Computational opposition analysis using word embeddings: A method for strategising resonant informal argument

1.1.Semiotics

Semiotics, often simply defined as the study of signs, has a long history dating back to antiquity. The Hippocratic tradition, for example, used the term σηµειον (“semeion” from which the term semiotics derives) to refer to signs in medical diagnosis – what we now call “symptoms”. Signs also have a place in the earliest history of argumentation. Aristotle used the term semeion, writing that the materials of enthymemes are “probabilities and signs” and that the syllogism is valid only for “necessary signs”. (Rhetoric, I, ii) [3].

Semiotics in the modern era derives largely from the work of two theorists: Ferdinand de Saussure, the Swiss linguist and instigator of structuralism; and Charles Sanders Peirce, the American polymath known as the father of pragmatism. While Saussure’s formality provides the bridge from linguistics to general sign analysis, Peirce is particularly relevant to the approach taken in this paper because he placed at the core of his inquiry the assertion that “all thought is in signs” [30, Vol. V, par. 254]. As a point of departure, Peirce’s ontology allows argument, and even argument schemes, to be viewed as evolving patterns of thought rather than fixed psychological capabilities. The mind from a Peircean perspective is something fallible that will resonate to its benefit or peril with its environment.

Marshall McLuhan, in his pioneering work on the semiotics of new media, expanded on Peirce. He urged us to pay attention to what modulates in our environments in order to understand what is happening in patterns of thought. A central tool for this effort in McLuhan’s later work was that of “figure” and “ground”. Borrowing from gestalt psychology, McLuhan maintained that we attribute our thoughts to only a few artefacts (the figure) while many others to which we have not paid attention are just as determinative (the ground). Hence, in his famous title, the medium (to which we rarely pay attention) is actually the message. In the case of informal argument, we theorise that subtle differences in the ground (typically words) unite and polarise allies and opponents on an issue, and that the figure of enthymemic or inferential structure that we retrospectively attribute – or to use Woods’ term “dress” [42] – sits upon this.

The most obvious use of ground modulation in persuasion is the tendentious choice of words, which was given explicit consideration by Aristotle:

Such modulation may have once been a minor rhetorical consideration. However, the rise in quantity and intensity of sign presentation the internet has made possible, and the processing pressure it puts us under, has unmasked – some might say, promoted [43] – its importance. In many cases of persuasion involving technology, for example, it may feel that we are not so much logically convinced as worn down or distracted into acquiescence. Technology has an indefatigable power to insist, and so any effects of ground modulation it employs (such as tendentious diction) are amplified.

To model the ground modulation structure of natural texts, we propose a computational approach to a common technique from applied semiotics that seeks to capture such patterns in ground and figure formations: opposition analysis.

1.2.Opposition analysis

In professional contexts, the semiotic tool most often employed to bring formality to questions of persuasion is opposition analysis (OA), whose history back to Saussure [7] is précised with a focus on computational semiotics in [4]. In establishing a visual format for OA, Greimas’ work on the semiotic square has been particularly influential [5,13,14]. In commercial practice, simplified versions of the semiotic square are more commonly employed, at least in initial surveys. For example, marketing consultants often begin by using OA to locate a product along two axes of “binary” opposition. Such a preliminary analysis of soft drink brands from [29] is offered in Fig. 1. It plots four well-known brands on axes of mature versus trendy, and individualistic versus social.

Fig. 1.

A simple binary opposition analysis of soda brands ([29] p. 133).

The axes for such OA diagrams are normally selected impressionistically based on:

Perhaps to enhance their impact in presentations to decision makers, the results of even fine-grained OA are commonly distilled to just two orthogonal axes and additional dimensions are smuggled in using colour differentiation and quadrant labels. This is illustrated in Fig. 2, which is an OA used, according to its authors, to model gaps in the children’s activity book market and set the basic parameters for the bestselling Unbored series. By plotting colloquial labels for characters in children’s lives along axes of silly↔earnest and authoritarian↔permissive, the authors arrived at meta-labels for each of the quadrants to represent the types of purchasing parents (“Wacky”, “Sherpa”, “Cajoling”, “Particular”). In doing so, they identified a gap in the market (the empty circle in the top right quadrant) which they then sought to fill with a duly customised product. Clearly this analysis is much more nuanced and purposeful than that in Fig. 1 due to the extra dimensions embroidered over the main axes.

Fig. 2.

An opposition analysis used in a project leading to the Unbored children’s activity books undertaken by Glenn and Leone [11, 14m25s].

