1
Explainable Text Classification in Legal Document Review
A Case Study of Explainable Predictive Coding
Rishi Chhatwal, Esq.
Legal
AT&T Services, Inc.
Washington DC, USA
Peter Gronvall
Data & Technology
Ankura Consulting Group, LLC
Washington DC, USA
Nathaniel Huber-Fliflet
Data & Technology
Ankura Consulting Group, LLC
Washington DC, USA
Robert Keeling, Esq.
Complex Commercial Litigation
Sidley Austin LLP
Washington DC, USA
Dr. Jianping Zhang
Data & Technology
Ankura Consulting Group, LLC
Washington DC, USA
Dr. Haozhen Zhao
Data & Technology
Ankura Consulting Group, LLC
Washington DC, USA
Abstract— In today’s legal environment, lawsuits and
regulatory investigations require companies to embark upon
increasingly intensive data-focused engagements to identify,
collect and analyze large quantities of data. When documents
are staged for review – where they are typically assessed for
relevancy or privilege – the process can require companies to
dedicate an extraordinary level of resources, both with respect
to human resources, but also with respect to the use of
technology-based techniques to intelligently sift through data.
Companies regularly spend millions of dollars producing
‘responsive’ electronically-stored documents for these types of
matters. For several years, attorneys have been using a variety
of tools to conduct this exercise, and most recently, they are
accepting the use of machine learning techniques like text
classification (referred to as predictive coding in the legal
industry) to efficiently cull massive volumes of data to identify
responsive documents for use in these matters. In recent years,
a group of AI and Machine Learning researchers have been
actively researching Explainable AI. In an explainable AI
system, actions or decisions are human understandable. In
typical legal ‘document review’ scenarios, a document can be
identified as responsive, as long as one or more of the text
snippets (small passages of text) in a document are deemed
responsive. In these scenarios, if predictive coding can be used
to locate these responsive snippets, then attorneys could easily
evaluate the model’s document classification decision. When
deployed with defined and explainable results, predictive
coding can drastically enhance the overall quality and speed of
the document review process by reducing the time it takes to
review documents. Moreover, explainable predictive coding
provides lawyers with greater confidence in the results of that
supervised learning task. The authors of this paper propose the
concept of explainable predictive coding and simple
explainable predictive coding methods to locate responsive
snippets within responsive documents. We also report our
preliminary experimental results using the data from an actual
legal matter that entailed this type of document review. The
purpose of this paper is to demonstrate the feasibility of
explainable predictive coding in the context of professional
services in the legal space.
Keywords- machine learning, text categorization, explainable
AI, predictive coding, explainable predictive coding, legal
document review
I. INTRODUCTION
In modern litigation, attorneys often face an
overwhelming number of documents that must be reviewed
and produced over the course of the matter. The costs
involved in manually reviewing these documents has grown
dramatically as more and more information is stored
electronically. As a result, the document review process can
require an extraordinary dedication of resources: companies
routinely spend millions of dollars sifting through and
producing responsive electronically stored documents in
legal matters [1].
Attorneys have responded to the exponential growth of
documents by employing machine learning techniques like
text classification to efficiently cull massive volumes of data
to identify responsive information. In the legal domain, text
classification is typically referred to as predictive coding or
technology assisted review (TAR). Predictive coding applies
a supervised machine learning algorithm to build a predictive
model that automatically classifies documents into
predefined categories of interest. Attorneys typically employ
predictive coding to identify so-called “responsive”
documents, which are materials that fall within the scope of
some ‘compulsory process’ request, such as discovery
requests, subpoenas, or internal investigations.
Over the past few years, the authors of this paper have
helped their clients reduce the cost of document review by
millions of dollars across dozens of matters using predictive
coding. These cost savings examples are not uncommon in
legal matters, and in fact the cost-savings aspect to this
technology has helped drive its adoption across the legal
community. While predictive coding is regularly used to
reduce the discovery costs of legal matters, it also faces a
2
perception challenge: amongst lawyers, this technology is
sometimes looked upon as a “black box.” Put simply, many
attorneys do not understand the underlying technology used
to generate predictive models and to rank documents as
responsive to a document request.
