support vector machines | e-Discovery Team

New Video to Introduce the TAR Course

April 23, 2017

We are continuing to upgrade the e-Discovery Team’s free TAR Course. The latest improvements include the addition of “homework assignments” to the first ten classes. These are challenging and add to the depth of the instruction. The homework includes both supplemental reading suggestions and exercises. We will add homework assignments to the last six classes soon. We also made a few minor revisions and additions to the written materials, but nothing substantial. Periodically we will add some more video content to the TAR Course. We started this weekend by adding a video to the first class:

Here is the list of all sixteen classes in the TAR Course.

First Class: Introduction
Second Class: TREC Total Recall Track
Third Class: Introduction to the Nine Insights Concerning the Use of Predictive Coding in Legal Document Review
Fourth Class: 1st of the Nine Insights – Active Machine Learning
Fifth Class: Balanced Hybrid and Intelligently Spaced Training
Sixth Class: Concept and Similarity Searches
Seventh Class: Keyword and Linear Review
Eighth Class: GIGO, QC, SME, Method, Software
Ninth Class: Introduction to the Eight-Step Work Flow
Tenth Class: Step One – ESI Communications
Eleventh Class: Step Two – Multimodal ECA
Twelfth Class: Step Three – Random Prevalence
Thirteenth Class: Steps Four, Five and Six – Iterate
Fourteenth Class: Step Seven – ZEN Quality Assurance Tests
Fifteenth Class: Step Eight – Phased Production
Sixteenth Class: Conclusion

Certification is not offered. Maybe someday. We created a test based on our TREC experiments that we may someday roll out.

Leave a Comment » | Lawyers Duties, Related Legal Webs, Review, Search, Technology | Tagged: best practices, document review, legal profession, legal search, machine learning, predictive coding, search, support vector machines, tar, technology | Permalink
Posted by Ralph Losey

Predictive Coding 4.0 – Nine Key Points of Legal Document Review and an Updated Statement of Our Workflow – Part Two

September 18, 2016

In Part One we announced the latest enhancements to our document review method, the upgrade to Predictive Coding 4.0. We explained the background that led to this upgrade – the TREC research and hundreds of projects we have done since our last upgrade a year ago. Millions have been spent to develop the software and methods we now use for Technology Assisted Review (TAR). As a result our TAR methods are more effective and simpler than ever.

The nine insights we will share are based on our experience and research. Some of our insights may be complicated, especially our lead insight on Active Machine Learning covered in this Part Two with our new description of IST – Intelligently Spaced Training. We consider IST the smart, human empowering alternative to CAL. If I am able to write these insights up here correctly, the obviousness of them should come through. They are all simple in essence. The insights and methods of Predictive Coding 4.0 document review are partially summarized in the chart below (which you are free to reproduce without edit).

1st of the Nine Insights: Active Machine Learning

Our method is Multimodal in that it uses all kinds of document search tools. Although we emphasize active machine learning, we do not rely on that method alone. Our method is also Hybrid in that we use both machine judgments and human (lawyer) judgments. Moreover, in our method the lawyer is always in charge. We may take our hand off the wheel and let the machine drive for a while, but under our versions of Predictive Coding, we watch carefully. We remain ready to take over at a moment’s notice. We do not rely on one brain to the exclusion of another. See eg. Why the ‘Google Car’ Has No Place in Legal Search (caution against over reliance on fully automated methods of active machine learning). Of course the converse is also true, we never just rely on our human brain alone. It has too many limitations. We enhance our brain with predictive coding algorithms. We add to our own natural intelligence with artificial intelligence. The perfect balance between the two, the Balanced Hybrid, is another of insights that we will discuss later.

Active Machine Learning is Predictive Coding – Passive Analytic Methods Are Not

Even though our methods are multimodal and hybrid, the primary search method we rely on is Active Machine Learning. The overall name of our method is, after all, Predictive Coding. And, as any information retrieval expert will tell you, predictive coding means active machine learning. That is the only true AI method. The passive type of machine learning that some vendors use under the name Analytics is NOT the same thing as Predictive Coding. These passive Analytics have been around for years and are far less powerful than active machine learning.

These search methods, that used to be called Concept Search, were a big improvement upon relying on keyword search alone. I remember talking about concepts search techniques in reverent terms when I did my first Legal Search webinar in 2006 with Jason Baron and Professor Doug Oard. That same year, Kroll Ontrack bought one of the original developers and patent holders of concept search, Engenium. For a short time in 2006 and 2007 Kroll Ontrack was the only vendor to have these concept search tools. The founder of Engenium, David Chaplin came with the purchase, and became Kroll Ontrack’s VP of Advanced Search Technologies for three years. (Here is an interesting interview of Chaplin that discusses what he and Kroll Ontrack were doing with advanced search analytic-type tools when he left in 2009.)

But search was hot and soon boutique search firms like, Clearwell, Cataphora, Content Analyst (the company recently purchased by popular newcomer, kCura), and other e-discovery vendors developed their own concept search tools. Again, they were all using passive machine learning. It was a big deal ten years ago. For a good description of these admittedly powerful, albeit now dated search tools, see the concise, well-written article by D4’s Tom Groom, The Three Groups of Discovery Analytics and When to Apply Them.

Search experts and information scientists know that active machine learning, also called supervised machine learning, was the next big step in search after concept searches, which are, in programming language, also known as passive or unsupervised machine learning. I am getting out of my area of expertise here, and so am unable go into any details, other than present the below instructional chart by Hackbright Academy that sets forth key difference between supervised learning (predictive coding) and unsupervised (analytics, aka concept search).

What I do know is that the bonafide active machine learning software in the market today all use either a form of Logistic Regression, including Kroll Ontrack, or SVM, which means Support Vector Machine.

e-Discovery Vendors Have Been Market Leaders in Active Machine Learning Software

After Kroll Ontrack absorbed the Engenium purchase, and its founder Chaplin completed his contract with Kroll Ontrack and moved on, Kroll Ontrack focused their efforts on the next big step, active machine learning, aka predictive coding. They have always been that kind of cutting edge company, especially when it comes to search, which is one reason they are one of my personal favorites. A few of the other, then leading e-discovery vendors did too, including especially Recommind and the Israeli based search company, Equivo. Do not get me wrong, the concept search methods, now being sold under the name of TAR Analytics, are powerful search tools. They are a part of our multimodal tool-kit and should be part of yours. But they are not predictive coding. They do not rank documents according to your external input, your supervision. They do not rely on human feedback. They group documents according to passive analytics of the data. It is automatic, unsupervised. These passive analytic algorithms can be good tools for efficient document review, but they not active machine learning and are nowhere near as powerful.

