Language use in the Jihadist magazines inspire and Azan
© Skillicorn and Reid; licensee Springer. 2014
Received: 24 January 2014
Accepted: 17 July 2014
Published: 24 September 2014
The language of influence or propaganda has been studied for a century but its predictions (simplification, deceptiveness, manipulation) can now be examined empirically using corpus analytics. Semantic models for intensity of belief and use of gamification as a strategy allow novel aspects of influence to be taken into account as well. We develop a semi-automated approach to assess the quality of the language of influence using semantic models, and singular value decomposition as a middle ground between high-level abstract analysis and simple word counting.
We then apply this approach in a significant intelligence application: examining the use of the language of influence in the jihadist magazines Inspire and Azan. These magazines have attracted attention from intelligence organizations because of their avowed goal of motivating lone-wolf attacks in Western countries. Our approach enables us to address questions like: How good are the authors and editors of these magazines at producing influential language (and so how great is the impact of these magazines likely to be)? How does this change with time, and as a reaction to world events, and what does this tell us about competence and strategic goals? What is the impact of changes in authorship?
KeywordsLanguage model Jihadist language Integrative complexity Gamification Deception Propaganda Corpus analytics
Inspire magazine is an online jihadist magazine, written in English. At the time of writing, twelve issues have appeared, the first in the middle of 2010 and the most recent in Spring 2014. The first eight issues were edited by Samir Khan, with substantial assistance from the charismatic preacher, Anwar al-Awlaki. Issue 8 did not appear until well after their deaths in 2011, but had clearly been written by them. Issues 9–12 are the product of different, so far unknown, editors.
The goal of Inspire appears to be to motivate lone-wolf attacks in Western countries by diaspora jihadists. Such potential attackers are assumed to have little Arabic, and so the magazine is written in English; to be familiar with high-quality publications, and so the magazine has high production values, including significant attention paid to layout and presentation; and to have had limited exposure to religion and particularly the religious language endemic in much of the Middle East. Inspire is thus much more “Westernized” than the typical discussions in, for example, Islamist forums. Intelligence agencies have taken Inspire’s role as a recruiting tool very seriously . Indeed, it is against the law to possess copies in the United Kingdom, and several people have been convicted for this offence; an Australian man has also been charged with a similar offence.
Azan magazine resembles Inspire in “look and feel” and appears to share many of the design elements, so it seems reasonable to assume that it was modelled on Inspire. At the time of writing, five issues have appeared, the first in March 2013 and the fifth early in 2014. Its focus is more on a South Asian (English-speaking) readership, but its goal is apparently also to encourage self-initiated attacks in Western countries. The magazine’s authorship is not known but the topics covered suggest concerns centered in Pakistan and Afghanistan, particularly a focus on the threat from drones. The textual content of issues of Azan is typically between three and four times longer than the textual content of issues of Inspire, and issues are appearing much more often so their author/editor is noticeably more productive.
The effect of these magazines is difficult to judge. Several lone-wolf attackers have had copies in their possession; but it is also clear that the magazines are widely read in their target communities. There has not been a noticeable uptick in lone-wolf attacks since they began appearing, so it is difficult to conclude that they have had much effect. We investigate whether magazines such as these meet their goals or not. We focus on the language patterns used in these magazines and what they can tell us about the aims and perceptions of the authors, and the constraints on their success.
A mapping of high-level descriptions of the language of influence or propaganda into hypotheses about the ways in which particular, measurable language elements are expected to be used in documents intended to influence.
A computational methodology for instantiating models of particular kinds of language in terms of extracted word frequencies and variational analysis using singular value decomposition. This enables model intensity of different documents to be compared, especially for each of the magazines across time, and hypotheses about expected language use to be validated or falsified.
This, in turn, allows questions about the potential success of these jihadist magazines to be answered: how well they are able to deploy influence (how dangerous these magazines are), and how they react to world events and their own successes and failures (how professional or strategic they are). We are also able to determine authorship of individual issues.
We conclude that both magazines deploy the language of influence fairly well, although Inspire seems to have difficulty staying “on message”. More worryingly, the intensity of jihadist language is rapidly increasing in Inspire.
