The US is looking to anticipate terrorist strikes by using IT to mine the enemy's online screeds.
'Facta, non verba,' goes the Latin saying. 'Deeds, not words.' But what if we could invert that and use writing or speech to predict action? What if we could use psychological and language theory with information technology to answer the more contemporary question, 'You can talk the talk, but can you walk the walk?'
Research recently completed by the US Department of Homeland Security (DHS) suggests this may be the case. It has sought to apply analysis and processing muscle to one of its most difficult challenges: identifying potentially violent terrorist groups and, beyond that, when their statements point to imminent attacks.
The project, which is currently awaiting funding for a second stage, was publicly disclosed at last month's annual meeting of the American Association for the Advancement of Science (AAAS) in Washington DC. All its contributors stressed that this is merely a beginning. There were, as we shall see, limitations in the data set and levels of nuance that the work did not address.
Nevertheless, significant differences emerged from a comparison of statements posted online by Central al Qaeda (AQ) – as controlled by Osama bin Laden and Ayman al-Zawahiri – and its sister group al Qaeda in the Arabian Peninsula (AQAP) with those from two established organisations with similar philosophical goals but which do not resort to violence, Hizb ut-Tahrir and the Movement for Islamic Reform in Arabia. Importantly, these differences were not confined solely to the theoretical or rhetorical content of the documents, but also their linguistic structure.
The DHS project used 320 documents from the four sources, with the important caveat that they were analysed in English translations, not the original Arabic. Three techniques were applied to the texts. Integrative complexity can broadly be seen as interpreting sentence structure in terms of the complexity of the thought expressed. Linguistic Inquiry Word Count (LIWC) addresses the types of word used. Frame analysis tries to identify more of the thematic content, breaking down texts into actors, subjects and the thought processes or emotions within the words sentences describe.
Similar techniques have been used to mine documents manually. The goal is to automate and speed up the process, and reduce human errors, through computing.
Integrative complexity is a well established concept that looks at the underlying structure of a statement's rhetoric rather than its content. It has undergone recent fine-tuning by a team from the Psychology Department at the University of Montana. That team's leader, Dr Luke Conway, also headed the DHS study group, and has proposed two subdivisions.
''Elaborative complexity' emerges when one singular point of view is defended in a complex manner. For example, consider the following statement: 'Peanut butter is delicious; it is creamy and sweet and makes for a healthy meal.' The viewpoint that peanut butter is great itself is never challenged or qualified, but rather is defended with two different dimensions (taste-related, health related),' Conway told the AAAS.
''Dialectical complexity', on the other hand, derives when a given topic is described from multiple points of view. Consider, for example, this statement: 'Peanut butter is delicious and creamy, but I don't like how it gets stuck on the roof of my mouth.' In this case, two dialectically-opposing views of peanut butter are presented, one positive and one negative.'
Any analysis of integrative complexity is, Conway acknowledged, 'very labour-intensive'. The work needs to be done by teams of human readers, who first grade basic complexity and then the two sub-constructs on a scale from 1 to 7, with those scores that most greatly diverge from the average being eliminated.
As a result, the DHS analysis confined itself to just five randomly selected paragraphs from each of the documents because of the time and budget available. The effort involved raises questions over the financial viability of the technique, particularly given that the two others are much more automated.
However, the approach still detected significant differences between the terrorist and non-terrorist groups. Overall, AQ and AQAP exhibited lower integrative complexity and, at a finer grain, elaborative complexity than the peaceful organisations. This does not fit the typical extremist 'prototype', Conway said.
'One speculative answer involves the sheer cognitive strain of engaging in the terrorist lifestyle,' he continued. 'It is possible that the lifestyle led by terrorists is simply more draining – planning for and executing attacks, hiding from authority figures – and as a result makes it more difficult for them to consistently produce complex statements.'
So, a metric begins to emerge for contrasting between violent and non-violent groups, but what about predicting when violent groups might be about to execute an attack? Here another interesting trend emerged; an increase in elaborative complexity, even though the greater mental effort likely to be involved might suggest the opposite.
Conway proposed a theory here that the increase 'is due to a kind of 'battening down of the psychological hatches' that occurs when people shift from trying to make a decision to the actual implementation of – and defence of – the decision'.
The LIWC technique has been developed by a team under Professor James Pennebaker at the University of Texas Psychology Department over the course of the last 15 years. It is with this approach that IT gets brought to bear as a text is put through computer analysis very quickly.
'Luke used human judges,' Pennebaker explains. 'A problem is that human judges don't agree. And if you put a series of depressing essays in front of them, they get depressed, and that affects their judgements.
'So this is a [computer] programme to analyse text and certain features of language. I could put all 320 documents through that in a few hours.'
Specifically, Pennebaker has found that what he calls 'function words' are very indicative of thought processes, and are also fruitful for this kind of detection as we cannot control them that strictly.
'In English there are about 500 common function words, but they make up more than half of what we say in everyday speech. And generally, you don't hear them; the brain doesn't process them,' he says.