Search Software Ghosts

Many of the software companies that made the multi-million dollar investments necessary to go to the next step and build document review platforms with active machine learning algorithms have since been bought out by big-tech and repurposed out of the e-discovery market. They are the ghosts of legal search past. Clearwell was purchased by Symantec and has since disappeared. Autonomy was purchased by Hewlett Packard and has since disappeared. Equivio was purchased by Microsoft and has since disappeared. See e-Discovery Industry Reaction to Microsoft’s Offer to Purchase Equivio for $200 Million – Part One and Part Two. Recommind was recently purchased by OpenText and, although it is too early to tell for sure, may also soon disappear from e-Discovery.

Slightly outside of this pattern, but with the same ghosting result, e-discovery search company, Cataphora, was bought by Ernst & Young, and has since disappeared. The year after the acquisition, Ernst & Young added predictive coding features from Cataphora to its internal discovery services. At this point, all of the Big Four Accounting Firms, claim to have their own proprietary software with predictive coding. Along the same lines, at about the time of the Cataphora buy-out, consulting giant FTI purchased another e-discovery document review company, Ringtail Solutions (known for its petri dish like visualizations). Although not exactly ghosted by FTI from the e-discovery world after the purchase, they have been absorbed by the giant FTI.

Outside of consulting/accountancy, in the general service e-discovery industry for lawyers, there are, at this point (late 2016) just a few document review platforms left that have real active machine learning. Some of the most popular ones left behind certainly do not. They only have passive learning analytics. Again, those are good features, but they are not active machine learning, one of the nine basic insights of Predictive Coding 4.0 and a key component of the e-Discovery Team’s document review capabilities.

The power of the advanced, active learning technologies that have been developed for e-discovery is the reason for all of these acquisitions by big-tech and the big-4 or 5. It is not just about wild overspending, although that may well have been the case for Hewlett Packard payment of $10.3 Billion to buy Autonomy. The ability to do AI-enhanced document search and review is a very valuable skill, one that will only increase in value as our data volumes continue to explode. The tools used for such document review are also quite valuable, both inside the legal profession and, as the ghostings prove, well beyond into big business. See e-Discovery Industry Reaction to Microsoft’s Offer to Purchase Equivio for $200 Million, Part Two.

The indisputable fact that so many big-tech companies have bought up the e-discovery companies with active machine learning software should tell you a lot. It is a testimony to the advanced technologies that the e-discovery industry has spawned. When it comes to advanced search and document retrieval, we in the e-discovery world are the best in the world my friends, primarily because we have (or can easily get) the best tools. Smile.

Search is king of our modern Information Age culture. See Information → Knowledge → Wisdom: Progression of Society in the Age of Computers. The search for evidence to peacefully resolve disputes is, in my most biased opinion, the most important search of all. It sure beats selling sugar water. Without truth and justice all of the petty business quests for fame and fortune would crumble into anarchy, or worse, dictatorship.

With this background it is easy to understand why some of the e-discovery vendors left standing are not being completely candid about the capabilities of their document review software. (It is called puffing and is not illegal.) The industry is unregulated and, alas, most of our expert commentators are paid by vendors. They are not independent. As a result, many of the lawyers who have tried what they thought was predictive coding, and had disappointing results, have never really tried predictive coding at all. They have just used slightly updated concept search.

Alternatively, some of the disappointed lawyers may have used one of the many now-ghosted vendor tools. They were all early version 1.0 type tools. For example, Clearwell’s active machine learning was only on the market for a few months with this feature before they were bought and ghosted by Symantec. (I think Jason Baron and I were the first people to see an almost completed demo of their product at a breakfast meeting a few months before it was released.) Recommind’s predictive coding software was well-developed at the time of their sell-out, but not its methods of use. Most of its customers can testify as to how difficult it is to operate. That is one reason that OpenText was able to buy them so cheaply, which, we now see, was part of their larger acquisition plan culminating in the purchase of Dell’s EMC document management software.

All software still using early methods, what we call version 1.0 and 2.0 methods based on control sets, are cumbersome and hard to operate, not just Recommind’s system. I explained this in my article last year, Predictive Coding 3.0. I also mentioned in this article that some vendors with predictive coding would only let you use predictive coding for search. It was, in effect, mono-modal. That is also a mistake. All types of search must be used – multimodal – for the predictive coding type of search to work efficiently and effectively. More on that point later.

Maura Grossman Also Blows the Whistle on Ineffective “TAR tools”

Maura Grossman, who is now an independent expert in this field, made many of these same points in a recent interview with Artificial Lawyer, a periodical dedicated to AI and the Law. AI and the Future of E-Discovery: AL Interview with Maura Grossman (Sept. 16, 2016). When asked about the viability of the “over 200 businesses offering e-discovery services” Maura said, among other things:

In the long run, I am not sure that the market can support so many e-discovery providers …

… many vendors and service providers were quick to label their existing software solutions as “TAR,” without providing any evidence that they were effective or efficient. Many overpromised, overcharged, and underdelivered. Sadly, the net result was a hype cycle with its peak of inflated expectations and its trough of disillusionment. E-discovery is still far too inefficient and costly, either because ineffective so-called “TAR tools” are being used, or because, having observed the ineffectiveness of these tools, consumers have reverted back to the stone-age methods of keyword culling and manual review.

Now that Maura is no longer with the conservative law firm of Wachtell Lipton, she has more freedom to speak her mind about caveman lawyers. It is refreshing and, as you can see, echoes much of what I have been saying. But wait, there is still more that you need to hear from the interview of new Professor Grossman:

It is difficult to know how often TAR is used given confusion over what “TAR” is (and is not), and inconsistencies in the results of published surveys. As I noted earlier, “Predictive Coding”—a term which actually pre-dates TAR—and TAR itself have been oversold. Many of the commercial offerings are nowhere near state of the art; with the unfortunate consequence that consumers have generalised their poor experiences (e.g., excessive complexity, poor effectiveness and efficiency, high cost) to all forms of TAR. In my opinion, these disappointing experiences, among other things, have impeded the adoption of this technology for e-discovery. …

Not all products with a “TAR” label are equally effective or efficient. There is no Consumer Reports or Underwriters Laboratories (“UL”) that evaluates TAR systems. Users should not assume that a so-called “market leading” vendor’s tool will necessarily be satisfactory, and if they try one TAR tool and find it to be unsatisfactory, they should keep evaluating tools until they find one that works well. To evaluate a tool, users can try it on a dataset that they have previously reviewed, or on a public dataset that has previously been labelled; for example, one of the datasets prepared for the TREC 2015 or 2016 Total Recall tracks. …

She was then asked by the Artificial Lawyer interviewer (name never identified), which is apparently based in the UK, another popular question:

As is often the case, many lawyers are fearful about any new technology that they don’t understand. There has already been some debate in the UK about the ‘black box’ effect, i.e., barristers not knowing how their predictive coding process actually worked. But does it really matter if a lawyer can’t understand how algorithms work?

The following is an excerpt of Maura’s answer. Suggest you consult the full article for a complete picture. AI and the Future of E-Discovery: AL Interview with Maura Grossman (Sept. 16, 2016). I am not sure whether she put on her Google Glasses to answer (probably not), but anyway, I rather like it.

