INTRODUCTION
eDiscovery is the process of identifying, preserving, collecting,
reviewing, and producing to requesting parties electronically stored
information that is potentially relevant to a civil litigation or
regulatory inquiry. Of these activities, the review component is by
far the most expensive and time consuming [
One of the most eficient approaches to TAR in recent years
involves a combined human-machine (IA, or intelligence
amplification) approach known as Continuous Active Learning (CAL) [
Recently in the legal technology sector, another seeding approach that does not rely on human assessment of the review collection but is based on artificial intelligence (AI) methods and derived from documents outside the collection has been gaining momentum. For this technique, which is often referred to as “portable models”, and known in the wider machine learning community as transfer learning, initial seed documents are selected not via human input, but by predictions from a machine learning model trained using documents from prior maters or related datasets. Portable models take a pure AI approach and eschew human knowledge in the cold start seeding process.
Notwithstanding the benefits asserted by the proponents of portable models as a seed-generation technique, we are aware of no formal or even informal studies addressing the overall impact of portable model seeding on the eficiency of a TAR review relative to 2
Separate and apart from the inherent value of an assessment of the impact of portable models on TAR, there are two principles atendant to the creation of portable models that serve as a further motivation for this study: (1) the increased regulatory pressure to maintain personal privacy; and (2) the growing need for stringent cyber security measures. Consideration of both principals is generally recognized as an essential step in the development and utility of modern AI applications, given their breadth and proliferation.
Recent years have seen an increased scrutiny from EU and United
States regulatory agencies. Data collection and reuse is under heavy
examination as regulators seek to minimize data collection and
maximize privacy and security. Portable models are a form of data
reuse; the models would not exist were it not for the original data.
As such, there are rights and obligations around the use of the
data that goes in to training portable models, and a strong need for
clearer assessments of risk when porting models. As Bacon et al
noted [
Naturally, most companies that invest in building an ML model are looking for a return on their investment.
From a financial perspective, such companies focus on using the IP laws and related IP contract terms, such as IP assignments and license grants, to maximize their control over the ML model and associated input and results. Data protection laws can run counter to these objectives by imposing an array of requirements and restrictions on the processing of various types of data, particularly to the extent they include personal information. The interplay between these competing considerations can lead to interesting results, especially when a number of diferent parties have a stake in the outcome.
hTe second, perhaps more important challenge with respect to
portable models is the possibility of data leakage. In recent years,
computer security and machine learning researchers have increased
the sophistication of membership inference atacks [
chine learning model served on a [portable model] service. The atacker feeds each record into the model.
Based on the confidence score the model returns, the
atacker tunes the record’s features and reruns it by
the model. The process continues until the model
reaches a very high confidence score. At this point,
the record is identical or very similar to one of the
examples used to train the model. After gathering
enough high confidence records, the atacker uses the
dataset to train a set of “shadow models” to predict
whether a data record was part of the target model’s
training data. This creates an ensemble of models that
can train a membership inference atack model. The
final model can then predict whether a data record was
included in the training dataset of the target machine
learning model. The researchers found that this atack
was successful on many diferent machine learning
services and architectures. [
Carlini et al [
One such risk is the potential for models to leak details from the data on which they’re trained. While this may be a concern for all large language models, additional issues may arise if a model trained on private data were to be made publicly available. Because these datasets can be large (hundreds of gigabytes) and pull from a range of sources, they can sometimes contain sensitive data, including personally identifiable information (PII): names, phone numbers, addresses, etc., even if trained on public data. This raises the possibility that a model trained using such data could reflect some of these private details in its output.
