Did Hillary Clinton’s PR team solicit fake Amazon book reviews for ‘What Happened’? Part #2

In the first post of this two-part linguistic investigation, we set up an unsupervised analytical approach; factor analysis to identify latent dimensions of linguistic variation in the ‘What’s Happened’ reviews then feeding these dimensions into a cluster analysis in order to identify a small number of distinct text types. We know that the reviewing patterns for ‘What Happened’ displayed ‘burstiness’ i.e. a high frequency of reviews within a short period of time (see Figure 1 below).  As Figure 2 below hypothesises, if there is a text type cluster that displays similar ‘burstiness’, we can infer that there there was probably some level of coordination of reviewing behaviour and identify linguistic features associated with less-than-authentic reviews.

Screenshot 2018-10-28 11.00.04

Figure 1: Quantity, frequency and rating of ‘What Happened’ book reviews in the first month after launch.

Screenshot 2018-10-27 20.29.15

Figure 2: Hypothesis for fake review detection using cluster analysis with time series. 

The factor analysis found four dimensions of language variation in the ‘What Happened’ reviews: Engagement; Reflection; Emotiveness; Commentary.

Dimenson 1: Engagement

One linguistic dimension of these reviews describes levels of Engagement. In engaging reviews, writers directly address (using ‘you’ pronouns) either the reader or Hillary Clinton. The style is conversational and persuasive with exclamations, questions and hypotheticals used to interact with the reader.

THANK YOU for telling your story Secretary Clinton! You have accomplished so much and are a genuine inspiration. If they weren’t so afraid of you, they wouldn’t work so hard to shut you up. Keep fighting and I will too!

It’s her side of the story. That’s what it claims to be, and that’s what it is. For those who don’t like it because you disagree with her, you’re missing the point. After reading it, did you get a better feel for who the candidate was, what she was thinking, and even what her biases were and are? If so, then the book does what it claims to do.

Non-engaging reviews are more  linguistically dense, using longer words and giving complex descriptions of the content.

The second chapter describes the days after the election, when she first isolated herself from the deluge of texts and emails from well-wishers. Eventually, however, she threw herself back into the fray, writing letters of thanks to supporters, attending galas, and spending time with her family.

Dimension 2: Reflection

A second linguistic dimension sees reviewers reflect on their personal experience of reading the book. This may include autobiographical elements, narratives related to the book purchase and reading occasions as well as feelings had while reading. The key linguistic features here are ‘I’-pronoun and past tense:

Like many other people, I wondered if this book would really be worth reading. I voted for Clinton but I wondered how much value there could be in her account of the 2016 Presidential election campaign. Luckily, this book is so much more. It hit my Kindle on Tuesday and as it happens I had three airplane flights (including two very long ones) on Wednesday and Thursday, so I made it my project for those flights. I didn’t have to force myself to keep going; once I started, her narrative and the force of her ideas and anecdotes kept me reading.

Dimension 3: Emotiveness

Reviews with a high Emotiveness score were extremely positive in their praise of the book and, especially, Hillary Clinton. This was signalled by use of long strings of positive adjectives that might reasonably be considered excessive:

A funny, dark, and honest book by one of the truest public servants of her generation. Her writing on her marriage was deeply heartfelt and true. The sad little haters will never keep this woman down, and history will remember her as a trailblazer and a figure of remarkable courage.

The People’s President, Hillary delivers her heartfelt, ugly cry inducing account about What Happened when she lost the Electoral College to the worst presidential candidate in modern history. Politics aside, America lost when they elevated Russia’s Agent Orange to the presidency. Think what you will, but America missed the chance to have a level headed, intelligent and resilient leader, and yes the first female president.

Hillary’s a smart, insightful, resilient, inspiring, kind, caring, pragmatic human being. This book is a journey through her heart and soul.

Dimension 4: Commentary

Reviews with high Commentary focused on Hillary Clinton and the other actors in the election story (high use of third person pronouns). The reviews analyse and evaluate Clinton’s perspective and explanation of what happened in 2016, in a conversational manner much like a TV commentator or pundit.

I disagree with the reviewers who says Hillary doesn’t take responsibility for her mistakes. She analyzes all the reasons she thinks she lost the election–yes, she talks about Russian interference, malpractice by the FBI, and false equivalence by the mainstream press IN ADDITION TO missteps she thinks she made. My own take is that she doesn’t pay enough attention to the reasons why Bernie Sanders was able to command so strong a following with so few resources; but that is part and parcel of who she is.

Historical memoir from the first female candidate for a major political party…a unique perspective and platform to write from. She does recount her successes as well as her failures…she was mostly shut down during the campaigns by repetitious questions and by over-coverage of Trump by the media. She is intelligent and well-informed and states her case without self-pity.

Having the identified these four linguistic functions in the ‘What Happened’ reviews, the trick is to see how they combined to form clusters of review text types – and whether any one of these clusters is more strongly correlated with the high frequency and early reviews.

As Figure 4 shows, hierarchical cluster analysis identifed four review text types: ‘Tribute’ reviews, the largest cluster, have high Emotiveness; ‘Pundit’ reviews have high levels of Commentary and Engagement; Content descriptive’ or ‘spoiler’ reviews talk about what’s in the book in an objective manner i.e. without Reflection or Engagement; ‘Experiential’ reviews narrate the writer’s personal Reflection on the experience of reading the book.

Screenshot 2018-10-28 13.44.25.png

Figure 4: 4-Cluster solution with mean factor loadings, interpretations and percentage of total reviews.   

