Litigation support and electronic discovery services company D4 has hosted a Web-based electronic discovery platform for area law firms and corporate clients for several years, and is now poised to provide that service to Canadian clients.

Read full article

The application of Predictive Coding can significantly reduce costs, dramatically reduce review time and increase quality for document review. Yes—cheaper, faster and better, all three. Computerized technology has greatly impacted the efficiencies of document review over the years. Keyword Searching was among the first techniques and while according to many studies this approach only ranges from 20% to 40% for responsiveness precision (100% precision would be all of the documents that truly are responsive), Keyword Search remains as the most common approach used today for reducing the number of documents to be reviewed. Keyword search was further advanced through ontologies which, in their simplest form, is an exhaustive keyword search where a set of “but not” terms are used to disambiguate over-inclusive keywords. For example, an ontology would state “include (keyword1, keyword2, keyword3, etc.) but not (excludeword1, excludeword2, excludeword3, etc.)”. Ontologies can also include proximity limiters (i.e. Tom w/2 Groom) as well as incorporate conceptual relationships. Studies have shown ontologies can improve responsiveness precision to 65% to 90%. In the past, ontologies had to be developed by highly skilled linguists working with counsel who have intimate knowledge of the legal issues of the case. This was expensive and time consuming. Predictive
Coding changes all that.

What is Predictive Coding?
Predictive Coding combines the efficiencies of a computerized sampling system with a human “expert.” The human interacts with the system by making “yes/no” calls to a question against a series of controlled samples of 40 documents at a time. Questions can be “Is this document responsive?” or “Does it pertain to this specific issue?” or “Is this document privileged?”, etc. The system builds an ontology in the background as it learns from the expert and presents subsequent samples. Normally after 25-40 iterations (1,000 to 1,500 documents), the system has sufficiently built the ontology to the point where it can “predict” what the human will choose as “affirmative” in the sample they are reviewing. Once it accurately predicts over a series of consecutive samples, the ontology is considered “statistically stable” and can be applied to the rest of the collection.

Predictive Coding is built upon a well established modeling framework called Predictive Analytics which is a type of a Support Vector Machine. Predictive Analytics encompasses a
variety of techniques from statistics, data mining and game theory that analyze current and historical facts to make predictions about future events. In business, predictive models exploit patterns found in historical and transactional data to identify risks and opportunities. Models capture relationships among many factors to allow assessment of risk or potential associated with a particular set of conditions, guiding decision making for candidate transactions. Predictive Analytics is used in actuarial science, financial services, insurance, telecommunications, retail, travel, healthcare, pharmaceuticals and now is being applied to document review via Predictive Coding. One of the most well-known applications is credit scoring, which is used throughout financial services. Scoring models process a customer’s credit history, loan application, customer data, etc., in order to rank-order individuals by their likelihood of making future credit payments on time. A well-known example would be the FICO Score.

The output of Predictive Coding for document review is a “Relevance Score” based on a scale of 0 to 100. Once the system has been trained and the model is statically stable, the rest of the collection can be scored based on the underlying ontology. That score can then be used to identify non-responsive documents as well as prioritize review towards the documents with the highest scores. Some Predictive Coding systems such as Equivio-Relevance provide interactive graphical tools to aid case management in determining the approach for specific relevance score zones. See example below.

Use Cases and Workflows:
Listed below are some sample use cases. There are many variations to these and the specific
workflow depends on specific case variables but these should provide guidance for
consideration.

1. Early Case: It is early in the case cycle and counsel hasn’t started the development of list of keywords. You anticipate a high volume of ESI that will need to be collected in order to find documents relevant to the case. Predictive Coding can be used on a set of ESI for a few key custodians which will result in:

a. finding key documents early in the process – may help determine if the case should settle or if there are sufficient facts to pursue

b. identifying the initial set of keywords that could then be used in negotiations with opposing and/or modified and used as a keyword filter to limit subsequent ESI collection.

