Advances in socio-semantic network analysis: discussion panels


Table of contents

Organizers

The organized session

Acknowledgements

Panel descriptions

Panel 1 – Seeking the atomic unit of cultural meaning

Panel 2 - Innovations in socio-semantic network analysis

Summaries

Sunbelt 2018, Utrecht

Panel 1: Summary of discussions

Panel 2: Summary of discussions

Sunbelt 2019, Montreal

Panel 1: Summary of discussions

Panel 2: Summary of discussions


Organizers

The organized session

At the name suggests, our session focuses on advances in socio-semantic network analysis. This emerging framework, where semantic and social network data are being jointly appraised, still poses a number of theoretical and methodological challenges.

In our organized session, we aim at addressing some of these challenges. Presenters and interested participants generally address theoretical, methodological or empirical issues related to socio-semantic networks, such as:

In recent years, our organized session also created a space for discussion (panels), where interested participants and session presenters were invited to discuss topical issues in socio-semantic network analysis. The purpose of these panels is to stimulate open discussion on relevant issues.

This document contains the descriptions of the discussion topics proposed by the organizers, as well as the summaries of the very engaging discussions these topics generated at Sunbelt 2018 in Utrecht and Sunbelt 2019 in Montreal.

 

Acknowledgements

We would like to thank all our session presenters and discussion panel participants, in Utrecht and Montreal, for joining and contributing to these very engaging and thought provoking discussion.


Panel descriptions

Panel 1 – Seeking the atomic unit of cultural meaning

Chairs: Nikita Basov and Camille Roth

This panel will discuss theoretically informed ways to capture cultural (shared) meaning at the elementary level. Perhaps, a bag of words is not enough. But are word associations or grammars sufficient? Should (2-mode) links between individuals and words also be accounted for, as the duality perspective suggests? If so, do we also need to account for social ties, because meaning is something inscribed and articulated in those, as proposed by relational sociology? Or perhaps, we do not care about social ties because interaction has little to do with sharing of meaning, as systems theory argues? Or should we account for the absence of interpersonal ties to encapsulate switching between relational contexts as the source of meaning?

Panel 2 - Innovations in socio-semantic network analysis

Chairs: Iina Hellsten and Adina Nerghes

Socio-semantic network analysis not only applies current methods in SNA and semantic network analysis, but also aimed at developing innovations in how to approach the social actors and the semantic units for the analysis. In this panel we will discuss the merits and limits of such innovations. Notably, we will discuss different, feasible ways of analyzing 2-mode networks while preserving the richness of such data; advances in 2-mode network measures (e.g., clustering or betweenness centrality); the question of bipartite relationships and their meaning to the analysis of socio-semantic networks; the current state and future aims of 2-mode network theory development. Several approaches to analyzing socio-semantic networks were discussed.


Summaries

Sunbelt 2018, Utrecht

By Adina Nerghes

Panel 1: Summary of discussions

The first panel started with a short presentation by Camille Roth. Camille recalled that socio-semantic networks were initially understood as a symmetric construct, while the social and the semantic sides usually refer to distinct ontologies. On the social side, there is little ambiguity in the definition of nodes (i.e. actors: usually individuals, sometimes organizations or groups) or links (denoting either interaction such as discussion, collaboration, friendship,  or affiliation, such as co-work, co-attendance, co-membership).

On the semantic side, the situation is fuzzier: nodes may typically denote hand-made labels, keywords, or automatically-extracted words, lemmas, and n-grams. Yet, the recent NLP literature makes it increasingly easy to extract more sophisticated and, more importantly, non-atomic semantic concepts from text corpuses, such as topics (which may in turn be represented either as partitions or distributions on words) or claims (which may appraised be as subject-verb-object predicates, or as sets of versions of a given utterance).