So, we have these four review text types…do any of these correlate with the bursty reviewing patterns identified? Figure 5 below shows that the actual linguistic pattern of ‘What Happened’ reviews appears to correlate with the burstiness pattern; a large proportion of the first day reviews are Tribute reviews and most of this review type occurs within the first week before tailing off during the rest of the month. The fact that no other review type is particularly time sensitive suggests that, at the very least, Tribute reviews are correlated with early reviewing and are potentially evidence of  coordinated recruitment of Hillary Clinton’s ‘fans’ as book reviewers.

What Happened Cluster Time Series

Figure 5: Distribution of ‘What Happened’ review text types during first month following book launch, compared to hypothetical deceptive and non-deceptive distributions. 

If Hillary Clinton’s PR team did solicit positive reviews in the early days of the book launch, perhaps it is not surprising; they would have been responding to an extensive negative campaign against her book which included manipulating review helpfulness metrics (i.e. massive upvoting of low-rated reviews) as well writing fake negative reviews.

From an investigative linguistic perspective, this analysis shows that: a) suspicious activity can be detected using linguistic data as well as network or platform metadata; b) unqualified praise and intense positive emotions are deception indicators in the online review genre; and c) cluster analysis is an effective way of recognising linguistic deception features in an unsupervised learning setting.

Did Hillary Clinton’s PR team solicit fake Amazon book reviews for ‘What Happened’? Part #1.

September 12, 2017, was the launch day for Hillary Clinton’s autobiographical account of the 2016 election she lost to Donald Trump, definitively entitled ‘What Happened’. By midday 1669 reviews had been written on Amazon.com. By 3pm over half of the reviews, all with 1-star ratings, had been deleted by Amazon and a new review page for the book had been set up. After Day 1, ‘What Happened’ had over 600 reviews and an almost perfect 5 rating. What happened?!

What Happened Indeed

Figure 1: Genuine support or fake reviews? Hillary Clinton’s ‘What Happened’ Amazon rating 1 day after launch (and after all the negative reviews were deleted )

There were good reasons to view the flood of negative reviews as suspicious. Only 20% of the reviews had a verified purchase and the ratio of 5-star to 1-star reviews – 44%-51% – was highly irregular; the vast majority of products reviewed on Amazon.com display an asymmetric bimodal (J-shaped) ratings distribution (see  Hu, Pavlou and Zhang, 2009), in which there is a concentration of 4 or 5 star reviews, a number of 1-star reviews and very few 2 or 3 star reviews.  The charts in Figure 2 below, originally featured in this QZ article,  show the extent to which ‘What Happened’ was initially a ratings and purchase pattern outlier.

atlas_HJnVJAL5Z@2xatlas_ryGnJaL5W@2x

Figure 2: Two charts indicating the unusual reviewing behaviour for ‘What Happened’. Source: Ha, 2017 

Faced with accusations of pro-Clinton bias as a result of deleting only negative reviews, an Amazon spokesperson confirmed that the company, in taking action against “content manipulation”, looks at indicators such as the ‘burstiness’ of reviews (high rate of reviews in a short time period) and the relevance of the content –  but doesn’t delete reviews simply based on their rating or their verified status. (Hijazi, 2007).

It would appear that Amazon have taken on board the academic literature suggesting that burstiness is a feature of review spammers and deceptive reviews (e.g. this excellent paper by  Geli Fei, Arjun Mukherjee, Bing Liu et al. ) and that it is right to interpret a rush of consecutive negative reviews close to a book launch as suspicious.

But what about the subsequent burst of 600+ positive reviews? One might expect the Clinton PR machine to mobilize its own ‘positive review brigade’ in anticipation of , or in response to, a negative ‘astroturfing’ campaign against her book. One could even argue that it would be foolish not to manage perceptions of such a controversial and polarising book launch. If positive review spam is identified, should it also be deleted?

I tracked the number of Amazon reviews of ‘What Happened’ for a month after its launch on the new ‘clean’ book listing (the listings have since been merged but you can see my starting point here). Figure 3 below shows clear signs of ‘burstiness’; the rate of reviewing decreases exponentially over the first month even while the rate of 5-star reviews remained consistently high.

Reviews by day

Figure 3: Number and frequency of ‘What Happened’ reviews in the first 30 days following its launch and deletion of negative reviews. 

So, it is perfectly legitimate to ask whether the ‘What Happened’ reviews were manipulated through ‘planting’ of ‘fake’ 5 star reviews written for financial gain or otherwise incentivised e.g. in exchange for a free copy of the book, which would circumvent Amazon’s Verified Purchase requirement. With my investigative linguist hat on, I’m wondering if there are any linguistic patterns associated with this irregular – and potentially deceptive – behaviour? (If there are, these could be used to aid deception detection in the absence of – or in tandem with –  non-linguistic ‘metadata’.)

A line of fake review detection research has confirmed linguistic differences between authentic and deceptive reviews, although the linguistic deception cues are not consistent and vary depending on the domain and the audience (see my brief overview in this paper). Since we don’t know the deception features in advance and no ground truth has been established (i.e. we don’t know for sure if there was a deception), I’m going to use two unsupervised learning approaches appropriate for unlabeled data: factor analysis, to find the underlying dimensions of linguistic variation in all the reviews, followed by cluster analysis to segment the reviews into text types based on the dimensions with the hope of finding specific deception clusters.

If there is a text cluster that correlates with ‘burstiness’ – i.e. occurs more frequently in the reviews closest to the book launch date and/or occurs repeatedly within a short time frame – then that would suggest there are specific linguistic styles and/or strategies correlated with this deceptive reviewing behaviour. The existence of such a distinct deception cluster would strongly suggest that Clinton’s PR team gamed the Amazon review system (understandably, in order to counter the negative campaign against the book).  Alternatively, different reviewing strategies might be distributed randomly across the review corpus and unrelated to its proximity to the book launch date. This would weaken the argument that linguistic variation in the reviews is a potential deception cue. The two scenarios are illustrated in Figure 4 below:

Deception cluster hypothesis

Figure 4: Hypothetical illustration of how review text types (clusters) might be distributed over a 30 day period in the case of astroturfed fake reviews (top) or genuine positive reviews (bottom). 

