A text data mining approach to the study of emotions triggered by new advertising formats during the COVID-19 pandemic

Companies invest large amounts of resources in advertising with the primary goal of making their products known and thus increasing their economic performance in terms of revenue. In a broader sense, advertising also helps build brand image and awareness. Advertising awareness can be described by whether consumers remember a brand’s advertising (Vaughan et al. 2016). Brand awareness refers to whether consumers remember or can recognize a brand, or simply whether consumers know about a brand (Keller 2013). It is also related to the identification with a particular company and affects the impact of advertising on consumers by influencing their reactions and perceptions (Bhattacharya and Sen 2003). The effects on the recipient of the advertisement include both affective and cognitive reactions. The first concerns preference for a particular product due to sentimental attachment, while the cognitive response concerns perception and memorization of the advertising content associated with a brand.

Cognitive response can be used to measure the effectiveness of an advertising campaign, as well as its ability to impress, be understood, and be remembered by its audience. Awareness measurements are often used in market research as a measure of brand performance and marketing effectiveness (Romaniuk et al. 2004). The most commonly used indicators related to brand and advertising awareness are: (i) recognition, i.e. the ability of consumers to identify a particular brand by its attributes over another; (ii) recall, i.e. how memorable an advertisement is to the audience; (iii) awareness, i.e. the percentage of target customers or accounts that show awareness (either with or without help). Advertising campaigns also aim to maximize “brand lift" increasing the perception and awareness of a product or brand over time in terms of recall and familiarity. The impact of an advertising campaign in terms of these attributes is typically measured directly, through consumer interviews and surveys, to understand and monitor the public’s awareness of a brand or product and to assess its commercial positioning and competitive strength in the relevant market.

In addition to television advertising, which is an important channel through which brands communicate with customers (Bruce et al. 2020), the media channel that has seen an advertising explosion in recent years is undoubtedly the Web (Fuxman et al. 2014). See, for example, Calder et al. (2009) for an examination of how consumer engagement on a website can increase the effectiveness of advertising. In recent years, there has also been an explosion of Web Analytics techniques aimed at gathering information valuable for understanding how the website works from the perspective of the user and the company (Xun 2015). All companies monitor their websites and the behaviour of users accessing the website based on a set of parameters automatically recorded by the provider or set by the company itself: the number of hits on each page of the website, the time spent on the website, the keywords entered through the search engines, the entry and exit pages. The same type of monitoring can be performed in real time on smartphones, tablets and social media platforms, opening up new interesting markets for advertising. The biggest change the Internet has brought, however, is represented by social media platforms, which have reshaped the way consumers create and distribute content through their ability to connect users with each other (Peng et al. 2018). Marketing communications through these platforms help companies reach new customers and increase demand for their products through user connections (Stephen and Toubia 2010; Schweidel and Moe 2014). On the Twitter platform, where the data for this analysis was collected, users retweet the content they read about an ad and build a bridge between the company that published the ad and new customers by sharing the ad with their followers (Gong et al. 2017).

These trends accelerated rapidly after the outbreak of the health emergency associated with the COVID-19 pandemic, which changed the everyday behaviour of citizens around the world and, consequently, the relationship between consumers and brands. On the one hand, consumers had to change their shopping habits, as they could no longer leave the house and shopped online whenever possible. On the other hand, brands had to rethink both their sales techniques and the way they advertised in order to appeal to audiences in difficult conditions. During the lockdown, messages of encouragement and solidarity were sent in all communication channels and even in advertising, as a kind of collective embrace that underlined the strength of the Italian people to defy adversity and emphasize those values that might have been taken for granted before, such as love, family, solidarity between people. Initially, there were only a few companies in Italy that designed advertising campaigns specifically for that time of crisis, and as the weeks went by, there were more and more. Some well-known companies in the field of telecommunications launched ad hoc commercials based on the message of shortening the distances to friends and relatives through video calls, phone calls, the Internet, the possibility of watching a movie or the favourite series TV, sharing emotions with family and friends. Advertisers directed messaging efforts toward risk communication strategies such as safe health and wellness practises (Deng et al. 2020).

