TREC and the gold standard for document review

Ron Friedman recently blogged an excellent critique of TREC Legal Track’s effort to objectively assess eDiscovery document review practices. Like Ron, I commend TREC Legal Track while wishing to offer comments that may contribute to their success. Like me, Ron is an attorney with long experience working in the technology sector, although for comparison with his math background I can only claim four years of college courses concerning statistical methods for assessing human behavior.

Benchmarking is valuable almost everywhere.

Benchmarking is valuable almost everywhere.

I strongly recommend reading Ron’s post for the benefit of his insights, whether or not you are already familiar with TREC Legal Track. I’d also like to offer my own observations about TREC Legal Track’s finding of low consistency between document classification decisions made by subject matter experts, who are spoken of as “gold standard” reviewers, and ordinary legal document reviewers. (In TREC Legal Track’s study, ordinary reviewers were 2nd and 3rd year law students. In real life the subject matter expert role is played by in-house or outside counsel, while much of the actual review work is performed by contract or outsource attorneys.)

Generally speaking, quality control processes involve benchmarking against some standard. Mechanical processes can be meaningfully benchmarked by physically sampling output (this is the essence of Six Sigma, in particular). For example, as machine parts come off an assembly line, samples can be selected and measured and the variance between their actual size and target size monitored not only to detect defects but to flag the processes responsible for defects. Human processes can also be benchmarked in a variety of ways. (This is in part the province of ITIL, the “Information Technology Information Library,” and the basis for the idea of “service level agreements”.) For example, those responsible for a customer service center may track the number of issues handled per hour, the type of issues handled, the number of resolutions or escalations per issue, revenue gained or lost per issue, etc.

Unfortunately, “responsiveness” and “privilege” are not only somewhat subjective in document review, standards for responsiveness and privilege will vary from case to case. For this reason standards need to be developed “on the fly” for each case, and these standards will by necessity be arbitrary (aka subjective) to some degree even if consistently applied. The good news is that the latest generation of document clustering software applications incorporate tools for developing consistent document review standards on the fly. Through an iterative feedback loop, the humans educate the machines to look for documents with certain characteristics, while the machines force the humans to refine their conception of responsiveness and privilege to a degree that the machine can reliably model it. After enough iterations have passed and the machine has reached some measurable standard of consistency, the humans can step back and let the machine do the rest of the review work. The machine does it more consistently than human reviewers could themselves, and at a much lower cost.

With document review the very idea of defining a “gold standard” for classification is less useful than it sounds. For instance, even if a panel of leading legal scholars could be formed for each eDiscovery matter, the mere fact that someone legitimately may be called a leading scholar doesn’t mean that their views will be consistent with anyone else’s — just well reasoned. But a “gold standard” is not what’s important here. What’s important is that in each case the attorneys responsible for responding to a document request do everything they can to carefully define and consistently enforce reasonable document review standards. This is what the current crop of document clustering applications are intended to do. That is the current model, anyway. I don’t pretend to be able to name the vendors who can or cannot deliver on this promise, although I think this will be the number one question in eDiscovery technology before long.

UPDATE: I discuss TREC’s role in forumulating new legal procedural rules for e-discovery in a later blog post, Catch-22 for e-discovery standards?

Reusing document clustering categories to spend less on eDiscovery?

After drafting a blog post about mass data sampling and classification in the “cloud,” I became curious about the potential for reusing categories developed in eDiscovery sampling and classification projects as “seeds” for later projects. For further insight I turned to Richard Turner, Vice President of Marketing at Content Analyst Company, LLC, a document clustering and review provider for eDiscovery.

schl¸sselBruce: I wonder to what extent document categories that are created using document clustering software when reviewing documents for eDiscovery can be aggregated across multiple document requests and/or law suits within the same company. Can previously developed categories or tags be reused to seed and thus speed up document review in other cases?

Richard: Regarding the notion of aggregating document categories, etc., it’s something that’s technically very feasible. And it could greatly speed document review if categories could be used to “seed” new reviews, new cases, etc. Here’s the challenge: we have found that most of the “categories” developed by our clients start-out case specific, and are too granular to be valuable when the next case comes along. It also hasn’t seemed to matter whether categorization was being used by a corporate legal department or an outside counsel – they’re equally specific.