We assume that intelligence analysts are assisted by automated or semi-automated analytic methodologies to carry out tasks such as extracting text and applying various language models to it, a task we carry out by hand. We make no special effort to handle issues that arise from the conversion process from Portable Document Format to text, or those that arise from automated part-of-speech labelling and word frequency extraction, on the assumption that a scalable version of this methodology should be robust in the presence of these issues.
There is a long history of modelling the language of influence or, more pejoratively, propaganda, and these analyses provide indications of what to expect of the language patterns of documents whose primary purpose is influence. Martin  defines propaganda this way: “an organized deliberate attempt to influence many people, explicitly or implicitly”. It is clear that Inspire and Azan meet these criteria. The essence of propaganda is that it is intended to be opinion forming, and the essence of its mechanism is that it is manipulative. Martin suggests that propaganda is characterized by: simplification, attractive visuals, deceptiveness of some kind, and psychological techniques that weaken or bypass rationality. Work in corpus analytics has made it possible to instantiate these characterizations in terms of particular sets of words and how they are used in documents and corpora.
We apply models with a semantic component: Rayson’s models of informative language and imaginative language , Suedfeld’s model of integrative complexity , Pennebaker’s model of deception , Koppel’s model of jihadist language , and a model for gamification developed by the authors. Gamification is being used as a strategy for motivation in business settings, but it has only just begun to be considered as a potential strategy for influence or propaganda.
Previous analysis of the approach and content of Inspire has focused on four questions: how does the content address the avowed purpose; what mechanisms are used to increase the “stickiness” of the message; what choice model for the readers is assumed; and are there contradictions inherent in the goal of the magazine? Azan has not yet received much academic attention, but the same four questions invite answers.
There are three audiences who appear to be addressed, at least implicitly, by these magazines. The most obvious and important are those sympathetic to jihadist ideology who are capable of being convinced to carry out attacks. Content for this audience consists of coercion towards carrying out attacks, and information about targeting and methodologies to support this. The second audience is the jihadist community itself who need to be shown that steps are being taken, who can point to the magazine as something to be proud of, and who can use it as a tool in their own recruiting. Content for this audience might be framed as morale building. The third audience is the population of Western countries at large. Content for this audience is intended to create terror by suggesting the presence of large numbers of motivated lone-wolf extremists in their midst, hard to identify and ready to strike anywhere without detectable precursors. A detailed analysis of the content of Inspire (up to Issue 7) was carried out by Ford . She argues that the messaging in Inspire is consistent with themes in other Al Qaeda communication, emphasizing: a clash of civilizations, Muslim unity, the necessity of (violent) jihad, religious justification, Al Qaeda’s superiority, and glorification of martyrdom and hero worship. Interestingly, issues of Azan have also (somewhat plaintively) asked for help from a fourth audience, anyone with knowledge of drone operations to help them develop countermeasures.
The second aspect is how the authors attempt to make their messages “sticky”, that is both standing out in the noisy media environment of the West, and compelling to the intended audiences. One major part of this is the use of narrative, both narratives of past successes, and biographies of previous heroes and martyrs of the movement. This is intended to create identification, to make abstract possibilities seem plausible and realistic, and perhaps even to create envy. Another part is the high production values of the magazine itself, making it more readable and serious-looking for Western eyes, used to the production quality of mass-market magazines. Considerable attention is paid to pictures and layout.
The contrast between Inspire and typical Islamist discussion forums, which also attempt to motivate attackers and attacks, is striking. There has also been research suggesting that crowdsourcing and gamification are being used in Inspire as a motivational strategy and to desensitize readers to the human consequences of violent attacks ,.
The third aspect is the choice model that is assumed by the writers and editors, that is what is supposed to go through the mind of a reader on the path towards willingness for violent action . A rational choice model is implausible since the intention is to influence individuals towards actions that will result in their deaths or long terms of imprisonment. Two choice models have been suggested. The first is bounded rational choice, in which the available choices are made to appear constrained, either by controlling the available information, or the amount of resource available to make a decision. In settings where the consequences to the individual of taking action are severe, for example in the military, appeals to honor and community are often made, and these magazines use the same strategy. The second choice model is situational action theory, which models the process as the infection of vulnerable individuals. There is some empirical evidence to support this – it explains the rates of lone-wolf attacks well – but it does not have much explanatory power as a way of predicting which individuals may become infected . Overall, those who are convinced to act in a violent way do not themselves enjoy the results of their actions. Accordingly, those who want to convince them must feel pressure to express the outcomes in abstract terms, and the goals in terms of what they are against, rather than what they are for.