The brands that have decided to change their approach were numerous and, to varying degrees, successful in creating empathy. For example, many car companies have moved in this direction: the FCA commercial titled Anthem to the Home or the Volkswagen brand, which used the hashtag #IoRestoInGarage (#Istayinthegarage). The Mazda video clarifies the shift in customer approach in the very first sentence: “This is not really a commercial, but a message of hope”. With a still image, a musical tapestry, poppy words and an absolutely minimalist style, the idea is to convey to customers that it is not about selling, but about being there, even if it still reminds them of the brand. Cosmetic companies have also released positive messages. A well-known makeup brand introduced the hashtags #torneràilsorriso (#smilewillcomeback) in its commercial with the song Ritornerai, showing a series of images of happy moments and memories of life before the emergency, such as visiting a city or going out with friends. It is therefore interesting to understand how consumers have responded to this change. In particular, we will study the patterns in Italian tweets in response to the aforementioned paradigm shift in advertising for some well-known brands in the food and consumer goods industries. This aspect makes the research design special and deviates from traditional approaches based on systematic reviews.

Using a corpus of tweets to analyze a particular research question requires that each tweet is represented in terms of a BOW model. In this simplified representation, both the grammar and the order of occurrence of words are irrelevant. It is only relevant whether a term occurs or not and how often it occurs in a text. Therefore, a number of preprocessing steps are required according to Natural Language Processing methods (NLP), detailed in the next section, to transform the raw text into a BOW representation (Manning and Schütze 1999; Manning et al. 2008). The final output is a vocabulary of terms V that forms the basis for further representations.

The simplest representation generates an indicator variable for each term in the vocabulary V, where 1 indicates the presence of the term in the tweet and 0 its absence (Manning et al. 2008; Wilbur and Kim 2009). This means that each document is represented as a binary vector \(e_{1:\vert V \vert } = (e_1, \ldots , e_{\vert V \vert })\) with dimensionality \(\vert V \vert\). The entire tweet corpus is therefore represented by a large sparse binary document term matrix (DTM) of dimension \(\vert {{{\mathcal {D}}}} \vert \times \vert V \vert\), where \({{\mathcal {D}}}\) denotes the set of retrieved tweets.

For a more structured representation, we can introduce the frequency of occurrence of terms in a document by using the vector space model (VS), where documents are represented as vectors in \({\mathbb {R}}^{\vert V \vert }\) (Salton et al. 1975). The DTM can be weighted by the term frequency \({\textsf {TF}}_{td}\) of the word \(t \in V\) in the tweet \(d \in {\mathcal {D}}\) and by the term frequency-inverse document frequency:which dampens the effect of local term-document counts. As usual, the inverse document count of the word t is defined as:where \({\textsf {DF}}_t\) is the number of documents in \({{{\mathcal {D}}}}\) that contain the term \(t \in \vert V \vert\) (document frequency). If the term t occurs in every document in the corpus, \({\textsf {IDF}}_t\) is 0. The fewer documents that contain the term t, the higher the value of \({\textsf {IDF}}_t\) (Nguyen 2014).

All data analysis in this Section was performed using R 4.1.2 (R Core Team 2021), along with a set of libraries specifically designed for NLP and text mining, most of which are part of the Data Science infrastructure tidyverse (Silge and Robinson 2016; Wickham et al. 2019). To better understand whether and how attitudes toward television advertising have actually changed, a preliminary descriptive analysis of the number of tweets posted and words used during the analysed period was conducted. This analysis is based on the rule that the higher the media exposure, the more a brand is appreciated, and that collective awareness, regardless of reputation, is in itself an indicator that can also become a decision factor for consumers.

It is worth noting that a study by Kantar, one of the world’s leading market research and marketing consulting firms, found that Internet browsing increased by 70% in the later stages of the crisis, viewing of TV increased by 63%, and social media use increased by 61% over normal use (Kantar 2022). Only a very small proportion of consumers expected advertising to stop. Kantar estimated that a six-month absence from TV would result in a decline in brand awareness of about 39%, further delaying recovery after the emergency. Consumers also expected that brands that continued to advertise would have a positive impact on society: a reassuring tone and messages that would provide encouragement to overcome the difficult situation. In addition, 75% of respondents believed that brands should not use the crisis to advertise their products. This shift in the characteristics of advertising at the time of the pandemic is also well documented in our data. Looking at the frequency distribution of the number of tweets posted in each of the three time periods considered, we find the following:These numbers are very significant from an awareness perspective: regardless of the content of the tweets and whether they expressed a positive or negative opinion, consumers tweeted 365% more during the lockdown than before the emergency.