2. Accelerated Review: Consider the relevance score distribution graph below. Using the relevance scores, a review team was able to divide the collection into three zones:

a. Zone 1 contains documents with relevance scores of 0 through 4. It contains 72% of the document population, but only yields less than 2% of the relevant documents. After a cursory review, the firm determined further review of the documents in this zone meets no criteria for reasonableness or proportionality of effort for a full initial review.

b. Zone 2 contains documents with relevance scores 4 through 37. It covers 3% of the population, but less than 1% of the relevant documents. The firm did some sampling, found very low yield of relevant documents – and decided that it was not worth the effort for a full initial review.

c. Zone 3 includes documents with relevance scores over 37. This zone covers 25% of the population, and contains 97.4% of the relevant documents. From the firm’s
point of view, this was a “must review” zone.

So, the firm was able to focus its initial review effort on just 25% of the population, and in so doing, they were able to identify and produce 97% of the relevant documents.

3. Review QC/Verification: When the review was complete, the firm turned to quality assurance. They set up a discrepancy matrix, comparing the relevance designations of the review team to the relevance scores regardless of zone as described above.

The graph above shows that there were 3,048 documents that the review team and Relevance agreed were responsive. Then there were 40,495 documents that the review team and Relevance agreed were not responsive.

Of particular interest to the firm were the 2,531 documents that the review team had marked as not responsive, but which Relevance scored as responsive. These documents represent potentially responsive documents that the reviewers may have missed. These 2,531 documents were submitted to a senior reviewer (so called “Oracle”) for second
pass review and verification. He found that almost 1,500 of these documents were in fact responsive. The responsive set increased from 4,624 to 6,000. That’s an additional
one-third on top of the original set that had been slated for production. The lead partner on the case saw this and his response was “My obligation is to make reasonable effort to
discover the responsive documents. If I’m not using this technology, I’m not fulfilling my obligation and I am open to risk to a claim from opposing counsel that we have not
disclosed all the relevant documents.” This firm has standardized on Relevance largely because of these risk considerations.

There are many variations of these use cases and workflows but they generally fall into the three categories shown above.

Summary
Predictive Coding goes beyond basic keyword searching. It is a powerful tool that uses well established predictive analytics and classification algorithms rather than just discrete keyword searches. Unlike keyword searches, Predictive Coding takes into account all the words in a document as well as words to exclude, along with the relationship of the words to one another to determine what is and what is not likely to be relevant. Predictive Coding incorporates human intelligence to leverage the results of review across large document populations. Predictive Coding can be used in a variety of workflows in several places along the EDRM lifecycle including Early Case, Accelerated Document Review and Review QC/Verification.

Last week, I presented a CLE to the San Diego Paralegal Association. During that presentation, one of the attendees posited a question concerning how conceptual engines cope with less than perfect data. The question, which was two-part, (paraphrasing) was as follows:

1. How does a conceptual engine cope with “dirty data”?
2. How does a conceptual engine deal with foreign language data, such as Mandarin, Russian, Spanish, German, etc?

So that we are all on the same page, there are a couple of points that provide background to my answer.

1. When the questioner was referring to “dirty data”, she was speaking of data that originated from a source that was older, likely had to be OCR’d and therefore, is less in quality than what one would find in pristine extracted text. We all know OCR is not perfect, but it is even more difficult to get good useable text from older paper documents being converted to a digital format.

2. It is also important to distinguish between languages and how they reside in electronic format. Languages such as Spanish, French & German are much simpler in form and can be accounted for in standard digital coding (ANSI). Mandarin, Russian, Arabic are more complex in the construction of the actual character, and as a result must be stored in a more complex form known as Unicode. In addition, Unicode has a larger data volume footprint than the standard ANSI characters.

Now that we’ve got the technical part out of the way, let’s talk about how conceptual engines work.