Examples of the kind of terms he is describing here are personal pronouns, articles, or words like 'but', 'upon' or 'by'. They do not include common verbs or nouns, however which he terms 'content words'.
An example of how 'function words' do disclose our processes comes in the structure of how we tell the truth. 'Someone telling the truth will use 'I' more and they'll also make distinctions: 'I did this, but I didn't do that,'' Pennebaker says. Personal pronouns, though, can also reflect far greater assertiveness.
The process of categorisation then goes through progressive levels of refinement, as the system's operators build a social-emotional dictionary that can be checked across a document by the software.
What the LIWC team discovered mirrored the kind of conclusions reached by the human integrative complexity analysis. At the highest level, the terrorist groups used fewer big words (those with more than six letters) than the peaceful ones. They also used significantly more social and emotional ones, again implying less cognitive complexity. And these trends accelerated in the run-up to an attack.
'In other research, we have found similar linguistic shifts in [former US President George W.] Bush around the time that the US went to war in Iraq. We find them in Hitler and in FDR [US Second World War president Franklin D Roosevelt],' says Pennebaker. 'It doesn't matter which team you're on, there are these shifts.'
If this sounds like using a lexicon as a blunt instrument, Pennebaker makes no apologies. 'You have to keep in mind, it's going to be pretty crude, but it's a pretty crude business we're in.' However, he also noted that the evidence so far can only offer a 'modest-to-weak' indication that the technique could be used predictively. Particularly, the nuances within words themselves meant that a sample needed to be checked in the original language against an Arabic dictionary.
Similarly, LIWC may be more powerful when used for 'live' discourse – speeches, even Twitter posts. 'I know how to cheat the system but I'm shocked at how lousy I am,' Pennebaker said. 'But it might work better on a phone tap or in a more spontaneous situation. We analysed Bush in a press conference and he was like an open book but in a [prepared] speech he was more opaque.'
Like integrative complexity, frame analysis is based on psychological theory of the 1970s but the work being undertaken by the team under Dr Antonio Sanfilippo at the Pacific Northwest National Laboratory (PNNL) is developing a computational strategy.
It essentially looks at how a source is using language to influence a target audience and how that audience responds. There are the issues and, then, the adjectives and verbs applied to the issues in a text. Using these, 'frames' identifying the tonality of the documents are developed.
'The automation here has to use natural language processing and text mining,' Sanfilippo explained, 'with the software performing tasks that recognises names, actors and objectives.'
Much like LIWC, the programme requires a lexicon of terms organised into classes. For this exercise, Sanfilippo's team used 90 per cent of the 320-document sample for training and 10 per cent for the analysis.
What it found was that the AQ and AQAP documents were more marked by terms that fell into four main categories: moral disengagement ('hate', 'fear', 'judge', 'criticise'); the violation of sacred values (particularly a sense of attacks on religious belief); social isolation ('confine', 'abandon', 'withdraw'); and violence and contention ('attack', 'fight', 'kill').
In essence, PNNL has begun to develop a linguistic, terrorist-specific thematic fingerprint – something closer to the science-fiction version of this kind of processing. However, the work on tracing back documents against 30, 60, 90 and 120 days from attacks was less conclusive in potentially predicting what happened, as it tried to identify 'contentious' frames and their increasing incidence within the sample. Accuracy for identifying a terrorist source was 78 per cent against 66 per cent for a violent act.
Again, more work, more samples and more material in the original language are all required. Also, the PNNL's earlier work shows just how subtle any of these processes will need to be.
It has also analysed 620 documents published by the Egyptian Muslim Brotherhood between 2005 and 2006. During that time, the organisation formally became the country's second largest party after a now widely discredited election. In that work, Sanfilippo and his fellow researchers found that the percentage of 'contentious' frames within the texts rose from 24 per cent to 35 per cent while 'cooperative' frames declined from 14 per cent to 10 per cent. The change was the result of heavy repression during this time by the Mubarak regime. However, this did not lead the Brotherhood to take violent action.
Five years later, it is widely seen as having played a deliberately low-key role in the early stages of Egypt's revolution, although its anger about the state of affairs in the country remains high. Thus, the need to distinguish between different types of organisation before apply any of these mining techniques is a prerequisite.
This leads to the question of civil liberties. Few would question that any security service needs to track and comprehend the potential sub-text in statements by known terrorist groups. Today, that task is undertaken by human intelligence analysts, but it is prone to error, costly and time-consuming. And often, the clock is ticking.
But who should be watched? Professor Pennebaker's team has conducted experiments with speed-dating where it claims to have had greater success predicting who would ultimately go out with whom than the participants themselves. Bravo, yet it is a little discomforting to know that that is possible.
By constructing its research in a two-tier way – first distinguish peaceful from violent groups and then track the violent – the DHS appears to be trying to pre-empt objections that it may acquire tools that give it too much of an insight into people in everyday life. And given that future funding for this specific project is currently on hold, it is moot to ask whether it would have the budget anyway to pry to that extent.
Nevertheless, automated information processing does potentially seem to offer some established psychological methods a new level of power.