Forensic Authorship Analysis and the World Wide Web

S. Larner

• Publisher: Palgrave Pivot
• Year: 2014
• Print ISBN: 9781137413741
• eBook ISBN: 9781137413758
• Topic: Languages & Linguistics
• Subtopic: Computer Science General
• Index: Languages & Linguistics

Implementing a novel method for identifying idiolectal co-selections, and taking the UNABOM investigation as a case study, this Pivot evaluates the effectiveness and reliability of using the web for forensic purposes.

1

Introduction: The UNABOM Investigation

Abstract: Increasingly, forensic linguists are using the web to generate evidence in cases of forensic authorship analysis. A striking example of this occurred during the trial of the Unabomber – a prolific serial bomber – when the web was searched to determine the distinctiveness of a set of idiolectal co-selections. However, to date, questions have not been asked about whether the web can be used reliably in forensic contexts. Therefore, using the Unabomber trial evidence as a case study, this chapter discusses the notion of idiolect and introduces research which explores two issues: (1) whether idiolectal co-selections can be used as a marker of authorship, and (2) whether the web is reliable enough to be used to produce forensic evidence.

Keywords: authorship attribution; idiolect; idiolectal co-selections; Unabomber; web corpus

Larner, Samuel. Forensic Authorship Analysis and the World Wide Web. Basingstoke: Palgrave Macmillan, 2014. DOI: 10.1057/9781137413758.0003.

Linguists are increasingly utilising the world wide web (henceforth “web”¹) as a corpus for research purposes (Volk, 2002) and, given that many linguists acknowledge corpus linguistics as a mainstream methodology (Lindquist & Levin, 2000), coupled with the importance of corpus linguistics methods in the field of forensic linguistics (Coulthard, 1994; Hänlein, 1999; Solan & Tiersma, 2004; Woolls & Coulthard, 1998), it stands to reason that forensic linguists increasingly turn to the web for investigative and evidential purposes. This is particularly the case for forensic authorship analysis – determining the author of a document whose authorship is contested, such as in detecting plagiarism and collusion, attributing a criminal text to an author from a list of potential authors, or profiling an unknown author based on linguistic characteristics. In such cases, the linguist may use the web to show the distinctiveness or rarity of particular words and phrases (Coulthard, 2004).

A striking example of this occurred during the trial of Theodore Kaczynski, a prolific American serial bomber. During the period of May 1978 to April 1995, a total of 16 bombing incidents occurred, initially targeted at individuals connected to universities and the airline industry. These specific targets led the FBI to codename the investigation UNABOM. As a result of the bombing campaign, three people were killed and many more were injured. In June 1995, The New York Times, The Washington Post, Penthouse, and Scientific American (as well as a Professor of Sociology at the University of California at Berkeley) received a manuscript – a terrorist’s manifesto purportedly written by the Unabomber – entitled An industrial society and its future. Along with the manuscript was a deal: if the manuscript was published in full, the bombing would stop. The Washington Post eventually published the manuscript in September 1995 (Fitzgerald, 2004).

Upon reading an internet version of the published manifesto, Linda Patrik became unnerved. Although she had never met her brother-in-law, there was something about the text that seemed familiar. She asked her husband, David Kaczynski, to read the publication and urged him to compare it with her brother-in-law’s writings. David sceptically complied, but started to suspect that his older brother, Theodore, may indeed be the Unabomber. The occurrence of one phrase in particular convinced him: cool-headed logicians. David recalled his brother “using that distinctive term on numerous occasions” (Fitzgerald, 2004: 208) and as a result, he contacted the FBI. To assist with the investigation, the Kaczynski family made available many documents known to have been written by Theodore for comparative analysis with the manifesto (p. 208).

James Fitzgerald of the FBI Behavioral Analysis Unit led a team of FBI agents and analysts in the comparative analysis of writings known to have been authored by the Unabomber (including so-called ruse letters which incited recipients to open accompanying bombing devices sent through the mail, ideological letters which outlined the Unabomber’s rationale for the bombing campaign, brokering letters in which the Unabomber tried to get the manifesto published, and of course the manifesto itself) with writings known to have been authored by Kaczynski (which included, amongst others, Kaczynski’s doctoral thesis, personal letters, and short stories) (Fitzgerald, 2004). In April 1996, after reviewing all of the evidence, including the report produced by the comparative analysis team, a federal judge signed a warrant to search Kaczynski’s cabin in Montana. The FBI arrested Kaczynski at his home, whereupon they found “a virtual treasure trove of evidentiary materials” (p. 215) including a fully assembled bomb and numerous bomb parts.