Many TAR offerings have a long way to go in achieving predictability, reliability, and comprehensibility. But, the truth that many attorneys fail to acknowledge is that so do most non-TAR offerings, including the brains of the little black boxes we call contract attorneys or junior associates. It is really hard to predict how any reviewer will code a document, or whether a keyword search will do an effective job of finding substantially all relevant documents. But we are familiar with these older approaches (and we think we understand their mechanisms), so we tend to be lulled into overlooking their limitations.

The brains of the little black boxes we call contract attorneys or junior associates. So true. We will go into that more throughly in our discussion of the GIGO & QC insight.

Recent Team Insights Into Active Machine Learning

To summarize what I have said so far, in the field of legal search, only active machine learning:

effectively enhances human intelligence with artificial intelligence;
qualifies for the term Predictive Coding.

I want to close on this discussion of active machine learning with one more insight. This one is slightly technical, and again, if I explain it correctly, should seem perfectly obvious. It is certainly not new, and most search experts will already know this to some degree. Still, even for them, there may some nuances to this insight that they have not thought of. It can be summarized as follows: active machine learning should have a double feedback loop with active monitoring by the attorney trainers.

Active machine learning should create feedback for both the algorithm (the data classified) AND the human managing the training. Both should learn, not just the robot. They should, so to speak, be friends. They should get to know each other

Many predictive coding methods that I have read about, or heard described, including how I first used active machine learning, did not sufficiently include the human trainer in the feedback loop. They were static types of training using single a feedback loop. These methods are, so to speak, very stand-offish, aloof. Under these methods the attorney trainer does not even try to understand what is going on with the robot. The information flow was one-way, from attorney to machine.

As I grew more experienced with the EDR software I started to realize that it is possible to start to understand, at least a little, what the black box is doing. Logistic based AI is a foreign intelligence, but it is intelligence. After a while you start to understand it. So although I started just using one-sided machine training, I slowly gained the ability to read how EDR was learning. I then added another dimension, another feedback loop that was very interesting one indeed. Now I not only trained and provided feedback to the AI as to whether the predictions of relevance were correct, or not, but I also received training from the AI as to how well, or not, it was learning. That in turn led to the humorous personification of the Kroll Ontrack software that we now call Mr. EDR. See MrEDR.com. When we reached this level, machine training became a fully active, two-way process.

We now understand that to fully supervise a predictive coding process you to have a good understanding of what is happening. How else can you supervise it? You do not have to know exactly how the engine works, but you at least need to know how fast it is going. You need a speedometer. You also need to pay attention to how the engine is operating, whether it is over-heating, needs oil or gas, etc. The same holds true to teaching humans. Their brains are indeed mysterious black boxes. You do not need to know exactly how each student’s brain works in order to teach them. You find out if your teaching is getting through by questions.

For us supervised learning means that the human attorney has an active role in the process. A role where the attorney trainer learns by observing the trainee, the AI in creation. I want to know as much as possible, so long as it does not slow me down significantly.

In other methods of using predictive coding that we have used or seen described the only role of the human trainer is to say yes or no as to the relevance of a document. The decision as to what documents to select for training has already been predetermined. Typically it is the highest ranked documents, but sometimes also some mid-ranked “uncertain documents” or some “random documents” are added in the mix. The attorney
has no say in what documents to look at. They are all fed to him or her according to predetermined rules. These decision making rules are set in advance and do not change. These active machine learning methods work, but they are slow, and less precise, not to mention boring as hell.

The recall of these single-loop passive supervision methods may also not be as good. The jury is still out on that question. We are trying to run experiments on that now, although it can be hard to stop yawning. See an earlier experiment on this topic testing the single loop teaching method of random selection: Borg Challenge: Report of my experimental review of 699,082 Enron documents using a semi-automated monomodal methodology.

These mere yes or no, limited participation methods are hybrid Man-Machine methods, but, in our opinion, they are imbalanced towards the Machine. (Again, more on the question of Hybrid Balance will be covered in the next installment of this article.) This single versus dual feedback approach seems to be the basic idea behind the Double Loop Learning approach to human education depicted in the diagram below. Also see Graham Attwell, Double Loop Learning and Learning Analytics (Pontydysgu, May 4, 2016).

To quote Wikipedia:

The double loop learning system entails the modification of goals or decision-making rules in the light of experience. The first loop uses the goals or decision-making rules, the second loop enables their modification, hence “double-loop.” …

Double-loop learning is contrasted with “single-loop learning”: the repeated attempt at the same problem, with no variation of method and without ever questioning the goal. …

Double-loop learning is used when it is necessary to change the mental model on which a decision depends. Unlike single loops, this model includes a shift in understanding, from simple and static to broader and more dynamic, such as taking into account the changes in the surroundings and the need for expression changes in mental models.

The method of active machine learning that we use in Predictive Coding 4.0 is a type of double loop learning system. As such it is ideal for legal search, which is inherently ad hoc, where even the understanding of relevance evolves as the project develops. As Maura noted near the end of the Artificial Lawyer interview:

… e-discovery tends to be more ad hoc, in that the criteria applied are typically very different for every review effort, so each review generally begins from a nearly zero knowledge base.

The driving impetus behind our double feedback look system is to allow for training document selection to vary according to the circumstances encountered. Attorneys select documents for training and then observe how these documents impact the AI’s overall ranking of the documents. Based on this information decisions are then made by the attorney as to which documents to next submit for training. A single fixed mental model is not used, such as only submitting the ten highest ranked documents for training.

The human stays involved and engaged and selects the next documents to add to the training based on what she sees. This makes the whole process much more interesting. For example, if I find a group of relevant spreadsheets by some other means, such as a keyword search, then, when I add these document to the training, I observe how these documents impact the overall ranking of the dataset. For instance, did this training result in an increase of relevance ranking of other spreadsheets? Was the increase nominal or major? How did it impact the ranking of other documents? For instance, were emails with a lot of numbers in them suddenly much higher ranked? Overall, was this training effective? Were the documents in fact relevant as predicted that moved up in rank to the top, or near top of probable relevance? What was the precision rate like for these documents? Does the AI now have a good understanding of relevance of spreadsheets, or need more training on that type of document? Should we focus our search on other kinds of documents?

You see all kinds of variations on that. If the spreadsheet understanding (ranking) is good, how does it compare to its understanding (correct ranking) of Word Docs or emails? Where should I next focus my multimodal searches? What documents should I next assign to my reviewers to read and make a relevancy determination? These kind of considerations keep the search interesting, fun even. Work as play is the best kind. Typically we simply assign the documents for attorney review that have the highest ranking (which is the essence of what Grossman and Cormack call CAL), but not always. We are flexible. We, the human attorneys, are the second positive feedback loop.