Tramer et al [
Given the potential dangers associated with modern AI applications such as portable models, we therefore ask: Do portable models provided a cognizable sustained advantage over human augmented IA processes suficient to warrant their use in the face of privacy and cybersecurity concerns? If not, perhaps the safer and more appropriate approach is to continue using traditional human-driven techniques. 3
RELATED WORK
hTe key foundation in our investigation is the observation that the
current state-of-the-art document review TAR process is based on
continuous active learning (CAL) [
Review workflows that are based on CAL have what might be called a “just in time” approach to prediction. Rather than atempting to induce a perfect model up front, CAL workflows dynamically adjust as the review continues. Often this means that early disadvantages, and even early advantages, wash out in the process. For example [citation anonymized for review] found that four searchers each working independently to find seed documents found diferent and diferent numbers of seeds. But after separately using each seed set to initialize a CAL review, approximately the same number of documents needed to be reviewed to achieve high recall. This study asks similar questions in the context of portable models—whether there is a significant improvement in review eficiency when using portable models relative to traditional, non-AI techniques.
Another common portable model theme is the claim that the
more historical data they are trained on, the beter their predictions
will be. While that may be true in some instances, it may not be in
others. What constitutes privileged documents in one mater might
have a diferent set of characteristics as privileged documents in
others mater. What constitutes evidence of fraud, or sexual
harassment in one mater might be diferent than in other maters. No
amount of “big data” gathered from dozens (hundreds? thousands?)
of prior maters and composed into a monolithic portable model
may be relevant to the current problem if the paterns in the current
problem don’t match the historical ones. Therefore a question that
every eDiscovery practitioner should be asking herself is where
the best source of evidence for seeding the current task lies. As [
We engage three primary research questions. The first question level-sets the value of the pure AI (portable model) approach. The second two questions compare the portable model approach to a human-initiated process.
RQ1 Does a CAL review seeded by a portable model outperform (at high recall) one seeded randomly RQ2 Do portable models initially find more relevant documents than does human efort RQ3 Does a CAL review seeded by a portable model outperform (at high recall) one seeded by human efort
When atempting to consider these questions in general, issues naturally arise: What portable models are we talking about? Trained on what data? And how close was that data to the target distribution? And what humans seeded the comparison approach? And what was their prior knowledge of the subject mater? hTese questions mater, and while we cannot answer them for every possible training set and human searcher, we have structured the experiments in such a way as to give the most possible “benefit of the doubt” to the portable model, and the least possible benefit to the human searcher. Thus if there are significant advantages of portable models over human efort, these should be most readily apparent when portable models are given the most afordances and humans the least.
hTe primary manner in which portable models are given an advantage is that we train them on a set of documents that is drawn from the exact same distribution as the target collection to which they will be applied. In practice, portable models are never given this advantage. Prior cases in eDiscovery are not always exactly the same. Diferent collections, even from the same corporate entity, exhibit diferent distributions, especially as employees and business activities change and evolve over time. Naturally, the more diferent the source distribution, the less efective portable models will be when applied to a new target collection. However, by holding the distribution the same, this gives us an upper bound on portable model efectiveness and establishes a strong baseline against which the human efort can be compared.
At the same time, the human efort is minimized. As will be
described in more detail below, a small team of human searchers
worked for a collective total of approximately half an hour per
topic. None of the humans were experts in any of the topics, nor
did anyone have recent prior knowledge on the topics, as the events
in this Jeb Bush TREC collection [
hTus, our experiments consist of a comparison between portable models trained in the best possible light vs human efort that is kept at a minimum. We do this because the core concept of portable models is that they will be suficiently broad in scope so as to be able to identify relevant documents in a collection that contains documents of a similar content and context to those on which they were trained. (“Relevance” here refers to the notion of “what is desired”, be it some sort of topical similarity such as age discrimination or fraud cases, or something like privilege.) That distributional similarity is not always guaranteed, and in fact it can be dificult a priori to know whether you a portable model has been trained on data similar enough to be useful. By using documents intentionally drawn from the exact same distribution, we are able to show an upper bound on portable model efectiveness. In practice, portable
Total Rel Richness Queries
Minutes Total Docs
We test these research questions using the TREC 2016 total recall
track document collection, topics, and relevance judgments [
