Large language models for sentiment analysis of newspaper articles during COVID-19: The Guardian (2024)

2.1 Recurrent Neural Networks

Unlike traditional feedforward neural networks, RNNs possess recurrent connections, enabling them to capture temporal dependencies within sequential data [50, 51]. RNN has significant advantages in handling sequential data and capturing temporal dependencies. The recurrent connections in RNN enable it to capture temporal dependencies within the data, thereby better understanding the contextual information of the sequences making them applicable to NLP [52]. Also, the recurrent structure enables the model to memorise information from previous time steps and propagate it into future steps. In terms of data input, RNN has a certain degree of flexibility and can handle input sequences of various lengths.

LSTM networks [27] provide an extension to simple RNNs for modelling long-term dependencies in sequence data [26], making them applicable to NLP problems. The networks address the issue of vanishing gradients [53] encountered by simple RNNs. LSTM networks require substantial training time and computational resources, given the complex structure. LSTMs demand significant computational power and memory, especially when training on large datasets [54]. Additionally, LSTMs are susceptible to overfitting, particularly in scenarios with relatively small data volumes, although this issue can be mitigated through gularization methods such as dropouts [55]. Despite these limitations, due to their ability to capture long-term dependencies and generalization performance, LSTMs remain an effective tool for NLP.

2.2 Transformer model

The Transformer model is an advanced RNN effective in processing sequential data, such as text or time series information [56].The Transformer model utilises an attention motivated by human behaviour and utilises the bidirectional LSTM network [57]. The attentional mechanism is crucial for understanding the key points in sentences or paragraphs and provides better contextual understanding. Training on large datasets also aids the model in understanding rare words or expressions [58]. Additionally, the Transformer model can be fine-tuned for specific tasks and the accuracy can be improved based on large-scale datasets for specific tasks [59].However, particularly large Transformer models require significant computational resources for training, including considerable time and memory resources, which makes the cost of maintaining these models quite high [60]. Furthermore, fine-tuning Transformer models on small datasets may lead to overfitting, where the model excessively learns specific patterns in the training data at the expense of its generalisation ability [61].

2.3 BERT and RoBERTa models

The BERT is a pre-trained LLM based on the Transformer model[30] and capable of understanding the bidirectional context of each word within the text. This is distinctly different from conventional unidirectional or fixed-direction language models.BERT operates in two main phases: pre-training and fine-tuning. During pre-training, it engages in two types of self-supervised learning: masked language modelling (MLM), which involves predicting tokens for randomly masked positions in the input, and next sentence prediction (NSP), which determines if two segments of text are sequentially related. For fine-tuning, which tailors the model for specific tasks, additional fully connected layers are appended above the final encoder output. The inputs can be prepared by tokenising text into subwords, then merging three types of embeddings - token, positional, and segment - to form a standardised vector.

The RoBERTa model [32] is a pre-trained LLM that enhances performance and robustness based on BERT. It adopts the ”robust optimization BERT method” in its design. In comparison to the original BERT model, RoBERTa shows several advantages and drawbacks.RoBERTa benefits from a larger training corpus, containing extensive text data from web pages, forums, books, and more, thereby enhancing the model’s performance.The model employs a stricter masking strategy (dynamic masking) by replacing all words in the input text with ”[MASK]”, facilitating more effective utilization of training data information.Furthermore, RoBERTa adopts a deeper network architecture, which contributes to further improving the model’s performance.Finally, it demonstrates superior performance across various natural language processing tasks, including the GLUE (General Language Understanding Evaluation) [62] and SuperGLUE [63] task sets.However, there are drawbacks to consider since RoBERTa requires a larger volume of text data and a deeper network structure in comparison to BERT, resulting in longer training times.The substantial model size of RoBERTa poses challenges in terms of resource requirements during deployment.