Conceptual engines, such as Relativity Analytics, treat files as nothing more than a big bag of words. It does not read and understand the data in the way that a human could read and understand it. Rather, it looks at each file to find recurring words or combinations of words. The engine then compares files to determine if they contain similar, recurring or combinations of words.

Think of it like this; the engine is actually looking to see if Bag of Words A contains similar words, or combinations of words to Bag of Words B. If they do, then they are conceptually related to each other. If they do not, then they are not conceptually related to each other. Yes, it is more technical than that, but that is the best way to visualize how these tools function.

Conclusion: So what does that mean for the questioner?

Well, while the text of the files is vitally important to the conceptual engine, that doesn’t mean a conceptual engine would be useless on her data.

First, if the data is “dirty”, you’d want to try to avoid including that “dirty” data in the set of data used to train the engine. Instead, try to use the data with the most pristine text as the training set for the engine. You may not generate refined buckets of documents on the back end of the effort for the “dirty” data, but you may be able to make easier work of it, depending on the quality of text in each individual file.

Second, when it comes to foreign language data, the language is not superbly important. Rather, the words, or combination of words is what is important. The engine can tell you if files are related – even if you can’t read the language. For the Unicode languages such as Mandarin, I’d expect the system to take longer to process and group the data, but it should still provide you with some workable buckets of data.

A lesson in what not to do using Voom vs. EchoStar as an example.

In Zubulake, the court stated that “once a party reasonably anticipates litigation, it must suspend its routine document retention/destruction policy and put in place a litigation hold to ensure the preservation of relevant documents.”

In 2005, Voom and EchoStar entered into a 15 year affiliation agreement. The agreement required EchoStar to include Voom channels as part of its packaged services. Fast forward a few years and EchoStar wanted out because they did not feel they were getting the best deal.

In the summer of 2007, EchoStar sent Voom a letter stating its desire to audit Voom’s records to see if it was in breach of the contract, which would allow an out for EchoStar. It wasn’t until January of 2008 that EchoStar officially terminated the agreement and Voom filed suit the next day. During this time frame EchoStar made three big mistakes. First, it did not implement a litigation hold; second, it allowed custodians to self-select “relevant” documents; and third, it allowed automated deletions to occur after the self selection.

The Plaintiffs moved for spoliation and the court, citing Zubulake, concluded that EchoStar should have reasonably anticipated litigation no later than June 20, 2007, the date its counsel sent written notice to Voom that it may be in violation of the contract.

The court ultimately imposed the sanction of precluding the defendant from calling its expert at trial and from introducing his expert report.

The headlines in the blogosphere about this opinion are talking about the Zubulake precedent and the question of when a party must begin to preserve ESI. This court clearly says that the preservation duty arises when one can reasonably anticipate litigation. Subjective, yes, and based to some extent on the facts of each case, but certainly before a complaint is filed. The opinion is an interesting read and I highly recommend it be read in its entirety.

Perhaps more interesting than the WHEN question is the WHO question. Who is responsible for determining possible relevance and thus preservation? Is it the custodian or legal counsel? Look at these statements taken from the opinion:

1. The court observed that in addition to failing to preserve electronic data upon reasonable anticipation of litigation, no steps whatsoever had been taken to prevent the purging of e-mails by employees during the four-month period after commencement of the action. EchoStar continued to permanently delete employee e-mails for up to four months after the commencement of the action, relying on employees to determine which documents were relevant in response to [*5]litigation, and to preserve those e-mails by moving them to separate folders. As the court put it: “EchoStar’s purported litigation hold failed to turn off the automatic delete function and merely asked its employees — many of whom, presumably were not attorneys — to determine whether documents were potentially responsive to litigation, and to then remove each one from EchoStar’s pre-set path of destruction.”

2. Regardless of its nature, a hold must direct appropriate employees to preserve all relevant records, electronic or otherwise, and create a mechanism for collecting the preserved records so they might be searched by someone other than the employee.

3. In certain circumstances, like those here, where a party is a large company, it is insufficient, in implementing such a litigation hold, to vest total discretion in the employee to search and select what the employee deems relevant without the guidance and supervision of counsel.