Human efort, aka manual seeding, was done with a small team of four searchers. For each topic, two of the searchers were instructed to run a single query and code the first 25 documents that resulted from that query. The other two searchers were given more interactive leeway and were instructed to utilize as many searches and whatever other analytic tools (clustering, timeline views, etc.) as they wanted, with a goal of working for about 15-30 minutes and stopping once they had tagged 25 documents. This was not strictly controlled, and some reviewers worked a few minutes longer, some a few minutes shorter. And some marked a few more than 25 documents, and some a few less, as is to be expected in normal, “in the moment” flow of knowledge work. Table 1 contains the manual efort statistics — with the total number of queries, total number of minutes, and total unique documents tagged as either relevant or non-relevant — for each topic. On average, the human reviewers worked for 31.4 minutes, issued 9.9 queries, and coded 64.3 documents so the overall efort was done at a fairly high pace and was relatively minimal in comparison to the size of the collection. 5.2
In order to compare portable models against both random and
human-seeded techniques in RQ1 through RQ3, there needs to be
a collection on which the portable model can be trained, separate
from the collection on which it and the comparative approaches
are deployed. We will refer to these two collections as “source” and
“target’, respectively. For the reasons enumerated in Section 4, we
carve out the portable model training source collection from the
same distribution as the target collection, and do so by selecting
documents at random. For a given topic, we:
(1) Shufle the collection randomly
(2) Split the collection into k groups
(3) For each group:
(a) Use that group as the portable model training “source”
collection S
(b) Train a model M using every document in S
(c) Use the remaining groups as the “target” collection T
(d) Select manual (human) seeds H by intersecting all found
docs (see Table 1) with T
(e) Select random seeds R from T until five positives
examples are found
(f) Selected portable seeds P from the top of the M-induced
ranking on T in an amount equal to jH j
(g) Use the appropriate seeds to run each experiment RQ1
through RQ3
(4) Average results across all k groups for the topic, but do not
average across topics
hTe specifics of step (3g) depends on the research question being
tested. For example, for RQ1, R and P are each (separately) used to
seed a continuous active learning (CAL) process. For RQ3, H and
P are used. Other than the diferent seedings, these CAL processes
are run exactly as in [
While selecting a source collection for portable model training from the same distribution as the target collection already ofers great advantage to the predictive capabilities of a portable model, i.e. puts it above where it would likely perform in more realistic scenarios, we extend this advantage even further by giving the model larger and larger source collections on which to train. We compare three primary source/target partitions: 20/80, 50/50, and 80/20, with k=5, k=2, and k=5, respectively. (In the 80/20 case, steps (3a) and (3c) are reversed, with the current group used as the target collection and the other groups used as the source collection.) The reason for the 20/80 partition is that the eDiscovery problem is a recall-oriented task. The larger the review population, aka the target collection, the more realistic the CAL process is likely to be. However, the disadvantage is that only 20% of the TREC collection is be used for training the portable model. The 80/20 partition reverses the balance: 80% of the collection is used to train the portable model, but only 20% of the collection is available to simulate the CAL review, which can be problematic for especially sparse topics. The 50/50 partition splits the diference.
Comment: An astute observer may find slight fault with the structure of this experimental setup, in that there is a small amount of knowledge overlap between the source and target partitions when doing human seeding. Specifically, the human searchers originally searched across the entire collection rather than across split collections. It is possible that a document found by a human searcher that ended up in a source partition has led the human to issue a query that found more or beter documents that ended up in the target partition. Thus even though the human-found documents in only the target partition are used to seed a CAL process (Step 3d, above), the existence of some of those seeds could have been influenced by knowledge of documents in the source partition. We note this issue and make it explicit, but do not think that it afects the overall conclusions of the experiment. One reason is that even if humans had some knowledge of documents in the source partition when finding the documents in the target partition, the portable model M is given knowledge of every document, positive and negative, in the source partition. Table 3 shows the raw number of positive documents used for training in the source partition, and it swamps the documents that the humans would have looked at in their short search sessions. Thus, one can think of any overlap during human seed selection as the background knowledge that human would likely already be expected to possess when working in a real scenario. E.g. an investigator working on detecting fraud or sexual harassment likely has some implicit background knowledge of fraud or sexual harassment. 6 6.1
RQ1: Portable- vs Random-Seeded Recall hTe results for our first question are found Table 2. Under the rubric of symmetry, the results are expressed in terms of raw percentage point (not percentage) diferences between the precision achieved at 80% recall for the portable model P-seeded review versus a random R-seeded review, and averaged across all partitions for each topic. Positive values indicate beter portable model performance; negative values the opposite. Nearly universally, P-seeding outperforms random seeding; the p-value under a binomial test is < 0.00001.
