Anywhere but Here
Enron’s Emails in the Midst of a Crisis
Corey Taylor
1,2
, Richard Leibbrandt
1
and David Powers
1
1
School of Computer Science, Engineering and Mathematics, Flinders University,
PO Box 2100, Adelaide 5001, South Australia
2
Institute of Pharmaceutical Chemistry, Philips-University Marburg, Marbacher Weg 6, 35032 Marburg, Germany
Keywords: Sentiment Analysis, Conditional Random Field, Topic Model, Information Retrieval, Text Processing,
Enron.
Abstract: The emotional states of employees under stress are likely to manifest themselves in ways other in inter-
personal interactions, namely, emails. The Enron Email Corpus was mined by both supervised and
unsupervised methods to determine the degree to which this was true for Enron employees whilst the
corporation was under investigation. Changes in language patterns were then compared against the timelines
of the investigation. The method as described validates both the use of a subset of a very large corpus and
the use of tagging methods to understand the patterns in various phrase types as used by Enron employees.
1 INTRODUCTION
There is now considerable evidence that perceived
stress at work is widespread and associated with ill
health at work. Work stress has been reported as
highest in the middle-aged, educated to degree level
and those separated or divorced (Smith, 2001).
Additionally, employees affected by work-related
stress are often underwhelmed by their employers’
attempts to address the issue. It has been reported
that employees in smaller organisations (i.e., less
than 100) rated their work environments
significantly higher than larger organisations on job
satisfaction and also seems relatively clear that the
level of trust between management and workers,
joint resolution on ways to manage stress and
organisation commitment to deal with issues causing
stress are crucial for the management of workplace
negativity (Buys et al., 2010) This is a finding
replicated elsewhere with similar conclusions
regarding the effect of workplace stress on a
worker’s health, well-being and productivity, even
with highly professional environments such as a
university (Shikieri et al., 2012).
Coupled with this are the effects of a singular
stressful event upon companies in changing worker
attitudes toward work and how that can manifest
itself in non-work-related behaviours. Particularly in
the short-term after an event, behaviours such as
‘communal bereavement’ become apparent (Hurley-
Hanson, 2011). ‘Incivility’ and mistreatment
generally has also been reported to impact on
organisational mood and subsequent behaviour as
persons within an organisation express their
displeasure with each other and, left unchecked, can
result in significant numbers of persons within an
organisation displaying negative affect and
subsequent impact on performance and the health of
the organisation generally. This can result in
withdrawal from participating in organisational
activities or more direct behaviour (e.g. undermining
a fellow employee, petty behaviours designed to
humiliate them). Generally, colleagues will not
overtly confront their co-workers but will find
expression of their displeasure in other media,
namely, email (Pearson et al., 2001; Walinga et al.,
2013). An exploration of an email corpus to discern
whether (and by how much) workers’ attitudes
manifest themselves in their emails is therefore
warranted.
The only significant public email corpus
available to explore the emotional content of emails
is the Enron Email Dataset and, thus far, analysis of
the emotional content of text in emails is non-
existent, the research mainly focussed on testing
researchers’ machine learning models and
algorithms. Researchers have mainly concentrated
their efforts on analysis of network clustering and
247
Taylor C., Leibbrandt R. and Powers D..
Anywhere but Here - Enron’s Emails in the Midst of a Crisis.
DOI: 10.5220/0005174502470253
In Proceedings of the International Conference on Agents and Artificial Intelligence (ICAART-2015), pages 247-253
ISBN: 978-989-758-074-1
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
their evolution into social communities. One study
found via directed graph/Singular Value
Decomposition (SVD) methods that clustering of
this nature tends to reflect seniority and pay
structure (Chapanond et al., 2005; Carley et al.,
2005). Analysis of network changes during the
course of the investigation has also been conducted
and from a study was reported that organisational
communication, even the sample that exists in email,
changes not only in volume but in who is
communicating with whom during periods of
organisational change and crisis and this reflects
changes in communication norms and the way in
which groups present themselves (Diesner et al.,
2005). Another study, again using SVD methods,
found that word use within the Enron dataset is
correlated to function within the organisation and
changes in word usage can reflect the increased
influence of ‘key players’ in major company events
(Keila et al., 2005).
