#### Maria Gabriella Grassiaa , Marina Marinoa , Rocco Mazzab , Michelangelo Misuracac , Agostino Stavoloa <sup>a</sup> Department of Social Sciences, University of Naples "Federico II", Naples, Italy. **Topic modeling for analysing the Russian propaganda in the conflict with Ukraine**

**Topic modeling for analysing the Russian propaganda in the conflict with Ukraine**

<sup>b</sup> Department of Engineering, University of Campania "Luigi Vanvitelli", Caserta, Italy. <sup>c</sup> Department of Business, Administration and Law, University of Calabria, Cosenza, Italy. Maria Gabriella Grassia, Marina Marino, Rocco Mazza, Michelangelo Misuraca, Agostino Stavolo

## **1. Introduction**

The conflict between Ukraine and Russia is changing Europe, which is facing a crisis destined to reshape the internal and external relations of the continent, shifting international balances. As the war in Ukraine continues, Russian propaganda about the conflict evolves. Modern propaganda can be seen as an attempt to influence opinion through the communication of ideas and values of a specific persuasive purpose (Abd Kadir et al. 2014).

A political organisation would want to convince people to concur with the message presented and accept it as their own beliefs, rejecting other point of view. It has been argued that practically all governments use forms of propaganda to bolster their support from other nations and citizenry (Pratkanis and Aronson, 1991).

For this reason, we present the preliminary results of an analysis conducted on the content of online newspapers used as propaganda tools by the Russian government. The selected newspapers create and amplify the narrative of the conflict, conveying information filtered by the Kremlin to advance Putin's campaign on the war. The goal of the work, therefore, is to understand the communication strategies that the Russian press used to motivate and justify the conflict in Ukraine and what types of information are disseminated by the selected newspapers. In this regard, through a Symmetric Non-Negative Matrix reduction factorization technique (symNMF), we extracted the main themes found in Russian newspaper articles to identify the topics used for propaganda.

## **2. Non-negative matrix factorization**

Non-negative matrix factorization (NMF) is a dimension reduction method to uncover latent low-dimensional structures in high-dimensional data (Kim and Park, 2008). NMF is an unsupervised approach in that the low-rank factor matrices are constrained to have only nonnegative elements (Kuang et al. 2015). So, the basis vectors of the matrix are represented as a linear combination of vectors with positive coefficients.

Nonnegativity improves interpretations of the information extracted from a given data matrix, allowing a better understanding of the results obtained from the analysis process. This is in contrast to dimensionality reduction techniques that rely on the singular value decomposition (SVD) method, such as principal component analysis (PCA). One of the major problems with PCA is that the basis vectors have positive and negative components, and the data are represented as a linear combination of these vectors with positive and negative coefficients (Pauca et al. 2004). This is because the principal components are orthogonal, implying the presence of some negative values. Factors obtained from the NMF, on the other hand, are positive vectors and better approximate the data, but are not necessarily orthogonal (Casalino et al. 2016).

Given a **X** matrix of size *m × n*, the decomposition of **X** into a matrix **W** of size *m × k* (called the base matrix) and a matrix **H** of size *k × n* (called the encoding matrix), such that their product approximates the matrix **X**:

$$\mathbf{X} \approx \mathbf{W} \mathbf{H} \tag{1.1}$$

Maria Gabriella Grassia, University of Naples Federico II, Italy, mariagabriella.grassia@unina.it, 0000-0002-7128-7323 Marina Marino, University of Naples Federico II, Italy, marina.marino@unina.it, 0000-0002-0742-5912 Rocco Mazza, University of Naples Federico II, Italy, rocco.mazza@unina.it, 0000-0002-4901-5225 Michelangelo Misuraca, University of Calabria, Italy, michelangelo.misuraca@unical.it, 0000-0002-8794-966X Agostino Stavolo, University of Naples Federico II, Italy, agostino.stavolo@unina.it, 0000-0001-5890-2195

Referee List (DOI 10.36253/fup\_referee\_list)

FUP Best Practice in Scholarly Publishing (DOI 10.36253/fup\_best\_practice)

Maria Gabriella Grassia, Marina Marino, Rocco Mazza, Michelangelo Misuraca, Agostino Stavolo, *Topic modeling for analysing the Russian propaganda in the conflict with Ukraine*, © Author(s), CC BY 4.0, DOI 10.36253/979-12-215-0106-3.43, in Enrico di Bella, Luigi Fabbris, Corrado Lagazio (edited by), *ASA 2022 Data-Driven Decision Making. Book of short papers*, pp. 245-250, 2023, published by Firenze University Press and Genova University Press, ISBN 979-12-215-0106-3, DOI 10.36253/979-12-215-0106-3

where **W** e **H** are both non-negative matrices. The product **WH** is an approximate factorization of rank at most *k*. Generally, the *k* rank of the two matrices **W** and **H** is assumed to satisfy that *k* ≪ *min {m,n}* (Gaujoux and Seoighe, 2010). The value of the parameter *k* identifies the numbers of factors to be used to explain data (Casalino et al. 2016). Matrix multiplication can be implemented as computing the column vectors of **X** as linear combinations of the column vectors in **W** using coefficients given by columns of **H**. Each column of **X** can be computed as follows:

$$\mathbf{x}\_{i} = \mathbf{W} \mathbf{h}\_{i} \tag{1.2}$$

where xi is the column vector of the product matrix **X** and hi is the vector of the matrix **H.** 

