While our focus in previous articles has been on using Large Language Models (LLMs) like GPT and Claude 2 to review and analyze discovery documents, there is more they can do. In this article we will focus on hearing and deposition transcripts. As you will quickly see, LLMs can analyze and report on information in transcripts, providing page and line numbers for their references.
Let’s take a look and see what they can do.
Our Test Transcript
For our exercise today, we are using a transcript for the continued deposition of Joseph Nadeau. He testified in a civil matter involving the Air Force and EPA that we found in the public domain. The transcript is 75 pages long, which seemed to be reasonable for a first test.
We are using Claude 2 for our transcript analysis. Claude 2 is a leading competitor to GPT 4, but offers a larger context window which we need in order to load the full transcript. Specifically, Claude allows us to submit up to 100,000 tokens (about 75,000 words). In contrast, GPT 4 offers a maximum of 32,000 tokens but the version generally available at present is about 8,200 tokens (about 6,000 words). The transcript we are examining contains just under 27,000 tokens (about 20,000 words). We could break up the transcripts into sections that will fit into GPT 4’s context window but that isn’t necessary to demonstrate how useful these LLMs can be with transcripts.
In this case we have loaded the entire transcript into Claude 2. That allows us to work directly with the full text of the document rather than working from a summary or analyze the information piecemeal.
Initial Questions
Let’s start by asking a few introductory questions.
All of the responses came directly from Claude and are unedited.
That gives us an idea about deposition contents and at least a starting description of the underlying litigation.
Creating a Deposition Summary
Most lawyers ask a junior lawyer or legal assistant to summarize a deposition transcript once received. The summaries come in different styles but the goal is to record the important statements or admissions made by the witness. These are often done with an eye toward admissions that might be important for trial or further depositions, rather than less important information like the witnesses’ educational background. But some lawyers want everything summarized.
Let’s see how Claude does summarizing this transcript.
Note that the page and line numbers are correct, and as are printed in the margins of the transcript (which starts on printed page 146, and has 25 printed line numbers per page).
Here is a second try at the summary, this time without focusing on our trial topics:
We can ask Claude to go into more detail on any of these points. Here is an example.
Asking Further Questions
Rather than create summaries, we can simply ask Claude questions about the testimony. Here are several examples.
Using LLMs to Analyze Transcripts
Did Claude do a good job in summarizing the deposition and answering questions about Ralph Nadeau’s Testimony? You be the judge.
Back in the 80s and 90s, John Tredennick spent many hours with a dictaphone, and later a computer, summarizing depositions in order to prepare for trial. No doubt associates and legal assistants are doing similar work today, taking many hours to work through lengthy transcripts. From there, the trial team has to search deposition summaries, or click through search hits from the original transcripts, to find the information they need for direct or cross examination. The cost to do this work can run to the tens of thousands of dollars.
In contrast, Claude answered our questions and created summaries in minutes, at a cost that is surely less than $10 for these reports. Should legal professionals use LLMs like Claude to help with transcripts? Our answer is a resounding yes.