The fourth aspect is the inherent contradiction in producing documents intended to convince others to carry out violent acts at a cost to themselves, an argument that has been made in detail by Ramsay . If it is acceptable to contribute to jihad by writing for and producing magazines, then it must also acceptable for magazine readers to contribute to jihad in less-lethal ways, perhaps recruitment and fund raising. This may go some way to explaining why the effects of Inspire and Azan have apparently been so modest.
Lemieux et al. have described Inspire from the perspective of the Information, Motivation, and Behavior (IMB) framework. This framework suggests that causing others to act requires providing them with information, providing them with motivation, and teaching them behavioral skills that will be required for the desired actions. They show that the issues match this framework quite well – for example, the most famous article in Inspire, “Build a bomb in the kitchen of your Mom” is a recipe book for building a certain kind of IED. The IMB framework may be used intuitively, but it is also possible that Khan and al-Awlaki had some informal psychological background and deployed this strategy consciously.
Informative language intensity will be high as the magazines try to train their readers in both targeting and techniques. (This is slightly unexpected: conventional theories of propaganda tend to emphasize transmission of ideas, rather than of actionable knowledge about, for example, bomb making).
Imaginative language intensity will be high to increase emotional drivers of motivation, and perhaps as a way to bypass rationality.
Integrative complexity will be consistent with other kinds of Western magazines – the goal is not to provide a textbook; but neither is it to become too overtly simplistic.
Deceptive language intensity will be high because of the inherent difficulty of convincing others to carry out actions that are not in their individual interest, and because influence tends to become manipulative.
Jihadist language intensity will be high, reflecting the mindset of the authors.
Gamification language intensity will be non-negligible as as a way to drive action via rewards from the process itself, and to provide emotional distance between acts and consequences.
We investigate empirically whether these hypotheses are supported. The temporal variation in intensity of these semantic models is of particular interest for two reasons: it provides us with insight about how author/editor strategy changes in relation to what happens in the outside world; and it allows us to gauge whether, for example, the danger posed by these magazines is increasing or decreasing.
Empirical evidence from Islamist forums has shown that high levels of jihadist language tends to be associated with low levels of deception ,. The validation of the hypotheses relating to jihadist and deceptive language intensity will shed light on whether these magazines resemble forum language or not.
Since authorship is fairly reliably detectable from text, we can also investigate whether magazine issues are written by the same author(s). We expect the first 8 issues of Inspire to display consistent language; and we can address questions of authorship for the remaining issues of Inspire and all of the issues of Azan.
A language model measures some high-level property of the documents in a corpus by counting the occurrence frequencies of words that are associated with that property. At its simplest, the total number of occurrences of a given set of words can be computed for each document and treated as a score – for example, counting the number of adjectives in a document to measure its `colorfulness’.
We compute a number of such measures for the magazine issues in what follows. Given a set of relevant words, either explicitly as a list, or by capturing all occurrences of a given part-of-speech, we compute a document-word matrix. Each row corresponds to a magazine issue, each column corresponds to a word of the particular language model, and the entries describe the frequency of each such word in each document.
The entries in each row of such a matrix could simply be summed to approximate the intensity of the language in each document. However, frequencies are necessarily larger in longer documents, so the rows must at least be normalized by dividing by the total number of word occurrences (that is, the length of the document), converting word frequencies into word rates.
In most language models, individual words are not equally significant markers of the property being considered, but summing across rows implicitly assumes that they are. Instead, we refine the analysis of a language model by transforming the space spanned by the words into a lower dimensional one, and embedding the (rows corresponding to the) documents as points in this space. If the language model describes only a single-factor property, this embeds the documents in a 1-dimensional space with documents that have the greatest intensity of the language model at one end, and those with the weakest intensity at the other. In other words, the embedding creates a ranking of documents by the intensity of the property. Even when the language model consists of more than one underlying factor, a ranking can be constructed by projecting onto a a vector that captures variation along more than one axis.
A representation of the documents (and the words) in a 2- or 3-dimensional space allows the results to be visualized, and similarities and differences exposed. When there is a time sequence to the set of documents, as there is for the issues of these two magazines, we can also glean information from the trajectory over time of the intensity of each kind of language.
