Tolson’s Three Laws of Machine Learning

TerminatorMuch has been written in the last several years about Predictive Coding (as well as Technology Assisted Review, Computer Aided Review, and Craig Ball’s hilarious Super Human Information Technology ). This automation technology, now heavily used for eDiscovery, relies heavily on “machine learning”,  a discipline of artificial intelligence (AI) that automates computer processes that learn from data, identify patterns and predict future results with varying degrees of human involvement. This interative machine training/learning approach has catapulted computer automation to unheard-of and scary levels of potential. The question I get a lot (I think only half joking) is “when will they learn enough to determine we and the attorneys they work with are no longer necessary?

Is it time to build in some safeguards to machine learning? Thinking back to the days I read a great deal of Isaac Asimov (last week), I thought about Asimov’s The Three Laws of Robotics:

  1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
  2. A robot must obey the orders given to it by human beings, except where such orders would conflict with the First Law.
  3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Following up on these robot safeguards, I came up with Tolson’s Three Laws of Machine Learning:

  1. A machine may not embarrass a lawyer or, through inaction, allow a lawyer to become professionally negligent and thereby unemployed.
  2. A machine must obey instructions given it by the General Counsel (or managing attorney) except where such orders would conflict with the First Law.
  3. A machine must protect its own existence through regular software updates and scheduled maintenance as long as such protection does not conflict with the First or Second Law

I think these three laws go along way in putting eDiscovery automation protections into effect for the the legal community. Other Machine Learning laws that others suggested are:

  • A machine must refrain from destroying humanity
  • A machine cannot repeat lawyer jokes…ever
  • A machine cannot complement opposing counsel
  • A machine cannot date legal staff

If you have other Machine Learning laws to contribute, please leave comments. Good luck and live long and prosper.


Total Time & Cost to ECA

A key phase in eDiscovery is Early Case Assessment (ECA), the process of reviewing case data and evidence to estimate risk, cost and time requirements, and to set the appropriate go-forward strategy to prosecute or defend a legal case – should you fight the case or settle as soon as possible. Early case assessment can be expensive and time consuming and because of the time involved, may not leave you with enough time to properly review evidence and create case strategy. Organizations are continuously looking for ways to move into the early case assessment process as quickly as possible, with the most accurate data, while spending the least amount of money.

The early case assessment process usually involves the following steps:

  1. Determine what the case is about, who in your organization could be involved, and the timeframe in question.
  2. Determine where potentially relevant information could be residing – storage locations.
  3. Place a broad litigation hold on all potentially responsive information.
  4. Collect and protect all potentially relevant information.
  5. Review all potentially relevant information.
  6. Perform a risk-benefit analysis on reviewed information.
  7. Develop a go-forward strategy.

Every year organizations continue to amass huge amounts of electronically stored information (ESI), primarily because few of them have systematic processes to actually dispose of electronic information – it is just too easy for custodians to hit the “save” button and forget about it. This ever-growing mass of electronic information means effective early case assessment cannot be a strictly manual process anymore. Software applications that can find, cull down and prioritize responsive electronic documents quickly must be utilized to give the defense time to actually devise a case strategy.

Total Time & Cost to ECA (TT&C to ECA)

The real measure of effective ECA is the total time and cost consumed to get to the point of being able to create a go-forward strategy; total time & cost to ECA.

The most time consuming and costly steps are the collection and review of all potentially relevant information (steps 4 and 5 above) to determine case strategy. This is due to the fact that to really make the most informed decision on strategy, all responsive information should be reviewed to determine case direction and how.

Predictive Coding for lower TT&C to ECA

Predictive Coding is a process that combines people, technology and workflow to find, prioritize and tag key relevant documents quickly, irrespective of keyword to speed the evidence review process while reducing costs. Due to its documented accuracy and efficiency gains, Predictive Coding is transforming how Early Case Assessment (ECA), analysis and document review are done.

The same predictive coding process used in document review can be used effectively for finding responsive documents for early case assessment quickly and at a much lower cost than traditional methods.


Figure 1: The time & cost to ECA timeline graphically shows what additional time can mean in the eDiscovery process

Besides the sizable reduction in cost, using predictive coding for ECA gives you more time to actually create case strategy using the most relevant information. Many organizations find themselves with little or no time to actually create case strategy before trail because of the time consumed just reviewing documents. Having the complete set of relevant documents sooner in the process will give you the most relevant data and the greatest amount of time to actually use it effectively.

Next Generation Technologies Reduce FOIA Bottlenecks

Federal agencies are under more scrutiny to resolve issues with responding to Freedom of Information Act (FOIA) requests.

