Analysis of communication patterns with scammers in Enron corpus Dinesh Balaji Sashikanth Master of Science in Computer Science School of Informatics and Computing Indiana University,Bloomington-47405,USA Abstract Beginning in the late 1990s, Enron execu- tives such as Jeffrey Skilling and Andrew This paper is an exploratory analysis into Fastow initiated a campaign to hide business fraud detection taking Enron email corpus losses from company stockholders and the as the case study. The paper posits con- general public. Enron was subsequently grant- clusions like strict servitude and unques- ed – government deregulation. As a result of tionable faith among employees as breeding grounds for sham among higher this declaration of deregulation, Enron execu- executives. We also try to infer on the na- tives were permitted to maintain agency over ture of communication between fraudu- the earnings reports that were released to in- lent employees and between non- vestors and employees alike. fraudulent-fraudulent employees 2.1 Deceitful Practices 1 Rationale Enron and other energy suppliers earned prof- its by providing services such as wholesale This document provides an in-depth scrutiny trading and risk management in addition to of communication patterns between venal and building and maintaining electric power plants, genuine employees in the Enron email corpus. natural gas pipelines, storage, and processing This project work is for the partial fulfillment facilities. Instead of reporting the profit on the for course B659 Detecting latent properties in trading and brokerage fees, Enron chose to re- text. I am thankful to Professor Markus Dick- port the entire value of each of its trades as inson, Associate Professor in Linguistics, Indi- revenue. ana University for providing valuable In the beginning, the company listed the cost suggestions and guidance. of supplying energy and the revenue it ob- 2 Introduction tained from that. But after Skilling joined the company, he adopted the mark to market ac- Enron was created in 1985 through the merger counting practice. Mark-to-market accounting of two natural gas companies at the behest of requires that once a long-term contract was Houston executive Kenneth Lay. Enron was an signed, income is estimated as the present val- American energy conglomerate based in Hou- ue of net future cash flow. ston, Texas. During the 1990s, it was consid- Enron used special purpose entities—limited ered one of the most powerful and partnerships or companies created to fulfill a successful corporations in the world. Lay re- temporary or specific purpose to fund or man- mained the chief executive of Enron through- age risks associated with specific assets. out its existence The special purpose entities were used for emails easily available for study , as such col- more than just circumventing accounting con- lections are typically bound by numerous pri- ventions. Enron's balance sheet understated vacy and legal restrictions which render them its liabilities and overstated its equity, and its prohibitively difficult to access. earnings were overstated. Enron disclosed to its shareholders that it had hedged downside 3 Previous work risk in its own illiquid investments using spe- Some of the well-known research works on cial purpose entities. However, investors were this corpus were: Research on dynamics of the oblivious to the fact that the special purpose structure and properties of organizational entities were actually using the company's own communication network as a function of time stock and financial guarantees to finance these (Diesner et al 2005), detecting community hedges. This prevented Enron from being pro- structure based on link mining (Qian et al tected from the downside risk. Notable exam- 2006). ples of special purpose entities that Enron From the linguistics perspective-work on for- employed were JEDI, Chewco, Whitewing, mal communication as a factor of social dis- and LJM. tance, relative power and weight of imposition (Peterson et al 2011), Recall enhancing meth- 2.2 Consequences ods for named entity recognition with Enron email as case study (Minkov et al 2005) and The scandal, revealed in October 2001, even- exploring the use of word ‘virtual’ in Enron tually led to the bankruptcy of the Enron Cor- corpus to establish the polysemic usage in con- poration. the company's stock price, which temporary contexts (Greg et al 2009) achieved a high of US$90.75 per share in mid- 2000, plummeted to less than $1 by the end of November 2001. Enron's shareholders lost $74 4 Research Motivations billion in the four years before the company's bankruptcy ($40 to $45 billion was attributed The primary motivation for this research was to fraud).As Enron had nearly $67 billion that to study and detect patterns in fraudulent it owed creditors, employees and shareholders communication by using existing research in received limited, if any, assistance aside from deception detection and author