For the semantic language models, we create two artificial documents to represent both strong and weak language patterns. Before computing the SVD, word frequencies are converted to z-scores. Therefore we create an artificial document (matrix row) all of whose entries are +1 and another all of whose entries are −1. These represent documents that appear as if they use every word of the language model at one standard deviation above (resp. below) the mean usage rate in each corpus. These documents do not participate in the computation of the SVD, but are inserted into the resulting space by postmultiplying them (as row vectors) by V∗S−1/2. The resulting positions approximate where a document with intense use of the language model being considered would lie (indicated by an “H”), and where one with weak use of the language model would lie (indicated by an “L”). This is useful in two ways. First, the line joining “L” to “H” indicates the direction in which the language intensity is increasing, and so orients the roles of the other points (documents). Second, it allows us to claim that levels of particular language are high when many of the journal issues fall near or above the “H” level, since this means that many words of the corresponding model must occur more than one standard deviation above the mean, even within such a small corpus.
Document-word matrices were extracted from versions of the magazines converted from Portable Document Format to plain text. This conversion, especially for such visually rich documents, introduces a number of artifacts. We remove only the most obvious of these, since a typical intelligence analysis pipeline would typically not have the resources to clean the text files extensively. Enough investigation was carried out to suggest that the remaining artifacts have negligible effects on the results.
Issue 3 – bombs aboard cargo planes, claim of bringing down a UPS plane in Dubai that is widely believed to be untrue;
Issue 5 – Arab spring;
Issue 7 – 9/11 anniversary.
The results presented in this section for each semantic model are plots with one point corresponding to each document, placed at coordinates derived from the first three columns of the U matrix. These three columns define a three-dimensional space in which the cloud of points is centered around the origin (because of the normalization to z-scores). Its axes are linear combinations of all of the word frequencies; these are not directly interpretable and so we omit them from the images, and the space is isotropic so that distances in all directions are equally meaningful. This space captures the variation among documents derived from the variations in the way that they use the set of words of each model. Distance from the origin is a surrogate for significant variation, and points in the same direction from the center use the set of words with similar frequencies. Thus closeness corresponds to semantic similarity. For issues of Azan and for the first 8 issues of Inspire which were written by the same authors, we connect the points corresponding to issues in time order so that variation with time (trends) can be seen.
Informative language Rayson’s informative language model  represents how a document succeeds in conveying ideas in a rational fashion; in our context, how well a document represents a call to action based on convincing arguments. It is based on the rate of use of nouns, adjectives, prepositions, and conjunctions.
The points marked as high (“H”) and low (“L”) intensity are the inserted artificial documents – documents that use every word of the model at one standard deviation above or below the mean rate. The line between these points defines the gradient from lower to higher intensity and so shows that informative language intensity increases from right to left. Issues 3 and 7, which we earlier commented were triumphalist, are also least informative, which seems consistent.
These results show that informative and imaginative language are not necessarily antithetical – it is possible for a document to be high in both, in just one, or in neither.
Integrative complexity The integrative complexity model , measures the intellectual complexity of the content of a document. It can be further subdivided into two aspects: dialectic complexity, the awareness that concepts can be considered from more than one point of view, and elaboration, the awareness that concepts can contain more than one simple idea . Integrative complexity is measured on a 7 point scale; its usefulness has been considerably advanced by the development of an automatic scoring algorithm which we use here .
Integrative complexity scores for randomly chosen articles in mainstream magazines
Decreases in first-person singular pronouns (“I”, “mine”);
Decreases in exclusive words, words that signal increasing refinement of the thought being expressed (“but” and “or”);
Increases in negative emotion words (“hate”, “angry”); and
Increases in action verbs (“go”, “make”).
Because intensity of deception is signalled by decreases in the frequencies of some words, word rates are mapped to z-scores, and those in the first two categories are then multiplied by −1 so that increasing magnitude is consistently associated with increasing deception. A singular value decomposition is then applied to the resulting matrix.
Jihadist language Koppel et al. developed a language model that measures the intensity of jihadist language. Again, the model was developed empirically. The model words were originally in Arabic; we use English versions obtained using Google Translate. The resulting 85 English words have been shown in other contexts to produce a strong single-factor ranking of Islamic forum postings, suggesting that, although the construction process is somewhat cavalier, the results are usable ,.
