See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/350726160

Corpus approaches to forensic linguistics

Chapter · January 2021

DOI: 10.4324/9780429030581

CITATIONS READS

5 642

1 author:

David Wright
Nottingham Trent University
25 PUBLICATIONS   142 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Street Harassment of Secondary School Students View project

Lay-witness voice description View project

All content following this page was uploaded by David Wright on 08 April 2021.

The user has requested enhancement of the downloaded file.

Wright, D. (2021) Corpus approaches to forensic linguistics. In M. Coulthard, A. May and R. Sousa-Silva
(eds.), The Routledge Handbook of Forensic Linguistics (2nd edn). London: Routledge, pp. 611-627.

37
Corpus approaches to forensic linguistics
Applying corpus data and techniques in forensic contexts

David Wright

Corpora in forensic linguistics

Corpus linguistics is ‘the study of language based on examples of real life language use’
(McEnery and Wilson 1996: 1), with the examples collected, stored and analysed as a corpus
(pl. corpora). Corpora can run into millions or even billions of words, and therefore require
the use of specialised software to quantitatively and qualitatively analyse them. Corpus
linguistics is a set of methods and procedures that can be applied in the analysis of a range of
texts and contexts that forensic linguists may be interested in examining.
Since the advent of modern-day corpus linguistics, many fields have benefitted from its
ability to identify patterns in text, add evidence to support qualitative analyses and explore
large datasets in ways not previously possible. However, uptake in forensic linguistics has
been relatively slow. This is likely due to a number of factors, not least the fact that the types
of data that forensic linguists work with are often not in abundance. Whether it is courtroom
or police interview transcripts, or evidential texts such as text messages, emails, letters or
threats, the data (at least in most parts of the world) is scarce, and many researchers spend
years to source and collect precious datasets, often after developing close working
relationships with organisations or individuals who have access to data that are otherwise in
short supply.
Some of the earliest and most seminal work in forensic linguistics is corpus-based. In the
work which coined the term ‘forensic linguistics’, Svartvik (1968) used a corpus approach to
analyse a set of disputed witness statements in a murder case. Similarly, in his analysis of the
Derek Bentley statement, a watershed case for forensic linguistics, Coulthard (1994) used
specialised corpora of ordinary witness statements and police statements, along with the
much larger spoken element of the COBUILD corpus, to question the authorship of Bentley’s
disputed statement. With a few notable exceptions, including early adopters of corpus
techniques such as Kredens (2002) in authorship analysis and Cotterill (2003) and Heffer
(2005) in courtroom discourse analysis, there was relatively little corpus linguistic work in
forensic linguistics in the twenty years since Coulthard (1994). The second decade of the
twenty-first century, however, has seen a healthy increase in the amount of corpus-based
forensic linguistics work across a range of research areas. Such work has demonstrated three
affordances offered to forensic linguistics by corpus approaches: (i) they allow for new
perspectives on familiar genres, (ii) they offer possible solutions to methodological
challenges, and (iii) they open up brand new avenues for research.
In terms of new perspectives, corpora and corpus methods have allowed for new
approaches to familiar types of discourse in forensic linguistics. For instance, corpus
linguistics has expanded the horizons of work in legal language, and the collection and
analysis of large datasets of legal language have provided new insights to our understanding
of legal texts. An example is Goźdź-Roszkowski and Pontrandolfo’s (2017) collection of
research investigating phraseology, lexical bundles and formulaic sequences in monolingual,
multilingual and translated legislative and judicial discourses (see also Finegan and Lee;
McAuliffe, this volume). Similarly, qualitative work in courtroom discourse has been
augmented by the quantitative analysis made possible by corpus techniques, as in the
examination of important high-frequency individual words and their usage across single trials
or a number of trials (e.g. Tkačuková 2015; Szczyrbak 2016). In addition, corpora have been
built and analysed which comprise some of the more underexplored aspects of trial discourse,
such as closing arguments or summing up (e.g. Johnson 2014; Felton Rosulek 2015;
Matoesian and Gilbert, this volume). Corpus linguistics can also provide new solutions to
existing methodological challenges in forensic linguistics. One such area is statutory
interpretation, where corpus techniques provide a transparent and rigorous alternative to the
traditional reliance on judge’s intuition, dictionaries and etymology (e.g. Lee and Mouritsen
2017; Solan and Gales 2018; Gries, this volume). Corpus linguistics has also been harnessed
in authorship analysis research as a means by which to combine elements of existing stylistic
and stylometric approaches, in terms of supporting or explaining statistical results of
authorship attribution with a qualitative examination of author style (e.g. Wright 2017; Nini
2018). Meanwhile, Grant (2017) demonstrates how using a corpus method can be a valuable
strategy in determining the meaning of slang words. The building and analysis of certain
types of corpus have made available, for the first time, new datasets and investigations into
areas of forensic linguistics not previously possible. One notable example is the digitization
and online publication of The Old Bailey Proceedings (1674–1913) which has given rise to
new insights into the historical courtroom (e.g. Archer 2014; Johnson 2018). Similarly, there
now exist analyses of previously unexamined genres, such as malicious and threatening
communications (e.g. Gales 2015), trolling (Hardaker, this volume) and online grooming
(e.g. Chiang and Grant 2017).
Drawing on three case studies, this chapter aims to support the integration of corpora and
corpus techniques into forensic linguistics by discussing some of the important considerations
in building and designing forensic corpus data and demonstrating the ways in which
established corpus linguistic techniques can be used in the analysis of such data.