profiling. The severance from Enron. Many executives at En- Enron email corpus was adopted because it ron were indicted for a variety of charges and represented a true large-scale-real world com- some were later sentenced to prison. munication where the emails of employees were available for research to the public with- 2.3 Enron email corpus out any privacy concern and also the ease of labelling fraudulent employees using the The Enron Corpus is a large database of over available legal records data and extensively 600,000 emails generated by 158 employees of studying their communication patterns. But the Enron Corporation and acquired by this dataset proved to be an ill-fitting for the the Federal Energy Regulatory Commis- research (extensively addressed in section 5) sion during its investigation after the compa- because of which the primary goals of the re- ny's collapse. A copy of the database was search were readopted and are stated as fol- subsequently purchased and released lows (S1 Genuine employees and S2 by Andrew McCallum, a computer scientist at Fraudulent employees): the University of Massachusetts Amherst. The corpus is "unique" in that it is one of the only -(TASK 1/3) Analyze the communication pat- publicly available mass collections of "real" terns of S1 with S2( closed or open, transpar- ent or opaque, formal or casual so as to infer Some peculiar problems revealed well into re- the level of trust S1 had over S2 and the level search and these were as follows: Kenneth of transparency as perceived by S1 over S2) Lay’s sent messages were actually that from his assistant Rosalee fleming, Jeff Skilling’s -(TASK 2/3)How did S2 carry out their malice sent messages were that from his assistant practice.(To learn about their modus operandi Sherri sera. Over 30% of Jeff’s inbox messag- especially how open were they in communi- es were about a donation drive and how much cating these plans.) each employee pledged to contribute. David Delainey had only 50 messages in his inbox -(TASK 3/3)Communication network analysis and the rest 1350 messages were in the form of of S2 across 2 different time intervals (Learn discussion threads. about significant changes in the communica- Due to the aforementioned problems, data for tion structure when covert practices were at most of the tasks was collected manual- their peak and when these practices were be- ly(going over through the emails of these 4 ing exposed employees and removing emails not meeting the criteria).Also only the received emails for these four were considered as their sent mes- 5 Datasets-The problems and splits sages were not theirs or their primary response The dataset was acquired from CALO project was through discussion thread. Though ex- hosted by Artificial Intelligence Center of SRI treme care was taken in terms of data purity, international and maintained by Carnegie the unavailability of large corpus particularly Mellon University, Pennsylvania. suited for this task renders the research infer- (www.cs.cmu.edu/enron). The dataset com- ences questionable for extrapolation to general prised of 619,446 messages of 158 users. instances. Compiling from the internet, list of fraudulent employees revealed that only 4 out of the 158 For Task-1: 290 emails (103 Lay,43 JFor,93 employees were actually fraudulent. Data JSki, 51 DDel) were considered about the rest of the fraudulent employees were not made public. Hence our research fo- For Task-2: All of Inbox messages of the 4 cuses primarily on these 4 employees: Kenneth employees were considered Lay, Chairman; Jeffery Skilling CEO; John Forney, Executive for energy trade and David For Task-3: emails from Jeff (inbox, Delainey, Head for trade and retail energy notes_inbox, deleted items), Lay (inbox, units. note_inbox, deleted items), Forney (inbox, de- As the dataset was a real-time collection, there leted items), Delainey (all_documents, inbox, was also the problem of spams, forwarded notes_inbox, deleted items) were considered messages, discussion threads and mailing lists. While spams were outright not useful for this 6 Task-1/3 S1-S2 Communication type of analysis task, forwarded messages and discussion threads were rejected because they 6.1 Methodology were basically a series of messages where one party’s communication pattern was influenced The emails specifically for this task were by another. Similarly group mails were not hand-chosen. Although the expected analysis intended for a particular individual and hence should be on the 2 way communication, be- were rejected. cause of the constraints mentioned above, only S1’s communication with S2 was considered. This implies that only the emails in inbox of communication, it implies ambiguity or un- folders of fraudulent employees were taken natural communication pattern. into consideration. The emails considered were free of newsletters, discussion threads and Emotiveness: charity drive messages. 290 emails were ob- tained which were used for analysis.