In 2016 this group of authors performed a study to
address that challenge; we set out to share our understanding
of the underlying technology with the legal community,
through an illustrative research paper [2]. This research
began to bridge the gap between academic research and the
legal technology industry. Our view was that it is imperative
that attorneys can reference peer reviewed research that
thoroughly examines and measures the performance of the
fundamental components of commonly used text analytics
techniques. Without this knowledge, a legal team could find
itself spending far too much time and money on a document
review, to their client’s detriment, by ineffectively
implementing a predictive coding review protocol. A sound
understanding of predictive coding’s machine learning
algorithms and preprocessing parameters places attorneys in
a better position to control the costs and impact that it can
have on a legal matter.
Legal technology experts continue to perform research to
make predictive coding less of a “black box” in the minds of
the legal community, but there is still more work to do [2].
The research in this study addresses a still misunderstood
component of the predictive coding process: explaining how
responsive documents are classified. Attorneys typically
want to know why certain documents were determined to be
responsive by the model, but sometimes those answers are
not obvious or easily divined. In many instances, for
example, a predictive model could be inaccurate, resulting in
the model flagging irrelevant documents as highly likely to
be responsive. This can be confusing for an attorney,
especially if the text content of the document doesn’t appear
to contain obviously responsive content. An attorney with
extensive knowledge of the training documents can make an
educated guess as to why a document was highly-scored, but
it can be difficult to pinpoint exactly what text in a document
heavily influenced the high score. The focus of our research
in this paper was to develop experiments that could target
and examine the document text that the predictive model
identified and used to generate the high score.
The Artificial Intelligence (AI) community has been
researching explainable Artificial Intelligence since the
1970s, e.g. medical expert system MYCIN [3]. More
recently, DARPA proposed a new direction for furthering
research into Explainable AI (XAI) [4]. In XAI systems,
actions or decisions are human understandable – “machines
understand the context and environment in which they
operate, and over time build underlying explanatory models
that allow them to characterize real world phenomena.”
Similarly, in an explainable machine learning system,
predictions or classifications generated from a predictive
model are explainable and human understandable.
Interpreting the decision of the machine is more important
now than ever because, increasingly, machine learning
systems are being used in human decision-making
applications and machine learning algorithms are becoming
more complex.
Understanding a model’s classification decision is
challenging in text classification because of the factors a
model considers during the decision-making process,
including word volume within text-based documents, and the
volume and diversity of tokens established during the text
classification process. In the legal domain, where documents
can range from one-page emails to spreadsheets that are
thousands of pages long, the complexity of the models
creates challenges for attorneys to pinpoint where the
classification decision was made within a document.
We find that the easiest way to understand a model’s
document classification decision is by evaluating small text
snippets and identifying the ones that support the
classification under examination. For the purposes of this
paper, we considered a text snippet to be a small passage of
words within a document usually ranging from 50 to 200
words. In legal document review, a document can be
considered responsive when one or more of its text snippets
are responsive and contain relevant information. Therefore,
if predictive coding could locate these responsive snippets,
attorneys could easily evaluate the model’s document
classification decision. In this scenario, it would be simple to
explain why the model made its classification decision and
this would help further minimize the black box nature of
predictive coding. In addition to creating an explainable
result, explainable predictive coding could enhance the
overall document review process by reducing the time to
review documents. Consider a scenario where attorneys only
need to focus on the review of a responsive text snippet and
not the entire text of a document – this would significantly
speed up the review process and decrease the cost of the
review.
Explainable predictive coding is well-suited for the legal
document review process. It could also help improve
investigative scenarios by quickly pinpointing potentially
sensitive responsive content and enable a quick summary
review of a small number of high-scoring documents
snippets. Quickly understanding the content in a data set
equips attorneys with the information they need to
effectively represent their clients.
This paper proposes the concept of explainable predictive
coding and simple explainable predictive modeling methods
as an effective means to locate responsive snippets within a
responsive document. We report our preliminary
experimental results using the data from an actual legal
matter, now concluded. The purpose of the paper is to
demonstrate the feasibility of explainable predictive coding.