In preparation for the trial, Kaczynski’s defence team attempted to undermine the basis on which the search warrant had been obtained: “If they could get a ruling from the judge during the pre-trial stage that the search warrant was obtained by the FBI improperly, and that there was not enough probable cause to support the search, the entire case against Kaczynski could conceivably be dismissed” (Fitzgerald, 2004: 216). The comparative analysis team’s report came under scrutiny by the defence’s expert witness, Robin Lakoff, who outlined seven areas of error, and opined that the claims of common authorship between the known writings of the Unabomber and the known writings of Theodore Kackzynski were “untenable and unreliable at best” (Fitzgerald, 2004: 217). On the other hand, the prosecution’s expert witness, Donald Foster, concluded that “Fitzgerald had cautiously understated the case for common authorship” (Foster, 2001: 107). According to Coulthard (2000), Lakoff argued that many of the lexical items which were shared across both groups of texts could “quite easily occur in any argumentative text” (p. 281) and were therefore not indicative of common authorship between the Unabomber’s terrorist manifesto and Kaczynski’s known writings. In particular, 12 words and phrases were selected for exemplification: at any rate, clearly, in practice, gotten, more or less, moreover, on the other hand, presumably, propaganda, thereabouts, and lexemes derived from the lemmas argu and propos (p. 281).

To counter the proposal that such words and phrases could occur in any similar text, Coulthard (2000) reports that the web was searched and approximately three million documents were found which included at least one or more of the 12 lexical items. However, when the search was limited to finding only documents which contained all 12 lexical items, only 69 documents were identified and each of these documents was an online copy of the terrorist manifesto (p. 282). Coulthard concludes of this evidence that “a writer’s combinations of lexical choices are more unique, diagnostic or idiolectal than people have so far been willing to believe” (p. 282). Furthermore, Coulthard (2004) argues that this evidence is “a powerful example of the idiolectal habit of co-selection and an illustration of the consequent forensic possibilities that idiolectal co-selection affords for authorship attribution” (p. 433).²

In order to judge this evidence as “powerful”, two assertions must firstly be accepted: (1) that idiolectal co-selection – that is, words which taken in combination appear to characterise an individual author’s linguistic style – is a useful marker of authorship; and (2) that the web is a valid and reliable corpus for producing forensic evidence. As will become clear in Section 1.3, the aim of this research is to test empirically both of these assertions, and since testing both assertions rests on the notion of idiolect and the use of lexis as a marker of authorship, it is firstly necessary to discuss both before setting out the scope of this research.

1.1 Idiolect

Although the term idiolect was first coined by Bloch (1948), Sapir (1927) laid the groundwork in his discussion of the relationship between speech and personality. Sapir outlined five levels of speech that were indexical of individual personality including voice, dynamics, pronunciation, vocabulary, and style. Of these, vocabulary and style are the most relevant precursors to the concept of idiolect. Sapir argued of vocabulary that:

We do not all speak alike. There are certain words which some of us never use. There are other, favorite, words which we are always using ... Individual variation exists, but it can properly be appraised only with reference to the social norm. Sometimes we choose words because we like them; sometimes we slight words because they bore or annoy or terrify us. We are not going to be caught by them. All in all, there is room for much subtle analysis in the determination of the social and individual significance of words. (p. 903)

Here, Sapir clearly draws out the potential for individual variation at the lexical level (further considered in Section 1.2) and the complex relationship between the individual and society, which is further exemplified through his consideration of individual style:

We all have our individual styles in both conversation and considered address, and they are never the arbitrary and casual things we think them to be. There is always an individual method, however poorly developed, of arranging words into groups and of working these up into larger units. It would be a very complicated problem to disentangle the social and individual determinants of style, but it is a theoretically possible one. (pp. 903–4)

Some linguists have given a more prominent place to writing alongside speech than Sapir (e.g. Coulthard, 2004) whilst for others, the term “style” is instead a recognised term for “idiolect in writing” (e.g. Kredens, 2002). In the context of this research, idiolect should be understood to include written language.