We like to remain in charge of teaching the classifier, the AI. We do not just turn it over to the classifier to teach itself. Although sometimes, when we are out of ideas and are not sure what to do next, we will do exactly that. We will turn over to the computer the decision of what documents to review next. We just go with his top predictions and use those documents to train. Mr. EDR has come through for us many times when we have done that. But this is more of an exception, than the rule. After all, the classifier is a tabula rasa. As Maura put it: each review generally begins from a nearly zero knowledge base. Before the training starts, it knows nothing about document relevance. The computer does not come with built-in knowledge of the law or relevance. You know what you are looking for. You know what is relevant, even if you do not know how to find it, or even whether it exists at all. The computer does not know what you are looking for, aside from what you have told it by your yes-no judgments on particular documents. But, after you teach it, it knows how to find more documents that probably have the same meaning.

By observation you can see for yourself, first hand, how your training is working, or not working. It is like a teacher talking to their students to find out what they learned from the last assigned reading materials. You may be surprised by how much, or how little they learned. If the last approach did not work, you change the approach. That is double-loop learning. In that sense our active monitoring approach it is like continuous dialogue. You learn how and if the AI is learning. This in turn helps you to plan your next lessons. What has the student learned? Where does the AI need more help to understand the conception of relevance that you are trying to teach it.

This monitoring of the AI’s learning is one of the most interesting aspects of active machine learning. It is also a great opportunity for human creativity and value. The inevitable advance of AI in the law can mean more jobs for lawyers overall, but only for those able step up and change their methods. The lawyers able to play the second loop game of active machine learning will have plenty of employment opportunities. See eg. Thomas H. Davenport, Julia Kirby, Only Humans Need Apply: Winners and Losers in the Age of Smart Machines (Harper 2016).

Going down into the weeds a little bit more, our active monitoring dual feedback approach means that when we use Kroll Ontrack’s EDR software, we adjust the settings so that new learning sessions are not created automatically. They only run when and if we click on the Initiate Session button shown in the EDR screenshot below (arrow and words were added). We do not want the training to go on continuously in the background (typically meaning at periodic intervals of every thirty minutes or so.) We only want the learning sessions to occur when we say so. In that way we can know exactly what documents EDR is training on during a session. Then, when that training session is complete, we can see how the input of those documents has impacted the overall data ranking. For instance, are there now more documents in the 90% or higher probable relevance category and if so, how many? The picture below is of a completed TREC project. The probability rankings are on the far left with the number of documents shown in the adjacent column. Most of the documents in the 290,099 collection of Bush email were in the 0-5% probable relevant ranking not included in the screen shot.

This means that the e-Discovery Team’s active learning is not continuous, in the sense of always training. It is instead intelligently spaced. That is an essential aspect of our Balanced Hybrid approach to electronic document review. The machine training only begins when we click on the “Initiate Session” button in EDR that the arrow points to. It is only continuous in the sense that the training continues until all human review is completed. The spaced training, in the sense of staggered in time, is itself an ongoing process until the production is completed. We call this Intelligently Spaced Training or IST. Such ongoing training improves efficiency and precision, and also improves Hybrid human-machine communications. Thus, in our team’s opinion, IST is a better process of electronic document review than training automatically without human participation, the so-called CAL approach promoted (and recently trademarked) by search experts and professors, Maura Grossman and Gordon Cormack.

Exactly how we space out the timing of training in IST is a little more difficult to describe without going into the particulars of a case. A full, detailed description would require the reader to have intimate knowledge of the EDR software. Our IST process is, however, software neutral. You can follow the IST dual feedback method of active machine learning with any document review software that has active machine learning capacities and also allows you to decide when to initiate a training session. (By the way, a training session is the same thing as a learning session, but we like to say training, not learning, as that takes the human perspective and we are pro-human!) You cannot do that if the training is literally continuous and cannot be halted while you input a new batch of relevance determined documents for training.

The details of IST, such as when to initiate a training session, and what human coded documents to select next for training, is an ad hoc process. It depends on the data itself, the issues involved in the case, the progress made, the stage of the review project and time factors. This is the kind of thing you learn by doing. It is not rocket science, but it does help keep the project interesting. Hire one of our team members to guide your next review project and you will see it in action. It is easier than it sounds. With experience Hybrid Multimodal IST becomes an intuitive process, much like riding a bicycle.

To summarize, active machine learning should be a dual feedback process with double-loop learning. The training should continue throughout a project, but it should be spaced in time so that you can actively monitor the progress, what we call IST. The software should learn from the trainer, of course, but the trainer should also learn from the software. This requires active monitoring by the teacher who reacts to what he or she sees and adjusts the training accordingly so as to maximize recall and precision.

This is really nothing more than a common sense approach to teaching. No teacher who just mails in their lessons, and does not pay attention to the students, is ever going to be effective. The same is true for active machine learning. That’s the essence of the insight. Simple really.

Next, in Part Three, I will address the related insights of Balanced Hybrid.

To be Continued …

6 Comments | Lawyers Duties, Review, Search, Technology, VENDORS | Tagged: best practices, borg, computer assisted review, document review, education, legal profession, legal search, machine learning, predictive coding, robots, science, search, support vector machines, tar, technology, too much information, vendors | Permalink
Posted by Ralph Losey

Predictive Coding 4.0 – Nine Key Points of Legal Document Review and an Updated Statement of Our Workflow – Part One

September 11, 2016

This blog introduces the e-Discovery Team’s latest insights and methods of document review.

We call this Predictive Coding 4.0 because it substantially improves upon, and replaces the methods and insights we announced in our October 2015 publication – Predictive Coding 3.0. In that two-part blog we explained the history of predictive coding software and methods in legal review, including versions 1.0 and 2.0. Then we described our new version 3.0 in some detail. Since that publication we have developed more enhancements to our methods, including many new, innovative uses of the predictive coding features of Kroll Ontrack’s EDR software. We even developed some new features not related to predictive coding. (Try out the new Folder Similar search in EDR for example.) Most of our new insights, just like our prior 3.0 version methodologies, can also be used on other software platforms. To use all of the features, however, the software will have to have bona fide active machine learning capacities. Most do not. More on that later.

These improvements naturally evolved to a certain degree as part of the e-Discovery Team members normal work supervising hundreds, maybe even thousands of document review projects over the past year. But the new insights that require us to make a complete restatement, a new Version 4.0, arose just recently. Major advances were attained as part of an intensive three months of experiments, all conducted outside of our usual legal practice and document reviews. The e-Discovery Team doing this basic research consisted of myself and several of Kroll Ontrack’s top document review specialists, including especially Jim Sullivan and Tony Reichenberger. They have now fully mastered the e-discovery team search and review Hybrid Multimodal methodologies. As far as I can see, at this point in the race for the highest quality legal document review, no one else comes even close to their skill level. Yes, e-discovery is highly competitive, but they trained hard and are now looking back and smiling.