1.3.Motivations for computational automation of OA

The centrality of OA to the theoretical position taken in 1.1, and the widespread availability of digital texts containing informal argumentation, makes a computational approach to OA an appealing avenue for investigation. While semiotics has taken an interest in computation [1,8,9,20,25], apart from [4], little work has been done in computational treatment of opposition analysis. To our knowledge no automated means of generating OA diagrams has been published. Moreover, while automation of a manual practice would be a valuable contribution in itself, we discern two other motivations.

Firstly, there is an opportunity to perform OA computationally in much higher dimensions than currently possible using impressionistic, “culture surfing”, manual methods. In Fig. 2, for example, the use of different colours for plotted values is used to include a third variable. One might consider the quadrant labels such as “Wacky” a fourth, but three dimensions is normally as far as manual OA goes. A general procedure for n-dimensional OA seems absent from semiotics. To view soft drink brands or activity books from a different perspective (rich versus poor, for example) it would seem necessary to rethink and redraw the entire analysis. In addition, a computational procedure for rotating oppositions between axes would, at the very least, be useful for revealing relations “on screen” that are not immediately obvious and help generate new ideas and understandings for persuasion in a particular field.

Secondly, computational treatment of OA can increase precision. As noted, OA is often reduced to binary OA in which objects are not plotted with specific values but, as in Fig. 1, simply allocated a quadrant. This not only affects the precision of analysis but makes longitudinal or real time updating an onerous task. It would be difficult, for example, to check on a weekly basis how near or far from each other Coke or Pepsi are on the mature↔trendy axis of Fig. 1 and tune online advertising accordingly. Moreover, the consistency of judgements at different times currently has no statistical rationale. With the streams of data now available to influencers from social media and microblogging, this capability would seem achievable and have obvious relevance and value.

For the reasons developed above – automation, multidimensionality, precision, novel visualisation, and time series updating – this paper contributes an approach to computational opposition analysis (COA) that it is hoped will be of use to those seeking to persuade.

2.1.Artefacts as proxies for thought

Summarising for present purposes, finite semiotics grounds the notion of opposition in the more obvious and ontologically tenable concept of oscillation in thought: the tendency of our thoughts to vary in regular patterns over time [38]. It asserts that this patterning arises in response to the cumulative effects of our environments, which are typically rich in human artefacts11 evolved to play a disciplinary role in thought. Proof of this is trivial: it would be difficult to argue, for example, that learning to read in childhood does not pattern our thoughts so that later we react to text in an email. The recurrent presentation of signifiers via artefacts correlates with recurrence and homogenisation of thought across a population. Ubiquitous artefacts and their patterning, therefore, can be considered proxies for thought; and to study thought we need to study this patterning. This motivates us to prefer in our approach the living term proxy to the moribund and text-centric term corpus in referring to artefacts from which our COA approach will be derived.

Although the practice of regarding artefacts as proxies for thought is rarely formalised, it is of course not new: we do this every day when we read an email from someone and believe we know their thoughts (or at least some refracted version of them). Even the artefacts of antiquity are still regularly taken as proxies for thought: we believe we know something about the thoughts of Aristotle, for example, by reading works attributed to him. Such habits of interpretation of proxies, once formalised, can be extrapolated from the artefacts the digital age now supplies in abundance. Our thinking around “gun control”, for example, may result from actual experience with guns, but also from cumulative exposure to television and movies, comments seen and made on social media, reports in the news, academic study and so on. In this paper, we contend that an analysis of the signifiers contained in such artefacts, taken as a proxy, can be used to derive a high-dimensional, population-level oppositional “shape” for thought about an issue.

2.2.Oscillation in thought as the basis of opposition

Finite semiotics does not propose that the shape of any particular formation of thought such an argument is static – that each side is locked in its own immutable, irreconcilable thought patterns and never the twain shall meet. Instead it views opposition as the result of systematic deviations from the common ground of thought that necessarily engender the issue. To illustrate, consider these two concocted, but perhaps not atypical, statements on gun control:

Figure 3, adapted from [38], illustrates how the patterns of thought implied by these statements engender both the issue and its oppositions. The two agents making the statements think about “gun control” using some common concepts such as “fear”, and “grow up”. However, they deploy these in quite different ways that give rise to oppositions. Some of these deployments may be incidental (e.g., “gun control” versus “guns are controlled”), while others, which might be seen as “field-defining” in Toulmin’s [39] terms, are likely to recur often in discourse on the issue (e.g., “women” and “men”, or “murder” and “death”).

Fig. 3.