Missing from the literature is analysis of
emotional and attitude content changes in the email
of employees. As suggested, the real moods of
employees may manifest themselves via indirect
communications such as email. The purpose of this
study was therefore to extract this information from
emails using both supervised and unsupervised
methods, with some basic assumptions, and to link
these patterns to investigation timelines. It was
hoped that changes in the use of language reflecting
employee’s moods would reflect the very public and
stressful nature of the investigation into and
subsequent financial collapse of the Enron
Corporation. Computational methods used to extract
this meaning from the corpus were a Conditional
Random Field used for tagging phrases of interest
and Probabilistic Topic Modelling to extract
semantic data from emails.
1.1 Conditional Random Field
A Conditional Random Field (CRF) is an undirected
discriminatively-trained graphical model, a special
case of which is a linear-chain CRF which is trained
to maximise the conditional probability of a label
sequence given an input sequence (Schneider, 2006).
This CRF has an exponential form allowing it to
integrate many over-lapping non-independent
features and avoiding the label-bias problem
(Le-Hong et al., 2006).
Let x = x
1
...x
T
be an input sequence and y = y
1
...y
T
be a corresponding state (or label) sequence. A
CRF with parameters = {λ, . . .} defines a
conditional probability for y given x to be:
1
1
1
(|) exp ( , ,,)
T
kk t t
tk
x
P
yx f y yxt
Z
(1)
where Z
x
is a normalisation constant that makes the
probabilities of all label sequences sum to one,
f
k
(y
t−1
,y
t
,x,t) is a feature function, and λ
k
is a learned
weight associated with f
k
. A feature function
indicates the occurrence of an event consisting of a
state transition y
t−1
→ y
t
and a query to the input
sequence x centered at the current time step t.
Inference in CRFs is done by finding the most
probable label sequence, y
∗
, for an input sequence,
x, given the model in Equation 1:
y
*
arg ma
x
y
P
(
y
|
x
)
(2)
During training, the weights λ
k
are set to maximise
the conditional log-likelihood of a set of labelled
sequences in a training set D = {(x
(i)
, y
(i)
) : i = 1, . . . , M}:
LL(D) log P
( y
( i)
, x
( i)
)
k
2
2
k
2
k
i1
M
(3)
k
f
k
(y
t 1
(i )
, x
(i)
,t) log Z
x
(i )
k
t 1
T
i1
M
k
2
2
k
2
k
The term
k
k
k
2
2
2
is a Gaussian prior used to
penalise the log-likelihood in order to avoid over-
fitting and
2
k
is a variance. Maximisation of
Equation 3 corresponds to matching the expected
count of each feature according to the model to its
adjusted empirical count:
2
() () ()
1
2
11
(,,,)
MT
iii
k
kt t
it
k
fy y x t
(4)
() ' ' ()
1
1' 1
('| ) ( ,, ,)
MT
ii
kt t
iy t
Py x fy yx t
The term
k
2
k
2
is used to discount the empirical
feature counts. Finding the parameter set that
maximises the log-likelihood in Equation 3 is done
using an iterative procedure called limited-memory
quasi-Newton (L-BFGS). Since the log-likelihood
function in a linear-chain CRF is convex, the
learning procedure is guaranteed to converge to a
global maximum.
1.2 Topic Model
Given the definition that a document is a mixture of
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
248
topics, that P(z) is a distribution of topics over a
document and P(w|z) represents the probability
distribution of words w given topic z, a model
specifying the following distribution over words can
be formed. As
P
(z
i
j
) is the probability that the jth
of T topics was sampled for the ith word token and
P
(w
i
| z
i
j
) is the probability of w
i
for topic j:
P(w
i
) P(w
i
| z
i
j)P(z
i
j)
j 1
T
(5)
To obtain topic and word information from the topic
model, we must first extract the topic-word
distributions,
, and topic distributions
. As this is
computationally expensive, we must sample from
the posterior distribution over z using a Gibbs
Sampling algorithm.