hTis is a wholly expected result. Closer examination of the simulated review orderings shows that in low richness domains most of the precision loss comes not from the CAL iterations, but from the larger number of documents needed to find enough positive ones to start ranking. Note that random seeding on the target partition also outperforms fully linear review by an average of 12.2, 10.6, and 5.6 percentage points on the 20/80, 50/50, and 80/20 partitions, respectively. So even random seeding of CAL is beter than no CAL at all. Topic 20/80 Partition 50/50 Partition precision 80/20 Partition
hTus in answer to the question: Does P-seeding produce an eficacious result, the answer is yes. However, the more important question is not whether portable models are useful; it is whether they are useful relative to other reasonable, simpler, or less risky alternatives. For that we turn to the remaining research questions. 6.2
hTe results for our second questions are found in Table 3 under the Target Portable and Target Manual columns for each partition group. For example, in the 20/80 partition, where 80% of the collection is used as the target collection and on average across all 5 folds, on topic 403 human efort found 18.4 positively-coded seed documents whereas the portable model M found 41.6 at the same level of efort (i.e. at 60 documents, as per Table 1). On topic 421 under the 20/80 partition, humans found an average 10.4 documents and M found 3.2.
hTe average number of positive training documents in the source partition, i.e. the data on which M is trained, is shown. The number of negative training examples is the remainder of the fold. Averages across all 34 topics are shown at the botom of the table, as is a binomial p-value.
hTese results show that when 20% of the collection is used to train M (20/80 parition), even though that data is literally from the same distribution as the target fold, the various M are able to find seed documents at a rate no beter than a small amount of human efort. There is only a diference of 0.2 documents across all topics, and while there is some variation between topics the diferences are not statistically significant (p=0.303). As the training partition increases, and 50% then 80% of the collection is used to train each M, so too does the ability of the model to find more seed documents. On the 50/50 partition M finds on average 3.5 more documents than the human at the given efort level, and on the 80/20 partition it finds an average 1.5 more documents. Both results are statistically significant.
When the number of seeds is normalized per fold and topic by the number of total seeds found, i.e. the positive seed precision, the manual efort has an average precision of 66.0% across all folds, whereas M precision is 64.2%, 76.5%, and 78.1%, respectively across 20/80, 50/50, and 80/20. That is, even though the average number of documents that M finds on the 50/50 partition is larger (3.5) than on the 80/20 partition (1.5), the later partition is smaller. The actual precision goes up slightly.
hTus in answer to the question: Do portable models initially find more relevant documents than does human efort, the answer is mixed. When given 20% of the collection for training, they do not. When given 50% or 80%, they do. However, the improvement is modest: a few percentage points, or a few extra documents. 6.3
hTe results for our third and final question are also found in Table 3 under the precision columns. Again, in the interest of symmetric magnitudes, precision is the percentage point diference between P-seeded versus H -seeded CAL. While again there is some variation across topics, on average on the 20/80 partition, P-seeding is 0.4 percentage points worse, while on the 50/50 and 80/20 partitions P-seeding is 0.6 and 1.2 percentage points beter. However, none of these results are statistically significant (p=0.303).
Furthermore, when we look at the raw document count diference between the two conditions (not shown in the table) another story emerges. In the 80/20 partition, on those topics for which Pseeded CAL is beter, is it beter on average by 186 total documents. Where H -seeded CAL is beter, it is beter by 896 documents. On the 50/50 partition, P- versus H -seeding is 271 vs 743 documents beter, and on the 20/80 partition it is 166 versus 655 documents. hTere is no consistent advantage of either approach over the other, but the negative consequences of H seeding seems to be far smaller than those of P seeding.
hTe reason for the topical sort order across all tables should now
become clear: All tables in this paper are sorted by the precision
of the 80/20 partition. This seems to be the partition for which
portable models are the strongest; they have the most training data.