3.1 Data

3.2 N-gram analysis

N-gram is a commonly used for NLP that reviews contiguous N-elements (usually words or characters) in text data [68]. The N-gram model assumes that each element in the sequence depends only on its previous element, and utilises this local dependency to make predictions or analyze data. The advantages of the N-gram model include simplicity, computational efficiency, and its ability to capture local information [69]. However, the N-gram model also has some limitations, such as its inability to capture long-distance dependencies, the need for large amounts of training data, and challenges in handling rare combinations. The application of N-gram models in news media analysis is extensively documented, reflecting the diversity and practicality of this method in real-world research. Lyse and Andersen [70] used N-gram in news media analysis to rank word sequences from the Norwegian Newspaper Corpus to assess the propensity co-occurrence of binary and triplets. Furthermore, N-gram-based text analyses were used for selected Persian [71], German [72], and English newspapers [73]. We will use N-grams to provide analysis of the data in association with sentiment analysis by LLMs.

3.3 Fine-tuning BERT and RoBERTa models

Given the diverse range of topics and writing styles found in our news articles dataset, fine-tuning the BERT and RoBERTa models are crucial so that they can effectively provide sentiment analysis with the 10 sentiments available in the SenWave training data [64].However, fine-tuning BERT also encounters numerous challenges. First of all, the fine-tuning process requires setting up a suitable environment to enhance efficiency and effectiveness [74]. Furthermore, news articles often present lengthy content, which will surpass BERT’s token limit. Segmenting these articles and aggregating sentiment scores poses another challenge, necessitating careful attention to maintain context and coherence [75]. Moreover, the presence of sarcasm and nuanced sentiments in news articles can perplex the model, which will potentially hinder accurate interpretation. Therefore, handling specialised terminology, proper nouns, or novel words absent from BERT’s vocabulary is important for maintaining sentiment analysis accuracy [76]. Therefore, addressing these challenges needs a multifaceted approach involving meticulous data preprocessing, domain-specific fine-tuning, and appropriate parameter setting. These measures collectively ensure the accurate and reliable sentiment analysis of news articles using BERT.

3.4 Framework

Our framework for using deep learning for sentiment analysis of newspaper articles includes several consecutive steps, as shown in Figure 1. The major components include: 1. Guardian data selection; 2. data cleaning; 3. bigram and trigram visualisation; 4. SenWave dataset cleaning; 5. pre-trained model calling; 6. model training and testing; 7. document-level sentiment analysis; 8. multi-label sentiment classification; 9. visualisation and comparison.

Our framework uses two data sets, including the Guardian articles and the SenWave dataset. We use the SenWave dataset for refining the LLM models. We first obtain the Guardian dataset from Kaggle and filter it based on publication dates and section names. We extract articles tagged (sections) as Australian news, UK news, World news, Business, Politics and Opinion. Note that the Opinion section encompasses local (UK) and international contributors. We believe that these sections are more important in sentiment analysis during COVID-19 since we cover the country of source of the Guardian, the world, and Australia which is geographically far but politically aligned with Great Britain.

In data processing (Step 2), we remove the items such as hashtags, stop words and web hyperlinks from the Guardian data using NLTK (Natural Language Toolkit) [77] tools. In order to obtain more meaningful bigrams and trigrams, and subsequent sentiment analysis, we also remove stop words based on categories, such as general terms, as shown in Table 2.

In Step 3, we segment the Guardian data into words and character sequences. In each Guardian article, starting from 2018, we generate all possible bigrams and trigrams for each quarter (data spanning every three months). We then visualise the bigrams and trigrams and analyse word combinations and frequencies to gain insight into article characteristics across different periods (quarters) covering COVID-19.

In Step 4, we extend our data cleaning process to the SenWave dataset, which consists of tweet texts, unlike the formal media articles from The Guardian. Given the dynamic nature of social language, we must account for abbreviations, emoticons, and other informal elements. To accomplish this, we follow the methodology outlined by Chandra and Krishna [10] for processing the SenWave dataset, ensuring it is ready for subsequent model training and testing. It’s worth noting that within the SenWave dataset, the term ”official report”, while typically considered a topic, is classified as a sentiment. This distinction is noteworthy, especially considering that in our analysis of The Guardian articles, sections discussing ”official reports” comprise a significant portion of the dataset.

In Step 5, we leverage pre-trained BERT and RoBERTa models via Hugging Face ⁴⁴4Hugging Face: https://huggingface.co/models. The Hugging Face is a leading company in the field of artificial intelligence dedicated to providing open-source NLP tools [78]. One of their flagship products is the Transformer model library, which provides a variety of pre-trained deep learning models, including BERT and RoBERTa. Hence, we can access these state-of-the-art models and fine-tune them for our sentiment analysis tasks to ensure optimal preparation.