For this purpose, the collection of documents by
a set of word indices w
i
and document indices d
i
, for
each word token I is represented. The Gibbs
sampling algorithm considers each word token
within the text collection in turn, and estimates the
probability of assigning the current word token to
each topic, conditioned on topic assignments to all
other word tokens. From this conditional
distribution, a topic is sampled and stored as the new
topic assignment for this word token. This
conditional distribution is P(z
i
=j | Z
-i
, w
i
, d
i
, · ),
where Z
i
= j represents the topic assignment of token
i to topic j, Z
-i
refers to the topic assignments of all
other word tokens, and “·” refers to all other known
or observed information such as all other word and
document indices w
-i
and d
-i
, and hyperparameters α
and β (Steyvers et al, 2006).
P(z
i
j | z
i
,w
i
,d
i
,)
C
w
i
j
WT
C
wj
WT
W
w1
W
C
d
i
j
DT
C
d
i
t
DT
T
t1
T
(6)
Here C
WT
and C
DT
are matrices of counts of words
with dimensions WxT and DxT respectively;
C
w
i
j
WT
contains the number w
j
of times word w is assigned
to topic j, not including the current instance i and
C
d
i
j
DT
contains the number of times topic j is
assigned to some word token in document d, not
including the current instance i.
Each Gibbs sample consists the set of topic
assignments to all N word tokens in the corpus from
a single pass through all documents. During the
‘burn-in’ period, the Gibbs samples have to be
discarded because they are poor estimates of the
posterior. After the burn-in period, the successive
Gibbs samples begin to approximate the target
distribution (i.e., the posterior distribution over topic
assignments).
From this process, estimation of
and
is
relatively straightforward. They correspond to the
predictive distributions of sampling a new token of
word ifrom topic j, and sampling a new token (as of
yet unobserved) in document d from topic j. They
are also the posterior means of these quantities
conditioned on a particular sample z.
i
j
C
ij
WT
C
kj
WT
W
k 1
W
(7)
i
d
C
dj
DT
C
dj
DT
T
k 1
T
(8)
2 METHODS
The corpus used for analysis was the Enron Email
Dataset, originally purchased by Leslie Kaelbling
and made publicly available for download by the
Federal Energy Regulatory Commission on
Carnegie-Mellon University’s servers.
1
The dataset
contains ~500K emails from senior management
executives at the Enron Corporation collected during
the time of investigation, covering approximately
2000-2004.
Emails in this ~2.2Gb dataset are in MBOX
format and arranged by email folders (e.g. inbox,
sent items, personal folders). As many of these
folders contain emails that are not relevant to this
study (e.g. spam) or contain significant amounts of
duplicated text (e.g. nested replies in long
discussions), the decision was made to only include
emails in users’ ‘sent items’ for analysis. The
rationale for this decision was that items such as
spam, mailing list digests, etc. were far less likely to
be in a person’s sent items. Pre-processing scripts
stripped out email headers, quoted and forwarded
messages, HTML and formulaic block text (e.g.
signatures). The resultant dataset contained ~90K
emails split into months for tagging.
The MAchine Learning for LanguagE Toolkit
(MALLET) was used for computation of the Enron
dataset. MALLET is a Java-based package for
statistical natural language processing, document
classification, clustering, topic modelling,
1
http://www.cs.cmu.edu/~./enron/
AnywherebutHere-Enron'sEmailsintheMidstofaCrisis
249
information extraction, and other machine learning
applications to text.
2
A modified version of the
SimpleTagger class was used for tagging of emails
with other methods written to instantiate, train and
query topic models.
A linear-chain Conditional Random Field (CRF)
was instantiated and trained in binary (on/off)
fashion on the following types of phrases:
1) Slacker language – 24 phrases
2) Aggressive language – 26 phrases
3) Panic language – 14 phrases
In each case, the most ‘descriptive’ word was chosen
as the word to tag as one of the above, the rest of the
words in a given sentence tagged with MALLET’s
default ‘O’. Example sentences for training the CRF
were obtained from one month of email data and
manually tagged commensurate with the above
speech categories. For example, “let’s do lunch” was
tagged as ‘slacker’ language, “I need it ASAP”
tagged as ‘aggressive’ language and “we have to get
onto this immediately” as ‘panic’ language.
Sentence-ending full stops were removed from each
text file to avoid tagging of each phrase type only in
the instance where it is used as a completed
sentence. The frequency of each tagged phrase type
in a given month of emails was then calculated.
Topic models were trained on 100 topics using
each month’s emails as its corpus. Alpha (i.e. sum
over topics) and beta (e.g. single dimension of
Dirichlet prior) hyperparameters were set to 0.01.
