Large document review projects can maximize efficiency by employing a two-filter method to cull documents from costly manual review. This method helps reduce costs and maximize recall. I introduced this method, and the diagram shown here illustrating it, at the conclusion of my blog series, Introducing “ei-Recall” – A New Gold Standard for Recall Calculations in Legal Search – Part Three. I use the two-filter method in most large projects as part of my overall multimodal, bottom line driven, AI-Enhanced (i.w. – predictive coding) method of review. I have described this multimodal method many times here, and you will find summaries of it elsewhere, including my CAR page, and Legal Search Science, and the work in progress, the EDBP outlining best practices for lawyers doing e-discovery.
My two-filter method of course employs deduplication and deNisting in the First Filter. (I always do full horizontal deduplication across all custodians.) Deduplication and deNisting are, however, mere technical, non-legal filters. They are already well established industry standards and so I see no need to discuss them further in this article.
Some think those two technical methods are the end-all of ESI culling, but, as this two-part blog will explain, they are just the beginning. The other methods require legal judgment, and so you cannot just hire a vendor to do it, as you can with deduplication and deNisting. This is why I am taking pains to explain two-filter document culling, so that it can be used by other legal teams to reduce wasted review expenses.
This blog is the first time I have gone into the two-filter culling component in any depth. This method has been proven effective in attaining high recall at low cost in at least one open scientific experiment, although I cannot go into that. You will just have to trust me on that. Insiders know anyway. For the rest, just look around and see I have no products to sell here, and accept no ads. This is all part of an old lawyer’s payback to a profession that has been very good to him over the years.
My thirty-five years of experience in law have shown me that most reliable way for the magic of justice to happen is by finding the key documents. You find the truth, the whole truth, and nothing but the truth when you find the key documents and use them to keep the witnesses honest. Deciding cases on the basis of the facts is the way our system of justice tries to decide all cases on the merits, in an impartial and fair manner. In today’s information flooded world, that can only happen if we use technology to find relevant evidence quickly and inexpensively. The days of finding the truth by simple witness interviews are long gone. Thus I share my search and review methods as a kind of payback and pay it forward. For now, as I have for the past eight years, I will try to make the explanations accessible to beginners and eLeet alike.
We need cases to be decided on the merits, on the facts. Hopefully my writing and rants will help make that happen in some small way. Hopefully it will help stem the tide of over-settlement, where many cases are decided on the basis of settlement value, not merits. Too many frivolous cases are filed that drown out the few with great merit. Judges are overwhelmed and often do not have the time needed to get down to the truth and render judgments that advance the cause of justice.
Most of the time the judges, and the juries they assemble, are never even given the chance to do their job. The cases all settle out instead. As a result only one percent of federal civil cases actually go to trial. This is a big loss for society, and for the so-called “trial lawyers” in our profession, a group I once prided myself to be a part. Now I just focus on getting the facts from computers, to help keep the witnesses honest, and cases decided on the true facts, the evidence. That is where all the real action is nowadays anyway.
By the way, I expect to get another chance to prove the value of the methods I share here in the 2015 TREC experiment on recall. We will see, again, how it stacks up to other approaches. This time I may even have one or two people assist me, instead of doing it alone as I did before. The Army of One approach, which I have also described here many times, although effective, is very hard and time-consuming. My preference now is a small team approach, kind of like a nerdy swat team, or Seal Team Six approach, but without guns and killing people and stuff. I swear! Really.
Some Software is Far Better than Others
One word of warning, although this method is software agnostic, in order to emulate the two-filter method, your document review software must have certain basic capabilities. That includes effective, and easy, bulk coding features for the first filter. This is the multimodal broad-based culling. Some of the multiple methods do not require software features, just attorney judgment, such as excluding custodians, but other do require software features, like domain searches or similarity searches. If your software does not have the features that will be discussed here for the first filter, then you probably should switch right away, but, for most, that will not be a problem. The multimodal culling methods used in the first filter are, for the most part, pretty basic.
Some of the software features needed to implement the second filter, are, however, more advanced. The second filter works best when using predictive coding and probability ranking. You review the various strata of the ranked documents. The Second Filter can still be used with other, less advanced multimodal methods, i.e. keywords. Moreover, even when you use bona fide active machine learning software features, you continue to use a smattering of other multimodal search methods in the Second Filter. But now you do so not to cull, but to help find relevant and highly relevant documents to improve training. I do not rely on probability searches alone, although sometimes in the Second Filter I rely almost entirely on predictive coding based searches to continue the training.
