Framing in Communication: From Theories to Computation

Framing is a dynamic process that allows for the use of multiple frames within a single discourse unit. This presupposes that frames must undergo certain changes and/or interact with each other. The goal of our working group was to identify the types of interactions that happen between and within frames, so-called ‘framing moves’, and to examine the way they occur, i.e., patterns of framing dynamics.

While our working assumption is that similar framing dynamics can be found in both dialogues and monologues, we focused on the dialogical structures to narrow the scope of the research. We embed patterns of the framing dynamics in the dialogue protocols offered by [2] in order to put some constraints on them. Our account of framing dynamics is constrained by dialogue type, i.e., within an instance of a given type (say, a negotiation dialogue) each of the framing moves is available according to protocol and constraints that are proper to the dialogue type.

We applied the concepts of framing moves to a real-life example dialogue. We chose an excerpt previously analyzed in terms of blends (defined as integrated mental spaces recruiting conceptual input from different input spaces) by [1] to test the applicability of the framing moves. The dialogue is taken from Loveline, a call-in radio show in North America that gives listeners medical and relationship advice⁶⁶6http://kroq.radio.com/shows/ (offline). In the following dialogue, the hosts are reacting to the caller’s concern about getting two orgasms in a row:

The example illustrates almost all of the identified framing moves. Adam begins the dialogue with the frame of a miracle which is then reframed by Dr. Drew who brings up the idea of a message. After the two humorous comments, Adam in a more serious manner starts a new frame of the health issue which is then taken on by Dr. Drew when he responds to the question. Adam goes back to the humorous approach with a reframe and elaboration when he talks about the experience of the caller as a Holy Grail after which Dr. Drew suggests an alternative view on the experience by calling it a purgatory thus reframing again. Dr. Drew switches the frame completely by comparing the caller, who will be trying to achieve something impossible, to Sisyphus. The dialogue excerpt is concluded with Adam merging the previously mentioned frames of the Holy Grail by referring to the caller’s experience as something ‘incredible’ with the frame of Sisyphus (‘you end up chasing’) into a new frame of Moby Dick.

Understanding the framing dynamics in dialogues can help readers reflect on the (non) intentional use of framing at a structural level beyond the content itself and can help us understand the effect of the interplay between multiple frames in a text. Our preliminary work provides a starting point for a comprehensive study of the dynamics of framing. We presented seven crucial components to assess the underlying dynamics. These framing moves were extracted in an inductive manner using two prototypical examples. We plan to extend this process by conducting an annotation study on different datasets (e.g., politics, humour, etc.) with multiple annotators to create a robust scheme in order to test our presented framework on a large scale. Given the close connections to related subfields such as dialogue protocols, future work will further strengthen our proposed framework’s unique position and novelty within the research landscape.

Constructing high-quality datasets for frame identification is an essential step in developing and evaluating corresponding computational models. To build a framing dataset, different questions can be asked such as: What are the possible typologies of frames? How to outline clear and practical annotation guidelines? And how to ensure the feasibility of the annotation given the complex nature of the task? Here, we discuss a pilot annotation study that seeks to deliver some preliminary answers to the questions above. We strive to propose frame typology in a relevant and appealing manner to the Natural Language Processing (NLP) community and back this up with real-world examples and a small dataset.

Generally speaking, developing an appropriate typology of frames requires a thorough understanding of framing and a solid theoretical ground for modelling it. Though this was the primary goal of one of the working groups in the workshop, we had limited time for collaborating and sharing the needed knowledge. Nevertheless, our pilot study illustrates several observations:

• $\mathrm{■}$

  Fame annotation is, in most cases, quite difficult and time-consuming. We have noticed that expert annotators are necessary for frame annotation, at least in the earlier stages of annotation.

• $\mathrm{■}$

  Some articles are substantially easier for identifying their frames. We assume that making the frames explicit and easy to grasp vs. making them implicit and subtle may be a strategy of the authors to convey the main message (e.g., persuade the audience).

• $\mathrm{■}$

  The frame can be represented in various forms; for instance, by a simple key phrase (e.g., fairness and equality) or by a more complex discourse relation (e.g., the contrast relation between fairness and security).

• $\mathrm{■}$

  Some articles attempt to address different frames and focus on one for them, while some completely ignore all but one of the frames.

• $\mathrm{■}$

  Frame identification, similar to many tasks in NLP but to a greater extent, is subject to the author’s intent and readers’ interpretation.

• $\mathrm{■}$

  Frames can be established in diverse elements of the articles, including their headlines, images with their captions, and lead paragraphs.