However, the network formalism works best for atomic items and it is unclear whether concepts generally lend themselves neatly to typical network analysis tools. For instance, what could or should ‘betweenness’ mean in a semantic network, what could a link between two claims refer to ? Camille eventually asks whether the semantic side should be considered as a network at all, instead of simply focusing on the attribution of concepts to actors. In this respect, we could envision socio-semantic networks as semantically enriched social networks, whereby the actors (and their relationships) would be complemented with a rich semantic description, in a manner similar to that of the social semantic web where semantic networks also feature social attributes.

Camille’s presentation was followed by Jan Fuhse talking about the ‘atomic’ unit, Interactions, and relations. Jan’s presentation focused on three main issues:

  1. The idea of the atomic unit in semantic networks as words and how words are not physical and that words acquire meaning through relations to each other. If we think of symbols and the relations between them, a symbol is a relation because a symbol acquires its meaning through relations to other symbols. Thus, meaning is part of the symbol and not something that lies outside of the symbol. Because of this, symbols are different from human actors or organizational actors. Human actors do things alone and human actors have attributes, while symbols do not have that.

  1. The most important relations are those between symbols and actors. Co-occurrences between symbols are not always uniform but they are forged by actors. Actors are the ones connecting symbols. If we look at symbolic networks as two-mode networks of actors and symbols, we realize that the culture of discourse is not coherent or homologous, it is not something that is shared uniformly across the actors that are involved. Rather, it is contentious and dialogical. Thus, the basic unit we should be looking at is the link between the actor and the symbol.

  1. If we look at these links between actors and symbols, we construct things as if they were stable, as if an actor is in possession of the symbol but what we in fact observe is some kind of process. It is an event, when these connections are forged and they might not be there in the next moment. Thus, we have to switch from the possession perspective to the practice perspective or a process perspective and also look at the temporality of links that are forged and then forgotten.

After this short presentation, the discussion moved on to the ways in which we define actors and whether these actors are also symbolically constructed. Loet Leydesdorff proposed a scenario in which an actor is relevant because his position as member of parliament, and not because he is Mr. Smith.  

From here, the discussion moved towards the ways in which actors connect symbols and how framing and evaluations may affect the structure of such actor-symbol networks. While looking at the other symbols employed by actors may provide some evidence regarding their use, issues like irony or sarcasm remain open challenges.

Following this discussion, Jana Diesner raised two main issues:

  1. Contextualization: What is the frame? What is the meaning of the link? What is the directionality? Is there a temporal dependency? These are some of the things we should put a lot more effort in;
  2. How precise is this? How much can we trust this? Thus, a call for more work towards best practices in socio-semantic network analysis.

Nikita Basov proposed that in sociology there is a need for a diversity of approaches to problems, depending on the ways in which they are theorized.  Linkage to theory does not always allow us to conclude on best practices. He also suggested that two-mode networks may not be enough because symbols may be used differently by the same actors in different discursive contexts.

The first panel was concluded by Johanne Saint-Charles with a remark that the discussion topics of the panels – one more theoretical and one more applied – cannot be separated and with some thoughts on the importance of the traditional social networks and the importance of the different types of relationships between actors.

Panel 2: Summary of discussions

The second panel started with two short presentations on two approaches to socio-semantic networks. The first presentation, given by Loet Leydesdorff covered a theoretical point as well as a methodological one.

Firstly, the dynamics of meaning are very different from the dynamics of social relations. Loet proposes that we need a dualistic view.

Meaning is provided from the perspective of hindsight, so against the axis of time, which requires us to move from a monolithic to a humanistic perspective.

Secondly, if we organize multi-mode networks in the ‘full matrix’ approach, we can add many more attributes and types of relations than in the two-mode or multi-mode networks.

The second presentation, given by Gabriel Jaime Velez Cuartas covered a different approach to mapping of documents, as well as a more conceptual point on theories and definitions.

Firstly, Gabriel highlighted the issue that when we talk about networks, we are talking about definitions of sets of networks or vertices and that solving any kind of paradoxes requires us to start from the definitions. Thus, the way we select our network links and nodes is semantically defining our network. When we create networks, we are defining these networks. If we take a cybernetical perspective, we can understand these practices as selections and these selections become the starting point towards the understanding of meaning. Gabriel also proposed the term ‘a semantic construction of networks’ for his argument. Gabriel then went on to present his approach to mapping of documents, an approach showing how we can understand documents with different weights.