Three case study corpora

The three corpora drawn on in this chapter (Table 37.1) are ‘specialised’ corpora (Flowerdew
2004: 21) in that they all represent specific but different text-types, they each comprise data
from a particular discourse community, and they were all collected to help answer precise
research questions and aims. However, some readers may not consider these corpora to be
‘forensic’. They do not consist of legal texts nor are they of texts or genres at any stage of the
legal process, such as police interviews or courtroom talk, and they are not illegal or illicit
texts and they do not constitute evidence in criminal or civil proceedings. They are, however,
bound by their shared aim of improving ‘the delivery of justice through language analysis’
(MacLeod and Grant 2017: 173):
Table 37.1 Details of three case study corpora used in this analysis

Seduction Forum Adolescent Enron Email

Corpus (SFC) Harassment Reports Corpus
(AHR)

Method of Data scraped from web Dataset collected Data released online,
collection using Python script through a web-based collected and pre-
app accessed by processed
participants

No. of texts 25,788 posts 61 reports 63,369 emails

No. of tokens 26,527,412 1,512 2,462,151

Public Yes No Yes

domain

• The ‘Seduction Forum Corpus’ (SFC) was collected to examine the discourses of
resistance and sexual consent in a specialised online community, and to determine
whether language used within this community constitutes the incitement of violent
offences against women and girls.
• The ‘Adolescent Harassment Reports’ (AHR) corpus was gathered to provide some
initial insight into the street harassment experienced by young people aged 11 to 15 in
England, in terms of what happened and how they responded.
• The ‘Enron Email Corpus’ is used to develop new methods of authorship analysis
which attempt to combine quantitative and qualitative approaches to analysing
authorial style.

Despite their shared aims, the three corpora diverge from one another in various ways. The
differences in the means by which these data were collected, their size and composition and
the necessary ethical considerations represent important points when sourcing and building
forensic corpora.

Corpus collection

The internet provides a rich set of possible data sets for forensic linguists. The Seduction
Forum Corpus (SFC) and the Enron Email Corpus both already existed online, but were
sourced in different ways. The SFC is made up of an entire discussion forum from the online
‘Pick-Up Artist’ community, a community committed to studying and applying the art of
pickup and seduction. The forum is a popular one but has been given a pseudonym in this
chapter to protect the identities of the forum members (discussed in more detail below). The
data were collected from the forum using GNU wget web-scraping scripts, and saved as .json
and .txt files. Meanwhile, the Enron Email Corpus was originally released into the public
domain in 2003 by a federal judge as part of a database of 1.6 million documents following a
criminal investigation surrounding the bankruptcy of the company. The corpus used here
draws on that collected and prepared by Carnegie Mellon University (Cohen 2009) and
contains emails sent by 176 Enron employees and is stored as individual emails and author
sub-corpora in .txt format. Both SFC and the Enron corpus required some pre-processing and
‘cleaning’ before they were suitable for their respective (forensic) purposes. With both
corpora, duplicate texts, web-associated HTML tags and unwanted metadata were all
removed, leaving only the text and metadata that were relevant and useful for the analysis. In
contrast, the texts in the Adolescent Harassment Reports (AHR) corpus were collected
directly from adolescents who, as part of a project on street harassment (Betts et al. 2019),
were invited to report any experiences of harassment they had in a given six to eight-week
period by using a web-based app. The respondents were asked a series of multiple-choice
questions about their experiences and were given the opportunity to describe the event(s) in
their own words in a free-text comment box.

Corpus size

The corpora also vary in size. The SFC is the largest at over 26 million tokens, followed by
the Enron Email Corpus at almost 2.5 million, with the AHR corpus at only 1,500. The latter
may be seen to be pushing the boundaries of what can be considered a ‘corpus’ as it contains
fewer words than are demanded by many undergraduate essays. But given that a corpus is a
‘collection of texts’ stored electronically (Baker 2006: 48), the collection of 61 reports
satisfies this criterion. This does, however, raise an important question to consider when
building a corpus for forensic purposes – how big should a corpus be or, perhaps, how small
can it be? There is no straightforward answer to the optimum size of a corpus; rather, its size
will be determined by a combination of the research questions being asked of it and
restrictions of practicality. For example, the purpose of building the SFC was to observe
discourses within the Pick-Up Artist community, and therefore any corpus used needed to
represent at least one part of this community. At the same time, given its online nature, there
were no restrictions on how much could be collected. In contrast, the size of the Enron corpus
is limited to the number of emails in the original set released by the courts, and so the corpus
was collected in its entirety to obtain any and all available data for each author. The AHR
corpus was collected with the aim of learning as much as possible about children’s
experiences of harassment, and its collection was determined by practical restrictions such as
access to participants, the willingness of participants to volunteer their experiences and the
time-consuming nature of collecting such data.

Ethical considerations

Given the nature of the field, it is likely that forensic linguists will be attracted to data that
is sensitive in some way, and for which there are likely to be important ethical and privacy
considerations. This is true of all three case study corpora. First, and most straightforwardly,
while the names and company email addresses of employees in the Enron corpus are visible,
there is no other personally identifiable information included such as addresses or social
security numbers, and Enron employees were able to request the removal of any emails from
the dataset before it was released by FERC. At the time of collection, the forum from which
the SFC is taken does not require registration to view posts; it is entirely accessible for the
public to view online in the clear web (as opposed to the ‘dark’ web) and is indexed by major
search engines. Therefore, although the members of the forum cannot give informed consent
to the use of their posts, given that the forum is in the public domain it is not likely that they
expect their posts to be hidden or private. Nevertheless, in the preparation of this corpus for
analysis a series of additional steps have been taken to protect the identities of the forum
users. The name ‘Seduction Forum Corpus’ omits the actual name of the forum, posts will be
presented without usernames and any googleable verbatim quotes or extracts from the forum
will be avoided. Lastly, when collecting the AHR corpus, as with any fieldwork
methodology, consent was obtained from participants. Because the data was being collected
from people under the age of 18, consent was also obtained from the headteacher of the
schools that were part of the project, and letters were sent to the children’s parents to inform
them of the study and to invite them to inform the schools if they did not want their child to
participate in the research.

Applying corpus techniques

The core tools, techniques and principles of corpus linguistics can be applied to each of the
three case study corpora. The analyses here draw on five core principles of corpus linguistics:
keywords, collocates, concordances, word clusters and part-of-speech tagging. Each of these
techniques is applied to one of the three corpora to demonstrate how they reveal different
things about the data and can help provide answers to forensically-relevant research aims.
The corpus software used for this analysis is Wordsmith Tools version 7 (Scott 2016).