A good early definition for the discussion of idiolect is Hockett’s (1958): “the totality of speech habits of a single person at a given time constitutes an idiolect” (p. 321). Hockett’s definition raises two issues: potentially, one might need to observe and catalogue every single speech habit before one could fully characterise an individual’s idiolect, and secondly that idiolect will change over time. In so far as totality means complete and entire, Hockett appears to suggest that idiolect is the entire repertoire of speech habits available to a single person. However, it is impossible to collect a totality, although for Hockett’s purposes this would not have been an issue. In fact, in a later paragraph, Hockett notes that the entire idiolect cannot be observed, only examples of the linguistic output that it generates (1958: 322). In other words, rather than being able to observe the totality of habits, all that the linguist can observe is what a speaker or writer actually does at the particular point of observation.

The second implication of Hockett’s definition, that idiolectal features can only be described at a given time, implies that idiolect is organic and evolutionary in nature and will differ when observed at different times. This raises the question of by how much and whether the difference is significant. Related to this is the issue of the rate at which such change occurs, a question which so far has received no definitive answer with the exception of Bel et al. (2012) who show that the use of bigrams and trigrams do not vary substantially across a span of between six and ten years for individual authors (ages unknown), indicating that this feature remains sufficiently stable for this limited period of time at least. Corroborative evidence is provided by Barlow (2010) who found that bigrams were used consistently over the shorter period of one year in the spoken language of White House Press Secretaries. The problem with such a definition for forensic purposes is that not only would the idiolect of one individual differ from that of another (as assumed in authorship attribution, although cf. Grant (2010) for an alternative view on the importance of idiolect for forensic purposes), but would also be subject to variation between the same individual when observed at different points. This would make the comparison of documents in the forensic context very difficult because Known Documents (those documents whose authorship is attested, henceforth KD) are rarely authored at the same time as each other or as the Questioned Documents (those documents whose authorship is unknown and under suspicion, henceforth QD).

Sixty years later, Louwerse (2004) claimed that writers “implicitly leave their signature in the document they write” and that idiolects “are person-dependent similarities in language use” (p. 207). He explains that if idiolect exists, texts composed by one author will show more similarities in language than texts composed by different authors (p. 207). However, a potential problem arises in relation to Hockett’s definition. Louwerse states that similarities between texts produced by one author will be greater than texts produced by different authors. Hockett proposes that idiolect will change over time. Unless the individual signatures upon which Louwerse’s definition relies remain static, the similarities between two pieces of writing by the same individual at different times could be no greater than the similarities between two individuals with similar linguistic backgrounds (a common assumption, e.g. Loakes, 2006). It seems then that the temporal dimension could indeed be a confounding variable in forensic authorship attribution. Through examining a third definition of idiolect, a clearer picture may be gained.

Coulthard (2004) also says that “every native speaker has their own distinct and individual version of the language they speak and write, their own idiolect” and that “this idiolect will manifest itself through distinctive and idiosyncratic choices in texts” (pp. 431–2, original emphasis). The main difference here is between what Coulthard refers to as choice and what Hockett refers to as habit. Insofar as choice implies conscious decision, habit implies an involuntary behaviour pattern. If idiolect is based on habit, it is reasonable to argue that a person’s linguistic patterns will remain constant, until such a time when that habit is changed. In this scenario, texts produced during a period when the habit remains the same should be comparable. Choice, however, is more volatile and dependent on many extra-linguistic factors (e.g. mood of the individual, genre of the text, audience of the text, time available to compose the text, and indeed recency) as well as conscious attempts to disguise identity. As such, any features of language that are subject to choice could result in differences between texts produced by the same author, regardless of when they were authored.

These three definitions, somewhat representative of the many that could have been reviewed (e.g. Labov, 1972; Trudgill, 1974, 2003; Wardhaugh, 2006) capture between them the key issue for authorship attribution, namely, the extent to which an individual’s idiolect really is a reliable signature irrespective of stylistic choice and change over time. However, it is readily acknowledged that the theory of idiolect, to date, lacks empirical investigation (e.g. Kredens, 2001, 2002; Louwerse, 2004; Kniffka, 2007), and, as discussed above, the totality of linguistic habits for each person can never fully be observed. This point is echoed by Coulthard (2004) who says that “any linguistic sample, even a very large one, provides only very partial information about its creator’s idiolect” (p. 432), so it remains a largely theoretical notion. In light of its theoretical basis, it is now necessary ...

Table of contents

• Cover
• Title
• 1  Introduction: The UNABOM Investigation
• 2  The Web as Corpus and Authorship Attribution
• 3  Attributing Documents to Unknown Authors Using Idiolectal Co-selection and the Web
• 4  Attributing Documents to Candidate Authors Using Idiolectal Co-selection and the Web
• 5  The Suitability of the Web for Producing Forensic Evidence
• Bibliography
• Index