Everyone is talking about artificial intelligence but few understand how it can be used in legal research.
AI systems promise to scale a lawyer's capabilities and capacity. Lets walk through how ROSS' AI Search makes that a reality.
Currently, lawyers run into two issues when performing legal research:
ROSS helps you avoid both situations. To illustrate how, let's take a look at a simple example of a typical legal research question: “What is the standard for gross negligence in New York after 2004?”
Here’s what ROSS does to find the most relevant cases in a snap. Scroll to follow the process chronologically.
We can break down the ROSS AI Search process into three main categories.
When you submit your query, ROSS analyzes the words using our own proprietary Natural Language Processing algorithms. The algorithms understand the time period and jurisdiction of interest and automatically apply filters to focus your query on those places and dates.
In our example question, ROSS detects “New York” as a jurisdictional filter and will only retrieve cases from New York state and federal courts, the Second Circuit Court of Appeals and the U.S. Supreme Court. It also detects “after 2004” as a date range. ROSS then automatically applies those parameters to focus your search results.
Now that ROSS has identified the appropriate date and jurisdiction filters in your query, it will retrieve the passages from our comprehensive corpus of U.S. case law that are most similar to the meaning of your query.
We find those passages by using a combination of industry standard search functions and proprietary algorithms. For example, this diagram illustrates the citation graph analysis we use to retrieve seminal cases that might otherwise be missed.
You may have noticed that instead of looking for cases ROSS is looking for passages within cases. We do that because answers to legal research questions are found within passages of text, not the case as a whole.
We aim to find precise answers to questions and not just keywords within documents.
At this point, ROSS has retrieved relevant passages and cases that contain an answer to your question. However, they are not ordered according to relevance.
We’ll address that in the next step.
Once we’ve retrieved all the cases relevant to your query we use our industry-leading AI algorithms to rank them so that you see the best cases first.
ROSS does not limit your search by merely counting the keywords in your query. Emphasizing keyword matches is an outdated and inferior way to rank search results because it often misses the point and intent of your question.
It’s helpful to understand four capabilities of our AI technology that we've combined to give you the best search results every time.
A. Machine Learning
Machine learning is a way to teach computers how to learn for themselves. We are teaching ROSS to recognize patterns of context, syntax and meaning within legal documents.
Supervised learning is a machine learning technique that uses examples to teach a computer through a process called training. During the training process, ROSS finds the patterns in the training data that yield the best answers.
We’ve trained ROSS to calculate the likelihood that a case passage will answer your question by showing ROSS over one million examples of questions and answers developed by lawyers.
B. Grammatical Structure
Our AI Search understands grammatical structure and the way the meaning of sentences changes depending on the relationships between types of words. Our technology correctly identifies the parts of speech (noun, verb, adjective, etc) in each query and groups them into discrete phrases (e.g., gross is an adjective that modifies negligence.).
Let’s return to our example question:
“What is the standard for gross negligence?”
Lawyers know that gross negligence is distinct from mere negligence. Thanks to machine learning, ROSS understands that too.
When ROSS is scanning the corpus, it understands multi-word terms of art such as “gross negligence”, “absolute discharge” and “fraudulent conveyance.”
ROSS takes the relationship between words in your query and looks for passages in cases that have a similar structure.
Our AI Search understands that "the boy loves a girl" does not have the same meaning as the "the girl loves a boy" even though they have exactly the same keywords. That's because the grammatical relationships are different.
A legacy search sees each word in isolation and simply looks for those words in different cases.
D. Word Embeddings
ROSS uses word embeddings as one of the key features of its AI Search strategy
Word embedding is a mathematical technique for placing words in a mutli-dimensional space so that words that are similar in meaning are closer together. This allows us to find words that are similar to or related to the words in your query and compare them to those in cases.
For example, lawyers come to understand intuitively that “duty” and “negligence” are related words. ROSS uses that relationship to broaden its understanding beyond simple keywords. If you use “duty” in a query ROSS may, depending on other factors, return cases that use the word “negligence” but not the word “duty”, because the word embeddings indicate that those two words are closely related.
ROSS’s ability to see patterns that a lawyer might not intuitively perceive allows ROSS to bring a deep richness to its responses to your queries. ROSS is specially trained on legal documents to use word embeddings to recognize and understand the context, syntax and meaning of case law. By using word embeddings, ROSS matches cases to the intent and meaning of your query, rather than just the bare keywords.
E. Facts and Motions
ROSS’s proprietary retrieval and ranking algorithms help you find cases that keyword search platforms miss. But that’s not enough. We designed ROSS to consider additional context by matching the facts and procedural posture of your case. ROSS then helps you find the winning case by highlighting the additional context it has identified.
ROSS could not complete this task without machine learning. Since cases do not always contain clearly marked or expressed facts and motions, keyword searches generally miss them.
Once we complete our holistic set of feature analyses for each case passage we use another machine learned model to combine each feature into a single score. ROSS uses the score to determine the relevance of a case to your query and order your search results.