Keywords, collocates and concordances

Pick-Up Artists and their practices have been described as ‘a movement that teaches men to
assault and harass women’ (Ratchford 2017), and the techniques used in achieving sexual
‘success’ with women have been determined as moving from ‘a seduction script, focused on
conversation and comfort, to a more aggressive and coercive approach reflecting
characteristics of a rape script’ (Denes 2011: 418). The analysis of this corpus seeks to
identify whether any of the discourse present in SFC constitutes the incitement of sexual
violence against women.
At over 26-million words, finding an appropriate place to start with an analysis of the SFC
poses a formidable task. This is especially true for a corpus and a discourse community which
the researcher is not familiar or acquainted with. Doing a direct search for specific words (e.g
‘rape’) risks missing more important and pervasive words and themes in the data and, in this
case, overlooks the potentially indirect nature of inciteful language. A keyword analysis,
however, can be a valuable first step in probing very large corpora, providing an overview of
the lexical composition of the corpus and identifying words which are suggestive of
potentially meaningful patterns and that offer routes for further exploration (Archer 2009: 2).
Although a straightforward frequency list will invariably be dominated by grammatical
words, and even the most common lexical words may not reveal anything meaningful about
the discourse under examination (Baker 2006:123), a keyword analysis identifies words
‘whose frequency is unusually high in comparison with some norm’ (Scott 2016). These
words often reflect the main concepts, topics or themes in a text or corpus. Table 37.2 shows
the top 50 keywords emerging from the SFC when compared against the 450 million-token
Corpus of Contemporary American English (COCA) (Davies 2012).
A number of semantic and grammatical categories emerge from the top 50 keywords into
which almost all of the words can be grouped, and which help the analyst characterise the
nature and content of the corpus. First, a wide variety of pronoun forms appear as keywords,
including first person (I’m, I’ve, me, I’d, I’ll), second person (you’re, you’ll, you’ve) and third
person pronouns (it’s, she’s, her, he’s, they’re). The proliferation of personal terms is
reflective of the involved and personal narrative registers represented in online texts of these
kinds, particularly given that PUA forums are regularly used by members to share their tales
of sexual encounters with women and to give advice, as in (1).
Table 37.2 Top 50 ranked keywords in the Seduction Forum Corpus (using log-likelihood)

N key word Freq. % N key word Freq. %

1 i'm 81659 0.31 26 they're 11480 0.04
2 girls 107658 0.41 27 pussy 11723 0.04
3 it's 74304 0.28 28 sex 26460 0.10
4 girl 96673 0.36 29 fucking 14800 0.06
5 i've 43897 0.17 30 tinder 9960 0.04
6 you're 40562 0.15 31 he's 9157 0.03
7 game 78865 0.30 32 you'll 7881 0.03
8 her 267993 1.01 33 you've 7742 0.03
9 she's 31512 0.12 34 beta 8946 0.03
10 guys 53882 0.20 35 ass 11343 0.04
11 me 171663 0.65 36 dude 8941 0.03
12 my 185770 0.70 37 lol 7071 0.03
13 that's 26392 0.10 38 fucked 8179 0.03
14 fuck 28342 0.11 39 dating 11492 0.04
15 shit 29159 0.11 40 am 30738 0.12
16 women 75120 0.28 41 getting 29871 0.11
17 bang 21518 0.08 42 ltr 5815 0.02
18 i'd 17900 0.07 43 dudes 6370 0.02
19 chick 18914 0.07 44 banged 7001 0.03
20 get 108182 0.41 45 gonna 5726 0.02
21 guy 39365 0.15 46 what's 5780 0.02
22 chicks 16740 0.06 47 banging 6948 0.03
23 i'll 14742 0.06 48 pretty 20979 0.08
24 there's 12746 0.05 49 alpha 7575 0.03
25 date 24778 0.09 50 text 14369 0.05

(1) My first bit of advice is that you don’t show your feelings for this girl. If you really
want to do this river walk then fine, but you better get the kiss. I get the romantic vibe
you’re trying to create, but get her back to your place afterwards.

Second, besides the third person pronouns she and her, there are other keywords which refer
to females (girl(s), women, chick(s)) and other men (dude(s), guy(s), alpha, beta), which
reflect the purpose of the SFC. Finally, a semantic category of keywords which emerges
contains those words related to dating and sex, the main topic under discussion in the forum,
and these reflect a preoccupation with sexual conquest (game, fuck(ing), fucked, pussy, sex,
ass, bang(ing/ed), date, dating, tinder, ltr, text). Some of these terms are community-specific,
such as the acronym ltr which stands for long term relationship.
Keywords are generally content words (Baker 2006: 127), and so it is notable that her, a
function word, is the eighth keyword in SFC. Bearing in mind the aim of this analysis is to
examine potentially inciteful discourses, the markedly high frequency of her warrants closer
attention. A collocation analysis identifies which words commonly co-occur with a word
under examination, and analysing these collocates can provide an insight into the contexts in
which it is used and the discourses of which it is part. Her serves grammatically as both the
object in a clause (e.g. I saw her) and as a determiner within a noun phrase (e.g. I saw her
face). A frequency analysis of the collocates immediately before and after her reveals which
verbs most frequently affect her, and which nouns are most commonly premodified by her
(Table 37.3).
 Table 37.3 Most frequent L1 VERB and R1 NOUN collocates of her in SFC

L1 Collocates R1 Collocates
N Word Freq. N Word Freq.
1 tell 5231 1 number 3583
2 get 4787 2 friends 3147
3 told 4607 3 friend 2363
4 let 2944 4 face 1934
5 give 2741 5 place 1796
6 make 2625 6 ass 1624
7 see 2564 7 phone 1553
8 fuck 2499 8 body 1288
9 take 2272 9 pussy 1272
10 ask 2229 10 mind 1202
11 bang 1684 11 head 1139
12 asked 1678 12 life 1093
13 call 1575 13 hand 1072
14 met 1564 14 boyfriend 1006
15 have 1452 15 eyes 937
16 text 1358 16 hair 906
17 like 1257 17 mouth 865
18 keep 1129 18 name 818
19 took 1124 19 parents 700
20 want 1054 20 family 677