An axiom declares a salient property of a real-world phenomenon. A set of axioms, or axiomatic systems, inductively defines the phenomenon if it entails all basic properties of the phenomenon that are not implied by others known about it. Such a set of axioms is irreducible. It opens the door to theoretical analysis of the phenomenon, i.e., the derivation of theorems that govern it. If derived theorems can be verified by observation in the real world, this raises trust in the validity of the set of axioms. A set of axioms only serves as a theoretical model of a phenomenon, if its derivable theorems predict real-world observations with sufficient accuracy.

Axiomatization relieves us from having to define frames directly, as definitions of seemingly elusive concepts, and those of frames in particular, are notoriously subject to fierce debate. In contrast, basic observations of and about frames are much less subject to debate.

As a first step towards an axiomatization of frames, we state the following three axioms:

• $\mathrm{■}$

  Axiom 1. Frames exist.

• $\mathrm{■}$

  Axiom 2. Exposure to frames has measurable effects.

• $\mathrm{■}$

  Axiom 3. A frame can be defined by what belongs to the frame, or by what does not belong to the frame. We call the latter a co-frame (“frame dualism”).

These axioms capture fundamental prerequisites for an operationalization of frames.

Axiom 1 postulates the existence of frames in communication. Frames have a representation both in people’s minds as well as in communication media, language in particular. When a frame is adopted by people, they can do this more or less reflexively (i.e., knowing that they adopt a particular frame rather than another frame). Axiom 2 postulates that frames can have measurable effects in the real world. For instance, people may change their behavior as a result of adopting a frame. Or the presence of a frame in a given piece of writing may be noticed by them. Axiom 3 postulates that the definition of a given frame can be discerned by investigating its “boundary”, i.e., a frame can be discriminated from its surroundings. For example, words can be identified that have a clear connotation with the frame, or actions of people, or depictions of situations, etc. This renders frames also distinguishable from other frames, albeit interrelations between frames are not excluded. Everything that does not belong to a given frame is collectively referred to as its co-frame, thus inducing a kind of frame dualism. When dealing with an inventory of frames (i.e., string names of frames), given one frame $X$, all other frames from the inventory combined are its co-frame. A co-frame can also be described by a ranking of frames according to saliences of the frame in question. The most salient frame $Y$ of the co-frame $\overline{X}$ of frame $X$ is its main co-frame. We do not claim this set of axioms to be complete, i.e., there may be more axioms required to derive all properties of frames that have been previously observed.

From Axioms 1 and 2, it follows that the knowing or unknowing adoption of frames, and their possible measurable effects on people’s behavior induces what we call camps: We define a camp by those people who share a common understanding of / recognition of / reaction to exposure to frame $X$.

As a further consequence of Axiom 2, the measurability of a frame also opens the door to its quantification. In this regard, we hypothesized that the presence of a frame may vary in terms of how well it can be recognized, called “strength” in our discussions at the time. In hindsight, a better choice of terms would have been “perceptibility”, since this term comes with less ambiguous connotations.

We conducted a brief framing perceptibility user study to support Axiom 2. We extracted six tweets about the ongoing Russian invasion of Ukraine that are said to evoke the genocide frame. Given a pair of tweets, the 21 participants of the seminar present at the time were then asked to indicate ad hoc and independent of each other in which of the two the frame of genocide is more perceptible. In our words then, which tweet contains a stronger genocide frame. We asked for participants’ opinions about three pairs:

• $\mathrm{■}$

  Tweet A1 (weaker). Ex. 2:

  A genocide that didn’t happen; nuclear ambitions that Ukraine doesn’t have and a threat to Russia that does not exist. I have yet hear a single justification for the murderous invasion of this country that even remotely bears scrutiny. #Ukraine

• $\mathrm{■}$

  Tweet B1 (stronger). Ex. 8:

  How are Western leaders sleeping during this genocide? It was posted by a woman who recorded herself right after the attack #Ukraine

• $\mathrm{■}$

  Tweet A2 (weaker). Ex. Y:

  Putin claims he is attacking to eliminate #Ukraine’s ‘‘Nazi’’ government… headed by a Jewish president! Screw sanctions, Putin only cares about the price of oil. If Biden would end the insane embargo of Venezuela, oil prices would collapse and so would Putin’s killing spree.

• $\mathrm{■}$

  Tweet B2 (stronger). Ex. X:

  #Putin is committing mass murder in #Ukraine. Why are we not doing everything in our power to stop him?

• $\mathrm{■}$

  Tweet A3 (stronger). Ex. Z:

  US taxpayer $$$ will fund mass murder and ethnic cleansing in my country, Ukraine.