4. In this case, EchoStar’s reliance on its employees to preserve evidence “does not meet the standard for a litigation hold (see Pension Comm. of the Univ. of Montreal Pension Plan, 685 F Supp 2d at 473; see also Einstein v 357 LLC, 2009 NY Slip Op 32784[U] [Sup Ct, NY County 2009] [finding that the failure to suspend deletion policy or to investigate the basic ways in which e-mail was stored constituted a “serious discovery default” rising to the level of gross negligence or willfulness entitling party to an adverse inference;

The key point here is that counsel can’t just throw a command over the transom to the client’s employees telling them to pick out the relevant documents. Supervision and guidance are required.

OMAHA (KPTM) — When you give away your old computer or cell phone, you may be giving up more than you think.

A Fox 42 Special Report shows how easy it is for computer-saavy thieves and blackmailers to access and extract private information.

“People can get that information and if it gets in the wrong hands, it can be used against you,” said Peter Coons, of D4 Discovery, a company that specializes in computer forensics.

Experts at D4 Discovery analyzed discarded cell phones and computers that Fox 42 borrowed.

Coons found personal financial data, photographs, business emails and instances, he found data suggesting that the computer’s former owner may have been visiting online sex chat rooms.

Information such as that, Coons confirmed, could be used for blackmail.

Since 2005, more than 345 million documents containing sensitive personal information have been involved in security breaches, according to privacyright.org.

Among the four cell phones examined, Coons found text messages containing passwords from the phone service provider notifying the owner that the password had been changed. He also found lists of contacts on that phone.

“You know there can be some very sensitive information in texts as well,” Coons said.

Experts say, to protect your data you can buy self-encrypting hard drives or you can take your computer to professionals who can magnetically scrub your hard drive. However, the one sure-fire way of making sure no one can access your private information is to have your hard drive shredded.

EDT Recycling in Omaha uses a heavy duty shredding machine to do that. Hard drives are fed into the machine and within minutes the drive is reduced to mangled strips of metal, making data recovery impossible.

Brian Gobbels, EDT Recycling President, said professional recyclers will always issue a numbered Certificate of Disposal for every drive they destroy.

As mentioned in my last post, many law firms and eDiscovery service providers have in recent months extensively discussed social media discovery in a general fashion. Such dialogue is important and educational, but it is great to see the innovative service firm D4 highlighting their major league, battle tested social media discovery capabilities by outlining a case study where they delivered a huge win for their clients.

D4, renowned experts in computer forensics and eDiscovery consulting, recently tackled a case involving the capture of several dozen Facebook and LinkedIn accounts amounting to over 2 million items. Utilizing X1 Social Discovery, D4 rapidly collected and searched through these items to quickly pare down the item count to roughly 7,000. Once the social media data set was at 7,000 items, D4 exported the data into the review platform Relativity.

In addition to D4’s prowess and expertise, three critical factors served as game changers to allow for this exercise that would otherwise have been impossible without best practices technology. First, D4 was able to collect these 2 million items on a highly automated basis with X1 Social Discovery. If screen captures were used instead, it may literally take a year to capture that volume of data. Second, once captured those two million items were indexed at the point of collection and placed within a designated case allowing for searching, sorting and analysis in a single interface and case-centric workflow. D4’s efforts reveal how X1’s workflow and patented fast-as-you type search enables very effective early case assessment for social media.

Finally, X1 Social Discovery’s unique ability to capture several dozens of metadata fields for social media was extremely important to the case. D4 states:

“We were able to take advantage of the metadata fields captured by D4 during the collection process…The attorneys were able to go through the posts in less than two days and found a dozen or more posts that were critical to the case. In my opinion, without using the tools and expertise offered by D4 they never would have found the evidence.”

We agree. Look for an upcoming webinar with D4 and X1 Discovery highlighting this case in more detail as well as other recent successes from the front lines.