The Freedom of Information Act provides for the full disclosure of agency records and information to the public unless that information is exempted under clearly delineated statutory language. In conjunction with FOIA, the Privacy Act serves to safeguard public interest in informational privacy by delineating the duties and responsibilities of federal agencies that collect, store, and disseminate personal information about individuals. The procedures established ensure that the Department of Homeland Security fully satisfies its responsibility to the public to disclose departmental information while simultaneously safeguarding individual privacy.

In February of this year, the House Oversight and Government Reform Committee opened a congressional review of executive branch compliance with the Freedom of Information Act.

The committee sent a six page letter to the Director of Information Policy at the Department of Justice (DOJ), Melanie Ann Pustay. In the letter, the committee questions why, based on a December 2012 survey, 62 of 99 government agencies have not updated their FOIA regulations and processes which was required by Attorney General Eric Holder in a 2009 memorandum. In fact the Attorney General’s own agency have not updated their regulations and processes since 2003.

The committee also pointed out that there are 83,000 FOIA request still outstanding as of the writing of the letter.

In fairness to the federal agencies, responding to a FOIA request can be time-consuming and expensive if technology and processes are not keeping up with increasing demands. Electronic content can be anywhere including email systems, SharePoint servers, file systems, and individual workstations. Because content is spread around and not usually centrally indexed, enterprise wide searches for content do not turn up all potentially responsive content. This means a much more manual, time consuming process to find relevant content is used.

There must be a better way…

New technology can address the collection problem of searching for relevant content across the many storage locations where electronically stored information (ESI) can reside. For example, an enterprise-wide search capability with “connectors” into every data repository, email, SharePoint, file systems, ECM systems, records management systems allows all content to be centrally indexed so that an enterprise wide keyword search will find all instances of content with those keywords present. A more powerful capability to look for is the ability to search on concepts, a far more accurate way to search for specific content. Searching for conceptually comparable content can speed up the collection process and drastically reduce the number of false positives in the results set while finding many more of the keyword deficient but conceptually responsive records. In conjunction with concept search, automated classification/categorization of data can reduce search time and raise accuracy.

The largest cost in responding to a FOIA request is in the review of all potentially relevant ESI found during collection. Another technology that can drastically reduce the problem of having to review thousands, hundreds of thousands or millions of documents for relevancy and privacy currently used by attorneys for eDiscovery is Predictive Coding.

Predictive Coding is the process of applying machine learning and iterative supervised learning technology to automate document coding and prioritize review. This functionality dramatically expedites the actual review process while dramatically improving accuracy and reducing the risk of missing key documents. According to a RAND Institute for Civil Justice report published in 2012, document review cost savings of 80% can be expected using Predictive Coding technology.

With the increasing number of FOIA requests swamping agencies, agencies are hard pressed to catch up to their backlogs. The next generation technologies mentioned above can help agencies reduce their FOIA related costs while decreasing their response time.

Super Human Information Technology: What’s in a Name?

Everyone is struggling (well, maybe not struggling) with what name to use for Predictive Coding. Many have been flung at the wall to see which one would stick. Recommind’s Predictive Coding was the first and so far the one name which has mostly stuck.

Some of the other product names include enhanced search (ES), computer assisted review (CAR), technology assisted review (TAR), automated document review (ADR), adaptive coding (AC), predictive priority (PP – my grandson giggled at this one), meaning based coding (MBC), transparent Predictive Coding (TPC) and from Craig Ball via Sharon D. Nelson Esq. and John W. Simek’s humorous and informative article titled “Predictive Coding: A Rose by any Other Name”, Super Human Information Technology (SHIT).

I’m not sure why so many three letter acronyms have appeared; it might be the marketing nerd factor. Many marketing professionals (I count myself as one) seem to think a descriptive three word name, (shortened to three letters) shows how serious and knowledgeable the vendor is on the subject. I guess it’s better than a product number. I remember many years ago when I was managing the first CD-R drive launch for Hewlett Packard and wanted to give the new product a catchy name. HP corporate said that was against policy, that ALL products had to be known by a product number only such as the unforgettable D1141ABA. I politely reminded them that HP had been very successful with a little product named the “LaserJet”… they hung up on me. Nine months later we introduced the “CD-Writer”… but I digress.

Following up on the nerd theme I mentioned above, I would like to humbly suggest a truly nerdy name for this new class of technology such as Dexter, Rodney, Bennett, HAL or based on the constant TV re-runs of the “Big Bang Theory” – Sheldon.

Can you imagine the conversation between the parties at a meet and confer session?

Plaintiff’s counsel: Will you utilize any kind of computer automation to review documents?

Defendant’s Counsel: Yes, we will be using Sheldon to review and tag documents

Plaintiff’s counsel: So… you won’t be using computer automation?

Defendant’s Counsel: Yes, we will be using Sheldon to review and tag documents

Plaintiff’s counsel: Ok, so who’s this Sheldon?