• $\mathrm{■}$

  Tweet B3 (weaker). Ex. 3:

  Outrageous hypocrisy! This is the military who have committed human rights atrocities & genocide for decades in the name of ‘‘Burma’s sovereignty’’. Both regimes must be held accountable for all serious human rights violations. #Ukraine

Our basic operationalization of frame perceptibility (strength) to arrive at ground truth labels was this: We define the strength of a frame as the number of references to the frame. Applied to the tweets, this meant we counted the number of term occurrences which either directly refer to genocide, or have a connotation with the frame (highlighted bold), in context of what the tweet was intending to say. Tweet B3 referred to another genocide, not the one in Ukraine. The voting was as follows:

A1:  6 vs. B1: 15
A2:  6 vs. B2: 14 vs. Tie: 1
A3: 14 vs. B3:  6 vs. Tie: 1

This distribution of votes results in a Krippendorff’s Alpha of -0.0349, which indicates random inter-annotator agreement and negative results for our ad hoc experiment, despite the seeming tendency of the group towards the true answer. So, while the majority decision would have been correct in all three cases, no individual annotator performed consistently well. Two comments were given by annotators: “Framing does not have ‹strength›” and “‹Stronger›is the wrong conceptualization”, prompting a discussion and our change of terminology suggested above.

Axioms 2 and 3 imply that frames are discernible entities, an important prerequisite for any kind of operationalization of frames or framing. A frame provides a structure for perceiving and interpreting phenomena in a particular way. Elements of such structures can include scenes – which are themselves structures containing actors and things, and relations between actors/things – and answers to questions such as:

• $\mathrm{■}$

  What is going on?

• $\mathrm{■}$

  What is at stake?

• $\mathrm{■}$

  Who are the important actors? What are their roles? How do they relate to each other?

• $\mathrm{■}$

  What is the problem?

• $\mathrm{■}$

  What are possible solutions? What are criteria for ranking solutions?

• $\mathrm{■}$

  What can be expected to happen next?

In this regard, another more intricate operationalization of frame “perceptibility” (formerly “strength”) that we conceived of was that the perceptibility of a frame is reciprocal to the number of answers it gives to the aforementioned questions.

Two frames can be compared not only in relation to the number of answers they provide, but more generally in relation to all aspects of their structure. Given frames f1 and f2, we can ask:

• $\mathrm{■}$

  Does f1 answer (some of) the same questions as f2?

• $\mathrm{■}$

  Does f1 mention (some of) the same actors? If yes, are these actors given the same or different (complementary or opposite) roles?

• $\mathrm{■}$

  Does f1 posit (some of) the same problems as f2? If yes, does f1 posit the same or different solutions to these problems? If not, are the problems posed by f1 and f2 complementary or mutually exclusive or in opposition to each other, etc.?

Comparisons of this kind can be used to provide assessments of degrees of overlap, complementarity, or mutual exclusivity of two frames.

The suggested three axioms above can serve as base for a number of tasks, ranging from basic tasks where two tasks are compared to more complex tasks and applications. Computational systems performing those tasks could assist stakeholders from various fields (e.g., journalists, politicians, speech writers, marketeers, educators).

Task 1 is to measure the effect of framing with all other variables fixed, i.e. we assume two texts with the same frame. The goal is to judge, first manually, then by computer models, which text is stronger. We leave the exact definition of “stronger” open for now.

In Task 2, the goal is to identify whether two texts have the same frame. Our perspective differs from the existing approaches to frame identification such that we do not rely on an existing set of predefined frame types, rather the task is simplified to a binary decision comparing two instances.

The next task, however, makes further assumptions and thus does rely on the availability of frame types. Task 3 is therefore text labeling, i.e. given a single text, does it have frame $X$? For instance, does this text describe a war as a genocide?

While the previous tasks were inherently classification-oriented, Task 4 is a text generation one. We formulate it as a text rephrasing task. Namely, given a text, we want to reframe it using a frame $X$. The open question here remains, as in the majority of text generation tasks, how to objectively evaluate the functionality.

Our list of potential tasks also includes an analysis of co-frames, where we allow for multi-label framing of texts, or an ambitious task of identifying whether a text is intentionally framed.

Given the operationalization of framing through a clear task definition, we envision the following applications. First, paraphrasing and reframing can be tailored to specific needs with respect to the audience. As with any other text generation task, a potential dual-use must be taken into account. Second, a writing assistant actively supporting framing or helping reframing a message can be beneficial in the educational context. Finally, an automated tool that highlights frames in a piece of text can help the reader to reflect on the effects of framing.

Here we outline potential strategies for compiling a corpus of annotated data. Our main presupposition is that we do not define what frames are. We propose selecting reporting on events, as framing influences how we perceive them, such as natural disasters.

Let’s exemplify with a set of texts about a volcano eruption. We might observe that some of these events are labeled with extremely opposite frames, such as “disaster” and “tourist attraction”. Therefore, starting with a structured database of such events uniquely identified by location and date, we might be able to sample relevant texts from social media or news. This collection would allow us to bootstrap tasks one and two, and also come up with a set of disaster-specific frame types for task three. The reframing task might be constrained in such a way that we would allow annotators only minimal lexical changes that would result in a different frame.