The insights we gained, and the skills we honed, including speed, did not come easily. It took full time work on client projects all year, plus three full months of research, often in lieu of real summer vacations (my wife is still waiting). This is hard work, but we love it. See: Why I Love Predictive Coding. This kind of dedication of time and resources by an e-discovery vendor or law firm is unprecedented. There is a cost to attain the research benefits realized, both hard out-of-pocket costs and lost time. So I hope you understand that we are only going to share some of our techniques. The rest we will keep as trade-secrets. (Retain us and watch. Then you can see them in action.)

Mark Williams, CEO Kroll Ontrack

Kroll Ontrack understands the importance of pure research and enthusiastically approved these expenditures. (My thanks again to CEO Mark Williams, a true visionary leader in this industry who approved and supported the research program.) I suggest you ask your vendor, or law firm, how much time they spent last year researching and experimenting with document review methods? As far as we know, the only other vendor with an active research program is Catalyst, whose work is also to be commended. (No one else showed up for TREC.) The only other law firm we know of is Maura Grossman’s new solo practice. Her time spent with research is also impressive.

The results we attained certainly make this investment worthwhile, even if many in the profession do not realize it, much less appreciate it. They will in time, so will the consumers. This is a long term investment. Pure research is necessary for any technology company, including all companies in the e-Discovery field. The same holds true, albeit to a lesser extent, to any law firm claiming to have technological superiority.

Experience from handling live projects alone is too slow an incubator for the kind of AI breakthrough technologies we are now using. It is also too inhibiting. You do not experiment on important client data or review projects. Any expert will improvise somewhat during such projects to match the circumstances, and sometimes do post hoc analysis. But such work on client projects alone is not enough. Pure research is needed to continue to advance in AI-enhanced review. That is why the e-Discovery Team spent a substantial part of our waking hours in June, July and August 2016 working on experiments with Jeb Bush email. The Jeb Bush email collection was our primary laboratory this year. As a result of the many new things we learned, and new methods practiced and perfected, we have now reached a point where a complete restatement of our method is in order. Thus we here release Predictive Coding 4.0.

Our latest breakthroughs this summer primarily came out of the e-Discovery Team’s participation in the annual Text Retrieval Conference, aka TREC, sponsored by the National Institute of Standards and Technology. This is the 25th year of the TREC event. We were honored to again participate, as we did last year, in the Total Recall Track of TREC. This is the closest Track that TREC now offers to a real legal review project. It is not a Legal Track, however, and so we necessarily did our own side-experiments, and had our own unique approach different from the Universities that participated. The TREC leadership of the Total Recall Track was once again in the capable hands of Maura Grossman, Gordon Cormack and other scientists.

This blog will not report on the specifics of the 2016 Total Recall Track. That will come at a later time after we finish analyzing the enormous amount of data we generated and submit our formal reports to TREC. In any event, the TREC related work we did this Summer went beyond the thirty-four research topics included in the TREC event. It went well beyond the 9,863,366 documents we reviewed with Mr. EDR’s help as part of the formal submittals. Countless more documents were reviewed for relevance if you include our side-experiments.

At the same time that we did the formal tests specified by the Total Recall Track we did multiple side-experiments of our own. Some of these tests are still ongoing. We did so to investigate our own questions that are unique to legal search and thus beyond the scope of the Total Recall Track. We also performed experiments to test unique attributes of Kroll Ontrack’s EDR software. It uses a proprietary type of logistic regression algorithm that was awarded a patent this year. Way to go KO and Mr. EDR!

Although this blog will not report on our TREC experiments per se, we will share the bottom line, the take-aways of this testing. Not everything will be revealed. We keep some of our methods and techniques trade-secret.

We will also not be discussing in this multi-part blog our future plans and spin-off projects. Let’s just say for now that we have several in mind. One in particular will, I think, be very exciting for all attorneys and paralegals who do document review. Maybe even fun for those of you who, like us, are really into and enjoy a good computer search. You know who you are! If my recommendations are accepted, we will open that one up to all of our fellow doc-review freaks. I will say no more at this point, but watch for announcements in the coming year from Kroll Ontrack and me. We are having too much fun here not to share some of the good times.

Even if we did adopt 100% transparency on our methods, it would take a book to write it all down, and it would still be incomplete. Many things can only be learned by doing, especially methods. Document review is, after all, a part of legal practice. As the scientists like to put it, legal search is essentially ad hoc. It changes and is customized to fit the particular evidence search assignments at hand. But we will try to share all of the basic insights. They have all been discussed here before. The new insights we gained are more like a deepening understanding and matter of emphasis. They are refinements, not radical departures, although some are surprising.

Nine Insights Concerning the Use of Predictive Coding in Legal Document Review

The diagram below summarizes the nine basic insights that have come out of our work this year. These are the key concepts that we now think are important to understand and implement. [Just like the 8-Step Workflow diagram above, this, and other diagrams in this blog may be freely used with attribution. But please do not change anything without my permission. I am also happy to provide you with higher resolution graphics if needed for presentation or publication purposes.]

The diagrams above and following will be explained in detail throughout the rest of this multipart blog, as will the restated 8-Step Workflow shown at the top of the page. These are not new concepts. I have discussed most of these here before. I am confident that all readers will be able to follow along as I set forth the new nuances we learned.

Although these concepts are all familiar, some of our deepened understanding of these concepts may surprise you. Some were surprising to us. These insights include several changes in thinking on our part. Some of the research results we saw were unexpected. But we follow the data. Our opinions are always held lightly. I have argued both sides of a legal issue too many times as a lawyer to fall into that trap. Our thinking follows the evidence, not our preconceptions. That is, after all, the whole point of research. Schedule permitting, we are also happy to provide in-person or online presentations that explain these concept-summary diagrams. If retained, you can also see it in action.

Although these insights and experiments were derived using Kroll Ontrack EDR software, they are essentially vendor neutral. The methods will work on any full-featured document review platform, but especially those that includes bona fide active machine learning abilities, aka Predictive Coding. As all experts in this field know, many of the most popular document review platforms do not have these features, even those stating they use Analytics. Active Machine Learning is very different, and far more advanced than Analytics, the early forms of which were called Concept Search. This type of machine learning is passive and clearly is not predictive coding. It has its place in any multimodal system such as ours, and can be a powerful feature to improve search and review. But such software is incomplete and cannot meet the standards and capability of software that includes active machine learning. Only full featured document review platforms with active machine learning abilities can use all of the Predictive Coding 4.0 methods described here.

Sorry dear start-up vendors, and others, but that’s the truth. Consumers, you get what you pay for. You know that. Not sure? Get the help of an independent expert advisor before you make substantial investments in e-discovery software or choose a vendor for a major project. Also, if you have tried predictive coding, or what you were told was advanced TAR, whatever the hell that is, and it did not work well, do not blame yourself. It could be the software. Or if not the software, then the antiquated version 1.0 or 2.0 methods used. There is a lot of bullshit out there. Excuse my French. There always has been when it comes to new technology. It does, however, seems especially prevalent in the legal technology field. Perhaps they think we lawyers are naive and technologically gullible. Do not be fooled. Again, look to an independent consultant if you get confused by all the vendor claims.