3
Section I introduces this concept. Section II discusses
previous work in explainable text classification. We
introduce our explainable predictive modeling method in
Section III and report the experimental results in Section IV.
We summarize our findings and conclude in Section V.
II. PREVIOUS WORK IN EXPLAINABLE TEXT
CLASSIFICATION
Research in Explainable AI has focused on two main
areas of explainable machine learning: model-based
explanations and prediction-based explanations. In a model-
based approach, models are inherently explainable, and these
types of models use decisions trees and If-Then rules to
explain the results. Complex models, such as deep learning
models like multilayer neural network models, non-linear
SVM, or ensemble models are not directly human
understandable and require implementing a more
sophisticated approach to interpret the decision. With
complex models, the proxy model approach is applied to
create an explainable model that approximates the
predictions of a given complex model [5].
An alternative to generating explainable models is to
produce an explanation for each individual prediction
generated by a complex model. Generally, a prediction-based
explanation method provides an explanation as a vector with
real-valued weights, each for an independent variable
(feature), indicating the extent to which it contributes to the
classification. This approach is not ideal for text
classification, because of the high dimensionality in the
feature space. In text classification, a document belongs to a
category, mostly likely because some parts of the text in the
document support the classification. Therefore, a small
portion of the document text is often used as evidence to
justify the classification result in text classification.
Predictive coding is an application of text classification
to documents falling within the scope of a legal matter. Text
data (documents) are often represented using the bag-of-
words approach and characterized with tens of thousands of
variables (words and/or phrases). Due to the high
dimensionality, understanding the classifications made by
text categorization models is very difficult and creates an
interesting research opportunity. Recent research found that a
prediction-based approach is often used to identify snippets
of text as an explanation for the classification of a document.
A text snippet that explains the classification of a document
is called a ‘rationale’ for the document in [6].
Zaidan, et al. [6] proposed a machine learning method to
use annotated rationales in documents to boost text
classification performance. In their method, in addition to
full document class labels, human document reviewers also
highlighted the text snippets that explained why the
corresponding document belonged to the class. The labeled
documents together with their annotated text snippets were
used as training data to build a text classification model
using SVM. Their experiments showed that classification
performance significantly improved with annotated
rationales over the baseline SVM variants using an entire
document’s text.
Zhang, et al. [7] presented a Convolutional Neural
Network (CNN) model for text classification that jointly
employed labels on both documents and their constituent
sentences. Specifically, they considered scenarios in which
reviewers explicitly labeled sentences that support their
overall document classification. Their method employed a
two-level learning approach in which each document was
represented by a linear combination of the vector
representations of its component sentences. In the first level,
a sentence-level convolutional model is built to estimate the
probability that a given sentence supports a document-level
classification. Then, in the second level, the CNN model
leveraged the contribution of each sentence to the aggregate
document representation in proportion to these probability
estimates. Zhang, et al.’s approach was applied to five data
sets that had document-level labels and the requisite
sentence-level labels. Their CNN model experiments
demonstrated that their approach consistently outperforms
strong baselines. Moreover, their approach naturally provides
explanations for the model’s predictions because each
sentence is evaluated for its probability of supporting the
document-level classification.
In [8], Martens and Provost described a method in which
the explanation of a document classification is a minimal set
of the most relevant words, such that removing all the words
in the set from the document would change the classification
of the document. An algorithm for finding this minimal set
of words was presented and they conducted case studies
demonstrating the value of the method using two sets of
document corpora.
Lei et al. [9] proposed an approach to generate rationales
for text classifications. Their approach combined two
components, a rationale generator and a rationale encoder,
which were trained to operate together. The generator
specified a distribution over text fragments (text snippets) as
candidate rationales and the encoder decided the
classifications of candidate rationales established by the
generator. The proposed approach was evaluated on multi-
aspect sentiment analysis against manually annotated test
cases. The results showed that their approach outperformed
the baseline by a significant margin. The approach was also
successfully applied to a question retrieval task.