After Gabriel’s presentation, Iina Hellsten proposed a question, in line with the earlier points raised by Loet: If we think social networks develop over historical time while meaning can only be provided in hindsight, what does that mean methodologically for socio-semantic networks? This question encouraged a discussion on the dynamics of social relationships as well as the temporal characteristics of meaning. Jan Fuhse argued that meanings are instantiated, and so are social relationship. If we are interacting with each other, the meaning of the interaction is only later negotiated through follow up interactions. Thus, a social network should not be different in that regard because it is always recursive. Social relations are only built through meaning and social interactions taking place (which are meaning).  

This line of argumentation furthered the conversation into the kind of predictions can be made based on information from the past as well as predicting options not taken in social relations. While we can only mode what we can observe, we can hypothesize the options not taken which in turn will change our vision of reality changes - in line with Gabriel’s argument earlier -- and so different hypothesis come to the fore.  Loet then suggested that we are also looking at the intersubjective layer, the intersubjective intentionality.

This line on discussion was followed by Iina Hellsten, drawing on the first discussion panel, and raising the issue of what is the atomic unit of meaning and that there is something more to meaning than the words exchanged in a conversation. This would mean going from mapping co-occurring words to a more general level. A metaphor, for example, carries a lot more meaning than a lot of other co-occurring words in semantic networks. But that makes it difficult to define what the meaning unit and what the more general level should be. Not everything is as meaningful in social interactions and that raises the question of how we can best use socio-semantic networks to uncover the meaningful parts of the interactions.

This methodological turn in the discussion was continued by Johanne pointing towards that fact that  figuring out what the methods we employ are revealing and not revealing demands more attention. With most of the methods we have, we will not be able to reveal that words that are most impactful in terms of the ways in which social relationships are constructed through interactions.  However, we do have the tools and methods to understand many other things and to better the world. Is then the question: What are our methods doing and how can we tell what we’re missing? And how much we do we say about the non-relationship (even in social networks)?

To these points raised by Johanne, Loet pointed that what we are missing is the continuous generation of options in social relations because we are only looking at the instantiations and not at the options. However, the number of options can be very large in a theoretical sense. So should these options then be confined to those that are essential to the         actors involved?

The next issue discussed regarded the different layers that we are looking at and the ways in which we can operationalize them in a network form. For example, how do we structure and analyze a network that takes into account the relationships between actors (social), the relations between words (semantic) and the relations between actors and words (socio-semantic)? To this, the discussion concluded that the primary focus should be the question we are trying to answer. Loet suggested that we should perhaps aim to create an entropy measure, which will allow us to measure what has been realized and the redundancy of this process.

As a concluding remark for the second panel, Nikita Basov pointed to the fact that – just as in the past – what we do is still very different and our approaches to socio-semantic network analysis are still very different, which is a great feature of our small community.

Sunbelt 2019, Montreal

By Adina Nerghes

In Montreal, our two discussion panels focused on continuing the discussions triggered by the two topics introduced the previous year in Utrecht (see description here). Camille Roth chaired and opened both of these discussion panels with a brief introduction of the topics.  

Panel 1: Summary of discussions

The first discussion panel focused on the first topic discussed the previous years in Utrecht, the atomic unit of cultural meaning. Camille briefly introduced this topic and the two main issues: unicity and atomicity. Unicity relates to the idea of whether or not two entities using a similar concept refer to the same meaning, while atomicity refers to the unit we want to analyze (words, ngrams etc.) to characterize the concepts actors are associated with.