And on a related recent development, John Waters of Law Technology News published a review of X1 Social Discovery after his extensive testing, and the results are consistent with the scalable functionality leveraged by D4:

“In my tests, X1 (Social Discovery) was fast and responsive…The interface is clean and simple, with a logical layout. And the feature set is just about everything you could ask for as you dig into the burgeoning mounds of status updates, tweets, digital pictures, hash tags, comments, and links. I was impressed with X1′s ability to search globally across the public feeds of Twitter, Facebook, and LinkedIn with the X1 Search engine’s fast-as-you-type feature. The product’s ability to scale makes it flexible enough to allow e-discovery pros to cope with a massive and rapidly expanding source of potential evidence.”

The broader point in both of these examples is that X1 Social Discovery is enabling technology that allows law firms, consultants, and other practitioners to transition from just talking about social media discovery to winning big league cases with cutting edge technology.

Relativity is a fully featured, web-based e-discovery platform for review, analysis, and production. Users can create and automate custom review workflows, make use of a complete set of text analytics capabilities including computer-assisted review, and can build applications within the software to manage and search all types of case-related data and information. In addition, users gain access to a growing application ecosystem, which includes integrations and applications built by third-party developers and the advice@kCura team to extend Relativity’s capabilities and allow case teams to do more with the software.

“As a Relativity Premium Hosting Partner with 5-years of experience selling and supporting Relativity in the United States, it is a natural progression to expand our reach into Canada in order to provide a truly North American solution,” says John Holland, CEO of D4.

The technology will be hosted and managed by D4 in a highly-secure Canadian IT infrastructure, while drawing on D4’s experience to ensure clients receive the highest level of service and performance from day one. This gives D4 a unique North American-wide footprint and greater scalability for their customers.

In This Issue

• Social Media eDiscovery and Review Tools by Peter Coons
• What in “TARnation” is going on at LegalTech New York?! by Tom Groom
• Judge Timothy Hillman and the Boston ALSP discuss Rule 26(f) ‘Meet and Confer’ Strategies by Chuck Kellner
• eDiscovery In the News and On the Web

If you were even barely conscious at LegalTech New York this past week, you got the message that Technology Assisted Review (TAR), also known as, “Predictive Coding”, has come of age. The litigation support industry considers this a settled matter, a done deal, an approach which, if you have not yet used it, or at least wrapped your brain around it, you are behind.

SUPPORT

At LegalTech, several panels have identified Technology Assisted Review as “disruptive”. Those of you who have worked with the TAR technologies and workflows on active cases know that the impact on the cost of discovery is truly a breakthrough.

Proponents include a number of federal district court judges and magistrates who have already weighed in on other important aspects of the eDiscovery process. Judge Andrew Peck of the Southern District of New York, previously wrote an essay for Law and Technology News promoting the use of Technology Assisted Review.

Proponents also include litigators on both small and large cases, from corporate law departments and law firms, from small and large legal enterprises, from both sides of the case captions, and from the government, as well as the private sector. If you were not at LegalTech, take a look at the Conference Schedule rosters for this track here.

ARGUMENT

There are many sound bites. “This is a game changer”, says Tess Blair, a partner and leader of the eData practice at Morgan Lewis. “This is the best tool I’ve found so far in containing the cost of discovery,” says Ralph Losey, eDiscovery Counsel at Jackson Lewis and veteran eDiscovery teacher and blogger.

Are you pressing a button and then just reviewing what the computer tells you to review? No, says Bob Trenchard, litigation partner and chair of the eDiscovery committee at WilmerHale. Those decisions are built by those who know the most about the case. In the SEC practice, the government will likely still ask for huge volumes of ESI for quick delivery. In IP, the patent work that is very complicated, this allows the most knowledgeable and specialized across an entire collection for discovery response.

What about discovery requests that are aimed at driving up the costs of the case? “There is a gamed process in many litigations, in which eDiscovery is being used as a sword and a threat,” says David Kessler, litigation partner at Fulbright & Jaworski. This technology helps level the field to help the client make good decisions about the data.