This effect is also evident when we count the number of words tweeted daily. Figure 1 was created using whitespace tokenization of all tweets and counting the number of tweets used on each day. Finally, we plotted these data as boxplots by taking summary values for each of the three time periods considered. In the pre-emergency period, the average number of tweets per day in response to the brands’ commercials TV was 1140, while in the emergency period it increased to 4845. After the end of the most acute phase of the emergency, it decreased again, but without reaching the pre-crisis level (it averaged 2268 words).

In Fig. 1, the day 2020/05/07 was intentionally removed because it can be considered an outlier. The reason for this decision can be seen in the following graph (Fig. 2), which shows the historical series of the number of words tweeted daily. The outlier can be clearly seen frpm this figure. On May 7, the number of tweeted words exploded to an unprecedented level. To be exact, on May 7 we had exactly 32,496 words, of which 89.4% appeared in tweets that had the Lavazza brand as a reference tag (1060 tweets over 1199, with 29,046 words over 32,496). The hashtag #Lavazza reached the third position among the trends on Twitter in Italy that day and became a Trending Topic.

Lavazza’s commercial was released on May 7, 2020, under the title Good Morning Humanity. It was the starting point of a global communication campaign aimed at conveying a positive message through the words of Charlie Chaplin’s speech to humanity, presenting on the horizon a new world ready to welcome a rediscovered humanity. This aspect is well illustrated in the wordcloud in Fig. 3, which is based on the 30 most frequent words that appeared in tweets with Lavazza as the reference mark on May 7, 2020. As can be seen from the graph, the most frequently used tokens include positive adjectives such as beautiful, great or new, or words such as humanity and like and words with negative emotion polarity such as dictator and controversy. The most common word is charliechaplin, the absolute star of the commercial with his speech about humanity based on the famous movie The Great Dictator. The presence of terms with negative content is explained by the fact that this commercial has caused a wide debate among Twitter users. On the one hand, it was appreciated as a masterpiece that evokes positive feelings for humanity in crisis. On the other hand, it was strongly criticised because it was seen as a rhetorical display of good feelings and goodwill towards other people. However, it should be noted that Lavazza’s previous commercials were more humorous and focused on the goodness of the product, the opposite of this new commercial launched on May 7, 2020, which is based on the emotional message about the progress of humanity, revealing a very obvious paradigm shift in the structure of the message.

The wordcloud of the 30 most frequent terms over the entire period considered for the analysis is shown in Fig. 4. It is immediately apparent that the most frequently used words in the comments on the TV commercials were #Istayathome and #everythingwillbefine, but also emergency, which perfectly fits that moment, and are justified by the massive use of such hashtags to comment on or end any written conversation. In particular, #Istayathome was launched by the social account of the Ministry of Health, which called for a massive spread. The call was welcomed by many brands, which in turn influenced consumers.

The words Italy, thankyou and new also play an interesting role. The latter (new) seems to refer to the ‘new world’ described in the speech of the great dictator from the Lavazza advertisement mentioned above. Moreover, the new advertising approach of the different brands seems to have been perceived as such and to have evoked a sense of appreciation and gratitude (thankyou) and a sense of belonging (Italy) among consumers. The evolution of thankyou and Italy over time confirms their particular use in the context of this contingent situation, as they accounted for no more than 1% of the relative daily frequency of tweeted words before the lockdown. These dynamic effects can be well documented by looking at the time series of daily occurrence of a particular term. Figure 5 shows the graphs for the terms thankyou and Italy as examples. In particular, thankyou peaks at 40 occurrences per day in the middle of the lockdown period and then drops dramatically to its usual average level before the lockdown. The word Italy shows a very similar trend.

To further characterize the differences between the three selected subperiods, we created the TF-IDF weighted DTM, according to expression (1). The original matrix contained 39,610 terms with 204,894 non-zero entries and had an extreme sparsity of nearly 100%. To reduce sparsity, we removed all terms that occurred less frequently in the tweets, i.e. all terms \(t \in V\) for which \(\textsf {DF}_t < \vert {\mathcal {D}} \vert \times (1 - 0.99)\). The resulting matrix contained only 58 terms with 25,688 non-zero entries and an overall sparsity of 98%. The large reduction in the vocabulary of terms is not important at this stage, as we are only interested in the words that are at the top of the list in terms of their frequency of occurrence adjusted for inverse document frequency.