In Step 6, we employ the cleaned SenWave dataset for refining the pre-trained models (BERT and RoBERTa) using GLoVe embedding [79] for multi-label sentiment classification. We note that BERT and RoBERTa can be used without GloVe embeddings, but we used GloVe to follow the work of Chandra and Krishna [10].

We begin by setting parameters, including maximum token length, batch sizes, epochs, and the learning rate to tailor the training process. Using the BERT and RoBERTa Tokenizers from the Hugging Face Transformers library [80], we convert the tweets into a compatible format for processing. This format includes the necessary attention masks and token type (given by identity number (IDs)), ensuring the model can effectively learn from our data. We enhance the model with a dropout layer for regularisation and a linear layer for adjusting the output, thus preparing it for robust training sessions. Training involves iterating over batches, where the model computes losses and updates weights accordingly. After training, we evaluate the model’s performance using metrics such as Hamming loss, Jaccard score, label ranking average precision (LRAP) score, and F1 scores. These metrics provide critical insights into the model’s effectiveness in handling multi-label classification tasks, underscoring our comprehensive approach from data preparation to model training and evaluation in a deep learning framework.

In Step 7, we utilise the fine-tuned BERT and RoBERTa models to conduct sentiment analysis on selected articles from the Guardian.

In Step 8, we conduct a multi-label sentiment classification of the Guardian data across the different quarters by categorising sentiments such as optimism, anxiety, pessimism, and surprise. This approach allows for a more nuanced understanding of the emotional nuances expressed within the articles during different phases of COVID-19 that spans quarter before and during the COVID-19 pandemic, covering the beginning, lockdowns, vaccination, end of lockdowns and unrestricted travel, locally and internationally, which marks the end of COVID-19.

In the final step (Step 9), we employ various visualisation techniques, including bigram and trigram analysis for selected sentiments, sentiment polarity, sentiment distribution histograms, and heatmaps. We compare the different sections of The Guardian before and during the COVID-19 pandemic. Through these visualisations, we aim to illustrate the evolution of sentiments and topics over time.

4.1 COVID-19 - cases and deaths

We first present statistics about COVID-19 deaths in relation to the Australia and UK news sections of The Guardian, respectively. Figure 2 describes a comparative analysis of The Guardian article counts and number of death cases across nine quarters for the given years (from the first quarter of 2020 to the first quarter of 2022). We extracted the COVID-19 death data from the Australian Bureau of Statistics [81] and the UK’s Office for National Statistics [82]. In the case of Australia ( Figure 2 - Panel (a)), we notice the spikes in the COVID-19 death counts in the third quarter of 2020, the fourth quarter of 2021, and the first quarter of 2022 corresponds to an increase in the volume of news content, suggesting a correlation between the pandemic’s severity and media coverage. In contrast, the UK experienced significant rises in death rates in the second quarter of 2020 and the first quarter of 2021, without an increase in the news report, indicating a possible divergence between public health data and media response. This analysis sets the stage for a focused examination of the specified quarters in subsequent, more detailed quarterly analyses.

4.2 Data analysis using n-grams

Next, we use n-gram analysis to get an overview about the major topics expressed in the selected sections of The Guardian during the first phase of COVID-19.According to Figure 3-Panel (a), in the first quarter of 2020, the lexicon of global news was dominated by phrases deeply tied to the emerging COVID-19 pandemic, with ”public health”, ”confirmed case”, and ”coronavirus outbreak” leading the frequency plots in bigrams analysis. The trigrams (Figure 3-Panel (b)) show ”chief medical officer”, ”public health England”, and ”confirmed case coronavirus” emerged as top phrases in the World News section. The phrase ”confirmed case” in both bigrams and trigrams points to the critical importance of tracking the virus in the beginning of the pandemic. Similarly, the frequent appearance of ”personal protective equipment” in discussions signals the pivotal role of protective gear in controlling the outbreak, hinting at the global surge in demand for such equipment and the supply chain challenges that ensued. This lexical analysis of early 2020 news coverage offers a window into the world’s collective focus and concerns during the initial stages of the COVID-19 pandemic, revealing an acute awareness of the crisis’s impact on public health.