III. EXPLAINABLE PREDICTIVE CODING
The main goal of the explainable predictive coding
concept introduced in this paper is not to revisit underlying
enhanced predictive model performance metrics, such as
precision and recall. Rather, the goal is to provide additional
information (explanations) for the labeling of each
responsive document and to help attorneys more effectively
and efficiently identify responsive documents during legal
document review. As with many other explainable text
4
classifications approaches, we use the prediction-based
approach instead of the model-based approach. Also, we are
interested only in generating explanations for responsive
documents, therefore we focus on documents identified as
responsive.
Explainable predictive coding sets out to build a model to
estimate Pr(r = Rationale | x
，
y = Responsive) · Pr(y =
Responsive | x), where x is a document, y is the model-
labeled designation of the document (‘responsive’ or ‘not
responsive,’ for instance), and r is a text snippet from x. The
task consists of two separate subtasks:
• The first subtask is a traditional text classification
task where a text classification model is built to
identify responsive documents.
• The second subtask is to generate a text
classification model which will be used to identify
responsive rationales in each responsive document.
And it is the identification of the rationale that
underlies the quest for the ‘explainability’ that this
study seeks to develop.
Table 1 outlines the process for identifying rationales of
responsive documents. In the following subsections, we
describe two approaches in building models for identifying
rationales of responsive documents in detail. The first
approach uses the document model (trained using labeled
documents) to identify rationales, while the second one
builds a rationale model using a set of manually annotated
rationales identified by attorneys.
Table 1: The Process for Identifying Rationales
1. Train a Document Model
2. Train a Rationale Model
3. Use the Document Model to identify responsive
documents
4. Break each responsive document into a set of
overlapping text snippets with n words
5. Apply the Rationale Model to score all text snippets
of responsive documents
6. For each responsive document, select the top n
scored snippets as the identified rationale(s)
A. Document Model Method
In this section, we introduce a simple rationale generation
method. In this method, we used a set of training documents,
each of which was labeled as either responsive or not
responsive by attorneys, and also a set of documents to be
classified. As mentioned above, the approach consists of two
phases. In the first phase, as is typical in traditional
predictive coding, a model is generated using the set of
training documents and the model is deployed to identify a
set of potentially responsive documents, using the model’s
predictions. In the second phase, the same model is applied
to generate one or more rationales for each of the identified
responsive documents.
To generate the rationales for a responsive document, we
break the document into a set of small overlapping text
snippets. We then apply the document model to classify
these snippets, on the spectrum of highly likely to be
responsive, down to not responsive. As such, the model
assigns a probability score between 0 and 1 to each text
snippet. The n text snippets with the highest scores are
selected as the rationales for the classifications under
examination.
Determining the optimal size of the text snippets is a
difficult task because the size of the rationales is unknown in
advance of generating the rationale. To solve this problem,
we deployed an iterative approach to break documents into a
set of overlapping text snippets, as follows: we break a
document into relatively large snippets – for snippets
receiving large probability scores, we continue to break them
into smaller sizes. This process continues until probability
scores stop increasing.
B. Rationale Model Method
In predictive coding, training documents are manually
reviewed by attorneys and are coded as responsive or not
responsive. For this study, when a training document was
coded as responsive, we asked the coding attorney to indicate
why it was responsive, and to so indicate by highlighting the
text snippets that supported his/her coding decision. These
highlighted snippets – much like a training documents –
became the manually annotated rationales.
Similar to the document model method, a traditional
predictive coding model is trained in the first phase to
identify responsive documents. In the second phase, a
rationale model is trained to generate one or more rationales
for each identified responsive document. In training a
rationale model, we use annotated rationales of responsive
documents as responsive training rationales and a set of
randomly generated text snippets from non-responsive
documents as not responsive training rationales. To generate
rationales for a responsive document, we first break the
document into a set of overlapping text snippets and then
apply the rationale model to determine which of the text
snippets are rationales.
IV. EXPERIMENTS
In this section, we report our results on a large data set
from a real legal document review. We describe the data set
and the design of the experiments in Sections IV.A and IV.B.
Results are reported in Section IV.C.