How patterns of thought around an issue such as “gun control” imply oppositions. (Note: the space represented is not intended to be taken as two-dimensional. It is a flattening of a theorised high-dimensional space.)

The underlying notion of oscillation is that the thoughts of differently convinced agents will nonetheless partially overlap. A common issue-defining set of thoughts will occur, as will a number of common variants. This leads to complex potentials for thought about an issue to occur as transitions from relatively defining to relatively variant structures. The more stable of these potentials are what we identify in our theory as oppositions.

It should be noted that this conception is also commensurate with drift in patterns of opposition. As new artefacts emerge (seemingly unrelated animal rights advertising, for example) the set of defining and variant thoughts change their relations. If someone is convinced to become a vegan, for example, many of their associations about gun control may shift also. This is not to say they will necessarily change sides. They may reverse their position, or simply come to think about the issue according to a new set of oppositions.

In dealing with such a dynamic entity as thought, it is self-evident that oppositions may vary greatly in being persistent or transitory. Persistent oppositions are often of most interest to the analyst, and certain types are already enshrined under special categories as antonyms or synonyms. Such oppositions are not our primary concern and are viewed simply as a special case. Oppositions may be of a much less expected form depending on the dynamic contexts created by circumstance. From Fig. 3 it is possible, for example, to see how “sons” and “children” might be strongly opposed in regard to gun control but not in regard to the abortion debate, which has different commonalities. Insightful work in this vein on different contextual varieties of textual opposition and how they are created has been carried out by Jeffries [18], while van Belle [40] has provided cogent analyses of “multimodel media” based on Fahnestock’s view of Aristotelian antithesis [10, ch. 4].

In summary, finite semiotics proposes that, to make it amenable for computational analysis, the synchronic notion of opposition must be “de-reduced” into the diachronic thought variations or oscillations that constitute it. Those thoughts, in turn, should be sought in the patterns of signifiers found in artefacts that can be considered their proxies. In this way, instead of a static set of oppositions, a proxy can be viewed as a set of complementary oscillations that it provokes in an audience through time. This more dynamic view of opposition allows us to approach it with more nuance, remain sensitive to its complexity, and find ways to trace and predict pathways of thought that are useful for persuasion.

3.1.Proxies

It was argued in 2.1 that opposition analysis must begin with the selection of artefacts that can be used as proxies for the target thought. Such proxies must contain a sufficient number of signifiers and arise from a context in which argument is in evidence. An ideal source, therefore, is one in which texts representing polarised views are identified.

On the topic of gun control, we found two opposing sets of texts on the website ProCon.org [31], which we designated Pro and Con. In accordance with the background discussion above, the articles were taken as proxies for two coexisting types of prevalent thinking about gun control. The Pro side favour tighter gun control; the Con side favour less control. Each set contained fifteen texts for a total of thirty. According to the ProCon.org site, the texts are a “professionally researched and curated nonpartisan compilation of the best Pro, Con, or General Reference responses” [32] to the question “Should More Gun Control Laws Be Enacted?”, and are sourced from “websites, magazines, newspapers, libraries, transcripts, videos, interviews, emails, legislation, direct correspondence, and more” [32].

3.2.Step 1: Identifying relatively prominent signifiers

In two proxies selected to represent opposing sides, a number of common signifiers will inevitably be prominent. In gun control discourse, for example, one would expect to find many instances of gun, rights, crime and so on. In our oscillation-based approach to OA, however, it does not necessarily follow that common signifiers are most provocative of semiosis. In fact, the signifiers derived from common topic and relevance measures in our model more likely constitute the stable zero axis around which oscillation occurs. More germane to opposition in our model are signifiers that are relatively more prominent in one proxy than another. A simple way to identify these is to take the difference in term frequency (expressed as a fraction of total terms) between the two proxies.44 So, if the frequency of gun in the Pro proxy is 0.017 and in the Con proxy it is 0.014, then the difference of 0.003 will probably not mark this signifier as one of interest. But if the frequency of massacre in the Pro proxy is 0.017 and in the Con proxy it is 0.003, then the difference of 0.014 may indicate a noteworthy signifier and something that resonates and divides.

A general formula, therefore, to derive a metric for the relative prominence of a signifier in related proxies is to take the difference of its frequency in each proxy. Formally stated: for signifier s occurring in one or both proxies A and B with total numbers of signifiers TA and TB and frequencies of s, fA(s) and fB(s), the relative prominence rs can be calculated as:

For convenience only, with regard to the gun control proxies, we assign A to the Pro proxy and B to the Con proxy, and this leads conveniently to positive or negative weighting of r indicating whether the signifier is favoured in the Pro proxy (positive) or Con proxy (negative). The top ten prominent relative signifiers for the Pro and Con gun control proxies calculated using this formula are shown in Table 1.

