Following up on my chat a few months back with Randy Goebel, I convinced Randy to take time out of his busy schedule to chat with me about where he sees AI initiatives going in 2018 (both our own homegrown work, as well as broader macro trends), what his hopes are for our newly released (and free!) EVA platform, and what misconception he’d most like to address about AI. Randy is one of the foremost researchers in the machine learning and reinforcement learning space, and on top of being an all-around great guy, brings diverse experience in both academia and industry to our team. For our first time readers, seasoned AI vets and everyone in between, this is a great article on the state of the industry, and I hope you have as much fun reading it as we did writing it!

Off the jump, when it comes to the ROSS platform, what excites you about what already is built?

The quality and depth of analysis for question answering and information retrieval. As researchers working on NLP and legal informatics, we understand the challenges in building the underlying machine-learned models, and my team is pretty excited about what’s already been accomplished.

When it comes to the ROSS platform, what excites you about what is yet to be built?

The biggest challenges we tackle in NLP related to the AI-complete challenge of automated summarization at multiple levels of detail.  Multi-level summaries are necessary for multi-level explanations of legal reasoning (and, of course all other domains, e.g., finance, medicine). What excites us is how ROSS provides a stable platform of curation of legal texts, upon which we can experiment with new methods, and refine existing methods within a business model that provides incremental guidance in assessing priorities. Since the problems we tackle are typically motivated by general AI, any assessment of where immediate value lies helps us focus our own research resources.

As you know we recently released the EVA platform, completely for free, in order to further access to justice by putting cutting-edge AI legal technology into the hands of everyone in the industry. What are your thoughts on EVA?

EVA is the best demonstration of legal information access that we have seen. We use it as an example of the kind of platform we believe should exist for a variety of legal scenarios, not just for expediting legal research for law firms, but for access to law … sometimes referred to as access to justice … for all people.

Some folks in the legal industry with no experience deploying artificial intelligence noted that EVA seemed like it needed more time before deployment, can you explore why releasing an early product is especially important when it comes to AI?

Related to number 3 above, most AI scientists have long ago abandoned the idea that we could ever “get it right” in any meaningful way, so incremental improvements and feedback from users is an essential part of the overall process of developing and delivering value. One can’t move in a positive direction without starting somewhere, and EVA is clearly a great start.

Remember that NLP research is only beginning to scratch the surface of what can be done with the automated understanding of text, and reasoning therein. Our experience across a broad spectrum of domains is that the best way to guide progress is by building demand, and understanding priorities emerging from that demand.

Another misconception is that the EVA tool, and ROSS for that matter as well, will remain static and will never improve. Can you explain for folks not so familiar with AI how companies release products, and how their current frameworks support the future of their releases?

There are two kinds of improvement. One comes from the continuous incremental curation of legal information from all sources, which helps guide incremental improvements in particular domains. For example, in contract law, the curation of contracts and annotation of salient summaries will help the continuous adjustment of supporting the process of assessing contracts, e.g., for mergers and acquisitions.

The second kind of improvement arises when users begin to not only trust the support system they access, but are confident enough to drive it to the edge of what it can not do. From that demand side (versus data side), new process demands will emerge. For example, by classifying or clustering an accumulating corpus of questions, one should be able to target the development of appropriate target summary representations, so that a particular case or set of legal cases support an explanation for a judge or lawyer from one kind of representation, and another for the law student, and even another for the legal layperson.

What are you most excited about with respect to the future of the legal industry?

Once legal information representation and reasoning systems build confidence in discovery and research, one can image exploiting the same capabilities to help design legislation that is easier to understand and apply. So one doesn’t just use legal informatics to automate parts of what is currently the normal practice, but then provides legislators and regulators with tools to consider the creation of laws.

What is a misconception about AI and law you would like to clear up as we wrap things up?

The current biggest misconception arises because of the recent success of a handful of machine learning methods, especially deep learning. We tend to get excited about how easily one can engineer high accuracy classification systems, and then think that the whole world is a classification. We repeatedly fall into thinking characterized by popularity of method, so that, to use the hammer analogy, every problem looks like a nail.

AI contains machine learning as a proper subset, and machine learning contains deep learning as a proper subset. The misconceptions maintained by those who believe deep learning solves the AI problem will create an accumulation of unrealized expectations.