In the second quarter of 2020 (Figure 4), analysis of bigrams in Australian and UK news highlights key concerns related to the COVID-19 pandemic. We find that Australia focused on ”coronavirus pandemic”, ”public health”, ”new case”, and ”tested positive”, reflecting a close attention to the outbreak’s evolution and its health implications. In contrast, the UK emphasised ”social distancing”, ”mental health”, and ”public health”, pointing to a broader scope of concerns that included preventive measures and the psychological impact of the pandemic. This contrast reveals differing priorities, where Australia’s media concentrated on tracking and testing the virus, whereas the UK highlighted the social and mental health dimensions of the crisis. Through this lens, we can glimpse how each country’s media spotlighted shared and distinct aspects of the pandemic, underscoring diverse strategies and responses to the pandemic.

4.3 BERT and RoBERTa - model evaluation

We present the training (model refinement) performance metrics for both the BERT and RoBERTa models using the SenWave training dataset with 90:10 percent train and test data split. The given metrics encompass Hamming loss, Jaccard score, LRAP score, macro and micro F1 scores. These key performance metrics are given in the literature [83] and used in related work for multi-label sentiment classification [10]. They offer a comprehensive evaluation of the model’s effectiveness in multi-label sentiment classification. Hamming loss measures incorrect label predictions, where lower values signify better performance. The Jaccard score assesses the similarity between predicted and true labels, with higher values indicating better overlap. The LRAP score evaluates the ranking quality of predictions based on the average precision of true labels, with higher values indicating superior ranking performance. The F1 macro-score computes the average F1 score across all classes, reflecting precision and recall equally. Similarly, the F1 micro-score considers the overall F1 score, useful for addressing class imbalance. The higher values of the respective F1 scores signify better model performance.

We refine the respective models using the Senwave train dataset and report the performance on the test dataset. We present the results in Table 3, where RoBERTa exhibits marginally superior performance across various metrics, excelling in LRAP, Jaccard and F1 micro scores. This indicates a slightly better overall performance when compared to the BERT model.

4.4 Sentiments detected by LLMs

We next compare the sentiments detected by BERT and RoBERTa models.Figure 5 presents sentiment distribution within The Guardian (Australia, UK, World news and Opinion sections), spanning from January 1, 2018, to March 31, 2022. This provides the predominant sentiment categories as identified by BERT and RoBERTa offering insights into the overarching emotional tone of the media coverage.A significant finding from Figure 5 is the prevalence of ”official report” as the dominant sentiment across the dataset, regardless of the LLMs used. This observation aligns with the nature of news reporting, which is fundamentally centred on neutrality and objectivity. Certain news articles conveyed information from official statements and reports during COVID-19. Therefore, the prominence of ”official report” as a sentiment category underscores the adherence of the news media to these journalistic standards. Although the prevalence of ”official report” sentiment is expected given the context of news reporting, it presents a challenge for our analysis focused on capturing a broader spectrum of sentiments related to the COVID-19 pandemic. The overrepresentation of this category may potentially obscure the nuances of public sentiment and emotional responses to the pandemic as reflected in news coverage. Consequently, to address this issue and enhance the clarity of human sentiment expressed, we excluded ”official Report” from subsequent analyses.This will allow for a more nuanced exploration of the emotional landscape of news coverage during the specified period. We further aim to uncover more distinct emotional patterns and shifts that may provide deeper insights into the collective sentiment towards the pandemic, beyond the confines of official reporting. Finally, we notice that the ”official report”, ”annoyed,” and ”denial” make the major differences between the sentiment counts when comparing BERT with RoBERTa in Figure 5.

Now that we compared the BERT and RoBERTa models, we will present the rest of the analyses using RoBERTa as our designated LLM.