Contrary to what some vendors will tell you (typically the ones without bona fide predictive coding features), predictive coding 3.0, and now 4.0, methods are not rocket science. You do not have to be a TAR-whisperer or do nothing but search, like my A-team for TREC. With good software it is not really that hard at all. These methods do, however, require an attorney knowledgable in e-discovery and comfortable with software. This is not for novices. But every law firm should anyway have attorneys with special training and experience in technology and e-discovery. For instance, if you practice in the Northern District of California, an e-discovery liaison with such expertise is required in most cases. See Guidelines for the Discovery of Electronically Stored Information. Almost half of the Bar Associations in the U.S. require basic technology competence as an ethical imperative. See eg. ABA Model Rule 1.1, Comment [8] and Robert Ambrogi’s list of 23 states, and counting, that now require such competence. (My own law firm has had an e-discovery liaison program in place since 2010, which I lead and train. I am proud to say that after six years of work it is now a great success.) So no, you do not have to be a full-time specialist, like the members of my TREC e-Discovery team, to successfully use AI-enhanced review, which we call Hybrid Multimodal. This is especially true when you work with vendors like Kroll Ontrack, Catalyst and others that have teams of special consultants to guide you. You just have to pick your vendors wisely.

To be continued …

17 Comments | Review, Search, Technology, VENDORS | Tagged: best practices, computer assisted review, document review, legal profession, legal search, machine learning, predictive coding, science, support vector machines, tar, technology, truth, vendors | Permalink
Posted by Ralph Losey

Five Reasons You Should Read the ‘Practical Law’ Article by Maura Grossman and Gordon Cormack called “Continuous Active Learning for TAR”

April 11, 2016

There is a new article by Gordon Cormack and Maura Grossman that stands out as one of their best and most accessible. It is called Continuous Active Learning for TAR (Practical Law, April/May 2016). The purpose of this blog is to get you to read the full article by enticing you with some of the information and knowledge it contains. But before we go into the five reasons, we will examine the purpose of the article, which aligns with our own, and touch on the differences between their trademarked TAR CAL method and our CAR Hybrid Multimodal method. Both of our methods use continuous, active learning, the acronym for which, CAL, they now claim as a Trademark. Since they clearly did invent the acronym, CAL, we for one will stop using it – CAL – as a generic term.

The Legal Profession’s Remarkable Slow Adoption of Predictive Coding

The article begins with the undeniable point of the remarkably slow adoption of TAR by the legal profession, in their words:

Adoption of TAR has been remarkably slow, considering the amount of attention these offerings have received since the publication of the first federal opinion approving TAR use (see Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012)).

I remember getting that landmark ruling in our Da Silva Moore case, a ruling that pissed off plaintiffs’ counsel, because, despite what you may have heard to the contrary, they were strenuously opposed to predictive coding. Like most other lawyers at the time who were advocating for advanced legal search technologies, I thought Da Silva would open the flood gates, that it would encourage attorneys to begin using the then new technology in droves. In fact, all it did was encourage the Bench, but not the Bar. Judge Peck’s more recent ruling on the topic contains a good summary of the law. Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015). There were a flood of judicial rulings approving predictive coding all around the country, and lately, around the world. See Eg. Pyrrho Investments v MWB Property, EWHC 256 (Ch) (2/26/16).

The rulings were followed in private arbitration too. For instance, I used the Da Silva More ruling a few weeks after it was published to obtain what was apparently the first ruling by an arbitrator in AAA approving use of predictive coding. The opposition to our use of cost-saving technology in that arbitration case was again fierce, and again included personal attacks, but the arguments for use in arbitration are very compelling. Discovery in arbitration is, after all, supposed to be constrained and expedited.

After the Da Silva Moore opinion, Maura Grossman and I upped our speaking schedule (she far more than me), and so did several tech-minded judges, including Judge Peck (although never at the same events as me, until the cloud of false allegations created by a bitter plaintiff’s counsel in Da Silva Moore could be dispelled). At Legal Tech for the next few years Predictive Coding is all anybody wanted to talk about. Then IG, Information Governance, took over as the popular tech-child of the day. In 2015 we had only a few predictive coding panels at Legal Tech, but they were well attended.

The Grossman Cormack speculates that the cause of the remarkably slow adoption is:

The complex vocabulary and rituals that have come to be associated with TAR, including statistical control sets, stabilization, F1 measure, overturns, and elusion, have dissuaded many practitioners from embracing TAR. However, none of these terms, or the processes with which they are associated, are essential to TAR.

We agree. The vendors killed what could have been their golden goose with all this control set nonsense and their engineers love of complexity and misunderstanding of legal search. I have ranted about this before. See Predictive Coding 3.0. I will not go into that again here, except to say the statistical control set nonsense that had large sampling requirements was particularly toxic. It was not only hard and expensive to do, it led to mistaken evaluations of the success or failure of projects because it ignored the reality of the evolving understand of relevance, so called concept drift. Another wrong turn involved the nonsense of using only random selection to find training documents, a practice that Grossman and I opposed vigorously. See Latest Grossman and Cormack Study Proves Folly of Using Random Search For Machine Training – Part One, Part Two, Part Three, and Part Four. Grossman and Cormack correctly criticize these old vendor driven approaches in Continuous Active Learning for TAR. They call them SAL and SPL protocols (a couple of acronyms that no one wants to trademark!).

Bottom line, the tide is changing. Over the last several years the few private attorneys who specialize in legal search, but are not employed by a vendor, have developed simpler methods. Maura and I are just the main ones writing and speaking about it, but there are many others who agree. Many have found that it is counter-productive to use control sets, random input, non-continuous training with its illogical focus on the seed set, and misleading recall point projections.

We do so in defiance of the vendor establishment and other self-proclaimed pundits in this area who benefitted by such over-complexity. Maura and Gordon, of course, have their own software (Gordon’s creation), and so never needed any vendors to begin with. Not having a world renowned information scientist like Professor Cormack as my life partner, I had no choice but to rely on vendors for their software. (Not that I complaining, mind you. I’m married to a mental health counselor, and it does not get any better than that!)

After a few years I ultimately settled on one vendor, Kroll Ontrack, but I continue to try hard to influence all vendors. It is a slow process. Even Kroll Ontrack’s software, which I call Mr. EDR, still has control set functions built in. Thanks to my persistence, it is easy to turn off these settings and do things my way, with no secret control sets and false recall calculations. Hopefully soon that will be the default setting. Their eyes have been opened. Hopefully all of the other major vendors will soon follow suit.

All of the Kroll Ontrack experts in predictive coding are now, literally, a part of my Team. They are now fully trained and believers in the simplified methods, methods very similar to those of Grossman and Cormack, albeit, as I will next explain, slightly more complicated. We proved how well these methods worked at TREC 2015 when the Kroll Ontrack experts and I did 30 review projects together in 45 days. See e-Discovery Team at TREC 2015 Total Recall Track, Final Report (116 pg. PDF), and (web page with short summary). Also see – Mr. EDR with background information on the Team’s participation in the TREC 2015 Total Recall Track.

