How neural networks free legal research from the prison of keywords
Virtually all research tools – whether for legal research, eDiscovery review, document review, or anything else – are confined to indexes. Every word in every case or contract or email or whatever is fed into an index, and the search tool searches for that word or a combination of indexed words.
Without a doubt, keyword research is powerful. It has radically changed the way lawyers and legal professionals work. Over the years, research scientists have enhanced it with Boolean and natural language queries.
But the bottom line is that if the word you’re looking for isn’t in the index, you won’t get any results.
Why is this important? Because in legal research we should be able to research concepts and circumstances. We should be able to search for legal backgrounds or fact patterns even when the words don’t match.
This is the key word of the prison in which we remain locked up.
Well, your liberator is here, and it comes in the form of the neural network. It’s a technology that allows search queries to find highly relevant results, even when the results don’t even contain one of the search terms.
Not only do neural networks free search from the constraints of keywords, but they also humanize it, making search operations work more like our brains – hence the nickname “neural”.
At ILTACON last week, the most fascinating panel I attended was the one called, What is the natural treatment and how can I use it?during which the aforementioned Arredondo spoke, as well as panelists Damien RiehlVice President, Workflow and Litigation Analytics, at Fastcase; Scott Reents, senior counsel, data analytics and e-discovery, Cravath, Swaine & Moore; and Samantha SeatonKnowledge Management Advisor, Fisher Phillips.
The establishment of the panel led Riehl to argue that there is no “right” form of NLP for every forensic research purpose – that traditional, sometimes rule-based NLP is more accurate, and that neural networks are sometimes the best option.
Arredondo then made the case for neural networks, followed by Reents and Seaton to provide examples of innovative ways to use these research tools in their businesses.
In his speech, Arredondo called neural networks “one of the greatest advancements in the history of research.”
To be clear, Arredondo has a horse in this race. In June, his company Casetext launched AllSearcha research tool based on neural network technology.
“I consider this the most significant product launch in company history by far,” he told me at the time.
During the ILTACON panel, Arredondo gave some examples of the power of neural network research that – to be honest – I don’t have in my notes.
But in a previous article I wrote about Casetext technology, I included some examples Casetext had provided of how an earlier version of its neural network research was able to “understand” words and phrases. in context. For example, when this statement was entered as a query:
“Target employees were not paid while waiting for loss prevention inspections before leaving work.”
The Casetext tool returned the following statement from the case Frlekin c. Apple Inc. (9th Cir. 2020):
“Employees receive no compensation for time spent waiting and undergoing exit searches, as they must clock in before undergoing a search.”
Thus, the search tool understood that “without compensation” was the same as “without compensation”, that “loss prevention inspections” were similar to “exit searches”, and that “before leaving work” was similar to “must point”.
Casetext’s AllSearch goes beyond legal research to all types of legal documents, meaning it can be used for eDiscovery collections, contract review, file banks, litigation records , transcripts of depositions, expert reports or any other collection of documents.
New ways to search
At ILTACON, Arredondo said this effectively opens up two new and powerful ways to conduct research.
On the one hand, you can now simply write a complete sentence as a query, without worrying about whether the sentence contains specific keywords.
“For the first time, you can write whatever you want to write, and the search engine will come to you,” Arredondo said.
On the other hand, you can simply drag and drop any document and search based on its content.
A key feature of neural network technology is that it trains efficiently. You’ve probably heard of the gaming AI programs developed by Google’s DeepMind, starting with the original AlphaGo – which had to be trained on thousands of human games – up to AlphaGo Zero – which was trained only on basic rules of the game, without examples – then to AlphaZero – who received no training of any kind and mastered three different games in three days.
Training a neural network for forensic research is much the same, Arredondo said — you just let it keep playing the game, giving it sentences with missing words and letting it learn to fill in the appropriate word depending on the context.
For humans, it’s easy. Arredondo gave the example, “The man went to the store to buy a WHITE of milk.” We all know how to say “bottle” or “carton”, but not “beaver”.
If you change it to “The woman went to the store to buy a WHITE of milk”, the answer does not change. But if you say, “The man went to the store to buy a BLANK beer,” the answer is there.
“A very, very important breakthrough has happened,” Arredondo said.
Examples and alternatives
In their panel portions, Reents, Cravath’s data analytics and e-discovery attorney, and Seaton, Fisher Phillips’ knowledge management attorney, offered real-life examples of using Casetext’s AllSearch product. with powerful effects – Reents for reviewing e-discovery and Seaton for reviewing collective bargaining collections.
In Seaton’s example, the company created an ACA database and then offered its customers access to the database. For enterprise customers, the benefit of neural network technology was that they did not need to know how to perform sophisticated searches or even have prior knowledge of contracts to perform effective searches in the database.
For his part, Riehl argued that NLP tools fall on a spectrum, grouped between “traditional” symbolic AI using rule-based approaches and the new generation of neural networks and machine learning tools. depth.
In this spectrum, symbolic AI tools are often better at finding specific legal terms and concepts because legal terms are unambiguous, while neural networks may be better at finding facts that involve more ambiguous concepts such as sentiment or actions, Riehl said.
The right tool often depends on the task at hand, he said. NLP can be used for:
- Search for law or facts.
- Label, classify and structure data such as queries and contracts.
- Perform analysis on the data.
- Generate first drafts of documents, sentences or quotes.
The task, Riehl explained, will determine the choice of tool, with structured AI outperforming neural networks in terms of precision and recall in some cases.
He gave the example of how Docket Alarm uses NLP to mark questions and documents for analysis, which I talked about in this article and which Riehl discussed in our recent LawNext podcast interview.
Key points to remember
The panelists offered two key points about NLP:
- Recent breakthroughs in applying neural networks to language have led to explosive gains, with profound and immediate implications for the legal profession.
- Often rule-based NLP (symbolic AI) is more efficient/accurate.
Both of these may be true, but it seems indisputable that neural network technologies are the biggest game-changer in law, freeing us all from keyword prison.