The results from the deception and jihadist language models are similar to previous analysis of Islamist forum posts ,. There it was argued that a reasonable explanation was that high levels of jihadist language were plausibly associated with high sincerity, and so low levels of deception. The recent issues of Inspire, which have especially high levels of jihadist language, have levels of deception that are in the mid-range, or perhaps even a little lower than that, suggesting that the same relationship may be present.
Gamification language Gamification has been used in many different settings, for example in business and learning, to increase the attraction of the desired process by adding elements of a game (goals, competition) to them. There are signs that a similar approach has begun to be used in jihadist propaganda as a way to change the calculus of carrying out an attack whose consequences are substantial, by focusing on the process rather than the consequences. In other words, gamification attempts to motivate individuals to take actions towards proximate goals, and perhaps in competition with others, providing notional real-world goals while at the same time blurring or down-playing the real-world consequences associated with achieving them.
We constructed a gamification lexicon using the commonest words from McGonigal’s book  extended with words from some popular gamification web sites.
We now revisit our hypotheses, given the results of the empirical analysis. We justify claims of the form “intensity with respect to this particular language model is high” by appealing to the placement of the artificial documents in relation to the magazine issues. In particular, the document indicated by the point labelled “H” represents a document containing occurrences of every word of the model at rates one standard deviation above the mean. Such a document represents extremely high intensity, especially for the informative, imaginative, and gamification models where the 1000 most-frequent words are used. Thus a magazine issue that lies above the level of the “H” document must use the corresponding language at extremely high levels.
Informative language intensity will be high. This is supported; Azan is clearly better in this dimension, and Inspire wanders from informative to triumphalist several times.
Imaginative language intensity will be high. This is partly supported; most issue lie just below the “H” marker. The issues of Inspire that were lowest on informative language also have low levels of imaginative language.
Integrative complexity will be consistent with other magazines. This is supported; levels of complexity vary but remain in the general range of mass-market Western magazines. This is counter-evidence to the perception that Islamist propaganda is necessarily simplistic.
Deceptive language intensity will be high. This is partially supported – most issues of Inspire are in the mid-range, with issues of Azan noticeably higher.
Jihadist language intensity will be high. This is partially supported. Again, the issues of Azan have higher intensities than the older issues of Inspire, but the three most recent issues of Inspire show the highest levels overall.
Gamification language intensity will be non-negligible. Some issues have very high levels of gamification, but others have very low levels, and there is no temporal pattern. The results suggest that gamification is being attempted, but with unsure technique leading to inconsistent levels.
The results based on these semantic language models are consistent with expectations of propaganda and the language of influence. Thus jihadists have adopted the playbook of governments, as they try to influence readers to adopt a jihadist ideology and, more importantly, to act on it.
Authorship is detectable. We can certainly discriminate the first 8 issues of Inspire from Issues 9–12, although Issues 3 and 7 are substantially different from the other earlier ones. The early issues may have been a collaborative effort between Khan and al-Awlaki, with the balance shifting depending on subject matter. All of the issues of Azan appear to be the work of a single author (despite the different author names associated with individual articles).
Azan issues are more consistent than those of Inspire, which shows a tendency to focus opportunistically on the issues of the moment, rather than on a strategic plan to convince and develop lone-wolf attackers. Azan’s admitted small number of issues have been focused and consistent, which speaks to tight editorial control, and intentionality. The early issues of Inspire were not nearly as focused or consistent; rather they tend to use current events as a jumping off point for each issue. This may be, to some extent, deliberate but it suggests the lack of a strategic and intentional view of the goals of the magazine. The more recent issues have been different from one another, but the common structures relating them in the various figures suggest that there is some deeper consistency that is now in play. This may some background figure who is exercising a kind of pseudo-editorial control, or it may reflect the homogeneity of the community from which it emanates.
It has been suggested that competition is one of the factors that might cause Inspire to become more strident . The intensity of Inspire is visibly increasing, but there has only been minimal overlap in time of the two magazines, so it is too early to examine this conjecture.
We have demonstrated a methodology that can be applied to a corpus in at least a semi-automated way to extract data relevant to a set of language models; and to visualize the properties of the documents in the corpus with respect to these language models. Consideration of similarity/variation and temporal changes allow us to see aspects of the authors and of their mindsets.