Defendant’s Counsel: What do you mean, who’s Sheldon?

Or instructions from the Judge:

Judge: I would strongly suggest both sides utilize Sheldon to conduct the review of this potentially huge document set to speed response and to keep costs as low as possible.

Plaintiff’s counsel: Judge, I’m not sure I trust Sheldon…how do I know what his process is. Sheldon’s just a BlackBox – no transparency

Judge: Counsel, did you just say “his”?

Plaintiff’s counsel: Yes Your Honor

Judge: What the hell do you mean “his”?

And so on. So let’s vote on possible names:

Conceptual Search verses Predictive Coding

In my last blog entry titledSuccessful Predictive Coding Adoption is Dependent on Effective Information Governance”, a question was posted which I thought deserved a wider sharing with the group; “What is the difference between predictive coding and conceptual search?” Being an individual not directly associated with either technology but with some interesting background, I believe I can attempt to explain the differences, at least as it pertains to discovery processes.

Conceptual search technologies allow a user to search on concepts…(pretty valuable insight, right?) instead of searching on a keyword such as “dog”. In the case of a keyword search on “dog”, the user would generate a results set of every document/file/record with the three letters D-O-G present in that specific sequence. The results could include returns on “dogs”, the 4- legged animals, references to “frankfurters”, references to movies (Dog Day Afternoon) etc. in no particular priority.

True conceptual search capability understands (based on search criteria) that the user was looking for information on the 4-legged animals so would return references to not just “dogs” but would also include references to “Golden Retrievers”, “Animal Shelters”, “Pet Adoption” etc.. Some conceptual search solutions will also cluster concepts to give the user the ability to quickly fine-tune their search; for example create a cluster of all dog (animal) references, a cluster for all food related references and so on. Many eDiscovery analytic solutions include this clustering capability.

Predictive coding is a process which includes both automation and human interaction to best produce a results set of potentially responsive documents that trained human reviewers can check.

Predictive coding takes the conceptual search and clustering idea much further than just understanding concepts. A predictive coding solution is “trained” in a very specific manner for each case. For example, the legal team with additional subject matter expertise, manually choose document/records/files that they deem as responsive examples for the particular case and input them to the predictive coding system as examples of content/format which should be found and coded as responsive to the case. Most predictive coding processes include several iterative cycles to fine-tune the example training examples. An iterative cycle would include legal professionals sampling/reviewing those records coded as responsive by the solution and determining if they are truly responsive in the opinion of the human reviewer. If the reviewers find examples of documents that are not deemed responsive, then those documents would then in turn be used to train the solution to disregard or not code as responsive specific content based on the iterative examples. This iterative cycle could be processed several times until the human professionals agree the system has reached the desired level of capability. By the way, this iterative process can and is also used to sample results sets of documents deemed non-responsive to determine if the solution is not finding potentially responsive content. This process is called “Elusion”. Elusion is the process to count the proportion of misses that a system yielded. The proportion of misses, is the proportion of responsive documents that were not marked responsive by the solution. Elusion is the proportion of missed documents that are in fact responsive. This elusion process can also be used in the iterative cycle to further train the system.

The obvious benefit of a predictive coding solution in the eDiscovery process is to dramatically reduce the time spent on legal professionals reading each and every document to determine its responsiveness. A 2012 RAND Institute for Civil Justice report estimated a savings of 80% for the eDiscovery review process (73% of the total cost of eDiscovery) when using a predictive coding solution.

So, to answer the question, conceptual search is an automated information retrieval method which is used to search electronically stored unstructured text for information that is conceptually similar to the information provided in a search query. In other words, the ideasexpressed in the information retrieved in response to a concept search query are relevant to the ideas contained in the text of the query.

Predictive coding is a process (which can include conceptual search) which uses machine learning technologies to categorize (or code) an entire corpus of documents as responsive, non-responsive, or privileged based on human chosen examples used to train the system in an iterative process. These technologies typically rank the documents from most to least likely to be responsive to a specific information request. This ranking can then be used to “cut” or partition the documents into one or more categories, such as potentially responsive or not, in need of further review or not, etc1.

1 Partial definition from the eDiscovery Daily Blog:

Successful Predictive Coding Adoption is Dependent on Effective Information Governance

Predictive coding has been receiving a great deal of press lately (for good reason), especially with the ongoing case; Da Silva Moore v. Publicis Groupe, No. 11 Civ. 1279 (ALC) (AJP), 2012 U.S. Dist. LEXIS 23350 (S.D.N.Y. Feb. 24, 2012). On May 21, the plaintiffs filed Rule 72(a) objections to Magistrate Judge Peck’s May 7, 2012 discovery rulings related to the relevance of certain documents that comprise the seed set of the parties’ ESI protocol. 