Figure 4 contrasts the top twenty relatively prominent signifiers in the Pro and Con gun control proxies, of which 13 are favoured in the Pro proxy and 7 in the Con. The dominance of the Pro side in the figure could suggest it exhibits a higher degree of polarisation that the Con proxy. In other words, it may indicate that it makes greater use of its preferred terms.

Fig. 4.

Bar chart of the top twenty relative prominent signifiers in the Pro and Con gun control proxies.

3.3.Step 2: Identifying and ranking possible opposition pairs by scoring on a measure of relative prominence and contextual similarity

As opposition must occur within the ordering restrictions of the code in which signifiers are embedded (in this case, the syntax of English), analysis of word context can be useful in determining which relatively prominent signifiers are most in opposition. At present one popular means of doing this is using word vectors,55 a popular implementation of which is “word2vec” [23]. Such models may be applied either internally by training on the proxy data provided it is sufficiently large, or externally by invoking a comparison with a large external proxy to provide a baseline of syntagmatic restrictions.

To illustrate the logic of the opposition ranking procedure, consider an example: if the Pro gun control proxy suggests the prominence of the signifier death relative to the Con proxy, and the Con proxy suggests the prominence of life relative to the Pro proxy, we can quantify this as an opposition (that is, as a defining substitution) by referring to the contexts in which each occurs in each proxy. If death occurs in one where life occurs in the other there is reason to think the signifiers are in opposition. For small proxies, this contextual similarity can be normalised by reference to a large, heterogenous external proxy. So, if the signifiers death and life commonly occur in similar contexts in broader discourse, we can theorise that both are possible in each Pro or Con context and that the use of one rather than the other (their relative prominence) tends to differentiate the two proxies. In other words, signifiers scoring highly on relative prominence and contextual similarity are excellent candidates to be strongly opposed and to occupy the ends of an axis of opposition in our COA.

The above logic can be formalised by stating that the opposition score O for two signifiers s1 and s2 is a linear combination of the differences in the relative prominence of the signifiers and a measure of their contextual similarity sim(s1,s2):

The constants k1 and k2 might be used to explore ratios of relative prominence and contextual similarity. For present purposes, if we assume these are both 1, the expression becomes:

It should be noted that the inclusion of the two relative prominence measures in this formula keeps commensurate contextual similarity and frequency of occurrence. Signifiers of great similarity in the formula will not be considered highly opposed if their relative prominence is low.

3.4.Step 3: Plotting the proxies on a classic opposition quadrant diagram

Having arrived at a computational means to identify oppositions in the proxies, it remains to plot them on a traditional OA diagram. The obvious means to do this is to discount occurrences of one signifier in a proxy with occurrences of the other signifier in the proxy. Thus, an axis score for each proxy can be derived by subtracting the normalised term frequencies of the two signifiers making up the axis. For example, for the woman↔man axis, the Pro proxy has a normalised term frequency of 9/1490 for woman and 0/1490 for man, giving it a plot value of 9/1490 (0.006040268). The Con proxy has a normalised term frequency of 1/1525 for woman and 6/1525 for man giving it a plot value of −5/1525 (−0.003278689). Formally, then, for a proxy A, its plot value v on the axis s1↔s2 is the frequency of s1 in A minus the frequency of s2 in A over the total signifiers TA:

To create a traditional OA diagram, the top two scoring axes (woman↔man and death↔murder) in the gun control proxies can be combined as shown in Fig. 5. Observe that the score for the Con proxy on death↔murder, while considerably lower than the Pro score, is positive due to the prevalence of death in the Con proxy. This is a common occurrence in OA diagrams as evidenced in Fig. 2. It may become highly relevant in analyses involving larger numbers of proxies.

Fig. 5.

A computationally generated opposition analysis diagram showing the gun control Pro and Con objects on the two highest scored axes.

Due to their computational provenance, these diagrams could be generated in as many dimensions as required. This would allow multidimensional exploration of the proxies using visualisation technology (such as Google’s TensorBoard, for example). Further inspection could be performed to relate axes of opposition on other purposeful metrics than simple ranking. For example, guided by the purpose of the analysis, the death↔life axis (which also ranked highly at 0.005150) could be substituted for death↔murder to further explore the relation of death to woman↔man. It would also be possible derive metrics that give a percentage of total opposition accounted for by a given combination of opposition pairs.