If you are using software without AI-enhanced active learning features, then you are forced to only use other multimodal methods in the second filter, such as keywords. Warning, true active learning features are not present in most review software, or are very weak. That is true even with software that claims to have predictive coding features, but really just has dressed-up passive learning, i.e. concept searches with latent semantic indexing. You handicap yourself, and your client, by continuing to use such less expensive programs. Good software, like everything else, does not come cheap, but should pay for itself many times over if used correctly. The same comment goes for lawyers too.
First Filter – Keyword Collection Culling
Some first stage filtering takes place as part of the ESI collection process. The documents are preserved, but not collected nor ingested into the review database. The most popular collection filter as of 2015 is still keyword, even though this is very risky in some cases and inappropriate in many. Typically such keyword filtering is driven by vendor costs to avoid processing and hosting charges.
Some types of collection filtering are appropriate and necessary, for instance, in the case of custodian filters, where you broadly preserve the ESI of many custodians, just in case, but only collect and review a few of them. It is, however, often inappropriate to use keywords to filter out the collection of ESI from admittedly key custodians. This is a situation where an attorney determines that a custodian’s data needs to be reviewed for relevant evidence, but does not want to incur the expense to have all of their ESI ingested into the review database. For that reason they decide to only review data that contains certain keywords.
I am not a fan of keyword filtered collections. The obvious danger of keyword filtering is that important documents may not have the keywords. Since they will not even be placed in the review platform, you will never know that the relevant ESI was missed. You have no chance of finding them.
See eg, William Webber’s analysis of the Biomet case where this kind of keyword filtering was use before predictive coding began. What is the maximum recall in re Biomet?, Evaluating e-Discovery (4/24/13). Webber shows that in Biomet this method First Filtered out over 40% of the relevant documents. This doomed the Second Filter predictive coding review to a maximum possible recall of 60%, even if was perfect, meaning it would otherwise have attained 100% recall, which never happens. The Biomet case very clearly shows the dangers of over-reliance on keyword filtering.
Nevertheless, sometimes keyword collection may work, and may be appropriate. In some simple disputes, and with some data collections, obvious keywords may work just fine to unlock the truth. For instance, sometimes the use of names is an effective method to identify all, or almost all, documents that may be relevant. This is especially true in smaller and simpler cases. This method can, for instance, often work in employment cases, especially where unusual names are involved. It becomes an even more effective method when the keywords have been tested. I just love it, for instance, when the plaintiff’s name is something like the famous Mister Mxyzptlk.
In some cases keyword collections may be as risky as in the complex Biomet case, but may still be necessary because of the proportionality constraints of the case. The law does not require unreasonably excessive search and review, and what is reasonable in a particular case depends on the facts of the case, including its value. See my many writings on proportionality, including my law review article Predictive Coding and Proportionality: A Marriage Made In Heaven, 26 Regent U. Law Review 1 (2013-2014). Sometimes you have to try for rough justice with the facts that you can afford to find given the budgetary constraints of the case.
The danger of missing evidence is magnified when the keywords are selected on the basis of educated guesses or just limited research. This technique, if you can call it that, is, sadly, still the dominant method used by lawyers today to come up with keywords. I have long thought it is equivalent to a child’s game of Go Fish. If keywords are dreamed up like that, as mere educated guesses, then keyword filtering is a high risk method of culling out irrelevant data. There is a significant danger that it will exclude many important documents that do not happen to contain the selected keywords. No matter how good your predictive coding may be after that, you will never find these key documents.
If the keywords are not based on a mere guessing, but are instead tested, then it becomes a real technique that is less risky for culling. But how do you test possible keywords without first collecting and ingesting all of the documents to determine which are effective? It is the old cart before the horse problem.
One partial answer is that you could ask the witnesses, and do some partial reviews before collection. Testing and witness interviews is required by Judge Andrew Peck’s famous wake up call case. William A. Gross Constr. Assocs., Inc. v. Am. Mfrs. Mut. Ins. Co., 256 F.R.D. 134, 134, 136 (S.D.N.Y. 2009). I recommend that opinion often, as many attorneys still need to wake up about how to do e-discovery. They need to add ESI use, storage, and keyword questions to their usual new case witness interviews.