From an intelligence perspective, this methodology allows us to draw conclusions about the changing editors at Inspire after the deaths of its founding editors; and the pool of authors and editors associated with Azan. We are also able to judge, to some extent, the sophistication of the authors and editors as they deploy persuasion techniques, visible via some of these language models. We are also able to judge intensity of properties such as informativeness, imaginativeness, deception, and (most significantly perhaps) jihadist language.
- G Weimann, Lone wolves in cyberspace. J. Terrorism Res. 3(2) (2012).Google Scholar
- R Martin, Propaganda and the Ethics of Persuasion, 2nd edn. (Broadview Press, 2013).Google Scholar
- P Rayson, A Wilson, G Leech, in Twenty First International Conference on English Language Research on Computerized Corpora. Grammatical word class variation with the British National Corpus sampler, (2000), pp. 295–306.Google Scholar
- Suedfeld P, Tetlock P: Integrative complexity of communications in international crises. J. Conflict Resolution 1977, 21: 169–184.Google Scholar
- Newman ML, Pennebaker JW, Berry DS, Richards JM: Lying words: predicting deception from linguistic style. Pers. Soc. Psychol. Bull 2003, 29: 665–675. 10.1177/0146167203029005010View ArticleGoogle Scholar
- M Koppel, N Akiva, E Alshech, K Bar, in Proceedings of the IEEE International Conference on Intelligence and Security Informatics (ISI 2009). Automatically classifying documents by ideological and organizational affiliation, (2009), pp. 176–178.Google Scholar
- SA Ford, Inspiring a narrative: a content analysis of Al Qaeda’s English-Language Inspire magazine. Master’s Research Essay, Carleton University (2012).Google Scholar
- EF Reid, AQAP Inspire magazine: implications for crowdsourcing and gamification. Technical report, Research Paper, National Intelligence University (NIU), Washington, D.C. (2012).Google Scholar
- EF Reid, in IEEE International Conference on Intelligence and Security Informatics (ISI). Crowdsourcing and gamification techniques in inspire (aqap online magazine), (2013), pp. 215–220.Google Scholar
- G Ramsay, Targeting, rhetoric and the failure of grassroots jihad. J. Terrorism Res. 3(1) (2012).Google Scholar
- J Woo, J Son, H Chen, in Proceedings of 2011 IEEE International Conference on Intelligence and Security Informatics, ISI 2011. An SIR model for violent topic diffusion in social media, (2011).View ArticleGoogle Scholar
- Lemieux AF, Brachman J, Levitt J, Wood J: Inspire magazine: a critical analysis of its significance and potential impact through the lens of the information, motivation, and behavioral skills model. Terrorism Pol. Violence 2014, 1(1):1–18.Google Scholar
- DB Skillicorn, in Foundations of Open-Source Intelligence FOSINT 2012. Lessons from a jihadi corpus, (2012).Google Scholar
- DB Skillicorn, in Proceedings of KDD 2010, Workshop on Intelligence and Security Informatics. Applying interestingness measures to Ansar forum texts, (2010), pp. 1–9.Google Scholar
- GH Golub, CF van Loan, Matrix Computations, 3rd edn. (Johns Hopkins University Press, 1996).Google Scholar
- P Suedfeld, PE Tetlock, S Streufert, in Motivation and Personality: Handbook of Thematic Content Analysis. Conceptual/integrative complexity (Cambridge University Press, 1992), pp. 393–400.Google Scholar
- Conway III LG, Thoemmes F, Allison AM, Towgood KH, Wagner MJ, Davey K, Salcido A, Stovall AN, Dodds DP, Bongard K, Conway KR: Two ways to be complex and why they matter: implications for attitude strength and lying. J. Pers. Soc. Psychol. 2008, 95(5):1029–1044. 10.1037/a0013336View ArticleGoogle Scholar
- L Conway III, KR Conway, LJ Gornick, SC Houck, Automated integrative complexity. Pol. Psychol. (2012). in press.Google Scholar
- J McGonigal, Reality Is Broken (Penguin Books, 2011).Google Scholar
- DB Skillicorn, C Leuprecht, in Proceedings of Foundations of Open Source Intelligence, at Advances in Social Network Analysis and Modelling (ASONAM). Improving the language of influence (ACM & IEEE, 2013), pp. 1028–1033.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.