The idea itself had merit, so we tossed it around with our Product Solutions Architects, and they came up with several observations. First of all, the categories people develop are driven by their need to solve a specific eDiscovery challenge, i.e. documents that are responsive to the case at hand. Second, when the next issue or case comes along, they naturally start over again, first by identifying responsive documents and then by using those documents to create categories – any “overlap” is purely coincidental. Finally, to develop categories that were really useful across a variety of issues or cases, they would need to be fairly generic and probably not developed with any specific case in mind.

I think that’s very hard to do for a first or even second-level review – it’s not necessarily a natural progression, as people work backwards from the issues at hand. Privilege review, however, could be a different animal. There are some things in any case that invoke privilege because of the particulars of the case, for example, attorney-client conversations which are likely to involve different individuals in different litigation matters. There are other things that could logically be generic – company “trade secrets” for example would almost always be treated as privilege, as are certain normally-redacted items such as PII (personally-identifiable information). Privilege review is also a very expensive aspect for eDiscovery, since it involves physical “reads” using highly-paid attorneys (not something you can comfortably offshore). Could “cloud seeding” have value for this aspect of eDiscovery? It’s an interesting thought.

Cloud-seeding: SaaS data classification via Panda Security’s new anti-virus offering

Panda Security recently released (in beta form) what it claims is the first cloud-based anti-virus / anti-malware solution for Windows PCs. Not only does it sound like a clever tool for data loss prevention, but it demonstrates another way in which information service providers can aggregate individual user data to develop classifications or benchmarks valuable to every user, a mechanism I’ve explored in previous blog posts.

In essence, every computer using Panda’s Cloud Antivirus is networked together through Panda’s server to form a “collective intelligence” for malware detection and prevention. Here’s how it works: users download and install Panda’s software – it’s a small application known as an “agent” because the heavy lifting is done on Panda’s server. These agents send reports back to the Panda server containing information about new files (and, I presume, related computer activity which might indicate the presence of malware). When the server receives reports about previously unknown files which resemble, according to the logic of the classification engine, files already known to be malware, these new files are classified as threats without waiting for manual review by human security experts.

Security Camera
Sampling at the right time and place allows proactive decision making.

For example, imagine a new virus is released onto the net by its creators. People surfing the net, opening emails, and inserting digital media start downloading this new file, which can’t be identified as a virus by traditional anti-virus software because it hasn’t been placed in any virus definitions list yet. Computers on which the Panda agent has been installed begin sending reports about the new file back to the Panda server. After some number of reports about the file are received by Panda’s server, the server is able to determine that the new file should be treated as a virus. At this point all computers in the Panda customer network are preemptively warned about the virus, even though it has only just appeared.

According to Panda’s April 29, 2009 press release:

Utilizing Panda’s proprietary cloud computing technology called Collective Intelligence, Panda Cloud Antivirus harnesses the knowledge of Panda’s global community of millions of users to automatically identify and classify new malware strains in almost real-time. Each new file received by Collective Intelligence is automatically classified in under six minutes. Collective Intelligence servers automatically receive and classify over 50,000 new samples every day. In addition, Panda’s Collective Intelligence system correlates malware information data collected from each PC to continually improve protection for the community of users.

Because Panda’s solution is cloud-based and free to consumers, it will reside on a large number of different computers and networks worldwide. This is how Panda’s cloud solution is able to fill a dual role as both sampling and classification engine for virus activity. On the one hand Panda serves as manager of a communal knowledge pool that benefits all consumers participating in the free service. On the other hand, Panda can sell the malware detection knowledge it gains to corporate customers – wherein lies the revenue model that pays for the free service.

I have friends working in two unrelated startups, one concerning business financial data and the other Enterprise application deployment ROI, that both work along similar lines (although neither are free to consumers). Both startups offer a combination of analytics for each customer’s data plus access to benchmarks established by anonymously aggregating data across customers.

Panda’s cloud analytics, aggregation and classification mechanism is also analogous to the non-boolean document categorization software for eDiscovery discussed in previous posts in this blog, whereby unreviewed documents can be automatically (and thus inexpensively) classified for responsiveness and privilege:

Deeper, even more powerful extensions of this principle are also possible. I anticipate that we will soon see software which will automatically classify all of an organization’s documents as they are created or received, including documents residing on employees laptop and mobile devices. Using Panda-like classification logic, new documents will be classified accurately whether or not they are of an exact match with anything previously known to the classification system. This will substantially improve implementation speed and accuracy for search, access control and collaboration, document deletion and preservation, end point protection, storage tiering, and all other IT, legal and business information management policies.