Interviews do help, but there is nothing better than actual hands on reading and testing of the documents. This is what I like to call getting your hands dirty in the digital mud of the actual ESI collected. Only then will you know for sure the best way to mass-filter out documents. For that reason my strong preference in all significant size cases is to collect in bulk, and not filter out by keywords. Once you have documents in the database, then you can then effectively screen them out by using parametric Boolean keyword techniques. See your particular vendor for various ways on how to do that.
By the way, parametric is just a reference to the various parameters of a computer file that all good software allows you to search. You could search the text and all metadata fields, the entire document. Or you could limit your search to various metadata fields, such as date, prepared by, or the to and from in an email. Everyone knows what Boolean means, but you may not know all of the many variations that your particular software offers to create highly customized searches. While predictive coding is beyond the grasp of most vendors and case managers, the intricacies of keyword search are not. They can be a good source of information on keyword methods.
First Filter – Date Range and Custodian Culling
Even when you collect in bulk, and do not keyword filter before you put custodian ESI in the review database, in most cases you should filter for date range and custodian. It is often possible for an attorney to know, for instance, that no emails before or after a certain date could possibly be relevant. That is often not a highly speculative guessing game. It is reasonable to filter on this time-line basis before the ESI goes in the database. Whenever possible, try to get agreement on date range screening from the requesting party. You may have to widen it a little, but it is worth the effort to establish a line of communication and begin a cooperative dialogue.
The second thing to talk about is which custodians you are going to include in the database. You may put 50 custodians on hold, and actually collect the ESI of 25, but that does not mean you have to load all 25 into the database for review. Here your interviews and knowledge of the case should allow you to know who the key, key custodians are. You rank them by your evaluation of the likely importance of the data they hold to the facts disputed in the case. Maybe, for instance, in your evaluation you only need to review the mailboxes of 10 of the 25 collected.
Again, disclose and try to work that out. The requesting party can reserve rights to ask for more, that is fine. They rarely do after production has been made, especially if you were careful and picked the right 10 to start with, and if you were careful during review to drop and add custodians based on what you see. If you are using predictive coding in the second filter stage, the addition or deletion of data mid-course is still possible with most software. It should be robust enough to handle such mid-course corrections. It may just slow down the ranking for a few iterations, that’s all.
First Filter – Other MultiModal Culling
There are many other bulk coding techniques that can be used in the first filter stage. This is not intended to be an exhaustive search. Like all complex tasks in the law, simple black letter rules are for amateurs. The law, which mirrors the real world, does not work like that. The same holds true for legal search. There may be many Gilbert’s for search books and articles, but they are just 1L types guides. For true legal search professionals they are mere starting points. Use my culling advice here in the same manner. Use your own judgment to mix and match the right kind of culling tools for the particular case and data encountered. Every project is slightly different, even in the world of repeat litigation, like employment law disputes where I currently spend much of my time.
Legal search is at core a heuristic activity, but one that should be informed by science and technology. The knowledge triangle is a key concept for today’s effective e-Discovery Team. Although e-Discovery Teams should be led by attorneys skilled in evidence discovery, they should include scientists and engineers in some way. Effective team leaders should be able to understand and communicate with technology experts and information scientists. That does not mean all e-discovery lawyers need to become engineers and scientists too. That effort would likely diminish your legal skills based on the time demands involved. It just means you should know enough to work with these experts. That includes the ability to see through the vendor sales propaganda, and to incorporate the knowledge of the bona fide experts into your legal work.
One culling method that many overlook is file size. Some collections have thousands of very small files, just a few bits, that are nothing but backgrounds, tiny images, or just plain empty space. They are too small to have any relevant information. Still, you need to be cautious and look out for very small emails, for instance, ones that just says “yes.” Depending on context it could be relevant and important. But for most other types of very small files, there is little risk. You can go ahead a bulk code them irrelevant and filter them out.
Even more subtle is filtering out files based on their being very large. Sort your files by size, and then look at both ends, small and big. They may reveal certain files and file types that could not possibly be relevant. There is one more characteristic of big files that you should consider. Many of them have millions of lines of text. Big files are confusing to machine learning when, as typical, only a few lines of the text are relevant, and the rest are just noise. That is another reason to filter them out, perhaps not entirely, but for special treatment and review outside of predictive coding. In other projects where you have many large files like that, and you need the help of AI ranking, you may want to hold them in reserve. You may only want to throw them into the ranking mix after your AI algorithms have acquired a pretty good idea of what you are looking for. A maturely trained system is better able to handle big noisy files.