Figure 6 displays the distribution of emotions within Australian, UK, and World news sections, providing an intriguing perspective on how location-specific contexts influence sentiment portrayal. Despite the smaller dataset for the UK, with a thousand fewer articles than in Australia, an interesting pattern emerges. The sentiment ”sad” is more pronounced in UK coverage, which could potentially reflect the higher COVID-19 mortality rates experienced in the UK compared to Australia during the pandemic since the data also covers the COVID-19 period. Moreover, it’s noteworthy that sentiments such as ”optimism”, ”thankful”, and even ”joking” are more prevalent in the UK news compared to Australia. Both UK and Australia have reported far less negative sentiments (sad, annoyed, anxious, and denial) when compared to the World News section. This phenomenon might indicate a cultural or editorial leaning towards maintaining a semblance of hope and gratitude, even when faced with dire circ*mstances. The higher occurrence of these positive emotions, despite the adversity, aligns with our previous observations of an increase in such sentiments during the pandemic period. Furthermore, it could also indicate a political bias of The Guardian in reporting negative news for the rest of the world, and more positive news about the Western world (UK and Australia).

In Figure 7, we filtered articles from the Australia news section spanning January 1, 2018, to December 31, 2019, each of which contained sentiments such as ”sad”, ”annoyed”, and ”denial”. We selected this based on the three emotions with the highest proportion in 2018 and 2019 in Figure 19. Through this analysis, we discovered that these trigrams encompass various topics, including economics, politics, and military affairs.

4.5 Sentiment polarity analysis

In the global impact of COVID-19, media reporting on pandemic-related information was influenced by various factors, including the government response measures, and public opinion. Consequently, media coverage would exhibit varying sentiments depending on the nature and subject of the news coverage. Analysing changes in sentiment polarity in Guardian articles before and during COVID-19 can provide an insight of the sentiments in the different phases.

We assigned weights to each sentiment to compute our custom sentiment polarity scores for the Guardian articles using RoBERTa, as presented in Table 4. These scores range from -1 to 1, where a positive score denotes favourable sentiment towards the statement, and a negative score suggests adverse sentiment.

Sentiment polarity score in these sections of The Guardian increased significantly from 2019 to 2020, at the start of COVID-19, as shown in Figure 8. According to The Guardian (2020) [84], this surge can be attributed to increased empathy, positive media narratives and people’s ability to adapt. People may write about solidarity, hope, and inspiration despite challenges, creating an upward trend of overall mood improvement.

According to Figure 8, in the overall analysis, the polarity of pre-COVID-19 and post-COVID-19 articles tends to lean towards negative sentiment, which may be attributed to the higher proportion of negative sentiment in the weights we assign. We focus on data from the first quarter (January - March) 2018 to the first quarter of 2022, including Australia and the UK sections. During this period, articles in all three sections of The Guardian showed swings in emotional polarity, reflecting the emotional ups and downs experienced by individuals during the COVID-19 pandemic. As shown in Figure 9, the sentiment polarity of The Guardian is heavily skewed toward the negative, indicating the prevalence of emotional distress among individuals in the COVID-19 era.

Looking at the Australian section of The Guardian in Figure 8, the polarity showed a downward trend from the second quarter of 2020 to the fourth quarter of 2020, possibly due to the rapid increase in global epidemic cases during this period. Sentiment polarity then continued to swing downward towards the last quarter of 2021. This quarterly decline may be related to Australia’s surge in new cases over the same period and current socioeconomic conditions. Instead, the rebound in sentiment polarity in the UK in the first quarter of 2022 may be attributed to the stabilisation of the pandemic in the country as shown in Figure 2. The overall trend in the global sentiment polarity fluctuated downward throughout the COVID-19 period, particularly evident in the first quarter of 2022, suggesting a potential negative impact of the epidemic on global emotional well-being. The regional differences in sentiment polarity values may be associated with factors such as epidemic response measures, change in vaccination rates, and economic recovery efforts in each region.

4.6 Sentiments differences pre-COVID-19 and during COVID-19

We need to have an understanding of the impact of COVID-19 on the style of the news coverage by The Guardian and this can be done by comparing the sentiment analysis results pre-COVID-19 and during COVID-19.Figures 10 and 11 present sentiment distribution and correlation using the RoBERTa model, spanning across the pre-pandemic and pandemic periods. It is evident from the analysis that both periods are predominantly characterized by negative sentiments. Specifically, ”denial”, ”annoyed”, and ”anxious” were the most frequently occurring sentiments detected by the BERT model. Conversely, RoBERTa identified ”annoyed” as the most common sentiment, followed closely by ”denial” and ”sad”. The onset of the pandemic brought about a noticeable increase in ”anxious” and ”sad” across both models, reflecting the global uncertainty and grief triggered by the crisis. This shift underscores the deep emotional impact the pandemic has had on societal sentiments, resonating through the tone of news coverage. Interestingly, despite the surge in negative sentiments, the models also registered an uptick in positive sentiments such as ”optimism” and ”thankful” during the pandemic.