Information subsequently extracted for each month
was the top-10 highest probability topics given
documents (i.e. emails) for the month and the top-10
words given topics in ranked order.
Counts of each phrase type per month were used
in Pearson correlations with the monthly closing
Enron stock price from that time. This data and
phrase count data as elucidated from the CRF was
added to a static plot of frequencies. The plot was
generated using the JChart2D toolkit.
3
3 RESULTS
Figure 1 is representative of the frequency of
slacker/aggression/panic phrases by month.
Additionally, the closing stock price at the end of the
month is plotted.
Output from the topic model was in the form of
topics with probabilities and words under those
2
http://mallet.cs.umass.edu
3
http://jchart2d.sourceforge.net/
topics in ranked order. Pearson correlation of phrase
type with stock price is not suggestive of a
relationship between ‘slacker’ and ‘aggressive’
phrases but a strong negative correlation with
‘panic’ phrases was evident.
Figure 1: Frequencies of language types and stock price
over time.
Table 1: Pearson correlations of counts of language types
with Enron stock price.
Phrase type R
2
Slacker language -0.032
Aggressive language -0.0004
Panic language -0.797
4 DISCUSSION
The frequencies of various language types may
possibly reflect events during the course of the
investigation into Enron. Highlighted examples as in
Figure 2 may be indicative of the following events
as detailed in court documents and re-published in
various books about Enron.
1: Spikes in ‘aggressive’ language during 2000 are
approximately prior to Enron’s all-time high stock
price of $90.56 recorded on August 23
rd
2000. This
corresponds to the activities of influential head of
trading Timothy Belden. Some of his activities were
found to have played a role in precipitating the
California energy crisis, a shortage of electricity
supply caused by illegal market manipulations,
eventually resulting in his conviction for conspiracy
to commit wire fraud. This was also reflected in
output from the topic model, for eample in June
2000, the word ‘California’ or other places in the
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
250
California area were used in 4 of the top-10 topics
for that month under topics about regulation and
energy.
Figure 2: Frequencies of language types and stock price
over time with highlighted events.
Table 2: Use of the word ‘California in topics – June 2000.
power gas california market price energy electricity prices
electric company
san francisco chronicle angeles diego los wine city
sgovenar@govadv.com times
california commission ferc energy utilities puc public file
order federal
call meeting conference week office number time sellers
discuss california
davis bill governor dow assembly california committee
energy jones state
2: A spike in ‘slacker’ language approximately
corresponds to executives such as Tom White
cashing out their stocks, retiring and making large
purchases.
3: Spikes in ‘panic’ language correspond to a
precipitous drop in Enron’s share price following on
from the beginning of an informal probe by the
Securities and Exchange commission commenced on
October 17 2001 and followed by Enron declaring
bankruptcy and massive layoffs on December 2
nd
2001.
4: Indicative spikes in ‘slacker’ and, subsequently,
‘aggressive’ language as resignations and senate
hearings involving executives such as Ken Lay
commence in early 2002. Other key events around
this time include the suicide of Cliff Baxter, former
vice chairman, and the release of The Powers Report
on Feb 2
nd
2002, a 218-page summary of internal
investigations into Enron’s collapse by the
University Of Texas School Of Law Dean William
Powers. This report publicly named Enron
executives and held them responsible for the
collapse of the company.
5: A spike in ‘aggressive’ language in early 2003
corresponds to announcement on March 19
th
2003
detailing that Enron will keep its North American
and international pipeline and power assets in an
attempt to emerge from bankruptcy. The topic model
for the month of February 2003 suggested increases
in the use of the word ‘Enron’, as it was in the top-
10 words for 4 of the top-10 topics for that month in
relation to topics such as trading and
communications. In most other months, ‘Enron’
appears either not at all or only in one of the topics
for that month.
Table 3: Use of the word ‘Enron’ in topics – February
2003.
call questions group issues comments credit enron 145 mark
discuss contact
ferc rto market transmission enron process meeting model
epsa filing
agreement draft master isda guaranty enron form letter send
doc
enron program entity kaiser trading product power tax
forward products
A precipitous drop in ‘aggressive language
corresponds to charges being laid against Chief
Financial Officer Andrew Fastow and his wife on
May 1
st
2003 for activities which resulted in
concealment of Enron’s massive financial losses.