The top 10 terms for each of the three time periods considered are shown in the bar graphs in Fig. 6. During the lockdown, as noted before, the most common words were hashtags of encouragement (#everythingwillbefine, #Istayathome, #wewillmakeit, #distantbutclose), which do not normally appear frequently in comments on television commercials. Not all terms that occurred most frequently during the lockdown were also frequently used after the lockdown. This initial rough dynamic analysis seems to indicate that the change in attitude toward advertising that occurred during the lockdown appears to be temporary. In fact, after the lockdown, we can observe a significant decrease in the frequency of terms that had characterized the lockdown period, and the reappearance of terms related to sports events (such as #naplesjuventus or #italiancupcocacola). It is very likely that in this phase commercials were again considered not as a moment of aggregation or expression of values of national solidarity, but as elements of disruption of television programs.

The model we used for the dynamic analysis is the following hierarchical Dirichlet-Multinomial multinomial mixture, introduced and studied in detail in the companion paper by Bilancia et al. (2022):where \(t = 1,2,\ldots , T\) indicates the time period. The p-dimensional vector, where \(p = \vert V \vert\):contains the counts of terms in V for the i-th tweet of the corpus used for the time period t. The hyperparameters \(\alpha ^{(t)},\theta ^{(t)} > 0\) are strictly positive real numbers and \({\mathbbm {1}}\) denotes a vector of all ones.

Each \(\beta _j^{(t)} \in {\mathbb {R}}^p\) is a vector of Multinomial parameters representing a topic, i.e., a probability distribution of the vocabulary of terms V. Different tweets may have different thematic content, i.e. terms or hashtags that appear in one tweet may be of little importance or not appear at all in another tweet. Since we do not know to which class each document belongs, the index of the row vector \(\beta _j^{(t)}\) corresponds to the index of the single element in the latent indicator vector \(z_i^{(t)} \in {\mathbb {R}}^k\) such that \(z_{ij}^{(t)}=1\) (all other elements are zero), while \(\lambda ^{(t)} \in {\mathbb {R}}^k\) denotes the mixing weights (Robert 2007). Since only one element of the vector \(z_i^{(t)}\) is non-zero, the Multinouilli distribution is a Multinomial distribution for a single trial.

The hyperparameters \(\theta ^{(t)}\) and \(\alpha ^{(t)}\) are assumed to be fixed and empirically determined. The exchangeable Dirichlet prior over the Multinomial parameters implicitly gives the data more weight in updating the posterior distribution of each \(\beta _j^{(t)}\) with an appropriate choice of the concentration parameter \(\theta ^{(t)}\). Similar considerations apply to the mixture weights \(\lambda ^{(t)}\). While Bayesian posterior inference is the natural framework for parameter estimation of the model proposed above, its applicability is complicated by the fact that the marginal distribution of the data cannot be written in closed form due to the high-dimensional integrations that enter the calculations. Therefore, we must resort to appropriate numerical methods for Bayesian estimation of the model parameters.

The approach we used for inference is that of variational inference (Blei et al. 2017), in which the joint posterior distribution is approximated by a probability distribution where the model parameters are assumed to be independent a posteriori. For simplicity, we suppress the explicit dependence on the time index t in this section. Moreover, the Multinomial parameters \(\beta _j\) are stacked by rows in the matrix \(\beta \in {\mathbb {R}}^{k\times p}\), where \(\beta = \{ \beta _{j\ell } \}\) for \(j=1,2,\ldots ,k\) and \(\ell =1,2,\ldots ,p\). The vectors \(z,\lambda\) and y collect all the remaining parameters and the sample observations.

In the synthesis, the starting point is the approximation of the posterior distribution \(p(\beta ,z,\lambda \vert y, \theta ,\alpha )\) by a variational distribution \(q(\beta ,z,\lambda \vert \nu )\), which itself depends on the variational parameters \(\nu\). Our optimization problem is:where the variational objective is the reverse Kullback-Leibler (KL) divergence between the posterior distribution and the variational distribution (Murphy 2013). The solution of (7) provides the best possible approximation to the intractable posterior distribution. The search for the optimal approximating distribution can be greatly simplified if we define the Evidence Lower Bound (ELBO) as follows:because minimizing the KL divergence with respect to the variational parameters is equivalent to maximizing the ELBO (which is a sum of linear terms when calculated in explicit form) with respect to the variational parameters (Zhang et al. 2019; Tran et al. 2021).