My prediction? Surely, Hillary Clinton’s PR team would not so be so brazen as to solicit fake positive reviews in bulk and in an organised fashion. Yes, there were a disproportionate number of reviews written in the first few days but I believe this was a spontaneous groundswell of genuine support. I do expect there to be a few different types of linguistic review style, reflecting the different ways in which books can be reviewed (e.g. focus on book content; retell personal reading experience; address the reader – these are some of the review styles I presented at the ICAME39 (2018) conference in Tampere). However, if the support is spontaneous I would expect these review styles not to be correlated with burstiness or other deceptive phenomena but to occur randomly throughout the month.

Check back here in a few days for Part #2:  Results and Discussion!

 

 

 

 

Genres of mass deception

In 2006, a federal court judged four of the ‘Big Five’ US tobacco companies – Phillip Morris, RJ Reynolds, British American Tobacco, Lorillard (sold to RJ Reynolds in 2014) – to have been operating for more than half a century as a de facto criminal enterprise guilty of racketeering. In November 2017, US tobacco companies finally issued, through national TV and print media, a series of statements correcting their sixty year deception of the American public. They had been appealing the original judgement for over ten years.

 

The ‘racket’ was the continued sale and marketing of tobacco products in full knowledge of their addictive properties and their causal connection to lung cancer.  Under the Racketeer Influenced and Corrupt Organisations (RICO) Act 1970, these tobacco companies were held to have defrauded smokers i.e. obtained smokers’ money by dishonest means, specifically “deceiving smokers, potential smokers, and the American public about the hazards of smoking and second hand smoke, and the addictiveness of nicotine” (United States vs. Phillip Morris et al, 2006, p4).

Below is a list of deceptions maintained by the ‘Big Tobacco’ enterprise:

  • false denial of the adverse health consequences of smoking
  • publicly denial that nicotine is addictive
  • concealment of research data and other evidence that nicotine is addictive
  • false denial of the manipulation of nicotine levels to create and sustain addiction
  • deceptive marketing and public statements suggesting ‘low tar’ cigarettes are less harmful than full-flavor cigarettes
  • false denial that their marketing targets youth
  • false and misleading public statements denying that environmental tobacco smoke (passive smoking) is hazardous to nonsmokers

How does one deceive the public for so long (linguistically)? My approach to this question is to first identify the agents and channels of deceptive communication. We know who received the deceptive messages but who were the senders? Did they use agents/messengers? What channels did they use?

The ‘Big Tobacco’ enterprise built an infrastructure of deception by establishing a number of front organizations i.e. groups that appear to independently support or be motivated by one particular purpose but are actually a ‘front’ for another group whose covert agenda they secretly serve. Front organizations are agent-messengers that appear to be senders. Examples of the forms that front organizations might take include think tanks, associations of consumers or workers, and single-issue interest groups.

Chief among these was the Tobacco and Industry Research Council (TIRC), founded in 1954, which later changed its name to the Centre for Tobacco Research (CTR). TIRC/CTR was established on the strategic recommendation of Hill & Knowlton, public relations counsel to the Big Tobacco enterprise. Backed by money from the Big Tobacco enterprise, it ran a multi-million dollar research programme providing substantial grants for ‘independent’ scientific research into the health effects of smoking. This produced a body of research that obfuscated the link between smoking and cancer and left their causal connection as an “open question”. TIRC/CTR also funded research that diverted discourse away from the dangers of smoking by suggesting alternative causes of cancer such as air pollution, diet and genetics.

This programme of decoy research clearly had a deceptive purpose. However, the research itself was not necessarily deceptive – it created doubt rather than false belief by challenging the anti-smoking research attracting the attention of the American government and health organisations at that time. Neither was this research directly responsible for the mass deception of the American public since the public was not the audience for scientific research.

In the 2006 judgement, Judge Kessler noted that The Big Tobacco front organisations disseminated ‘commentaries’ on both pro- and anti-smoking research through a variety of publication channels, including:

  • management commentaries in annual reports, read by business professionals
  • newsletters and booklets targeting medical professionals with favourable research summaries.
  • public statements, comments made by tobacco-friendly scientists discrediting research that linked smoking with cancer.

Whilst annual reports and newsletters helped communicate the tobacco deception to professional outgroups, as shown in Figure 1 below, press releases were the most influential channel for reaching the general public. Every publication and statement made by a scientist connected to the enterprise was accompanied by a press release. This would be sent out to thousands of editors and then transmitted to the general public through newspapers and popular magazines.

Thus, the press release plays a doubly deceptive role; it reports the deceptive framing of the discourse around tobacco/cancer research and then amplifies its interpretation though popular media. Indeed, PR agency Hill & Knowlton prided themselves on their ability to spin “obscure scientific reports favourable to the industry into headline news across the country”.

Deception Infrastructure S

Figure 1: Mass deception; infrastructure, channels and genres

Below (Figure 2) is one example of press release distortion of research in the tobacco domain.  This 1955 study from the British Empire Cancer Campaign (a forerunner of Cancer Research),  published in the British Medical Journal, reports a nuanced set of findings in relation to the carcinogenic properties of smoking. It reports findings indicating that: i) tar is not carcinogenic in mice,  ii) condensation from tobacco smoke is carcinogenic in mice and iii) carcinogenic effects vary between species so more research is needed.