For these activities, he subsequently served 6 years
in prison.
As the above is merely a descriptive analysis, the
magnitude and direction of any association is
unknown and conclusions drawn are entirely
subjective. However, it can be said with some
certainty that, to some degree, the emotional states
of employees in Enron were reflected in language
patterns they chose to use in sending emails. This is
broadly supportive of the contention as stated earlier
that the language used in emails by employees
working within a dysfunctional and uncivil work
environment, whilst perhaps responding in direct
ways too, will reflect both the emotional state of the
employee and the work environment in the use of
what is ostensibly a work tool.
The strong negative correlation between panic
phrases and stock price is also indicative of the
validity of the method. This provides a cautious
validation of the use of CRFs to tag phrases based
on semantic considerations, as opposed to syntactic
or for Named-Entity Recognition purposes, even
with a lightly-trained model comprised of only 24
‘slacker phrases, 26 ‘aggressive’ phrases and 14
AnywherebutHere-Enron'sEmailsintheMidstofaCrisis
251
‘panic’ phrases. This is as opposed to training sets
comprised of hundreds or sometimes thousands of
training phrases, as is more usually observed in the
use of CRFs. That the use of relatively broad
categorisations of phrases was able to approximately
reflect the timelines of the investigation into Enron
means the method could be extended in many ways.
There are many limitations with the approach as
detailed within this study. Selection of phrases
corresponding to the three categories studied was
entirely subjective and therefore there was a risk of
bias in model training. Additionally, the nature of
the corpus meant that although there were extensive
attempts to clean the dataset, many artifacts of email
in its raw form remain (e.g. spam, multiple quoting
biasing counts). The precise nature of the association
between phrase use and actual events in Enron’s
history can only be guessed at, more information
regarding the detailed course of events would be
required to validate the accuracy and sensitivity of
the association detailed here. The a priori nature of
CRF model training in this instance virtually
guarantees bias.
There are also general limitations in probabilistic
topics models which may affect inferred results;
topic models are prone to overfitting, as in, the mode
by which an individual document’s topic mixture is
established is not robust enough to handle the
addition of new documents to the trained corpus.
Related, the number of free model parameters
increases linearly with the number of training
documents, making re-training a computationally
expensive exercise.
Possible extensions to the software are many and
varied. Results from what was a relatively lightly
trained CRF model seemed reasonable but it was
trained only on binary data with a first-order model.
The use of higher-order model will likely increase
the precision of tagging of phrases as more
information about context is modelled. This may
also allow a finer grained model training of more
specific phrases as ‘slacker’, ‘aggressive’, etc. are
relatively broad terms for the language being
modelled.
Instead, sub-types of slacker/aggressive/panic
phrases could be tagged. The results of a topic
model could also be used to inform the tagging of
phrases rather than the a priori method as detailed in
this study. Ideally, a formal evaluation of tagging
predictive accuracy could be conducted on non-
Enron emails or on Enron emails with a k-folds
cross-validation methodology with attendant
measures of fit (e.g. positive predictive value).
5 CONCLUSIONS
The method as detailed provides a broad method for
the descriptive analysis of email data by tagging of
phrases that are semantically interesting. That this
exercise even broadly reflects the timeline of
investigation validates the use of both a sub-set of
the full Enron corpus as well as the method used to
tag information of interest. This is suggestive that
the performance of even a lightly-trained model may
be acceptable on a far smaller test set than would be
the case were it exhaustively trained on the full
Enron Email Dataset.
REFERENCES
Blei, D. (2012). Probabilistic Topic Models.
Communications of the ACM , 55 (4), 77-84.
Buys, N. M. (2010). Employees’ Perceptions of the
Management of Workplace Stress. International
Journal of Disability Management, 5 (2), 25-31.
Chapanond, A. K. (2005). Graph Theoretic and Spectral
Analysis of Enron Email Data. Computational &
Mathematical Organization Theory, 11, 265-281.
Chekina, L. G. (2013). Exploiting label dependencies for
improved sample complexity. Machine Learning, 91,
1-42.
Dahl, C. (2004). Pipe Dreams: Greed, Ego, and the Death
of Enron/Anatomy of Greed: The Energy Journal, 25
(4), 115-134.