The lexical verbs of which her is most commonly the object in SFC can be organised into
material processes (get, let, give, make, fuck, take, bang, met, keep, took), verbal processes
(tell, told, ask(ed), call, text) and mental processes (see, like, want). Some of these verbs are
unsurprising given the nature of the forum, as they relate to communicating, meeting or
sexual involvement with women and girls (e.g. call, text, met, like, fuck, bang). However,
investigating some others more closely reveals some more malicious discourse. The most
common material process get, for example, shows how the get + her collocation gives rise to
some particularly predatory discourses:

1 If I get her alone with me, it’s in the bag.

2 Always in busy places, the gym, get her drunk, get physical, get her home.
3 Even if she tells me she just wants to be friends, I try to get her drunk and isolated.
4 I was thinking go to the hotel, get her drunk and try and get her to an empty room.

These concordance lines show get + her in context and represent common patterns in the use
of this collocation in SFC. Namely, there is evidence here of forum members advising each
other to make women vulnerable before making sexual advances towards them by getting
them alone, drunk or both. Get her drunk and get her alone occur 69 and 42 times
respectively in SFC, suggesting that isolating and lowering the inhibitions of their female
‘targets’ is considered a helpful step in their pursuit of sexual gratification.
 Similarly, make as the sixth most frequent L1 verb collocate is used within some patterns of
unambiguous physical and sexual manipulation in which men advise each other to force
women to behave in certain ways for their own sexual benefit:
1 I'm sure I can make her do whatever I want with time.
2 Make her do the sluttiest, most degrading shit you can imagine when you see her.
3 Make her do anal or get the fuck out.
4 I usually make her suck my dick while she’s doing all that.
5 I get hard and make her suck my dick. Then I bang her as the television plays.
6 Then you can make her your sex slave, for you and only you.
7 But in the bed room you can make her your whore.

At the very least, this sample of concordance lines provides an insight into the male-
dominated, patriarchal and misogynistic ideologies that underpin much of the discourse in the
SFC. Forum members are encouraged to act as though women are their property, their ‘sex
slave’ and ‘whore’. Related to this are the examples from the corpus in which forum
members appear to be encouraging other members to make women and girls perform certain
sexual acts. Such acts are either packaged in vagueness, such as ‘make her do whatever I
want’ or can be explicit and specific such as ‘make her do anal’ and ‘make her suck my dick’.
The specific choice of the pattern make her is notable insofar as it denotes force and
exclusive male agency in causing something to happen, which in turn raises questions over
consent in such circumstances.
Shifting attention to the R1 collocates of her, in which her is a determiner, the most
dominant category of noun in SFC is words referring to various parts of the body: face, ass,
body, pussy, hand, eyes, hair and mouth. While attention may be more immediately drawn to
those sexualised body parts (i.e. ass, body, pussy), those which are less overtly related to sex
are in fact used as part of highly sexualised discourses:
1 Do me a favour and spit in her face if you ever get a chance.
2 Then put your dick in her face and tell her what you want her to do.
3 I shoved her face in my crotch and she gave me a blow job.
4 Make out and then push her head toward your cock.
5 Just shove your dick in her mouth and tell her whatever you think she wants to hear.
6 How about grabbing her by her hair and shoving your cock in her mouth?
7 I choked her and pulled her hair hard enough so that it really hurt.
8 Don’t forget to pull her hair and choke her.

Face, head, mouth and hair appear a combined total of 4,844 times as immediate R1
collocates of her. These words alone are relatively innocuous, but as even the very small
sample of concordance lines here show, they are used within wider discursive patterns in SFC
which show aggressive and abusive behaviours towards women and girls being
recommended and encouraged. This includes the suggestion of forcing women to engage in
sexual acts.
One final her body-part collocate worthy of note here is hand, which appears with a
frequency of 1,072. A concordance analysis reveals a pattern of behaviour reported by forum
members in which they force a woman’s hand to touch their penis, or wherein they advise
such a course of action in the pursuit of sexual intercourse:

1 You have to take her hand and put it on your cock.

2 I take her hand and put it on my dick. I basically push it as far as she will go.
3 just before she resists just take her hand and put it on your dick.
4 Ten mins later take her hand and put it on your dick.
Forcing someone to engage in sexual activity in this way, without consent, amounts to sexual
assault and therefore the encouragement to do just that amounts to incitement to commit a
sexual offence.
The forensically-motivated aim of this analysis was to identify whether there are discourses
present in SFC which represent incitement of violence against women and girls. Although it
has only been possible here to scratch the surface of a 26-million word corpus, the analysis
has demonstrated the ways in which keywords, collocations and concordances are useful
tools for the analyst in probing and navigating a large and unfamiliar dataset. Importantly, her
and subsequent collocates are unremarkable words at first glance, but a closer inspection of
how these words and collocates are used within this community has revealed patterns of
predatory and potentially abusive discourses emerging from the forum, glorifying sexual
aggressiveness, and in which violent sexual offences against women and girls seem to be
reported, encouraged and incited.

Part-of-speech tagging

Often the types of corpora that forensic linguists are able to collect and analyse are much
smaller than the 26-million words of the SFC. That is the case with the Adolescent
Harassment Reports (AHR) corpus, which contains 61 reports of harassment experienced by
young people aged between 11 and 15, and totals only 1,512 words. Even with such a small
amount of data, there are techniques from corpus linguistics that can expedite useful
analytical procedures that would be time-consuming to perform manually. Part-of-speech
(POS) tagging is one such technique, in which automated software is used to ‘tag’ each word
in a corpus for its part-of-speech (or ‘word class’). This subsequently renders the corpus
searchable by grammatical patterns as well as lexical ones, and because POS tags appear at
much higher frequencies than individual lexical items, this can be particularly useful for
smaller datasets with fewer overall tokens.
The AHR corpus was collected with the aim of understanding the types of street
harassment young people experience and how they respond to it. Central to this analysis,
therefore, are the verbs which are prevalent in the reports submitted. The AHR corpus was
tagged using the free-to-use browser-based CLAWS tagger (Garside and Smith 1997).
Despite having an accuracy rate of 96–97% (Garside and Smith 1997: 119), the tags assigned
to each word in the AHR corpus were manually checked for accuracy, given that some of the
reports included non-standard spellings. An example of a POS-tagged report is shown in (2)
(PNP stands for personal pronoun, CJC for conjunction, AV0 for adverb and PRP for
preposition; all the V codings are for verb forms).