And sorting by precision allows us to see where P-seeding vs
H -seeding each shine. To that end, we introduce one more metric
into the discussion: The WTF “inefectiveness” metric [
From this perspective, we see that where portable models perform the strongest, i.e. on the 80/20 partition where they are given 80% of the available positive documents, there are outliers in both directions. The top three P-advantage outliers show a 60.5, 12.0, and 5.7 percentage point diference. The top three H -advantage outliers show a 25.0, 12.6, and 11.1 percentage point diference. However, in terms of raw document counts these translate to a 670, 667, and 468 documents for the P-advantage, and 9249, 2006, and 892 documents for H -advantage. There appear to be fewer “WTFs” from H -seeding. 7
We have shown that a portable model can be useful. Certainly relative to linear review, and even relative to randomly seeded CAL workflows, taking a portable approach ofers a significant advantage. They are also marginally beter than humans when it comes to finding initial seed documents. When it comes to sustained advantage, i.e. precision at 80% recall, the advantages fade. There is no statistically significant diference in human vs portably seeded CAL workflows, and slight evidence that the outliers for the portable approach are worse.
We note also that porting models carries with it significant risk in the form of intellectual property rights, data leakage via membership inference atacks, privacy, and security. It is every party’s own subjective decision as to whether the advantages of portable models outweigh the challenges and risk. However, from the results in this study we would recommend continuing to invest in human-driven seeding (IA–intelligence augmentation) processes and not going all in on AI. At least relative to the topics studied in this paper, the modicum of efort required of the human are a fair trade relative to risk. Even when portable models are built on a corporation’s own data, and models are not swapped between diferent owners and therefore risk is lower, we do not yet find that the portable model provides a sustained advantage. 8
Certainly this is but one study and more studies with a wider range
of models, data collections, and human efort are needed. Perhaps
the humans could have done even beter if given more time, were
working on a domain in which they had specific expertise, or were
given more powerful analytics with which to find seed documents.
Conversely, portable models were given all possible advantages
in this experimental structure by building them on documents
drawn from the exact same distribution as the target collect, in
ever increasing amounts (20%, 50%, and 80%). It is not likely that
portable models will ever be trained on prior data as perfectly
similar to the target distribution. Therefore, portable models might
likely have performed much worse in realistic scenarios where the
source and target collections are further apart. E.g. when modeling
fraud or sexual harassment, does what constitute evidence of fraud
or harassment in one collection express itself the same way in
another collection? Future research is needed in three diferent
areas: (a) More and larger collections from similar but not identical
distributions on which to train models, or perhaps the “right” small
collections on which to train models, as per [
Beter and stronger baselines are not limited to more efective human efort. They also include other existing, common practices. For example, many companies dealing with sensitive information keep lexicons of search terms used to find sensitive information. While a lexicon could in some sense be thought of as an “unweighted” portable model, one diference is that it’s manually constructed, transparent, and can embed human intuition and paterns never seen in prior data, i.e. lexicons do not need to be trained. Another common approach for corporations with repeat litigation is the idea of a “drop in seed”. That is, instead of building large models based on huge datasets from all possible prior maters, some in the industry have developed the ad hoc practice of taking a few coded documents from previous maters, which maters are known to be similar to the current mater, and using those as the initial seeds on the target collection. This of course only works behind the firewall, as companies will not transfer documents to other companies. But given the security, privacy, and related membership inference atack risks of portable models, companies might not want to transfer their own models to a competitor, either. So in addition to comparing portable models against human seeding, they should be compared against lexicons and drop-in seeds. Perhaps these later approaches outperform both portable models and human-seeded approaches when considering the total cost of a review and not just the document count.
hTe cost of the portable model (vendor charge) versus the human approach (e.g. half an hour of searcher time) needs to be considered as well, and not just the cost of the subsequent document review. In short, this is a rich space for the exploration of tradeofs, risks, and advantages for various human-driven vs machine-driven eDiscovery processes.