Diesner, J. C. (2008). Conditional random fields for entity
extraction and ontological text coding. Computer and
Mathematical Organisation Theory, 14, 248-262.
Diesner, J. F. (2005). Communication Networks from the
Enron Email Corpus “It’s Always About the People.
Enron is no Different". Computational &
Mathematical Organization Theory, 11, 201-228.
Dreijer, J. H. (2013). Left ventricular segmentation from
MRI datasets with edge modelling conditional random
fields . BMC Medical Imaging, 13, 1-24.
El Shikieri, A. M. (2012). Factors Associated with
Occupational Stress and Their Effects on
Organizational Performance in a Sudanese University .
Creative Education , 3 (1), 134-144.
Hayashida, M. K. (2013). Prediction of protein-RNA
residue-base contacts using two-dimensional
conditional random field with the lasso . BMC Systems
Biology, 7 (Suppl 2), 1-11.
Hurley-Hanson, A. G. (2011). The Effect of the Attacks of
9/11 on Organizational Policies, Employee Attitudes
and Workers’ Psychological States. American Journal
of Economics and Business Administration, 3 (2), 377-
389.
Hutton, A. L. (2006). Crowdsourcing Evaluations of
Classifier Interpretability. AAAI Technical Report SS-
12-06 Wisdom of the Crowd, 21-26.
Jahanian, R. T. (2012). Stress Management in the
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
252
Workplace. International Journal of Academic
Research in Economics and Management Sciences , 1
(6), 1-9.
Keila, P. S. (2005). Structure in the Enron Email Dataset.
Computational & Mathematical Organization Theory ,
11, 183-199.
Klimt, B. Y. (2005). Introducing the Enron Corpus. 2.
Le-Hong, P. P.-H.-T. (2006). On the Effect of the Label
Bias Problem in Part-Of-Speech Tagging. Computing
and Communication Technologies, Research,
Innovation, and Vision for the Future, 103-108.
McCallum, A. (2002). MALLET: A Machine Learning for
Language Toolkit. http://mallet.cs.umass.edu .
Ozcaglar, C. (2008). CLASSIFICATION OF EMAIL
MESSAGES INTO TOPICS USING LATENT
DIRICHLET ALLOCATION. MSc. thesis, 1-37.
Pearson, C. A. (2001). When Workers Flout Convention:
A Study of Workplace Incivility. Human Relations ,
54 (11), 1387-1419.
Priebe, C. C. (2005). Scan Statistics on Enron Graphs.
Computational & Mathematical Organization Theory ,
11, 229-247.
Robinson, S. (2009). The Nature of Responsibility in a
Professional Setting. Journal of Business Ethics, 88,
11-19.
Schieman, S. W. (2006). The Nature of Work and the
Stress of Higher Status. Journal of Health and Social
Behaviour , 47 (3), 242-257.
Schneider, K.-M. (2006). Information extraction from
calls for papers with conditional random fields and
layout features . Artificial Intelligence Review, 25, 67-
77.
Smith, A. (2001). Perceptions of Stress At Work. Human
Resource Management Journal, 11 (4), 74-86.
Steyvers, M. G. (2006). Probabilistic Topics Models. In L.
Erlbaum, Latent Semantic Analysis: A Road to
Meaning (pp. 1-15).
Styler, W. (2011). The Enronsent Corpus. University of
Colorado - Boulder , 1-7.
Sun, C. G. (2007). Rich features based Conditional
Random Fields for biological named entities
recognition . Computers in Biology and Medicine, 37,
1327-1333.
Sun, X. H.-Y. (2011). Chinese New Word Identification:
A Latent Discriminative Model with Global Features .
Journal of Computer Science and Technology, 26 (1),
14-24.
Walinga, J. R. (2013). Transforming stress in complex
work environments Exploring the capabilities of
middle managers in the public sector . International
Journal of Workplace Health Management, 6 (1), 66-
88.
Wallach, H. (2001). Efficient Training of Conditional
Random Fields. Masters's thesis, 1-8.
Wu, J. L. (2010). Chunk Parsing and Entity Relation
Extracting to Chinese Text by Using Conditional
Random Fields Model. Journal of Intelligent Learning
Systems & Applications, 2, 139-146.
AnywherebutHere-Enron'sEmailsintheMidstofaCrisis
253