(2) They_PNP stopped_VVD and_CJC beeped_VVD and_CJC looked_VVD

then_AV0 waved_VVN at_PRP me_PNP

In the 60-tag set used in CLAWS, there are 25 different verb tags, all of which begin with
‘V’. Therefore, a wildcard query for ‘V*’ was run in Wordsmith Tools to identify all of the
verbs present in the reports. While the verbs in the reports which have subjects other than the
young people reflect the types of harassment they were victim to, the verbs for which they are
the subjects indicate what the victim themselves did before, during and after the harassment
incident. The corpus query found 137 different verbs in the corpus, totaling 341 tokens.
These verbs were then divided into those for which the young people themselves were the
subject and those for which they were not.
The verbs for which the young people themselves were not the subject were manually
categorised according to the different types of harassment to which they related (Table 37.4).
Note that the original spellings of words has been preserved in the table and in the reports
presented here. The harassment type most commonly reported in the free-text comments are
incidents involving people in cars beeping at the children, slowing down or stopping
(Example 3). These are closely followed by incidents with some verbal interaction between
the harasser and the child (4). This often involves the child being called names and men
initiating a dialogue with young girls, complimenting them and inviting them into their cars.
Next are types of harassment that do not involve any verbal interaction, in which children are
watched, stared at, and pictured/videoed by someone on a mobile phone (5), or smiled, waved
or pointed at (6). In a small number of cases, children are followed, chased and even cornered
by their harassers (7), while acts of physical aggression are relatively rarely reported by
children (8).

Table 37.4 Types of harassment expressed by verbs in the AHR corpus

Harassment type Freq. Verbs

Involving vehicle 29 beeped, stopped, slowed, turned (around), drove,
speeding, horned, honked, driving, cycled, curbed.

Verbal interaction 25 said, called, shouted, laughing, asked, whispered, told,

shouting, say, lafed, convincing, calling.

Being watched 15 looked, stared, watching, took, videoing, papped, staring.

Non-verbal interaction 12 smiled, waved, stuck, shrugged, showed, pulling (faces),

pointing, pointed, whistled.

Being followed 9 followed, follow, following, cornering, coming, chased.

Physicality 5 grabbed, yanked, threw, hit.

(3) we crossed the road and the guy beeped his horn and stuck his middle finger up.

(4) I was walking to my friends house and I got horned at and I looked and two guys
(men) asked me 'how old are you babe? It made me feel completely disgusted.

(5) I was with my sister and a friend, we were sitting at the back of the bus and was
stared at a while after this fat dark skin male took out his camera and started
videoing us we covered our faces and moved the [sic] reported his actions.

(6) A taxi driver was driving past then he slowed down and did a creepy smile at my
face.
 (7) I was in a shop in [the town] and this man started staring at me and my frinds [sic] and
when we were about to leave he started to get his things and he started to follow us
we ran away and we lost him.

(8) I was followed and once touched on my boobs and so people was laughing at me, I
felt scared about it. And I was scared that they would take me.

Although there is a growing understanding of the types of stranger or street harassment that
adult women are most often victims of (e.g. Kearl 2010), much less is known about the
harassment experiences of adolescents. The results from this analysis therefore, albeit of a
relatively small sample of reports, provide some important initial insights into the most
common patterns of street harassment that young people are victim of. Indeed, it is perhaps
notable that such clear patterns of harassment emerge even though the sample of reports is so
small.
At the same time, although there has been a good deal of research attention on the
interactional and behavioural techniques and strategies adult women use to cope with or
respond to stranger harassment (e.g. Fairchild and Rudman 2008), there is less clarity on
what resources young people rely on when faced with street harassment. An analysis of the
verbs present in the AHR corpus for which the young people themselves are the subject
provides some answers to this (Table 37.5).

Table 37.5 Verbs attributed to themselves in young people’s reports of harassment in AHR
corpus

Young person’s actions Freq. Verbs

Before event 28 walking, going, sitting, waiting, riding, left, leave,

crossed, talking.

Response to event 26 ran, walked, moved, went, run, move, lost, cycling,
covered, said, shouted, call, spoke, reported, rang.

The young people reported themselves as subjects of verbs both before and after the
harassment event. Those verbs reported before young people are harassed provide us with the
immediate setting of the incident, and they almost all see the young person (and often their
friends) walking or going to and from school, the shops, or each other’s houses, as in (9).

(9) I was walking to school and out of my bag and some man grabbed it

By contrast, there is more variation in verbs used in the reports after the harassment and these
reflect the young people’s responses to what they had experienced. Most often, they took
passive or evasive action, including running, moving or walking away (Example 10), which
is in line with existing research on women’s coping strategies for stranger harassment
(Saunders et al., 2017). However, in addition to these passive strategies, children also report
taking more active measures, including confronting their harassers (11), telling their parents
(12) and, in a very small number of cases, informing the police (13). Previous research has
found that these active strategies are less frequently employed by adults in response to
harassment for fear of escalating violence or being disbelieved.

(10) I was in a shop in [the town] and this man started staring a me and my frinds [sic] and
 when we were about to leave he started to get his things and he started to follow us we
ran away and we lost him.

(11) ther was car on rode and he wasent moveing so I went behind it to corss and it nearly
ran me over so I cross and sed what the hell man and he stered and lafed and drove
away.

(12) When me and my best mate was walking some man in a car went by us, turned
around and came back to follow us in his car. He said 'How old are you' my mate said
'why' and he said your beautiful so then I rang my mum and we tried running away.