This nuance is lost in the press release, which seizes on the first finding related to tar and extends it to smoking and tobacco in general. The press release, authored by Hill & Knowlton, draws it authority by reporting the statement made by the TIRC chairman Timothy Hartnett and uses repetition to reinforce its point three times in the first page – “Outstanding British scientists could not induce cancer in experimental animals with tobacco smoke derivatives” / “Experiments conducted at several leading British medical research institutions had failed to induce any cancers” / “18 month experiments fail to show any connection between cancer and smoking”.

 

Figure 2: Comparison of British Medical Journal article and TIRC press release

The fact that this repetitive message itself references a message that is practically self-authored (considering the close relationship between TIRC and Hill & Knowlton, who even shared the same office at one stage) is indicative of the low information quotient in this press release. Yet providing information is arguably the main purpose of the press release genre. What we have here, then, is a deficient or deviant genre communication – inauthenticity (deception) has compromised the integrity of the genre.

The same has been suggested about the annual reports produced by the TIRC, “which read much like industry position papers” (USA vs Phillip Morris et al, 2006, p58). The extract below, from the 1958 TIRC annual report, is illustrative:

A problem may well be obscured, and its solution delayed, by the soothing acceptance of an oversimplified and immature [tobacco theory] hypothesis. The proponents of the tobacco theory have generated increasingly intensive and extensive propaganda. As a result, a non-scientific atmosphere, conducive to prematurity, unbalance, and inadequacy of public judgement, has pervaded the whole field. The prohibition concept discounts or ignores all considerations of smoking benefits in terms of pleasure, relaxation, relief of tension or other functions.

Once again, in this case by allowing bias to enter annual reporting, a genre is compromised through performing a deceptive purpose. These two examples suggest that the tobacco deception was partly sustained by ‘genre fronts’ – communications that appear to serve one conventional purpose but in fact fulfill functions of another genre or are simply deficient.

Press releases are more central to mass deception than annual reports because they are more frequent and produced for immediate impact in popular media. Press release spin is picked up by editors and transferred to newspapers and magazines – sometimes wholesale, sometimes with additional fervour. This resulted in a news environment that actively facilitated the disinformation strategies of the Big Tobacco enterprise leading to a massively misinformed general public.

The selection of headlines, news articles and editorials in Figure 3 below reflect the general tenor of the ‘decoy discourse’ maintained by the tobacco industry, which involved: attacking the integrity of government scientists, casting doubt on environmental controversies such as the use of pesticides and climate change, and linking cancer to spurious causes. The confluence of tobacco advocates and climate change skeptics is a striking feature of this mass deception.

 

Figure 3: Selection of newspaper and magazine items used to support tobacco advocacy. Taken from ‘Bad Science: A Resource Book’

Newspaper items come in a variety of sub-genres, for example news articles whose purpose is to present factual information and editorials that present an opinion. Editorials themselves can be written by newspaper staff or invited writers as ‘op-eds’. Both of these sub-genres can be used for deceptive means i.e. the presentation of false facts in news articles and the advancement of an undisclosed agenda in the case of editorials. In such cases, the news genre loses its features of objectivity and transparency and becomes distorted as ‘fake news’.

Taking the ‘tobacco deception’ as a case study, it would seem that the infrastructure for maintaining this deception was built on the deviant use of a variety of genres – annual reports, press releases, newspaper/magazine items – that were connected by channels invisible to the general public i.e. front organizations and funded scientists. The schema presented above will be used in future posts to evaluate similar deception controversies in the environmental and health domains.

 

 

Who is ‘we’? Investigating pronouns for deception.

If you are ever arrested and asked to make a statement by the police on US soil, be careful with your pronouns. Law enforcement officers in the US are likely to have received training in analysing statements for deception from Mark McClish or Don Rabon, which means the pronouns you use will be inspected very closely. McClish and Rabon come in for a lot of stick from forensic linguists working in academia, due to the lack of citations and over-generalisations in their blogs and best-selling books. 

language police.png

But to be fair, many of the leading academics working on lie detection have said similar things about the use of first person pronouns being an indicator of veracity. Newman, Pennebaker and colleagues, in their seminal work ‘Lying Words: Predicting Deception From Linguistic Styles’ provided some empirical support for the correlation between low self-reference and deception; subsequently it has been found to broadly hold in online dating profiles, business communication and criminal narratives (but, interestingly, not the case in consumer reviews which use ‘reader engagement’ as a deception strategy).

It should be noted, however, that Newman and Pennebaker’s prediction rate was 67% – that is wrong 1 in 3 times. Consequently, simply counting the use of any first person pronouns is not by itself the magic cue for deception detection. There are a number of different first person pronouns – I, me, my, mine, myself, we, us, our, ours, ourselves. Not only do these have different strengths in terms of socio-psychological ideas of distance and commitment (compare ‘I was hit’ with ‘The car hit me’) but they work differently linguistically. For example,  ‘I’ and ‘me’ will correlate with verbs, ‘my’ with nouns; also ‘I’ correlates with stance and modal verbs (‘I thought’, ‘I tried’, ‘I would’) so is a more ‘active’ first person pronoun than the passive ‘me’.

The above also applies to ‘we’, ‘us’ and ‘our’. In addition first person plural pronouns have the additional pragmatic parameter of clusivity to distinguish who exactly is included in the ‘we’ (see Figure 1 below). And there is the linguistic phenomenon of nosism which includes royal and editorial ‘we’ (see Ben Zimmer’s excellent 2010 article to fully appreciate the complexities of ‘We’).

clusivity we.png

Figure 1: Referential parameters of ‘we’: inclusive (left), exclusive (right). By LucaLuca. Reproduced under Creative Commons licence

A case in point is ex-UK Prime Minister David Cameron’s response to questions about his alleged use of off-shore tax havens to avoid paying tax, as revealed in the Panama Papers leak. In April 2016, 10 weeks before the EU Referendum and his subsequent resignation, Cameron was taking questions about the upcoming referendum and speaking in support of remaining in the EU at a town-hall style Q&A event held at PriceWaterhouseCoopers’ Birmingham offices.