The discussion started with the issue of cultural elements as concepts or words on the semantic side. If two words are linked, we can derive an association and by collecting and grouping more such associations we can derive topics or themes. And while looking at concept associations in the context of the people or groups to understand meaning is not new, the question remains whether or not this is a proper practice to capture meaning. For example, we know that when a political artist speaks about political art and when a politician is speaking of political art their meanings will be very different. Thus, meaning is not only in the connections or in the groups of people but also in the characteristics of the people using them. As such, we should account for the interactions between people as well as the expressions of meanings. This brought us to another important question:  what is the sufficient elementary unit of meaning that we would be satisfied with?

From this question, the discussion moved to the idea that the elementary unit of meaning depends on the expected outcomes. Simplified, if the sharing of concepts leads to social connections, then we should look at how complex the structure of meaning should be in order to be able to relate it to the outcome we are interested in. However, defining outcomes is also problematic due to the different perspectives. Researchers, participants etc., have very different perspectives on what outcomes are.

The discussion on outcomes highlighted the different way of  approaching this type of research, whether from an inductive or a deductive perspective, as well as the potential need for a clear definition of what we understand cultural meaning or shared meaning to be before we start measuring it. When discussing definitions, the dependence between the social and the semantic was highlighted, as well as the danger of defining concepts as a function of observations. Thus, it is crucial to always determine and recognize the different roles and functions observations and data play in the different phases of the research process (e.g., generating theories, concepts, hypotheses, testing etc.). It was also highlighted that whether you take an inductive or a deductive approach to this type of research, the typical process (e.g., from theory to methods to data) rarely works when working with real-world data and that the actual research process generally becomes iterative.  

Here, the discussion turned to back to the idea of meaning and the dependence of this meaning on social actors and their context. The question once again became: What do we do when concepts in a semantic network have different meanings, depending on the social actors using them?

The discussion here highlighted three elements: semantics, meaning and context. First, it was suggested that in order to disambiguate the meaning of concepts we need all the semantics that stem from language. Second, it was suggested that the meaning of concepts is generated through interactions and outside the interaction. Thus, meaning depends not only on the ways people use words with each other, but also on how that word is used elsewhere. Third, the importance of context was brought into the discussion. It was suggested that co-occurrence makes it difficult to disambiguate meaning and that adding another layer of context can help. For instance, the history of the person using the words. However, the difficulty of including context, which is always changing, was also discussed. While there seems to be an assumption that because concepts are used together, they are shared, meaning is always changing and just because two people share a word does not necessarily mean they refer to the same meaning.

This led to yet another important question: how we can methodologically validate this type of structure between language and the people producing this language. This topic linked back to the importance and difficulty of including context and to the concept of shareness and the meaning behind sharing concepts. People always share language but the important issue is how people use concepts in specific contexts and whether we can link shareness of concepts to the history of the link between two people.

This first discussion panel concluded with yet another important issue, which we aimed to discuss during our second panel (see summary below): If we consider ties as containing the language and the history of the person (context), how can we operationalize this?

Panel 2: Summary of discussions

The second discussion panel focused on innovations in socio-semantic network analysis, especially with regard to social network analysis, following the second topic addressed in Utrecht in 2018. More specifically, on the issues arising from the application of classical social network analysis (SNA) tools to socio-semantic networks. As network researchers, when we see semantic networks we tend to naturally apply social network tools. It is also meaningful to apply such tools to semantic networks? What does clustering mean in semantic networks? What does it mean to speak of eigenvector centrality in semantic networks?

When we talk about socio-semantic networks, networks with actors and concepts, the same questions apply. For instance, is applying 2-mode network techniques to socio-semantic networks a valid approach? Is it good practice to look at socio-semantic networks as 2-mode networks and apply 2-mode network metrics? And at a higher level,  we have lots of SNA methods to determine communities, but what does that mean when we work with networks of actors and concepts?