(13) I was with my sister and a friend, we were sitting at the back of the bus and was
stared a while after this fat dark skin male took out his camera and started videoing us
we covered our faces and moved then reported his actions.

These findings provide new and valuable information on the types of street harassment young
people experience in England and how they tend to linguistically and behaviourally react and
respond. From a forensic linguistic point of view, this not only addresses an offence of
(linguistic) harassment that has traditionally been overlooked by the field, but in practical
terms it can inform education, policy and legislation regarding how best to protect vulnerable
victims from harassment. This was a small sample, but the use of POS tagging revealed
patterns in the discourse that were integral to providing initial answers to the research
questions posed, that is the types of harassment experienced and the responses to it. POS
tagging provides forensic linguists with an additional ‘layer’ of corpus-derived results that go
beyond straightforward lexical patterns and can be applied to a dataset of any size.

Collocation profiles and word clusters

Quantitative approaches to authorship analysis which rely on feature sets at the lexical level
distinguish authors and texts from one another by comparing the relative frequencies with
which they use particular words or word combinations. In turn, the attribution of questioned
texts to a given author is made on the basis of how closely the relative frequency of these
features in the questioned texts matches that of a given known or training sample. Such an
approach relies on a very narrow view of linguistic variation; the frequency with which an
author uses a particular word (or set of words) does not account for differences in how they
use those words or word combinations, which may be very distinctive. A corpus approach
can bring to light stylistic differences between authors that may be overlooked by purely
quantitative methods.
Content words have typically been avoided as style markers in quantitative, stylometric
authorship analyses, as they are generally considered to be indicators of topic rather than
authorial style. Despite the promising results of studies which have drawn on content words
as well as function words (e.g. Jockers and Witten 2010), function words have largely been
relied on for lexical-level authorship analysis. However, in a case of disputed email
authorship within a company, Coulthard (2013) reports that distinctive use of content words,
and in particular register-related content words, were useful in determining authorship.
Drawing on the Enron Email Corpus, the analysis which follows shows that by using a corpus
linguistic approach, content words can be central to the investigation of idiolect. The specific
word which is the focus of this analysis is deal.
The word deal is embedded in the community register of Enron, as it relates to the core
business of Enron as an energy trading company. Deal in this context refers to the verb sense:
‘to carry on commercial transactions; to do business, trade, traffic’, and the noun sense: ‘an
act of dealing or buying and selling; a business transaction’ (Oxford English Dictionary). The
importance of this word in Enron is attested by its frequency. Deal occurs 4,134 times,
accounting for 0.17% of the total 2,462,151 tokens and is used by 125 of the 176 authors.
Importantly for the purposes of this analysis, the proportion of 0.17% represents our ‘base
rate knowledge’ (Turell and Gavaldà 2013) of how frequently deal is used in the Enron
population under examination. For context, deal occurs 87,551 times in COCA, accounting
for only 0.02% of the 464,020,256 tokens.
Authors in the Enron corpus can be distinguished from one another on the basis of how
frequently they use deal. For instance, 42 of the 176 authors use deal with a higher relative
frequency than that of the 0.17% base rate for the corpus. Amongst those 42 are two traders,
Kate Symes and Daren Farmer. Farmer uses deal 286 times in an email set totalling 24,389
words (1.17%). Symes uses deal 682 times in a much larger total email set of 58,577 words
(1.16%). Both of these relative frequencies are much higher than the base rate for the corpus,
and are in fact the third and fourth highest relative frequencies in the corpus. If two authors
who use deal fewer than ten times in much smaller email sets are excluded, Farmer and
Symes become the two most prolific users of deal. In terms of frequency alone, there is not
much to distinguish between Symes and Farmer. Any approach to analysing authorial style
that relied solely on lexical frequency may not distinguish between these two authors, and
therefore would not be able to correctly assign any questioned text to either. However, an
analysis of how they use the word quickly reveals marked differences.
The ‘collocational profile’ (Sinclair 1996) of a word captures the words that are most
frequently found in its immediate environment within a given span (in this case a ten-word
span, five words to the left and five to the right). In order to observe differences in the ways
in which Farmer and Symes used deal, their collocation profiles can be compared (Tables
37.6 and 37.7).
The words that are highlighted in bold are those which do not appear in the collocation
profiles of the other author, and even the most cursory glance shows that the two authors tend
to use deal in distinctive ways when compared with each other. There is not scope to discuss
all the differences here, but a sample has been chosen for closer inspection.
One immediate and important difference between the two authors as revealed by their
respective collocation profiles is that, whereas Farmer typically uses deal as an object,
Symes’ use of deal is normally as part of a longer noun phrase. For instance, one of Farmer’s
patterns which contains distinctive collocates in L1 and L2 position is his repeated use of the
I have VERB + deal, which he uses 21 times in total, with created sitting in the verb slot most
commonly (n=15), as in (14).

(14) Megan,
I have created deal # 1075281 to cover a sale from Cleburne to ENA. I also
repathed Unify […]

The other six instances see adjusted (twice), taken, extended, rolled and updated filling the
verb slot. As the collocation profile shows, Farmer prefers to use some of these other verbs,
namely extended and rolled without the auxiliary, which are found five and six times
respectively, as in (15).