Figure 2 is a transcript of David Cameron’s response to an unexpected question from Sky News journalist Faisal Islam regarding the controversy over his connection to an offshore investment company (Blairmore Holdings) owned by Cameron’s late father. Cameron denied owning any shares or offshore investments but was roundly criticised for his evasive answer (Cameron restricts his answer to the present tense, despite Faisal Islam’s specific temporal reference to the past and the future – lines 4-5). Five days later, under public pressure to resign, Cameron was forced to admit that he had owned shares in his father’s business (which he sold at a profit shortly before taking office as Prime Minister).

Cameron Panama response 1

Figure 2: Transcript of David Cameron’s first public response on the Panama Papers allegations. Given during a Q&A with workers at accountancy firm PWC in Birmingham, 5 April 2016.

David Cameron’s use of pronouns is an excellent example of linguistic duplicity, shifting between inclusive ‘we’ (yellow), exclusive ‘we’ (red) as well as ‘I’, all in reference to himself; the identities referred to by ‘I’ are split between David Cameron as a UK citizen and his role as Prime Minister.

Furthermore, the scope of ‘we’ varies within the text and is sometimes unclear. In answer to a question about “you and your family” with regard to the financial business of one’s late father, one might expect ‘we’ to refer to some aspect of family. However, Cameron initially moves to include the whole audience (and viewers) in his personal financial affairs by referring to  ‘we’ as a nation with the contextual reference “our tax authority” and later “our own country” (line 8). In the middle of his response, a different (exclusive) ‘we’ appears mid-sentence – “I have a house, which we used to live in, which we now let out while we are living in Downing Street” (lines 16-17). There is no explicit reference to the scope of this ‘we’ but the reference to personal property ownership means one can assume that is not the national ‘we’.

If one assumes that it is the ‘we’ originally asked for in the reporter’s question – i.e. “you and your family” – then Cameron has violated the right frontier constraint (Webber, 1988), which stipulates that anaphoric elements such as pronouns are interpreted in ambiguous cases by reference to information at the end of the previous discourse unit i.e. the right frontier (for languages with left-to-right scripts). That Cameron returns to ‘we as a nation’ for the remaining text further highlights his dynamic use of pronominal reference.

The linguistic duplicity displayed by David Cameron above is in stark contrast to the language he used when owning up to his involvement with Blairmore Holdings in a hastily-arranged national TV interview. As the transcript shows, Cameron doesn’t use ‘we’ at all in response to similar questions. This case study shows that assessing veracity and potential deception by tracking pronoun use is valid but more complex than simply counting; the inherent capacity for linguistic duplicity is contained within a complex system of deceptive pragmatics.

What does honesty look like (statistically)?

Certain linguistic features (e.g  reference, modality) facilitate deception because they are malleable to context and flexible to interpretation. My first blog post showed that deceptive communication contains ‘outliars’, portions of texts with an unusually high concentration of these linguistic features; in the second post we saw that the linguistic hotspots where these features cluster can be taken as ‘points of interest’ worthy of further investigation. Of course, liars do not have a monopoly on the use of modals! Furthermore, truth-tellers can sometimes be mistaken for liars due to nervousness, fear of disbelief, or perceptions of powerlessness (known as the ‘Othello error’). So what does honesty (non-deceptive) communication look like?

sharapova mistake

In my Standford Decepticon 2017 conference paper I tested the ‘Outliar’ investigative linguistic methodology on honest admissions of doping – true confessions – by the following five sports persons and professionals:

true doping confessions.png

The Maria Sharapova case took the tennis world by surprise (she was the first high-profile female tennis player to fail a drug test). In 2016, Sharapova was banned from competition after testing positive for meldonium during the Australian Open in January of that year. Meldonium is a heart medication that was found by the World Anti-Doping organisation (WADA) to be particularly popular amongst sports persons from Russia and Eastern Europe, perhaps due to its ability to block the body’s conversion of testosterone to oestrogen. Having placed meldonium on a watch list in 2015, WADA had fully prohibited the substance from January 1 2016, two weeks before the Australian Open. Following the failed drug test, Sharapova admitted she had been taking meldonium as medication since 2006 and stated that she had negligently and inexcusably missed the communications from WADA prohibiting its use.

Linguistic analysis of the explanation Sharapova gave to fans via her Facebook page shows two ‘outliars’ at the beginning and end of the post (see Figure 4 below).

Sharapova outliar graph

[1] I want to reach out to you to share some information, discuss the latest news, and let you know that there have been things that have been reported wrong in the media, and I am determined to fight back. You have shown me a tremendous outpouring of support, and I’m so grateful for it.

[13] I have been honest and upfront. I won’t pretend to be injured so I can hide the truth about my testing. I look forward to the ITF hearing at which time they will receive my detailed medical records. I hope I will be allowed to play again. But no matter what, I want you, my fans, to know the truth and have the facts.

Figure 3: Outliar analysis of Maria Sharapova’s 2016 Facebook post and outlier extracts.

Sharapova begins her post by suggesting she has been a victim of unjust media coverage. It had been widely reported that she had received five ‘warnings’ about the upcoming change to the WADA regulations. Sharapova agreed that she had received newsletters with links to the WADA rule changes but argued that these were ‘communications’ rather than warnings through which one had to “hunt, click, hunt, click, hunt, click, scroll and read” in order to find information about the prohibition. Sharapova ends her post by strongly maintaining that she is being honest about her genuine mistake (of using Meldonium as medication after the ban).