Suppose we have *n* data points represented as the columns in *X = [X1, …, Xn],* and try to group them into k clusters. When **W** and **H** are subject to nonnegativity, it is possible to interpret the dimension reduction in (1.2) as clustering results: the columns of the first factor **W** provide the basis of latent k-dimensional space, and the columns of the second factor provide **H** the representation of *x1,…, x<sup>n</sup>* in the latent space. So, the cluster assignment for each data point is made by choosing the largest item in the corresponding column of **H** (Kuang et al. 2015).

The matrices **W** and **H** are found by solving an optimization problem defined with the Frobenius norm (a distance measure between two given matrices), the Kullback-Leibler (KL) divergence (a distance measure between two probability distributions), or other divergences.

The usual approach to NMF is to approximate **X** by calculating **W** and **H** to minimize the Frobenius norm of the **X** - **WH** difference, such that (Pauca et al. 2004):

$$\sum\_{i=1}^{n} \sum\_{j=1}^{m} (X\_{ij} - (\mathsf{W}H)\_{ij})^2 = ||X - \mathsf{W}H||\_F^2$$

The formulation in (1.3) has been applied to many clustering tasks in which the *n* data points are available in **X** and are used as an input. The relationship between the data points is represented as a graph, where each node corresponds to a data point and a similarity matrix An×n contains the similarity values between each pair of nodes (Moutier et al. 2021). The NMF is not a general clustering method that performs well in every circumstance, where the limitation can be attributed to its assumption on the cluster structure (Kuang et al. 2015). As we know, the goal is to approximate the original data matrix using a linear combination of basis vectors. When the underlying *k* clusters have nonlinear structure, NMF cannot find any *k* basis vectors that represent the clusters respectively.

So, it is used the SymNMF, the symmetric variant of the NMF, that handles symmetric matrices **A** as input. This method is based on a similarity measure between data points and factorizes a symmetric matrix containing pairwise similarity values into the product of a nonnegative matrix and its transpose (Jia et al. 2021). The factorization of **A** will generate a cluster assignment matrix that is nonnegative and captures the cluster structure inherent in the graph representation. Given an *n × n* symmetric matrix **A** and a reduced rank *k*, SNMF seeks to find the best factorization so that:

$$\mathbf{A} \approx \mathbf{H} \mathbf{H}^{\mathrm{T}} \tag{1.4}$$

where **H** can be viewed as the cluster indicator and **HT** the transpose matrix.

Compared with NMF, SymNMF concerns only the factorized similarity matrix **A** and doesn't consider whether the structure of the data is linear or non-linear. It can be regarded as a graph clustering method, and it is more effective for nonlinearly separable data than NMF (Kuang et al.2015). It has demonstrated to be a powerful method for data clustering (Jia et al. 2021), for learning topics in text mining (Yan et al.2013).

Also, SymNMF is related to spectral clustering, SC, and both share the same loss function only with different constraints (Ng et al. 2001) and it can directly generate the clustering indicator without post-processing, while SC needs extra post-processing, like K-means, to finalize clustering.

# **3. Methodology**

The proposed work shows preliminary results. Specifically, the analysis is carried out from March 2021, when the Russian military moved weapons and equipment into Crimea, to the end of March 2022, the day of the first negotiations in Istanbul. The selection of the newspapers is based on past study: the report "Pillars of Russia's Disinformation and Propaganda Ecosystem" produced by U.S. Department of State. According to the report, the journals cover various geographies, and they have their own target audiences. These newspapers are influenced by the Russian government and institutions, thus highlighting a Kremlin-driven information and regime interpretations given to the facts of war. The papers chosen are as follows:


We extracted 3,396 newspaper articles, and two of them were withdrawn because they were not written in English; so, we had 3,394 articles. As we know, textual data are unstructured, so it's necessary to perform some phases of pre-processing for having structured data. There are different steps:


The pre-processing phase returned a database composed by 40.360 tokens, 5010 types and 3394 documents. In this way, the term-document matrix indicates the number of occurrences of each term in the document. The dimension of term-document matrix is 5010x3394.