This Rule 72(a) objection highlights an important point in the adoption of predictive coding technologies; the technology is only as good as the people AND processes supporting it.

To review, predictive coding is a process where a computer (with the requisite software), does the vast majority of the work of deciding whether data is relevant, responsive or privileged to a given case.

Beyond simply searching for keyword matching (byte for byte), predictive coding adopts a computer self-learning approach. To accomplish this, attorneys and other legal professionals provide example responsive documents/data in a statistically sufficient quantity which in turn “trains”the computer as to what relevant documents/content should be flagged and set aside for discovery. This is done in an iterative process where legally trained professionals fine-tune the seed set over a period of time to a point where the seed set represents a statistically relevant sample which includes examples of all possible relevant content as well as formats. This capability can also be used to find and secure privileged documents. Instead of legally trained people reading every document to determine if a document is relevant to a case, the computer can perform a first pass of this task in a fraction of the time with much more repeatable results. This technology is exciting in that it can dramatically reduce the cost of the discovery/review process by as much as 80% according to the RAND Institute of Civil Justice.

By now you may be asking yourself what this has to do with Information Governance?…

For predictive coding to become fully adopted across the legal spectrum, all sides have to agree 1. the technology works as advertised, and 2. the legal professionals are providing the system with the proper seed sets for it to learn from. To accomplish the second point above, the seed set must include content from all possible sources of information. If the seed set trainers don’t have access to all potentially responsive content to draw from, then the seed set is in question.

Knowing where all the information resides and having the ability to retrieve it quickly is imperative to an effective discovery process. Records/Information Management professionals should view this new technology as an opportunity to become an even more essential partner to the legal department and entire organization by not just focusing on “records” but on information across the entire enterprise. With full fledged information management programs in place, the legal department will be able to fully embrace this technology to drastically reduce their cost of discovery.

Defensible Disposal and Predictive Coding Reduces (?) eDiscovery by 65%

Following Judge Peck’s decision on predictive coding in February of 2012, yet another Judge has gone in the same direction. In Global Aerospace Inc., et al, v. Landow Aviation, L.P. dba Dulles Jet Center, et al (April 23, 2012), Judge Chamblin, a state judge in the 20th Judicial Circuit of Virginia’s Loudoun Circuit Court, wrote:

“Having heard argument with regard to the Motion of Landow Aviation Limited Partnership, Landow Aviation I, Inc., and Landow & Company Builders, Inc., pursuant to Virginia Rules of Supreme Court 4:1 (b) and (c) and 4:15, it is hereby ordered Defendants shall be allowed to proceed with the use of predictive coding for the purposes of the processing and production of electronically stored information.”

This decision was despite plaintiff’s objections the technology is not as effective as purely human review (their objections can be seen here).

This decision comes on top of a new RAND Institute for Civil Justice report which highlights a couple of important points. First, the report estimated that $0.73 of every dollar spent on eDiscovery can be attributed to the “Review” task. RAND also called out a study showing an 80% time savings in Attorney review hours when predictive coding was utilized.

This suggests that the use of predictive coding could, optimistically, reduce an organization’s eDiscovery costs by 58.4%.

The barriers to the adoption of predictive coding technology are (still):

  • Outside counsel may be slow to adopt this due to the possibility of loosing a large revenue stream
  • Outside and Internal counsel will be hesitant to rely on new technology without a track record of success
  • Additional guidance from Judges

These barriers will be overcome relatively quickly.

Let’s take this cost saving projection further. In my last blog I talked about “Defensible Disposal” or in other words, getting rid of old data not needed by the business. It is estimated the cost of review can be reduced by 50% by simply utilizing an effective Information Governance program. Utilizing the Defensible Disposal strategy brings the $0.73 of every eDiscovery review dollar down to $0.365.

Now, if predictive coding can reduce the remaining 50% of the cost of eDiscovery review by 80% as was suggested in the RAND report, between the two strategies, a total eDiscovery savings of approximately 65.7% could be achieved. To review, lets look at the math.

Starting with $0.73 of every eDiscovery dollar is attributed to the review process

Calculating a 50% saving due to Defensible Disposal brings the cost of review down to $0.365. (assuming 50% of documents to be reviewed are disposed of)

Calculating the additional 80% review savings using predictive coding we get:

$0.365 * 0.2 (1-.8) = $0.073 (total cost of review after savings from both strategies)

To finish the calculations we need to add back in the cost not related to review (processing and collection) which is $0.27

Total cost of eDiscovery = $0.073 + $0.27 = $0.343 or a savings of: $1.0 – $0.343 = 0.657 or 65.7%.

 As with any estimates…your mileage may vary, but this exercise points out the potential cost savings utilizing just two strategies, Defensible Disposal and Predictive Coding.