One of the most exciting panels at the 2022 ILTA Conference (What is Natural Language Processing and How can I Use It?) showcased dramatic breakthroughs in neural net technology to natural language processing (NLP). Pablo Arredondo, Co-Founder / Chief Innovation Officer, Casetext provided an intensely passionate tutorial on the history of AI through the centuries. The launch of commercial research systems such as Lexis and Westlaw in the 1980’s dramatically transformed legal research by making cases and statutes keyword-searchable. Arredondo enthusiastically proclaimed that neural net technology can now “free research from the keyword prison.”
Let’s start with a definition. Natural Language Processing is a multi-disciplinary field concerned with the interactions between computers and human language. NLP is also generally considered a sub-field in the broader field of artificial intelligence. As with other areas of AI, there are two major approaches to natural language processing. Symbolic AI is a rules-based approach that remains largely tied to human experts and keywords (think chess grandmasters giving advice to a computer). With the other approach, machine learning, algorithms learn not from humans but from direct experience with training data (think of a computer teaching itself to play chess by playing over and over). Neural nets, a type of machine learning loosely inspired by the way biological neurons connect, represent the most powerful form of machine learning.
The application of neural nets to language involves setting up simple (for humans) language tasks such as “guess the missing word.” * The neural net starts in a random state and its guesses are laughably bad. But through a massive amount of trial and error and self-correction, the neural net “trains” itself to more accurately predict missing words. In the process, for reasons that still a bit of a mystery (or “black box” to some), the neural nets get quite “smart” about language. Smart enough to, in Arredondo’s words “have profound and immediate implications for the legal profession.”
The star of the show was Casetext’s AllSearch tool, which leverages neural nets to enable attorneys to search over their own documents. Two attorney panelists – Scott Reents, Lead Attorney, Data Analytics and E-Discovery at Cravath, Swaine and Moore and Samantha Seaton, Knowledge Management Counsel at Fisher Phillips – provided jaw-dropping examples of how neural nets are helping in the trenches of litigation.
Reents discussed how Casetext’s AllSearch tool enabled Cravath to more quickly find critical documents in a litigation. The ability of neural nets to take an entire sentence as a query enabled Cravath attorneys to search with more granularity; their ability to match relevant language even if articulated completely differently uncovered key evidence.
Reents described how Cravath was even able to turn written discovery directly into search queries. Interrogatory requests could simply be cut-and-pasted into the AllSearch engine to surface documents that would help respond.
Finally, Reents described using a different neural net to run sentiment-analysis on a document production. This tool quickly isolated the kind of “emotionally charged” statements so often correlated with useful evidence. One striking example he provided involved locating a relevant email where someone said “[like] bringing a knife to a gun fight.” Knives and gunfights aren’t “top of mind” when lawyers are reviewing internal corporate documents; this email would have been invisible inside the keyword prison.
Reents’ primary focus was on e-discovery, but he mentioned that Cravath associates were finding neural nets equally useful to search case law. Reents joked that the only complaint he heard was that Casetext’s Parallel Search (the neural net engine aimed at case law) sometimes made finding on-point authority “too easy.”
During her remarks, Samantha Seaton put the power of Allsearch into historical context. Back in the “bad old days”, lawyers had to compile and flag hundreds of documents, exhibits and statements, organized into massive printed binders. The lawyers’ brains were the neural nets and they were aided only with post-its and binder tabs. Seaton described being an associate tasked with trying to memorize all the important documents and statements in the weeks leading up to trial. If a sudden need emerged to impeach a witness, attorneys would have to hope someone had marked the relevant documents in a way that was easy to find. Seaton contrasted this with the power of neural nets, describing how loading up 39 days of witness testimony into AllSearch, and bringing it into the courtroom enabled lighting fast “on the fly” searches to locate contradictory evidence
Bespoke Client Repository
Seaton discussed how Fisher Phillips deployed AllSearch to help support an internal client need. The client had over many years amassed hundreds of union contracts covering a wide variety of jurisdictions. There was no standard terminology, so even if the documents could be scanned and keyword searched, similar contract terms could not be located easily. That is until Fisher Philips pointed neural nets at them, giving their client dramatically better accessibility to the information in the documents without the need for Boolean queries or even familiarity with legal terminology.
One application that has been widely embraced by litigators at the firm is using Allsearch to validate facts in a brief. According to Seaton, litigators can write a brief quickly, but they get bogged down when they have to locate every instance in a trial record that supports a given statement in the brief. This is particularly painful in class actions where there are thousands of pages of deposition transcripts. When the entire case record and document production is uploaded – this verification can be accomplished with lightning speed. Attorneys spend less time on the painfully slow process of keyword retrieval and can focus on higher value work.
The Next Frontier - Answering legal and factual questions
Seaton also provided a fascinating preview of an early stage prototype that Casetext gave her access to. The prototype leverages an approach called Fusion-in-Decoders and combines neural net search with neural net synthesis – enabling a much more robust form of question-and-answer to be run against case law and litigation records.
The powerful use cases amply proved the key takeaway of the panel. Neural nets, even in these early days, are demonstrating they can markedly improve the practice of law.
*For a more in-depth description of this technology see Arredondo and Qadrud-Din, "From Vellum to Vectors," AALL Spectrum, Vol. 25, No. 5 (May/June 2021)
Editor's Note: Want to compare more than one neural net search tool? Check out Syntheia's Text Search APIs, providing plain language neural net search functions combined with auto clause extraction for contracts: https://docs.syntheia.io. The neural network API is so new it doesn't yet have a discrete product name, but as soon as it does, you will be able to find it on Legaltech Hub, along with Casetext's AllSearch.