We specify the variational distribution \(q(\beta ,z,\lambda \vert \nu )\) using a classical mean-field approximation with independent components (Blei et al. 2017):where \(\phi _j \in {\mathbb {R}}^p\), \(\gamma _i,\eta \in {\mathbb {R}}^k\), and:An explicit expression of the ELBO for this choice of the variational distribution can be found in Bilancia et al. (2022). In addition, an iterative coordinate ascent algorithm is described to find the parameter values that optimize the ELBO and correspond to the best possible approximation to the joint posterior distribution of the model parameters. Taking advantage of the fact that the variational distribution of each \(\beta _j\) is Dirichlet, approximate posterior estimates of the probability distributions over the vocabulary of terms V can be computed using the following expression:and can be compared over time, as shown in the next subsection.

The sequence of time-indexed models, which have the hierarchical structure presented in Sect. 5.1, represents a seemingly unrelated set of models, since no mechanism is provided for dynamically modeling latent parameters over time. This possibility would be essentially redundant for our purposes, since the goal is to compare the occurrence probabilities of terms in a vocabulary V common to all periods without unnecessarily overparameterizing the model.

For this purpose, we divided the observation period from 2020/01/01 to 2020/06/30 into \(T=12\) subperiods of about fifteen days, as shown in Table 1. The frequency distribution of the number of tweets belonging to each period clearly shows that social media usage increased exponentially during the most acute phases of the lockdown, from 297 (1.42%) tweets in period 1 to 3757 (17.91%) in period 6. This last period corresponds to the most acute phase of the lockdown and emergency, when the stay-at-home mandate was extended to the entire national territory and all non-essential productive activities were stopped. The common vocabulary V was constructed based on the global document term matrix for all 20,982 tweets, taking only \(t \in V_{\text {full}}\) terms for which \(\textsf {DF}_t < \vert {\mathcal {D}} \vert \times (1-0.997)\), for a total of \(p = 449\) terms, with an overall matrix sparsity of 99%. This choice was empirically calibrated and was based on a tradeoff between the computational complexity of model estimation and the informativeness of the term vocabulary V.

We applied our model with \(k = 2\) and used the default values \(\alpha ^{(t)} = 1\) and \(\theta ^{(t)} = 5 / k\) for all t. This minimal choice of the number of topics is the simplest possible, avoids introducing unidentified components into the model, and assumes the existence of a main topic versus a background topic that collects the occurrence of terms that do not capture the dominant semantics of the time period we are considering. In this way, we can also filter out noise without violating the assumption that a short text such as a tweet cannot be considered as a mixture of latent topics, since each tweet is hard-clustered to a single topic. The choice of hyperparameters is neutral with respect to \(\alpha ^{(t)}\) and uninformative with respect to \(\theta ^{(t)}\). In the absence of additional information that would justify a different choice, they appear reasonable, leaving most of the responsibility for updating the posteriori distribution to the data.

We applied the algorithm for a total of 100 runs, starting from randomly chosen initial values. In Fig. 7 we show a subset of the trajectories of the coordinate-ascent variational inference (CAVI) algorithm. It is obvious that each run reaches a different stationary point, since we have many different local maxima in the ELBO. Better local optima lead to a variational approximation closer to the exact posterior (Plummer et al. 2020). Table 1 also shows the estimates of the mixture weights: As can be seen, the main component is often interchanged between \(\lambda _1^\star\) and \(\lambda _2^\star\). This phenomenon is a manifestation of the well-known label switching problem (Diebolt and Robert 1994). Since the likelihood of the model is a sum of terms invariant to each of the k! possible permutations of the summands, if we assume an exchangeable prior over the parameters, this invariance is inherited from the posterior distribution, which has k! symmetric modal regions corresponding to all possible permutations of the parameter labels. However, variational inference, like any optimization algorithm, depends on the initial conditions and focuses on one of the possible k! modes of the posterior surface depending on these conditions. Of course, this is not a drawback, because we know that a single mode contains all the information for exploring latent groups and estimating parameters, and the order of labelling is not relevant (Blei et al. 2017).