A. The Data Set
The data set was collected from an actual legal matter,
now concluded, that contains documents including email,
Microsoft Office documents, PDFs, and other text-based
documents. It consists of 688,294 documents manually
5
coded by attorneys as responsive or not responsive. Among
the 688,294 documents, 41,739 are responsive and the rest
are not responsive. For each of the responsive documents, a
rationale was annotated by a review attorney as the
justification for coding the document as responsive. In
practical terms, most rationales are continuous words,
phrases, sentences or sections from the reviewed and labeled
documents. A few rationales contain words that are
comments from the attorney and do not occur in the
documents. Some rationales may consist of more than one
text snippet, which occur in different parts of the document.
Annotated rationales have a mean length of 52 words,
with a standard deviation of 112.5 words. 97.5% of these
rationales have fewer than 250 words. To reduce the effect of
outliers, such as very long or very short rationales, we limit
our rationales to those with 10 or more words but fewer than
250 words and those that can be precisely identified in the
data set – resulting in 23,791 responsive documents with
annotated rationales in our population. These 23,791
documents established our responsive population, covering
57% of all the responsive documents in the above 688,294
population. Proportionally, we randomly selected 365,742
documents from the not responsive documents within the
688,294 population to define the not responsive population
in our experiments.
B. Experiment Design
The purpose of these experiments was to study the
feasibility of automatically identifying rationales for
responsive documents with and without annotated rationales.
We conducted two sets of experiments. In both sets of
experiments, we built two types of predictive models, a
document model and a rationale model. A document model
was trained using documents with responsive and not
responsive labels, whereas a rationale model was trained
using responsive and not responsive labeled text snippets. A
responsive labeled snippet was an annotated rationale, while
a not responsive labeled text snippet was a randomly selected
text snippet from a not responsive document. Not responsive
snippets could also be selected from responsive documents,
but we did not fully explore that within the confines of this
study. Rather, we adopted a random process to sample not
responsive snippets from not responsive documents. To
ensure not responsive snippets have similar parameters as
responsive snippets, we enforced two constraints in sampling
a not responsive snippet from a not responsive document: (1)
the length should be a random number between 10 and 250,
which is the same snippet length range as the responsive
snippet population; (2) the starting position of the not
responsive snippet should be a random position between zero
and the attorney reviewed document’s length, minus the
snippet’s length from the first constraint. One not responsive
text snippet was selected from each not responsive
document.
The first set of experiments evaluated the performance of
both the document and rationale models in classifying
annotated responsive rationales from not responsive snippets
randomly selected from the not responsive documents. In
these experiments, both document and rationale models were
evaluated using a test set comprised of annotated rationales
and randomly selected not responsive snippets. Precision and
recall were used as performance metrics. The first set of
experiments were performed to test the performance of the
models on the text snippets alone and provide multiple ways
to interrogate the results of the modeling methods.
The second set of experiments simulate a real legal
application scenario and apply both document and rationale
models to responsive labeled documents to identify
rationales that “explain” the models’ responsive decision. In
these experiments, a responsive document is divided into a
set of overlapping snippets. Then, the models are applied to
these snippets to identify rationales. We encountered the
following question: how should this study decide if a snippet
should be treated as a rationale or not? Our answer to that
question in basic predictive coding principles: one simple
way is to identify the snippet with the largest score in a
document and consider that as the rationale for the
document. Recall (the percentage of identified rationales)
was used as the performance metric. An annotated rationale
is correctly identified if it is included in the text snippet with
the largest score.
The machine learning algorithm used to generate the
models was Logistic Regression. Our prior studies
demonstrated that predictive models generated with Logistic
Regression perform very well on legal matter documents [2,
10]. Other parameters used for modeling were bag of words
with 1-gram and normalized frequency [2]. The results
reported in the next section are averaged over a fivefold
cross validation.
C. Results of the Experiments
Figure 1 details the precision and recall curves for the
document and rationale models in discriminating annotated
responsive rationales from not responsive text snippets. The
curves are the average of fivefold cross validation results.
Each of the five document models in the fivefold cross
validation was trained using an 80% random subset of the
23,791 responsive documents and 365,742 not responsive
documents, while each of the five rationale models were
trained using an 80% random subset of the 23,791 annotated
rationales and 365,742 not responsive text snippets. In each
fold, both the document and rationale models were tested
using a 20% random subset of the 23,791 annotated
rationales and 365,742 not responsive text snippets, i.e. on
average, 4,758 annotated rationales and 73,148 not
responsive text snippets. The responsive document rate is
6.5%.