File type is a well-known and often highly effective method to exclude large numbers of files of a same type after only looking at a few of them. For instance, there may be database files automatically generated, all of the same type. You look at a few to verify these databases could not possibly be relevant to your case, and then you bulk code them all irrelevant. There are many types of files like that in some data sets. The First Filter is all about being a smart gatekeeper.
File type is also used to eliminate, or at least divert, non-text files, such as audio files or most graphics. Since most Second Filter culling is going to be based on text analytics of some kind, there is no point for anything other than files with text to go into that filter. In some cases, and some datasets, this may mean bulk coding them all irrelevant. This might happen, for instance, where you know that no music or other audio files, including voice messages, could possibly be relevant. We also see this commonly where we know that photographs and other images could not possibly be relevant. Exclude them from the review database.
You must, however, be careful with all such gatekeeper activities, and never do bulk coding without some judgmental sampling first. Large unknown data collections can always contain a few unexpected surprises, no matter how many document reviews you have done before. Be cautious. Look before you leap. Skim a few of the ESI file types you are about to bulk code as irrelevant.
This directive applies to all First Filter activities. Never do it blind on just logic or principle alone. Get you hands in the digital mud. Do not over-delegate all of the dirty work to others. Do not rely too much on your contract review lawyers and vendors, especially when it comes to search. Look at the documents yourself and do not just rely on high level summaries. Every real trial lawyer knows the importance of that. The devil is always in the details. This is especially true when you are doing judgmental search. The client wants your judgment, not that of a less qualified associate, paralegal, or minimum wage contract review lawyer. Good lawyers remain hands-on, to some extent. They know the details, but are also comfortable with appropriate delegation to trained team members.
There is a constant danger of too much delegation in big data review. The lawyer signing the Rule 26(g) statement has a legal and ethical duty to closely supervise document review done in response to a request for production. That means you cannot just hire a vendor to do that, although you can hire outside counsel with special expertise in the field.
Some non-text file types will need to be diverted for different treatment than the rest of your text-based dataset. For instance, some of the best review software allows you to keyword search audio files. It is based on phonetics and wave forms. At least one company I know has had that feature since 2007. In some cases you will have to carefully review the image files, or at least certain kinds of them. Sorting based on file size and custodian can often speed up that exercise.
Remember the goal is always efficiency, and caution, but not over cautious. The more experienced you get the better you become at evaluating risks and knowing where you can safely take chances to bulk code, and where you cannot. Another thing to remember is that many image files have text in them too, such as in the metadata, or in ASCII transmissions. They are usually not important and do not provide good training for second stage predictive coding.
Text can also be hidden in dead Tiff files, if they have not been OCR’ed. Scanned documents Tiffs, for instance, may very well be relevant and deserve special treatment, including full manual review, but they may not show in your review tool as text, because they have never been OCR text recognized.
Concept searches have only rarely been of great value to me, but should still be tried out. Some software has better capacities with concepts and latent semantic indexing than others. You may find it to be a helpful way to find groupings of obviously irrelevant, or relevant documents. If nothing else, you can always learn something about your dataset from these kind of searches.
Similarity searches of all kinds are among my favorite. If you find some files groups that cannot be relevant, find more like that. They are probably bulk irrelevant (or relevant) too. A similarity search, such as find every document that is 80% or more the same as this one, is often a good way to enlarge your carve outs and thus safely improve your efficiency.
Another favorite of mine is domain culling of email. It is kind of like a spam filter. That is a great way to catch the junk mail, newsletters, and other purveyors of general mail that cannot possibly be relevant to your case. I have never seen a mail collection that did not have dozens of domains that could be eliminated. You can sometimes cull-out as much as 10% of your collection that way, sometimes more when you start diving down into senders with otherwise safe domains. A good example of this is the IT department with their constant mass mailings, reminders and warnings. Many departments are guilty of this, and after examining a few, it is usually safe to bulk code them all irrelevant.
To be continued. In the next part I will discuss the second filter and go deeper into AI-enhanced predictive culling. I will also discuss how tested keywords can also be used, if you do not yet have the skill-set or good software needed for predictive coding.