After this necessary technical introduction, we can examine the results by comparing the estimates of the probability distributions \(\beta _j^{(t)}\). In Fig. 8, we compared period 2 (from 2020-01-16 to 2020-01-31) with period 6 (from 2020-03-16 to 2020-03-31). In period 2, we have a dominant component (the relative mixing weight is about 92%) in which the top terms are consistent with the fact that the first news of virus circulation and health emergency in China spread at this time (the first two confirmed cases of SARS-CoV-2 infection in our country involved two Chinese tourists and are dated January 31, 2020). The secondary component shows that the top-10 terms mainly concern sports events such as the soccer championship. In period 6, the mandate to stay at home obviously led to much greater use of social media and TV. The top terms in component 1 are clearly related to the emotional dimension of gratitude and the desire for the emergency to end, which peaks in these days and then begins a slow descent. Note, on the other hand, that in the background component the references to sports events have completely disappeared (in Italy the soccer championship was canceled on March 9, 2020). On the contrary, it is worth noting the appearance of terms such as #mediaset (one of the main commercial TV operators) and #jingle, proving that TV commercials played an important role in conveying the widespread emotions during this period of the pandemic. These data, matched with raw frequency analyses, confirm that the role of advertising on TV did indeed change during the lockdown and was perceived as a vehicle for solidarity and positive emotions to bridge the worst period of the pandemic in our country.

However, this phenomenon is only temporary. In Fig. 9 we compare period 8 (from 2020-04-16 to 2020-04-30) with period 12 (from 2020-06-16 to 2020-06-30). In period 8, the lockdown was coming to an end. Beginning on May 3, 2020, there was an initial gradual easing of the lockdown such as the reopening of public parks, the resumption of food deliveries, and the resumption of various manufacturing activities. Only movements motivated by proven necessities were allowed (work, emergencies, and meetings with family), subject to the prohibition of gatherings with a minimum interpersonal distance of at least 1 meter. Looking at the distribution of the top-10 terms in both the main and background components, terms related to emergency, feelings of gratitude, and national cohesion continue to dominate. However, the reference to the media becomes weaker, with the term #social appearing only in the background component. The situation becomes even more evident in period 12, considering that since June 11, 2020, containment measures have been further relaxed and gambling and betting offices, theaters and cinemas, cultural and social centers were reopened (and since June 3, mobility between regions was fully possible again). This situation is reflected in the top-10 terms: although direct references to social media or Facebok are present, the most important term refers to a sports event, while those related to the dimension of emotion and gratitude have largely disappeared.

To further confirm this analysis, we have presented in Fig. 10 the estimates of the top-10 terms of period 6 for each of the 12 periods considered (using the main component). These estimates have been smoothed with the loess smoother, which is why they may be just below zero at the extremes of the time curve. It is quite clear that both the hashtags associated with emergency response and those associated with feelings of gratitude and national cohesion peak during the two main periods of lockdown (periods 6 and 7, from mid-March to mid-April 2020) and then decline rapidly as the summer period begins and containment measures are relaxed (with the notable exception of Italy, whose probability of occurrence remains roughly stable throughout the period considered). Thus, although the link between some frequently occurring terms and the new dimension of advertising campaigns launched during the most acute phase of the emergency is undeniable, it is equally obvious that this effect is temporary and does not correspond to a permanent paradigm shift in advertising.

To provide additional evidence, we plotted a similar graph in Fig. 11, but we used the top-10 terms with the highest average probability of occurrence (calculated as the average of the estimated probabilities of occurrence in each period) for the entire period considered. Note that most of the terms are the same as in the previous figure, or the temporals trends are qualitatively similar, confirming that the peak occurrence of the top terms is a transient phenomenon characterized by a bell-shaped temporal curve (in most cases). A noteworthy aspect is that the top term in this case is new, which has an unusual temporal trend with two peaks in the first and last parts of the curve. In the first phase, the spike in occurrence can certainly be related to the fact that the virus was initially named as the ‘new 2019 coronavirus’ (2019-nCoV) and then renamed to its current name SARS-Cov-2 in February 2020. On the contrary, as we have already noted, the second spike could apparently be related to the new world described by the Great Dictator in the influential Lavazza advertisement, or even to the new phase of the pandemic beginning on May 3, 2020, at the end of the most severe phase of the emergency.