## **4. Preliminary results**

In the final stage of the pre-treatment process we applied the documents and words matrix vector space model. Each word is considered a vector where each element *ai* represents the weight of that element within the individual document. In NMF, the term-document matrix is too sparse to estimate reliable arguments, so more stable and dense data are used.

According to Yan et al. (2013), for reducing the sparsity of term-document matrix, we created a co-occurrence matrix **W** composed of the vectors **wi**., whose elements *ai* represent the number of times each word pair *<***wi***,***wj***>* co-occurs within the same document. For each pair of vectors, we calculated the cosine measure, a similarity index that measures the similarity between two vectors of an inner product space. In this way we created the similarity matrix **S**. On this matrix was applied the SymNMF for identifying the main topics. According to this, we found five topics:


**Tab.1** – Topics extracted from SymNMF

Tab. 1 shows the terms associated with the topics extracted from the SymNMF. It is possible to define the topics that the Russian media used as motivations for the Ukrainian conflict.

The first topic excerpt refers to the threat posed by Ukraine. These newspapers present the conflict as a potential problem for relations with Europe and the West. There is an increasing urgency to seek common ground with the nation, declaring that it is responsible for all the casualties that are occurring. According to the Russian government, Ukraine is exaggerating the issue by not thinking of the collective good and making decisions that only sour the relationship with the Kremlin.

In "topic 2" reference is made to the war dimension. The terms given allow identification of the main arguments the Russians used to justify the invasion. The war is presented as a "humanitarian operation" that Putin undertook to liberate Kiev and Ukraine. Russian propaganda aims to present the Ukrainian people, not as victims but as perpetrators of their crimes and murders: the term "mercenary" refers to a narrative that Ukrainian soldiers murdered Danish mercenaries. This serves to dispel the idea of Ukrainians as a subjugated people. All words referring to the dimension of war emphasize the belligerent spirit of the population, which wants to get rid of the Russian "nearby" enemy even with the use of atomic bombs and munitions received by the U.S.

Related to this, there is "Topic 3" that highlights diplomatic-international relations. Especially, the topic appears to be declined in a general way but such that the dimension described can be interpreted. According to Russian media, President Putin has repeatedly proposed talks and negotiations and set out Russia's conditions, the first is that no NATO base be installed in Ukraine. The topic of international agreements turns out to be central as on the one hand the media talk about the Russian government's willingness to mediate with Ukraine, and on the other hand, they emphasize how the same nation wants to join NATO and improve the relationship with President Biden.

"Topic 4" identifies the economic motivations of the invasion. There is a common view that the Ukrainian government initiated the conflict to increase its economic and geopolitical power to expand into neighbouring countries and counter the Russian nation. In particular, the newspapers report a series of events related to the Ukrainian economy: the emigration of citizens to Poland to improve wage conditions, the loss of public money, and the resulting debts. The latter aspect is central as it predicts Ukrainian dependence on America. By this, it is narrated that European powers are also obliged to give money to Ukraine to pay for previous situations: for example, the Germans pay a debt to Ukraine for the occupation of Crimea.

The last topic refers to a philosophical-identarian dimension, in which Western and liberal values are criticized and exaggerated. The thoughts of numerous philosophers who present liberalism in a negative light are reported, stating that it "should be opposed to fascism" but it imposes itself on Western civilizations, taking on the same characteristics. For this reason, the U.S. is presented as a nation that does not want to assert other world powers by imposing its own economic and social vision. In this regard, the Russian media propose a vision of its nation as one that engages in the pluralism of ideas and goes to represent an alternative of freedom to the Western world. Numerous articles are referencing how the U.S. wants to change tradition and classical roots (e.g., Dante's works have been called politically incorrect and have undergone liberal cleansing) by going on to criticize the philosophers and thinkers of the time. On the other hand, Russia is presented as the guardian of "true and authentic" European values and not of "globalization" and "liberalism".

## **5. Conclusions**

As discussed earlier, the work is in the preliminary stage and the aim is to identify the themes that the Russian government used to narrate the conflict. For future developments, we are expanding the list of Russian information sources in order to conduct a more comprehensive analysis.

Russian Federation invests its propaganda channels and its intelligence services to conduct activities to support their information system, and it leverages outlets on news sites or research institutions to spread these narratives.

So, the Kremlin use these tactics as part of its approach to using information as a weapon. In this regard, the Russian government has issued a series of measures, ordering all media outlets to report on the invasion of Ukraine only through official state sources, blocking numerous sites for spreading unfounded news and threats of high treason. Russia's willingness to employ this approach provides it with three advantages. First, it allows for the introduction of numerous variations of the narratives, to fine tune their information narratives to suit different target. Second, it provides plausible deniability for Kremlin officials when they peddle different information, allowing them to deflect criticism while still introducing damaging information. Lastly, it creates a media multiplier effect that boost their reach and resonance.

## **References**

