growth. But it's challenging to support different languages if the team is unfamiliar with the language in question. Enterprises need to have an operational plan of how to support each language, including sourcing [foreign] language experts, before rolling out the application to a particular geographical area.

The Takeaways

AI agents are expected to interact intelligently with humans, and platforms are using voice, text, and vision as input methods. Amazon and other big players are shaping the future of the industry. For example, Jeff Bezos's company is incentivizing innovators to develop AI agents that can talk with humans about random topics.

But with enterprises, it's crucial to develop applications that can accomplish narrow, business-oriented tasks extremely well — if not near-perfectly. When conversational AI can help customers book flights or hotels, companies can see ROI on their machine-learning investment. Conversational AI for enterprises tends to work with smaller datasets. Companies like to protect their proprietary information.

A key challenge to perfecting AI for commercial use-cases is to overcome mistranscriptions of proper nouns, names, company-specific abbreviations, products, and similar terms. Developers can use various methods to accurately assess a user's query such as leveraging contextual information. Translation is another challenge. Companies should leverage foreign language experts to inform developers on nuances of non-native language.