The second set of experiments evaluated the document
and rationale models’ ability to identify rationales of
responsive documents. In these experiments, both the
document and rationale models were the same models that
6
were used in the first set of experiments. We use snippets of
the responsive documents as the candidate rationales. Each
responsive document was divided into a set of n (50, 100,
and 200) word snippets with n/2 words overlapping between
neighboring text snippets. Table 2 details the statistics of the
text snippets for each snippet setting. Both document and
rationale models were deployed to assign probability scores
to each snippet of a responsive document. M (1, 2, 3, 4, 5)
top scoring snippets were selected as the identified rationales
for each model. An identified snippet is a true rationale if it
overlaps with the annotated rationale identified by the
attorney reviewer.
Figure 1: Precision and Recall Curves for Rationale
Model and Document Model
Figure 1 shows that the rationale models performed quite
well. At 80% recall, the rationale models achieved 70%
precision. While performing less effectively than the
rationale models, the document models’ performance was
encouraging, considering the responsive document rate is
6.5%. At 75% recall, the document models’ achieved
precision greater than 25%, which is roughly four times
more than the responsive rate.
Table 2: Text Snippet Statistics
Snippet
Setting
Total
Number of
Snippets
Number of
Documents
Average
Number of
Snippets
50
933,997
23,791
39
100
473,181
23,791
20
200
242,578
23,791
10
Table 3 reports the recall (or percentage of responsive
annotated rationales successfully identified) of the document
and rationale models using different sizes of text snippets.
The rationale models successfully identified responsive
rationales for close to 50% of the responsive documents with
the top scoring text snippet for all sizes of snippets. For the
top three scoring snippets, the rationale models achieved
more than 70% recall. For the top five scoring snippets, the
rationale model achieved around 80% recall.
The rationale models performed better than the document
models for snippets with 50 words. Given that the average
rationale length in our experiment is 52 words (see Section
IV.A), rationale models performed quite well in identifying
snippets with a length similar to the rationales in the
population. Both model types achieved similar recalls for
snippets with 100 words, and the document models
performed better than the rationale models for snippets with
200 words. One explanation for this observation is that a
snippet rarely covers all words of the annotated rationale and
a snippet almost always includes words that are not in the
annotated rationale. Rationale models were trained using
annotated rationales, therefore they did not tolerate noise
(irrelevant words) very well. On the other hand, document
models were trained using the full document text, which
includes the words within the annotated rationales and also
words throughout the rest of the document, thereby allowing
it to be more tolerant of noise. Snippets with 200 words are
typically longer than the average true rationale containing 52
words, thus may include a significant number of irrelevant
words, which negatively impacts the performance of the
rationale models. We observed that recall performance for
the top scoring snippets degrades as the size of the snippets
increases for the rationale models. This is likely because
snippets with more words include more noise.
Table 3: Rationale Recall for Rationale and
Document Models
Number
of words
in
Snippet
Top K
Snippets
Rationale Recall
Rationale
Model
Document
Model
50
1
48%
44%
2
62%
56%
3
71%
65%
4
76%
71%
5
79%
75%
100
1
47%
51%
2
64%
64%
3
73%
73%
4
79%
78%
5
82%
82%
200
1
45%
60%
2
68%
73%
3
79%
81%
4
84%
86%
5
88%
89%
7
While the results reported in this case study just begin to
scratch the surface of how explainable predictive coding has
practical applications in legal matters, they are promising:
the results demonstrate that it is possible to build text
classification models to identify rationales automatically,
with or without the use of annotated rationales. In practical
terms, this means that legal teams could evaluate responsive
rationales generated by a Rationale Model or Document
Model to substantially reduce the number of words an
attorney must review to evaluate the responsiveness of a
document.
Rationale Model
With this model and given that the average word count of
a responsive document in this study is 970 words, an
attorney who is reviewing the top four 50-word snippets (125
to 250 words, accounting for the overlap among snippets)
could exclude, on average, 720 to 845 words from the review
of each document, while still achieving 76% recall.