These anomalous extracts are particularly emotional when compared to the main body of this post, in which Sharapova gives specific details about all the communications she did receive (see yellow highlighted text in Figure 4 below). There is a lot of literature that suggests specific details are a strong indicator of veracity in legal genres such as witness statements. (Professor Aldert Vrij’s research on Criteria Based Content Analysis is a good place to start.) These anomalous extracts could just be ‘Othello errors’ that are confusing emotional intensity for deception.

Sharapova FB 1cSharapova FB 2c

Figure 4: Maria Sharapova Facebook post, March 2016. Last accessed 21/7/2018

Accounting for the ‘Othello error’ is one reason a full ‘Outliar’ analysis uses an additional measure of language change within a text – intratextual language variation – when assessing text veracity. Texts can range from having a uniform style with consistent use of features throughout – a stable text – to displaying marked changes in language style at several points – variable or ‘spiky’ text.  Outliar captures this by summing the amount of change shown in a text.

Figure 5 is an example of this. It compares ‘Outliar’ analysis of Sharapova’s Facebook post (left) one of a Lance Armstrong TV interview in whch he falsely denied doping (see previous blog for more discussion of Armstrong’s deception). Visually, you can see that  Lance Armstrong’s language use displayes high variability in comparison to which Sharapova’s language is relatively stable.

Figure 5: Comparison of the ‘Outliar’ analysis of Maria Sharapova’s Facebook post (left) and Lance Armstrong’s ESPN interview (right) .

Figure 6 below shows a statistical measure of intratextual language variation for five false doping denials vs. five true doping confessions (see p7 here for the formula). It can be seen that the deceptive communications show more language change than the honest ones. So, combining outlier text detection with an overall measure of language variability can be helpful in distinguishing honesty from dishonesty. Frequent and marked language style change is a signal of potential deception.

intratext analysis edit 2

Figure 6: Analysis of intratextual variation. Y-axis = total intratextual variation measured as aggregate z-score for each text; X-axis represents ten texts in total –  five deceptive texts (false denials by: 1) Barry Bonds; 2) Linford Christie; 3) Lance Armstrong; 4) Alex Rodriguez; 5) Marion Jones) and five honest texts (true confessions by: 1) Maria Sharapova; 2) Dwain Chambers; 3) Victor Conte; 4) Floyd Landis; 5) Levi Leiphemer)

In Sharapova’s case, the tribunal were satisfied she had not intended to cheat (although she was found to have also taken the drug to enhance her performance) and her relatively light ban (reduced from two years to 15 months on appeal) reflected the fact that she had been negligent but not deceptive. I would argue that the (relatively) stable language of both Sharapova’s Facebook post and the initial press conference where she announced her drug test failure support the tribunal finding. The press conference video is below – judge for yourself.

 

 

The case of the asthmatic cyclists: deception detection as investigative linguistics

Did you know that Sir Bradley Wiggins, Sir Mo Farah and I-thought-he-was-a-sir Chris Froome all suffer from asthma? As do a number of other succesful sports persons accused of using performance-enhancing drugs?  What I call ‘investigative linguistics’ led me down this particularly rabbit hole. My investigative linguistic approach examines texts for ‘points of interest’ (POIs). It uses deception detection tools on communications with unknown veracity in order to automatically identify POIs. One benefit is that you can approach a topic without any prior knowledge or biases and quickly find avenues that are objectively worth exploring.

After analysing the known deceptions of Lance Armstrong (see previous blog), I collected a bunch of statements made by sports people admitting or denying their use of perfomance-enhancing drugs. Based on currently available evidence these statements were divided into three categories: a) false denials, b) true confessions and c) presumed-to-be-true denials (see my Stanford Decepticon 2017 presentation for full details). For the ‘true’ denials category, I picked five recent high-profile cases: two relating to the cycling controversy around Sir Bradley, Sir David Brailsford and Team Sky (video explainer), and three connected to the controversy around the infamous athletics coach Alberto Salazar and his Nike Oregon Project which engulfed Sir Mo Farah, the Canadian Cameron Levins and, indirectly, Paula Radcliffe MBE.

true denials dataset

It was a surprise to me that each of the interviews (graphed below) mentions asthma and related issues. If I had done some prior research I would have realised that the provision and use of asthma medication during and around major sport events was a key issue in sports doping. Still, it shows that deception detection techniques can be used for exploration i.e. to find the ‘points of interest’ worthy of further investigation. Furthermore, a ‘naive’ approach helps to avoid unconscious bias affecting the analysis

Slide1

(‘Outliar’ analysis of four ‘true denials’. Interview responses are represented as a time series on the x-axis (c.30-60 second chunks). Green shading indicates ‘outliar’ text. Asthma mentions marked with ∇)

In Bradley Wiggins’ Guardian interview (given to counter suggestions of illegal doping during the 2012 Tour de France), the analysis highlights inconsistencies in Wiggins’ stance towards his asthma allergy:

[14] I was paranoid about making excuses: “Ah, my allergies have kicked in.” I’d learned to live with this thing. It wasn’t something I was going to shout from the rooftops and use as an excuse and say, “my allergies have started off again”. That’s convenient isn’t it Brad, your allergies started when you got dropped.

[17] I didn’t mention it in the book. I’d come off a season of … I’d won everything that year. When I was writing the book I wasn’t sat there thinking, “I’d better bring my allergies up”. I was flying on cloud nine after dominating the sport all year. It wasn’t something that I brought to mind.

