As someone who has spent many years working in the semantic technology area of AI, I was recently asked by Dataversity to provide some predictions for their annual year-end article on what’s coming in this field in the year ahead. This post elaborates on the input I provided to Dataversity about where semantic AI may be headed in 2019 and beyond.
First, a little retrospective. Semantic technologies driven by natural language processing (NLP), both voice and text-based, have really exploded in the past year. In particular conversational capabilities, as utilized in bots, virtual assistants, customer care apps and other UX-intensive use cases now seem to be popping up everywhere. Although the use of semantics is most visible on the front-end of those systems for capabilities like natural language interpretation and understanding (NLI/U), there’s also a lot of semtech on the back-end, too. For example, semantic-based processing of some kind is likely taking place — in conjunction with statistical-based machine learning approaches and some scripted content — to formulate specific answers to questions or to provide summaries and other refinements to the otherwise raw results that are extracted directly from unstructured text sources.
As we head into 2019 I think these NLP-based conversational interactions will definitely continue to increase and improve. There’s actually been some backlash this past year against bots in particular, mainly because so many of them were poorly implemented or didn’t really provide users much value. So there may be a pause or slowdown specifically with bots in 2019 until vendors of bot technologies and the businesses that use bots can figure out how to mitigate or fix the current issues. I think the way to do that is to make the bots smarter in at least a couple of important aspects. First, bots need to be less rigid in how they interact with users, ideally allowing users to interact naturally as if they were communicating with another person. Currently, users too often have to adapt to the way the bots want to be interacted with, when it should be the other way around. Second, bots and other conversational systems also need to improve their understanding of what the users really want to know or what they are trying to do. That only comes through a better understanding of context and the deeper meaning behind the words the users are speaking or typing. Using ontologies or knowledge graphs as part of such applications offers a way to help in that regard.
I also think interactions with these conversational systems will start to move from primarily simple informational requests or single-point transactions to much more of a true transactional mode. By that I mean a series of smaller discrete transactions chained together in an intelligent and relevant work flow, something that would then culminate in the user achieving some bigger result that provides more user value. For example, planning dinner before you leave work involves deciding what to eat, whether to dine out, pick up take-away on the way home or arrange delivery once at home, and where to make the purchase. This in turn involves choices and trade-offs involving location, price, timing, and the range and quality of selections, perhaps also taking into account other events you may have on your calendar, as well as those of other affected people such as friends, roommates or family members.
Apart from audio and text-based natural language processing, another area where semantic AI technology is now beginning to be more widely applied is video. One major application area is of course security, in specific video surveillance. Whether for home, business or public security, analysis of surveillance video using sound patterns, object recognition and identification, and other forms of statistical and semantic AI is likely to increase significantly in the future. More generally, as video becomes the dominant medium for news and entertainment, marketing and advertising, and even retail, it also becomes a critical content asset to leverage more effectively. So as we head into 2019 there will likely be growing interest in analyzing videos and enriching them with semantic metadata so they can be better managed, used and reused in more targeted and effective ways.
As mentioned, some applications use a combination of statistical machine learning-based approaches complemented or augmented by a knowledge graph or other form of ontology or knowledge-based semantics. Google and Microsoft have been doing this for several years and more recently other major tech companies like Apple, Facebook and Amazon are, too. Though the use of this combined, hybrid approach to building intelligent systems is starting to become more common, it still more the exception than the rule. That may finally be changing as we head into 2019. One bellwether example I was recently surprised to discover is that Uber appears to be using knowledge graphs behind at least some of its applications, too. I don’t know the exact extent of use, but I’d imagine it is to better understand customers and the things they’re doing — in other words the user context — so Uber can be more proactive about meeting its customers’ needs. I don’t know whether this involves applying the knowledge graphs mostly to rear-view analytics as feedback to drive improvements in applications and algorithms, or whether some of the semantic processing is being done in real-time. The goal is certainly to be able to do that type of analysis in real-time or near real-time, as that enables more tailored user interactions, as well as opportunities to offer users additional related products and services at the moment when those are most relevant. This is as example of how intelligent applications, including assistants, could become more proactive and action-oriented in a transactional context. And that’s where the additional user value really starts to appear.
There are lots of challenges to overcome though before this type of capability becomes common place, and I think it will take much longer than just 2019 to address all of them. Not only do you need good algorithms and knowledge models, but those would be best deployed either on the device itself or at least near the edge. Going to a centralized cloud for complex transactions requires a lot of time and a lot of data movement back and forth. But putting that intelligence and associated processing capability at or near the edge, such as on a smart device like a phone or speaker, will also require more technology improvements. So I think semantic AI technology vendors, both from a hardware and software standpoint, have their work cut out for them in the next few years. That effort will be worth it though for more fully transactional assistants and other semantic applications that exhibit more realistic, human-like intelligence than the ones we’ve had up until now. I don’t know about you, but I’m anxious to see where this goes in 2019 and beyond!