(15) I rolled deal 150325 for the first 3 days of Jan. I expect this point to be
zero for the rest of Jan.
Table 37.6 L5 to R5 collocational profile for deal in Farmer’s data Enron Email Corpus
(produced using the ‘patterns’ feature in Wordsmith Tools Concord tool)

N L5 L4 L3 L2 L1 Centre R1 R2 R3 R4 R5
1 the the the the the deal ticket to the the the
2 is to to have this for for cover for of
3 deal we on to created to is this this deal
4 we you can on new with the is to in
5 to this of for to in you should in you
6 for should have of on has have for is to
7 have changed is pricing spot is should deal be on
8 you record deal change rolled and this sitara volume meter
9 volume and price corrected extended the in zero are mtr
10 volumes with and allocated term at under and gas flow

Table 37.7 L5 to R5 collocational profile for deal in Symes’ data Enron Email Corpus

N L5 L4 L3 L2 L1 Centre R1 R2 R3 R4 R5
1 to the to the the deal is been changed to the
2 the kate the on this has in the the to
3 thanks is just changed in numbers and to and it
4 this to is to peak entry is in it is
5 kate and and of of and to it is and
6 and this kate in last was be on kate with
7 in deal not entered forward number the mw thanks you
8 entered thanks i've is and in kate is as kate
9 deal mike this that that blotter that and at be
10 is was deal and correct to was deal deal for

These patterns are not found in Symes’ deal collocation profile and so serve to distinguish
Farmer from his colleague.
Meanwhile, although Farmer uses deal as the head of a noun phrase, he does so less
frequently than Symes. Most typically, he uses new deal (n=13) and spot deal (n=8).
However, the ADJECTIVE + deal pattern is far more characteristic of Symes’ collocational
profile. As shown in Table 37.7, peak deal (n=12), last deal (n=9), forward deal (n=9) and
correct deal (n=5) constitute her most frequent uses of deal but are not found at all in
Farmer’s emails (e.g. 16).

(16) This was a futures contract done by Matt Motley, and I entered the forward
deal. I wasn't aware that futures deals don't get confirmed. I only knew that
when the contract expires, I need to enter a forward deal in EnPower to hold
the trader's position […].

Difference in the collocational profiles of Farmer and Symes can also be observed in the most
frequent collocates to the right of deal in the authors’ data. There are major differences in the
R1 position, for instance, wherein the only collocate distinctive of Farmer in this position is
ticket, which he uses 14 times, as in (17).

(17) Nicole,
The correct price is 4.50. I changed the deal ticket.
This use of deal contrasts with Symes’; whereas Famer uses deal as a premodifier for ticket,
Symes’ most proliferate pattern is to use deal to premodify number(s), which she uses 57
times, but this pattern is not present at all in Farmer’s data (e.g. 18).

(18) Geir Solberg in Real Time is fixing this right now - I'll let you know the
new deal number in just a second.
Thanks,
Kate

Word strings and word n-grams are now commonly part of the stylometrist’s toolkit when
conducting statistical authorship analysis, and such chunking of text in this way will capture
an author’s patterns of use of particular words. However, the pre-occupation of such
approaches with frequency can overlook author-distinctive uses at the stylistic level. The
corpus-based concept of collocational profiles offers an alternative perspective on authors’
uses of a particular lexical feature; whereas the frequency with which a person uses a word
may not be evidentially significant, the ways in which they use that word may be. As the
analysis here has demonstrated, the collocational profiles of deal, an extremely common
word embedded within the register of this particular discourse community, can vary greatly
between authors (in this case, two people who have the same or a very similar job). Such an
approach can unlock the potential for distinguishing between authors using content words,
which, notwithstanding their inclusion in word strings, have been ignored by authorship
analysts.

Conclusion

The techniques demonstrated here are well-established in corpus linguistics and have been
applied in other fields of linguistics for decades. Depending on the nature of the data and the
aims of the research, they allow the forensic linguist to explore large, unfamiliar datasets,
they offer techniques for enriching smaller datasets to reveal otherwise hidden patterns, and
they provide a means by which quantitative and qualitative analyses can be combined.
However, the types of corpora forensic linguists are interested in are often hard to come by,
especially as the field remains pre-occupied with the analysis of ‘hidden’ genres related to the
legal process and criminal evidence. As the field expands, so too will the range of genres and
text-types that fall within the remit of forensic linguists whose aim it is to improve the
delivery of justice through language analysis. Indeed, the continued expansion of the field
may depend, at least in part, on the broadening of the types of data that are subjected to
forensic linguistic attention. None of the three corpora used in this chapter is ‘forensic’ in the
traditional sense; yet they all hold potential for improving the delivery of justice, whether by
gaining an understanding of the influence of online communication to incite violent offences
against women, by protecting young people from street harassment, or developing new
methods of authorship analysis to identify or eliminate suspect authors. The broadening of the
scope of forensic linguistics opens up the possibility of collecting new and diverse corpora
for forensic purposes and, as the recent surge in the use of corpus methods exemplifies,
adopting such methods offers exciting and previously unexplored directions for forensic
linguistics.
Acknowledgements

I would like to thank Ikechukwu Onyenwe for the technical support provided during the
collection and preparation of the Seduction Forum Corpus.

Further reading

Baker, P. (2006) Using Corpora in Discourse Analysis. London: Continuum.

Kredens, K. and M. Coulthard. (2012) ‘Corpus linguistics in authorship identification’, in P.
Tiersma and L. Solan (eds.) The Oxford Handbook of Language and Law, Oxford:
Oxford University Press, 504–516.
Larner, S. (2015) ‘From intellectual challenges to established corpus techniques: introduction
to the special issue on forensic linguistics’, Corpora, 10(2): 131–143.
Other chapters in this volume that draw on corpus methods and approaches are: Finegan,
Gries, Hardaker, and McAuliffe.