In these two extracts, asthma is simultaneously a big deal and a non-issue for Wiggins. While this does not in any way confirm deception or guilt, it does indicate a defensive stance that is worth investigating. This can be contrasted with Mo Farah’s discussion of his own asthma:

[4] “This picture has been painted of me. It’s not right. I am 100% clean. I love what I do. I want to continue winning medals. But I want people to know that I am 100%, I am not on any drugs, I am not on thyroids, I am not on any other medication. The only medication that I am on, I am on asthma and I have had that since I was a child. That’s just a normal use. I am on TUE [therapeutic use exemption] where you have … it’s just the normal stuff. And that’s it.” – Sky Sports interview, 2015

In contrast to Wiggins, Mo Farah volunteers information in a non-defensive fashion about his asthma and use of Therapeutic Use Exemptions (TUEs – a doctor’s note and prescription). Canadian runner Cameron Levins’ response to questions about his use of prescription drugs registers as more ‘interesting’ than Farah’s although not as high as Wiggins:

Interviewer: No prescription drugs?

Levins: I have some medication I take for my asthma, but that is something that is wrong with me. I’m asthmatic.

Interviewer: Was that before you came on with the (Nike Oregon) Project?

Levins: Yeah, I was dealing with it before I joined the project actually. A little bit after the London Olympics I started having quite a bit of difficulty with it. So it was before I joined the project.

Levins later goes on to explain that “adult onset asthma is pretty common”. Obviously, having asthma since childhood is easier to defend which may explain the higher ‘interestingness’ score of Levins response compared to Farah.

In a 2016 interview with Sky Sports news David Brailsford, director of Team Sky (whose riders included Bradley Wiggins and Chris Froome), offered the following highly ‘interesting’ reply when asked about his team’s covert use of Therapeutic Use Exemptions to obtain otherwise prohibited drugs:

[5] We’ve reviewed this over the years as we’ve moved forward. We have changed our policy, we’ve changed the way we do it, and in the future going forward, I think we’re going to take the next step, which has been debated on a wider basis across the whole of the TUE process, and look at having the consent of the riders to make all TUEs transparent

In this segment, Brailsford tries draw a line under anything that may have occurred in the past, using many words related to looking to the future. In no way presuming anything illegal on Brailsford or Team Sky’s part, previous policy would clearly be a ‘point of interest’.

So, this analysis suggests that cyclists use of asthma medication is more ‘interesting’ than that of athletes. (In Farah’s interview, the analysis flags his comments about missed drug tests rather than specific doping allegations.) Understanding the reasons for this can then provide a focus for further investigation. As Chris Froome’s recent successful appeal shows, the asthma issue may be due to faulty regulations based on models with a tendency to generate ‘false positives’ – itself a form of deviance if not deception.

froome inhaler

(picture © BBC/Getty Images, 2018)

For this naive analyst, investigative linguistics revealed an important connection between asthma and sports doping that is clearly ‘interesting’. Application of the same techniques to the domains of business, politics and finance will definitely be interesting…

 

 

 

 

 

 

 

 

 

 

 

 

 

Linguistic Pointing and Deception Detection

So here’s the thing. You can tell somebody is lying – or more correctly, deceiving – by the words they use. I’m not talking about gesture or disguise or other types of non-verbal deception. I mean when there are words and text – either written or spoken – those words will reveal deception if you know what to look for.

“Listen, nobody believes in doping controls more than me.” — Lance Armstrong

Does that mean it’s possible to detect deception by reading a text or transcript or listening to someone speak? Not exactly. Factors such as human truth bias and our reliance on heuristics to process information mean that judgement derived from our senses is not entirely reliable (although it can be improved by education and training).

Deception detection is possible by processing a text. Now, unless you are some kind of artificial intelligence, you will rely on an automated tool for computational and statistical analysis. Non-verbal deceptions such as credit card or other financial fraud are already detected using statistical algorithms and other data mining techniques. Advances in linguistics mean that texts can also be processed as data and then classified and grouped together – all without being read or heard.

There are different features of language that can be analysed. Words, word sequences, types of words, grammar, syntax and so on. These features can be analysed individually or as groups that represent underlying concepts (e.g. ‘certainty’ or ‘complexity’). You can analyse known true and deceptive texts for these language features, compare the frequencies and distributions, and find linguistic tendencies that correlate with deception and truth. But what are these linguistic features?

There are many sets of linguistic features that have been used for deception detection (see Hausch et al’s 2015 meta-analysis for a comprehensive list of experiments and linguistic features used). The features that are most effective are the ones that enable the linguistic act of deception.

Take the following:

– “ Car.”

By itself, this single common noun cannot be a lie. If I said ‘car’ and pointed to a bicycle or a phone then that could be a lie. There are linguistic different resources for ‘pointing’:

– “That is a car.”
– “I have a car”
– “My car”

In fancy linguistic terminology, ‘pointing’ is known as referential indexicality. Other types of ‘pointing’ can be to a particular place, period of time, event, assumption, thought and so on – even to the text itself. Drawing on the linguistic theory and influence of the Prague School, I use a set of these linguistic features to analyse texts for ‘textual hotspots’ – a linguistic equivalent of the non-verbal hotspots such as micro-expression, gesture and voice identified by the psychologist Dr Paul Ekman (on whom the TV show, ‘Lie To Me’ is based).

I have developed a tool for identifying these linguistic hotspots using anomaly detection techniques adapted from banking fraud. Research has shown that these linguistic deception features cluster together when deceptive language is being used. Ergo, anomalous clusters of linguistic features that point to deception are areas of potential deception. Not to steal Dr Ekman’s thunder, I am calling these anomalous textual hotspots ‘outliars’.