So now what? You’re no longer in the seminar room where proponents are preaching to the choir. You’re back in your office, a new case comes in, no matter what size, and you have choices.

If you have been at the leading edge without TAR, you’re going to slog through testing and revision of several iterations of search criteria before knowing how much you have to review, then you’re going to review your search hits, and then you’re going to test some of the non-hit collection to see what you left behind. You’re doing it this way because it is currently acceptable, and there’s not yet enough support for TAR or how to implement it.

If the experience and pronouncements of leading litigators and judges can’t carry the day, let’s look at how to promote the use of TAR. It might be helpful to break the discussion down as simply as possible to Technology, Workflow, Cost and Quality.

TECHNOLOGY

Without going into too much detail, the main difference between the old way and this way is the addition of a concept-based search engines. The essence of these is that they help you find subject matter beyond keyword search hits. They work in several different ways and may use different approaches and mathematical algorithms.

What is important to TAR is what they have in common: the ability to group together documents that have similar content or meaning. Many also have the ability to find documents that are subjectively, or conceptually, similar, even if they do not contain many, or any words in common. The use of concept-based search engines in TAR is to be able to propagate a document decision to large numbers of documents with similar content or meaning.

WORKFLOW

Like the iterative development of search terms, the TAR process contains both inductive and iterative workflow. No matter what software you’re using, the first part of the process requires that you evaluate some volume of documents for the attributes important to your project, e.g., responsive, privileged, and not responsive.

This initial coding needs to be done by people with the highest level of knowledge about what is important and relevant to the discovery response. Assisted by the technology described above, their coding is propagated “predictively” to parts of the collection that have not been reviewed.

Using methods of statistical sampling, the predictions are tested and refined, typically by a slightly larger team of reviewers, until the desired level of confidence is reached in thatthe collection is coded properly. The workflow includes quality assurance sampling of each attribute set: responsive, privileged, not responsive, etc.

COST

All this whizbang sounds expensive, but it’s NOT. In fact, it can save thousands to millions of dollars over even the best methods of review after keyword culling.

With a service provider, the premium to use TAR over the cost of traditional processing, search and hosting is minimal. In terms of historical eDiscovery industry pricing, the cost to deploy TAR technology and workflow is now about the same as costs for eDiscovery processing and hosting of about a year or two ago.

But the return on investment is massive. Proponents agree on their direct experience that savings have been huge. They have leveraged the knowledge of the people closest to the case and have propagated their review decisions across many times the number of documents actually reviewed. “The technology helps to drive economies of scale in small to medium-sized cases” says Kessler. “[TAR has] changed the way that we have priced our work. We can offer pricing in attractive ways to clients, flexibly and predictably for discovery, in ways that were not possible before,” says Blair.

QUALITY

Does quality suffer because you are not reviewing all the documents? At LegalTech’s “Man v. Machine” panel that included Ralph Losey, Judge Peck, and Maura Grossman from Wachtell Lipton, the panel pointed out that study after study proves that human review is not nearly as good as widely perceived and that both precision and recall are improved by the use of TAR technologies and workflows.

MOVING FORWARD

Judge Peck recalled on that panel a New Yorker comic by J.B. Handlesman, in which the lawyer says, “You have a pretty good case, Mr. Pitkin. How much justice can you afford?” to illustrate how eDiscovery has driven up litigation costs. All agree that technology-assisted review can and will continue to help level discovery to a more appropriate position in the process of leveling disputes. That panel quoted my favorite author, William Gibson, from an Economist article in December 2003, “The future is already here – it’s just not evenly distributed.”

“Predictive Coding” (future link to Recommind’s cease and desist letter), “Technology Assisted Review”, “Categorization”…whatever you want to call it…the new trend in e-discovery is trying to utilize technology that you’re probably already familiar with.

(Probably not as familiar as this guy, but this post should get you up to speed.)