We Agree to Disagree with Grossman and Cormack on One Issue, Yet We Still Like Their Article

We are fans of Maura Grossman and Gordon Cormack’s work, but not sycophants. We are close, but not the same; colleagues, but not followers. For those reasons we think our recommendation for you to read this article means more than a typical endorsement. We can be critical of their writings, but, truth is, we liked their new article, although we continue to dislike the name TAR (not important, but we prefer CAR). Also, and this is of some importance, my whole team continues to disagree with what we consider the somewhat over-simplified approach they take to finding training documents, namely reliance on the highest ranking documents alone.

Despite what some may think, the high-ranking approach does eventually find a full diversity of relevant documents. All good predictive coding software today pretty much uses some type of logistic regression based algorithms that are capable of building out probable relevance in that way. That is one of the things we learned by rubbing shoulders with text retrieval scientists from around the world at TREC when participating in the 2015 Total Recall Track that Grossman and Cormack helped administer. This regression type of classification system works well to avoid the danger of over-training on a particular relevancy type. Grossman and Cormack have proven that before to our satisfaction (so have our own experiments), and they again make a convincing case for this approach in this article.

Still, we disagree with their approach of only using high-ranking documents for training, but we do so on the grounds of efficiency and speed, not effectiveness. The e-Discovery Team continues to advocate a Hybrid Multimodal approach to active machine learning. We use what I like to call a four-cylinder type of CAR search engine, instead of one-cylinder, like they do.

High-ranking documents;
Mid-level, uncertain documents;
A touch, a small touch, of random documents; and,
Human ingenuity found documents, using all type of search techniques (multimodal) that seem appropriate to the search expert in charge, including keyword, linear, similarity (including chains and families), concept (including passive machine learning, clustering type search).

Predictive Coding 3.0 – The method is here described as an eight-part work flow (Step 6 – Hybrid Active Training).

The latest Grossman and Cormack’s versions of CAL (their trademark) only uses the highest-ranking documents for active training. Still, in spite of this difference, we liked their article and recommend you read it.

The truth is, we also emphasize the high-probable relevant documents for training. The difference between us is that we use the three other methods as well. On that point we agree to disagree. To be clear, we are not talking about continuous training or not, we agree on that. We are not talking about active training, or not (passive), we agree on that. We are not talking about using what they call using SAL or SPL protocols (read their article for details), we agree with them that these protocols are ineffective relics invented by misguided vendors. We are only talking about a difference in methods to find documents to use to train the classifier. Even that is not a major disagreement, as we agree with Grossman and Cormack that high-ranking documents usually make the best trainers, just not in the first seed set. There are also points in a search, depending on the project, where the other methods can help you get to the relevant documents in a fast, efficient manner. The primary difference between us is that we do not limit ourselves to that one retrieval method like Grossman and Cormack do in their trademarked CAL methodology.

Cormack and Grossman emphasize simplicity, ease of use, and reliance on the software algorithms as another way to try to overcome the Bar’s continued resistance to TAR. The e-Discovery Team has the same goal, but we do not think it is necessary to go quite that far for simplicity sake. The other methods we use, the other three cylinders, are not that difficult and have many advantages. e-Discovery Team at TREC 2015 Total Recall Track, Final Report (116 pg. PDF and web page with short summary). Put another way, we like the ability of fully automatic driving from time to time, but we want to keep an attorney’s learned hand at or near the wheel at all times. See Why the ‘Google Car’ Has No Place in Legal Search.

Accessibility with Integrity: The First Reason We Recommend the Article

Here’s the first reason we like Grossman & Cormack’s article, Continuous Active Learning for TAR: you do not have to be one of Professor Cormac’s PhD students to understand it. Yes. It is accessible, not overly technical, and yet still has scientific integrity, still has new information, accurate information, and still has useful knowledge.

It is not easy to do both. I know because I try to make all of my technical writings that way, including the 57 articles I have written on TAR, which I prefer to call Predictive Coding, or CAR. I have not always succeeded in getting the right balance, to be sure. Some of my articles may be too technical, and perhaps some suffer from breezy information over-load and knowledge deficiency. Hopefully none are plain wrong, but my views have changed over the years. So have my methods. If you compare my latest work-flow (below) with earlier ones, you will see some of the evolution, including the new emphasis over the past few years with continuous training.

The Cormacks and I are both trying hard to get the word out to the Bar as to the benefits of using active machine learning in legal document review. (We all agree on that term, active machine learning, and all agree that passive machine learning is not an acceptable substitute.) It is not easy to write on this subject in an accurate, yet still accessible and interesting manner. There is a constant danger that making a subject more accessible and simple will lead to inaccuracies and misunderstandings. Maura and Gordon’s latest article meets this challenge.

Take for example the first description in the article of their continuous active training search method using highest ranking documents:

At the outset, CAL resembles a web search engine, presenting first the documents that are most likely to be of interest, followed by those that are somewhat less likely to be of interest. Unlike a typical search engine, however, CAL repeatedly refines its understanding about which of the remaining documents are most likely to be of interest, based on the user’s feedback regarding the documents already presented. CAL continues to present documents, learning from user feedback, until none of the documents presented are of interest.

That is a good way to start an article. The comparison with a Google search having continued refinement based on user feedback is well thought out; simple, yet accurate. It represents a description honed by literally hundreds of presentations on the topic my Maura Grossman. No one has talked more on this topic than her, and I for one intend to start using this analogy.

Rare Description of Algorithm Types – Our Second Reason to Recommend the Article

Another reason our Team liked Continuous Active Learning for TAR is the rare description of search algorithm types that it includes. Here we see the masterful touch of one of the world’s leading academics on text retrieval, Gordon Cormack. First, the article makes clear the distinction between effective analytic algorithms that truly rank documents using active machine learning, and a few other popular programs now out there that use passive learning techniques and call it advanced analytics.

The supervised machine-learning algorithms used for TAR should not be confused with unsupervised machine-learning algorithms used for clustering, near-duplicate detection, and latent semantic indexing, which receive no input from the user and do not rank or classify documents.

These other older, unsupervised search methods are what I call concept search. It is not predictive coding. It is not advanced analytics, no matter what some vendors may tell you. It is yesterday’s technology – helpful, but far from state-of-the-art. We still use concept search as part of multimodal, just like any other search tool, but our primary reliance to properly rank documents is placed on active machine learning.

The Cormack-Grossman article goes farther than pointing out this important distinction, it also explains the various types of bona fide active machine learning algorithms. Again, some are better than others. First Professor Cormack explains the types that have been found to be effective by extensive research over the past ten years or so.

Supervised machine-learning algorithms that have been shown to be effective for TAR include:

– Support vector machines. This algorithm uses geometry to represent each document as a point in space, and deduces a boundary that best separates relevant from not relevant documents.