This also extends to visualisations of socio-semantic networks. For example, the frequent use of force-directed layouts when visualizing socio-semantic networks. What are the implications and the intuitions behind the application of such layout algorithms to socio-semantic networks, knowing that the original justification of such layouts comes from the fact that nodes should be represented close to one another if they share many common connections. When introducing concepts/words alongside actors, is this justification for force-directed layouts still valid? And can we think of different ways of visualizing these types of networks? For example, the use of matrix based representations of socio-semantic networks, which are not very common ways of representing 2-mode networks. Matrix based representations are a way of organizing columns and rows in such a way that sub-groups become visible. Thus, from a methodological perspective, what are the implications of applying classical SNA tools when dealing with socio-semantic networks?

Perhaps we should start simply by trying to understand what does it mean for two actors to be linked by the similarity of their discourse and what does betweenness centrality mean in such networks. And what does betweenness centrality mean in a semantic network? Is it a measure that speaks to grammar rules or language usage? So maybe what we need is a conversation on the meaning and testing of such measures.

The first idea here was that—specifically in semantic networks—the meaning of these measures depends in great part of how the networks are generated. For example, in the standard approach of removing most of the grammar in the textual data, betweenness centrality will be indicative of connective concepts, while in semantic networks generated from intact textual data measures such as betweenness centrality would speak more to the grammatical/syntactical role of words and concepts. The complexity of understanding these metrics rises even more when we add other layers of connections, such as social interactions between actors, links between actors and concepts etc.

The informativeness of betweenness centrality depends in great part on how the semantic networks are mapped, but this is also true for social networks. For example, betweenness is very informative when we talk about social networks of information flows but it becomes less informative when we talk about dyadic relationships. The same applies to semantic networks, in which we can’t read co-occurrences in semantic networks as sentences.

One of the opinions voiced in this discussion was that in the case of co-occurrence semantic networks, betweenness centrality does not really work. Another opinion was that if we want to use betweenness centrality and the networks are formed based on units of language and discourse, we first need to establish what the relationships between the elements of discourse are.

As the problem of betweenness centrality is a very complex one, a simpler example was suggested by one of the discussants. If we take a network of synonyms, in which the links are based on specific features that denote synonymity between concepts, and if we have a very specific way of expressing meaning through a concept, the shortest path in the network will pass through this particular concept.  

Another point of discussion here referred to the idea that maybe of measuring betweenness centrality of concepts in a semantic network, we should perhaps measure the betweenness centrality of larger units, such as topics (e.g., hashtags). Thus, the level at which we investigate semantic networks should not be single co-occurring words.

Another point that was brought into the discussion was that of comparisons. The use of betweenness centrality and other such SNA metrics can be useful when we are trying to compare social networks to semantic or socio-semantic networks and answer questions such as: how do they compare, how do they co-evolve or how do they influence each other. For these reasons, we would need to have comparable measures and this could be one of the reasons for applying SNA metrics in the semantic context. But then we come back to the idea that betweenness is mostly about path and directions, while semantic networks traditionally are non-directional. So perhaps betwenness centrality would become useful if and when we deal with directional semantic networks. But when talking about directionality in semantic networks, in most languages grammar dictates the ordering of the words in a sentence.

We can take any network metric and interpret it based on the kind of network we are applying it to. When applying such measures to semantic networks, their interpretation refers to meaning and that is acceptable. Network metrics do not have to mean the same thing, their interpretation is dependent on the context to which they are applied.  

Beyond betweenness centrality, there is possibly and even bigger blind spot that happens at the level of constructing semantic networks, which is the application of stop-word lists. Thus, we introduce a list of nodes to be removed from the network, unlike social networks where there is only one acceptable way of removing node, which is the ego when we deal with ego networks. Although there are justifications for such practices, by applying stop-word lists—including some of the most connective and/or central words—we are substantially changing the structure of our networks which will in turn have major implications for most of the metrics applied to such networks. Thus, it is difficult to apply the same criteria used in social networks even to the construction of the networks themselves.

Here, once again, the discussion moved towards the inclusion of context and the abstraction of analysis units to a higher level than words, for example key words.  

As all our other discussion panels, time proved to be insufficient to cover all the discussion-worthy issues in socio-semantic networks and we hope to keep the conversation going and to continue these fruitful discussions in future years at Sunbelt.