Have you ever rated movies in Netflix?

How about Pandora? (I know quite a few attorneys who use this web radio service to help power them through their doc review.)

(The slightest thing can set anyone off after they’ve been engaged in 4 straight hours of doc review.)

And I’m guessing you’ve made a few purchases through Amazon? So then you’re familiar with the wave of recommendations that aren’t particularly relevant because they’re all based off a recent gift purchase.

Dear Amazon,

That Dave Barry book wasn’t for me. It was for my Grandfather. I understand that Mr. Barry’s a prolific writer (and I’m counting on that so I’ll have something to give my Grandfather next Christmas), but please do not send me emails every time Dave Barry publishes something.

Thanks.

Computer Scientists generally refer to these sorts of processes as “recommender systems”, “recommender platforms” or “recommender engines”, and the name says a lot about the intended purpose: they’re designed to make predictions and recommendations about books, movies, music, etc. based on previously stated “preferences” or “ratings”. If you like something, it tries to bring up similar instances of that thing.

So you can probably already see where this is going…the e-discovery world is now applying this same technology to the arduous process of document review (in the hopes of making it less arduous).

How does it work?

It basically works the same way Neflix, Pandora and Amazon work, but instead of assigning a one star rating to “Analyze This” so that you’ll never have to see Robert DeNiro in a comedic role ever again (at least in the world of your Netflix queue…Hollywood still seems to think it’s a good idea), you’re making responsiveness calls on a sample set of data from your review set.

So, as you begin to mark the sample set for responsiveness, the system learns your preferences and then attempts to make judgments about the remaining data set based on these initial responsiveness calls. The larger the sample set, the better informed the system becomes. (It’s like assigning one star to “Meet the Fockers” so that Netflix definitely understands you don’t want to see DeNiro in a comedy while assigning five stars to “Raging Bull”, “Goodfellas” and “Heat” so that the system understands that you still like other movies that DeNiro was in.)

But how does it know which documents are similar?

Good question. This is where we get a little more acquainted with the world of our friend:

(Hey! Check out this cool sine wave I drew on my TI-92.)

Scientists and mathematicians have devised methods that computer systems can use to think about documents and their content in terms of concepts and meaning instead of just literal content. For example, you’re probably familiar with employing text searches to find relevant documents. And if you are, you’re also probably familiar with the limitations of this approach: searching for a specific term, or set of terms, will often yield irrelevant data, or it will fail to capture relevant material simply because a relevant document didn’t contain the term, or terms, listed in the search.

Review platforms, like Relativity, have started incorporating this sophisticated technology to allow the user to organize documents around the concepts and meanings expressed within the content of a document. (Explaining how these technologies work would require a much, much longer conversation…). This technology is how the software can then extrapolate from your initial coding set and generate coding “recommendations”/”predictions” for the remaining set of un-reviewed data.

From here, you can devise workflows that will allow you and a review team to focus on the most relevant data set while placing the less relevant material aside for another time.

Ultimately, the purpose of this technology isn’t to completely remove attorneys from the equation, but rather expedite the review process and hopefully save everyone time and money.

Wait! What about the defensibility of all this?

I’ll address that in more detail in a future blog post, but there are signs of support amongst judges and courts for the use of these technologies. One of the clearest endorsements came in an article written by Judge Andrew Peck, United States Magistrate Judge for the Southern District of New York.

(“In his free time, Judge Peck is a member of the Baker Street Irregulars and other Sherlock Holmes societies.”)

“Until there is a judicial opinion approving (or even critiquing) the use of predictive coding, counsel will just have to rely on this article as a sign of judicial approval.”

From “Search, Forward: Will manual document review and keyword searches be replaced by computer-assisted coding?”, Law Technology News (October 2011)

So for Judge Peck, when it comes deciding on whether or not to use “predictive coding”, the answer is elementary*…

*I apologize to anyone who was offended by this cheap Sherlock Holmes joke.