– Logistic regression. This algorithm estimates the probability of a document’s relevance based on the content and other attributes of the document.

Conversely Cormack explains:

Popular, but generally less effective, supervised machine-learning algorithms include:

– Nearest neighbor. This algorithm classifies a new document by finding the most similar training document and assuming that the correct coding for the new document is the same as its nearest neighbor.

– Naïve Bayes (Bayesian classifier). This algorithm estimates the probability of a document’s relevance based on the relative frequency of the words or other features it contains.

Ask your vendor which algorithms its software includes. Prepare yourself for double-talk.

If you try out your vendors software and the Grossman-Cormack CAL method does not work for you, and even the e-Discovery Team’s slightly more diverse Hybrid Multimodal method does not work, then your software may be to blame. As Grossman-Cormack put it, where the phrase “TAR tool” means software:

[I]t will yield the best possible results only if the TAR tool incorporates a state-of-the-art learning algorithm.

That means software that uses a type of support vector machine and/or logistic regression.

Teaching by Example – Our Third Reason to Recommend the Article

The article uses a long example involving search of Jeb Bust email to show you how their CAL method works. This is an effective way to teach. We think they did a good job with this. Rather than spoil the read with quotes and further explanation, we urge you to check out the article to see for yourself. Yes, it is an oversimplification, after all this is a short article, but it is a good one, and is still accurate.

Quality Control Suggestions – Our Fourth Reason to Recommend the Article

Another reason we like the article are the quality control suggestions it includes. They essentially speak of using other search methods, which is exactly what we do in Hybrid Multimodal. Here are their words:

To increase counsel’s confidence in the quality of the review, they might:

Review an additional 100, 1,000, or even more documents.

Experiment with additional search terms, such as “Steve Jobs,” “iBook,” or “Mac,” and examine the most-likely relevant documents containing those terms.

Invite the requesting party to suggest other keywords for counsel to apply.

Review a sample of randomly selected documents to see if any other documents of interest are identified.

We like this because it shows that the differences are small between the e-Discovery Team’s Hybrid Multimodal method (hey, maybe I should claim Trademark rights to Hybrid Multimodal, but then again, no vendors are using my phrase to sell their products) using continuous active training, and the Grossman-Cormack trademarked CAL method. We also note that their section on Measures of Success essentially mirrors our own thoughts on metric analysis and ei-Recall. Introducing “ei-Recall” – A New Gold Standard for Recall Calculations in Legal Search – Part One, Part Two and Part Three.

Article Comes With an Online “Do it Yourself” CAL Trial Kit – Our Fifth Reason to Recommend the Article

We are big believers in learning by doing. That is especially true in legal tasks that seem complicated in the abstract. I can write articles and give presentations that provide explanations of AI-Enhanced Review. You may get an intellectual understanding of predictive coding from these, but you still will not know how to do it. On the other hand, if we have a chance to show someone an entire project, have them shadow us, then they will really learn how it is done. It is like teaching a young lawyer how to try a case. For a price, we will be happy to do so (assuming conflicts clear).

Maura and Gordon seem to agree with us on that learn by doing point and have created an online tool that anyone can use to try out their method. In allows for a search of the Jeb Bush email, the same set of 290,099 emails that we used in ten of the thirty topics in 2015 TREC. In their words:

There is no better way to learn CAL than to use it. Counsel may use the online model CAL system to see how quickly and easily CAL can learn what is of interest to them in the Jeb Bush email dataset. As an alternative to throwing up their hands over seed sets, control sets, F1 measures, stabilization, and overturns, counsel should consider using their preferred TAR tool in CAL mode on their next matter.

You can try out their method with their online tool, or in a real project using your vendor’s tool. By the way, we did that as part of our TREC 2015 experiments, and the Kroll Ontrack software worked about the same as theirs, even when we used their one-cylinder, high ranking only, CAL (their trademark) method.

Here is where you can find their CAL testing tool: cormack.uwaterloo.ca/cal. Those of you who are still skeptical can see for yourself how it works. You can follow the example given in the article about searching for documents relevant to Apple products, to verify their description of how that works. For even more fun, you can dream up your own searches.

Perhaps, if you try hard enough, you can find some example searches where their high-end only method, which is built into the test software, does not work well. For example, try finding all emails that pertain to, or in any way mention, the then President, George Bush. Try entering George Bush in the demo test and see for yourself what happens.

It becomes a search for George + Bush in the same document, and then goes from there based on your coding the highest ranked documents presented as either relevant or non-relevant. You will see that you quickly end up in a TAR pit. The word Bush is in every email (I think), so you are served up with every email where George is mentioned, and believe me, there are many Georges, even if there is only one President George Bush. Here is the screen shot of the first document presented after entering George Bush. I called it relevant.

These kind of problem searches do not discredit TAR, or even the Grossman Cormack one-cylinder search method. If this happened to you in a real search project, you could always use our Hybrid Multimodal™ method for the seed set (1st training), or start over with a different keyword or keywords to start the process. You could, for instance, search for President Bush, or President within five of George, or “George Bush.” There are many ways, some faster and more effective than others.

Even using the single method approach, if you decided to use the keywords “President + Bush”, then the search will go quicker than “George + Bush.” Even just using the term “President” works better than George + Bush, but still seems like a TAR pit, and not a speeding CAR. It will probably get you to the same destination, high recall, but the journey is slightly longer and, at first, more tedious. This high recall result was verified in TREC 2015 by our Team, and by a number of Universities who participated in the fully automatic half of the Total Recall Track, including Gordon’s own team. This was all done without any manual review by the fully automatic participants because there was instant feedback of relevant or irrelevant based on a prejudged gold standard. See e-Discovery Team at TREC 2015 Total Recall Track, Final Report (116 pg. PDF), and (web page with short summary). With this instant feedback protocol, all of the teams attained high recall and good precision. Amazing but true.

You can criticized this TREC experiment protocol, which we did in our report, as unrealistic to legal practice because:

(1) there is no SME who works like that (and there never will not be, until legal knowledge itself is learned by an AI); and,

(2) the searches presented as tasks were unrealistically over-simplistic. Id.

But you cannot fairly say that CAL (their trademark) does not work. The glass is most certainly not half empty. Moreover, the elixir in this glass is delicious and fun, especially when you use our Hybrid Multimodal™ method. See Why I Love Predictive Coding: Making document review fun with Mr. EDR and Predictive Coding 3.0.

Conclusion

Active machine learning (predictive coding) using support vector or logistic regression algorithms, and a method that employs continuous active training, using either one cylinder (their CAL), or four (our Hybrid Multimodal), really works, and is not that hard to use. Try it out and see for yourself. Also, read the Grossman Cormack article, it only takes about 30 minutes. Continuous Active Learning for TAR (Practical Law, April/May 2016). Feel free to leave any comments below. I dare say you can even ask questions of Grossman or Cormack here. They are avid readers and will likely respond quickly.