I’ve tested this hypothesis on a number of known deceptions and the results, which are promising, were presented at the Decepticon 2017 conference held at Stanford University.  I chose statements made by sportspersons about the use of performance-enhancing drugs and doping because these are high-stakes deceptions and so more likely to leave linguistic traces.

Screenshot 2018-06-21 19.50.12

One of the most famous examples of doping deception is Lance Armstrong. How did Lance Armstrong successfully deceive so many millions for so long? Bullying, good lawyers and a fairytale narrative of cancer recovery and global charity certainly played their part. But the key to Armstrong maintaining this deception – through various testimony, interviews, biographies – was his sustained verbal performance.

One classic example is Armstrong’s 2005 interview with Bob Ley on the ESPN show ‘Outside the Lines’.’Outside the Lines’ is an investigative ESPN TV series that takes a critical look at American sports issues. This interview was conducted by the usual anchorman, Bob Ley. Armstrong was a year into his first retirement, after winning his 7th Tour de France in 2005 and had just been cleared of doping allegations after a lengthy trial. The show is renowned for its tough questioning and investigative slant, and Bob Ley did not hold back. Below is a transcript of the interview.

Figure 1 shows my ‘Outliar’ analysis of the responses. The  analysis picks out two linguistic hotspots of potential deception – sections 5 and 12. These are highlighted in the transcript but I’m going to lay them out here for analysis.

Screenshot 2018-06-23 21.02.48

Figure 1: Outliar analysis of Lance Armstrong’s interview responses. Armstrong’s interview responses are represented as a time series on the x-axis (c.30 second chunks). The y-axis measures the relative frequency of linguistic deception features; text segments scoring over 3.5 are recorded as anomalous (the Iglewicz-Hoaglin method).

Segment 5 contains the following extract. Bob Ley had asked whether it was true Armstrong had made a phone call to Dr Prentice Steffen threatening “to spend a lot of money to make your life miserable” if Steffen did not retract comments accusing Armstrong of doping [transcript lines 46-50].

ARMSTRONG: Not true. Steven, er Prentice Steffen I think was his name, was not part of the team when I was there, I hardly know him. The only interaction I ever had with him I think was when he was a team doctor with the Mercury cycling team and I helped one of their young riders I think get care for testicular cancer. That’s the first and only interaction I ever had with him.

In this outlier extract, Armstrong denies the accusation by distancing himself from Dr Prentice Steffen (Steven, Stefan, what was his name again?). Instead he foregrounds his charity work for an anonymous sick rider. The underlined sentence introduces three new referents – a young rider, Mercury and cancer.  Such a topic shift and introduction of third party issues is a pragmatic technique for diverting attention. The cluster of pronouns picked out by the analysis – ‘he’, ‘their’, ‘I’– facilitate the diversion and leave the final ‘him’ ambiguous (technically this ‘him’ should refer to the nearest qualifying noun i.e. ‘one of their young riders’).

In contrast,  the following extract from segment 8 is representative of Armstrong’s ‘baseline reading”. Bob Ley had asked whether it was true Armstrong had made a phone call to Greg Lemond, threatening to smear him: “I can produce 10 people that say you took EPO”:

With regards to Greg Lemond I have to say as a young guy, I did idolize him in 1989, I think we all remember that incredible story coming back after getting shot and winning the tour by 8 seconds, the smallest margin ever. I mean he was a guy that quite literally put all of us into cycling, because he was appealing to us at a young age. But, er, for a past champion and a great champion, one of the greatest athletes of all time, to be so involved in a case, I mean I ask you Bob, I ask the viewers, why would you be so involved?

In answering this, Armstrong appears to show rare humility; he acknowledges his own inspiration and even someone else’s achievements. However, a closer reading reveals a mocking tone in which Armstrong draws attention to the narrowness of LeMond’s victory – “winning the tour by 8 seconds, the smallest margin ever”. The transcript shows that taunting accusers is Armstrong’s baseline linguistic behaviour in this interview, which is why the reticent language used when discussing Dr Prentice Steffen in segment 5 stands out as deceptive.

The analysis also flags the following segment 12 as a linguistic hotspot of potential  deception. Here Ley has asked Armstrong about his attempts to shut down the World Anti-Doping Agency (WADA) investigation that was shining a light on doping in cycling and Armstrong at that time.

Now there’s two people involved in this process. There’s the athletes and there’s the people who police the athletes. And both of them have to be ethical. Listen, nobody believes in doping controls more than me. I’ve submitted to all of them, whether in competition or out of competition. Now listen, I’m not saying my best defence is I’ve never tested positive. All I’m saying is that the last few years when you were supposed to tell the investigators and the drug testers everywhere you were everyday of the year, I did it.

Here, rather than taunting accusers, Armstrong again points the linguistic finger. He insinuates that the drug testing process and its ubiquitous participants (‘people’, ‘investigators’, ‘testers’) may not be ethical and he portrays himself as a willing (and perhaps slightly persecuted) subject to the testing regime. However, with the assertion “nobody believes in doping controls more than me”, Armstrong leaks the fact that he has been expert at manipulating the drug testing system. He immediately realises this ‘slip’ and moves to deny its implicature that his “best defense is I’ve never tested positive”. The final ‘it’ is ambiguous and difficult to resolve – a deception strategy we also saw in the above segment 5.

There are more examples like this in my Decepticon 2017 Stanford presentation (including an interesting connection between doping and asthma!) so take a look at that if you are interested in more detail on the method (or write to me). But the real value of this method, I think, is as an investigative linguistic tool which can identify ‘points of interest’ and thus aid forensic and journalistic investigations. So future blog posts will probe the public statements and testimony related to the key events, scandals and crimes of this post-truth era.