Extrapolating these results across an entire set of responsive
documents, the average word savings in these scenarios is
17,129,520 to 20,103,395 words per experiment. Using the
average number of words per document to estimate the
document review savings, the use of the Rationale Model
could result in a savings of 17,659 to 20,725 documents —
or 74% to 85% of the responsive documents.
Document Model
The top five Document Model snippets achieve a similar
recall (75%) to that of the Rationale Model’s top four 50-
word snippets. In practice, a Document Model is easier to
implement when compared to a Rationale Model because
attorneys do not have to annotate rationales to create
responsive training data. Using the Document Model, an
attorney reviewing the first top 50-word snippet could
exclude, on average, 920 words from the review of each
document while still achieving 44% recall. These results are
particularly compelling because an attorney would only have
to evaluate nearly half of the responsive information in the
data set and also exclude a substantial amount of irrelevant
information. This approach would result an average savings
of 21,887,720 words when extrapolated across the entire
responsive document set for these experiments.
V. SUMMARY
Explainable AI has drawn the attention of the AI and
legal communities alike. It has proven that it is possible to
develop machine learning systems that are explainable, to
meet practical demands and to effect positive impacts on
legal outcomes. The authors believe that explainable AI has
the potential to significantly advance the application of text
classification in legal document review matters. This study
fortified our thinking that the concept of explainable
predictive coding has a promising future in the legal
document review realm. These results open the door to
conduct future studies in explainable predictive coding. We
plan to conduct more experiments on additional data sets.
Further, we plan to explore more advanced machine learning
technologies to continue evolving our understanding of
categorization rationales.
ACKNOWLEDGMENT
We thank Katie Jensen for helping us prepare the data set
and we thank the anonymous reviewers for their helpful
comments.
REFERENCES
[1] Nicholas M. Pace & Laura Zakaras, “Where the Money Goes:
Understanding Litigant Expenditures for Producing Electronic
Discovery,” RAND at 17 (2012).
[2] Chhatwal, R., Huber-Fliflet, N., Keeling, R., Zhang, J. and Zhao, H.
(2016). Empirical Evaluations of Preprocessing Parameters’ Impact
on Predictive Coding’s Effectiveness. In Proceedings of 2016 IEEE
International Big Data Conference.
[3] Buchanan, B. G., & Shortliffe, E. H. (1984). Rule-Based Expert
Systems: The MYCIN Experiments of the Stanford Heuristic
Programming Project. Boston, MA, USA: Addison-Wesley Longman
Publishing Co., Inc.
[4] Gunning, D. (2017). Explainable artificial intelligence (xai). In
Defense Advanced Research Projects Agency, DARPA/I20.
[5] Dong, Y., Su, H., Zhu, J., & Bao, F. (2017). Towards interpretable
deep neural networks by leveraging adversarial examples. ArXiv
Preprint ArXiv:1708.05493.
[6] Zaidan, O., Eisner, J., & Piatko, C. (2007). Using “annotator
rationales” to improve machine learning for text categorization. In
Human Language Technologies 2007: The Conference of the North
American Chapter of the Association for Computational Linguistics;
Proceedings of the Main Conference (pp. 260--267).
[7] Zhang, Y., Marshall, I., & Wallace, B. C. (2016). Rationale-
Augmented Convolutional Neural Networks for Text Classification.
In Proceedings of the Conference on Empirical Methods in Natural
Language Processing (pp. 795–804).
[8] Martens, D., & Provost, F. (2014). Explaining data-driven document
classifications. MIS Quarterly, 38, 73--100.
[9] Lei, T., Barzilay, R., & Jaakkola, T. (2016). Rationalizing Neural
Predictions. In Proceedings of the 2016 Conference on Empirical
Methods in Natural Language Processing (pp. 107--117).
[10] Chhatwal, R., Huber-Fliflet, N., Keeling, R., Zhang, J. and Zhao, H.
(2017). Empirical Evaluations of Active Learning Strategies in Legal
Document Review. In Proceedings of 2017 IEEE International Big
Data Conference.