References

Archer, D. (2009) ‘Does frequency really matter?’, in D. Archer (ed.) What's in a Word-list?:
Investigating Word Frequency and Keyword Extraction, London: Routledge, 1–16.
——(2014) ‘Historical pragmatics: evidence from The Old Bailey’, Transactions of the
Philological Society, 112(2): 259–277.
Baker, P. (2006) Using Corpora in Discourse Analysis. London: Continuum.
Betts, L. Harding, R. Peart, S., Sjolin Knight, C., Wright, D. and Newbold, K. (2019)
‘Adolescents’ experiences of street harassment: creating a typology and assessing the
emotional impact’, Journal of Aggression, Conflict and Peace Research, 11(1): 38–
46.
Chiang, E. and Grant, T. (2017) ‘Online grooming: moves and strategies’, Language and
Law/Linguagem e Direito, 4(1): 103–141.
Cohen, W. W. (2009). Enron Email Dataset [online]. Retrieved from
http://www.cs.cmu.edu/~enron/.
Cotterill, J. (2003) Language and Power in Court: A Linguistic Analysis of the OJ Simpson
Trial, Basingstoke/New York: Palgrave Macmillan.
Davies, M. (2012) The Corpus of Contemporary American English: 450 million words,1990-
present [online] https://www.english-corpora.org/coca/
Heffer, C. (2005) The Language of Jury Trial: A Corpus-Aided Analysis of Legal–Lay
Discourse, Basingstoke/New York: Palgrave Macmillan.
Coulthard, M. (1994) ‘On the use of corpora in the analysis of forensic texts’, Forensic
Linguistics. International Journal of Speech, Language and the Law, 1(1): 27–43.
—— (2013) ‘On admissible linguistic evidence’, Journal of Law and Policy, 21(2): 441–466.
Denes, A. (2011) ‘Biology as consent: problematizing the scientific approach to seducing
women’s bodies’, Women’s Studies International Forum, 34(5): 411–419.
Fairchild, K. and Rudman, L. A. (2008) ‘Everyday stranger harassment and women’s
objectification’, Social Justice Research, 21(3): 338–357.
Felton Rosulek, L. (2015) Dueling Discourses: The Construction of Reality in Closing
Arguments, Oxford: Oxford University Press.
Garside, R. and Smith, N. (1997) ‘A hybrid grammatical tagger: CLAWS4’, in R. Garside, G.
Leech and T. McEnery (eds.) Corpus Annotation: Linguistic Information from
Computer Text Corpora, London: Routledge, 102–121.
Goźdź-Roszkowski, S. and Pontrandolfo, G. (2017) Phraseology in Legal and Institutional
Settings: A Corpus-based Interdisciplinary Perspective, London: Routledge.
Flowerdew, L. (2004) ‘The argument for using English specialized corpora to understand
academic and professional settings’, in U. Connor and T. A. Upton (eds.) Discourse in
the Professions: Perspectives from Corpus Linguistics, Amsterdam/Philadelphia:
John Benjamins, 11–33.
Gales, T. (2015) ‘The stance of stalking: a corpus-based analysis of grammatical markers of
stance in threatening communications’, Corpora, 10(2): 171–200.
Grant, T. (2017) ‘Duppying yoots in a dog eat dog world, kmt: determining the senses of
slang terms for the courts’, Semiotica, 216: 479–495.
Jockers, M. L. and Witten, D. M. (2010) ‘A comparative study of machine learning methods
for authorship attribution’, Literary and Linguistic Computing, 25(2): 215–223.
Johnson, A. (2014) ‘“Dr Shipman told you that ...” The organising and synthesising power of
quotation in judicial summing-up’, Language and Communication, 36: 53–67.
—— (2018) ‘“How came you not to cry out?” Pragmatic effects of negative questioning in
child rape trials in the Old Bailey Proceedings 1730–1798’, in D. Kurzon and B.
Kryk-Kastovsky (eds.) Legal Pragmatics, Amsterdam/Philadelphia: John Benjamins,
41–64.
Kearl, H. (2010) Stop Street Harassment: Making Public Places Safe and Welcoming for
Women. Santa Barbara, C.A.: Praeger.
Kredens, K. (2002) ‘Towards a corpus-based methodology of forensic authorship attribution:
a comparative study of two idiolects’, in B. Lewandowska-Tomaszczyk (ed.)
PALC’01: Practical Applications in Language Corpora, Peter Lang: Frankfurt am
Mein, 405–437.
Lee, T. R. and Mouritsen, S. C. (2017) ‘Judging ordinary meaning’, Yale Law Journal,
127(4): 788–879.
MacLeod, N. and Grant, T. (2017) ‘“go on cam but dnt be dirty’: linguistic levels of identity
assumption in undercover online operations against child sex abusers’, Language and
Law/Linguagem e Direito, 4(2): 157–175.
McEnery, T. and Wilson, A. (2001) Corpus Linguistics: An Introduction, Edinburgh:
Edinburgh University Press.
Nini, A. (2018) ‘An authorship analysis of the Jack the Ripper letters’, Digital Scholarship in
the Humanities, 33(3): 621–636.
Ratchford, S. (2017) ‘I tried to find out if Pick Up Artists are still influential in 2017’, Vice,
25 August 2017. [online] available from:
https://www.vice.com/en_au/article/j55bxd/i-tried-to-find-out-if-pick-up-artists-are-
still-influential-in-2017
Saunders, B. A., Scaturro1, C., Guarino, G., and Kelly, E. (2017) ‘Contending with
catcalling: the role of system-justifying beliefs and ambivalent sexism in predicting
women’s coping experiences with (and men’s attributions for) stranger harassment’,
Current Psychology, 36(2): 324–338.
Sinclair, J. M. 1996. ‘The search for units of meaning’, Textus 9(1): 71–106.
Scott, Mike. 2016. Wordsmith tools version 7. Stroud: Lexical Analysis Software.
Solan, L. M. and Gales, T. (2018) ‘Corpus linguistics as a tool in legal interpretation’,
Brigham Young University Law Review, 2017(6), 1311–1358.
Svartvik, J. (1968). The Evans Statements: A case for Forensic Linguistics, Gotëborg:
University of Gothenburg Press.
Szczyrbak, M. (2016) ‘Say and stancetaking in courtroom talk: a corpus-assisted study’,
Corpora, 11(2): 143–168.
Tkačuková, T. (2015) ‘A corpus-assisted study of the discourse marker well as an indicator of
 judges’ institutional roles in court cases with litigants in person’, Corpora, 10(2):
145–170.
Turell, M. T. and Gavaldà, N. (2013) ‘Towards an indexof idiolectal similitude (or distance)
in forensic authorship analysis’, Journal of Law and Policy, 21(2): 495–514.
Wright, D. (2017) ‘Using word n-grams to identify authors and idiolects: a corpus approach
to a forensic linguistic problem’, International Journal of Corpus Linguistics, 22(2):
212–241.

View publication stats