In Figure 12, we can easily see that negative sentiments such as ”denial”, ”sad” and ”annoyed” are the most prominent in Australia, UK and World News sections, before and during COVID-19. Comparing the sentimental changes before and during COVID-19, the content of negative sentiments in Australia and the World News has increased compared to before the pandemic. On the contrary, in UK news, the proportion of negative sentiments during the pandemic has decreased.

4.7 The Guardian’s Opinion section

We transition to a focused examination of the sentiments from the Opinion section of The Guardian presented in Figure 13. We reintegrate the ”official report” sentiment which was previously omitted from earlier analyses to focus on actual sentiments. As mentioned earlier, the ”official report” category typically dominates due to the expected neutrality in news reporting. However, the Opinion section deviates from this trend, with ”annoyed” and ”denial” being the leading sentiments instead of ”official report”. This distinctive distribution aligns with the nature of opinion pieces, which often serve as a platform for expressing dissent and subjective viewpoints.The analysis of the Opinion section across the pre-pandemic and during-pandemic periods reveals surprising shifts in sentiment. Contrary to what might be expected, ”annoyed” and ”denial” sentiments exhibit a decrease during the pandemic, while ”anxious” and ”sad” sentiments show a justifiable increase, reflecting the global distress caused by the crisis. Concurrently, the increase in ”official report” sentiments during the pandemic is also understandable, as opinion writers often refer to official statements and statistics to ground their arguments amidst the unfolding events. Interestingly, the ”joking” sentiment has decreased suggest a shift towards more serious discourse in opinion pieces during the pandemic, indicating the gravity of the situation and possibly a change in the role of humour in public commentary. Although these shifts are noticeable, they are not marked by a significant magnitude. This could indicate that while the emotional tone of opinion pieces has shifted to accommodate the realities of the pandemic, the fundamental nature of the opinion section as a space for personal expression remains intact. It may also suggest that opinion writers and readers have maintained a level of emotional consistency, despite the transformative global events.

4.8 Sentiment-based trigram analysis

We have seen the trend pre and during COVID-19 and different sentiments expressed, but this does not give an indication of why such sentiments were expressed. A way to know why the sentiments such as ”denial” or ”annoyed” were expressed would be by looking at the texts where the LLM detected such sentiments and providing trigram analysis. Hence, this is a way backward, where we look at sentiments detected and then analyse the text further for those sentiments. In Figure 2(a), a minor peak in the number of death cases in Australia during the early stages of COVID-19 is evident, occurring in the third quarter of 2020. Guided by the prevalent emotions identified in Figure 19(c) during the same period, which notably included ”sad”, ”annoyed”, and ”denial”, we selected the Australia news section and extracted trigrams. Figure 14 presents phrases such as ”health human service” and ”hotel quarantine program” within these trigrams. ”Health human service” likely pertains to medical services during the COVID-19 crisis, while ”hotel quarantine program” likely relates to isolation policies implemented during the epidemic.

According to Figure 2(a), after a small peak in the third quarter 2020, Australia’s death toll stabilised until the third quarter 2021, hence we present Figure 15 using the same method, which also contains ”sad”, ”annoyed” and ”denial”; we notice that there are no COVID-19 related trigrams except ”local health district” and ”chief health officer”. This may indicate that media and public attention may have shifted to other events, issues or topics, rather than focusing on coverage and discussion of the epidemic. Moreover, this can also indicate the bias in news reporting by The Guardian that gets more attention from grim news articles.

As the number of deaths continued to rise, the public attention has once again shifted towards COVID-19. Notably, in Figure 16, the reappearance of the trigrams ”rapid antigen test” and ”positive rapid antigen” suggests an increase in positive rapid antigen test results. This uptick reflects the escalating number of COVID-19 cases and the accelerated spread of the virus, signalling a renewed focus on the epidemic.

We can see from Figure 17(a) that before COVID-19, the World News section focused on more diverse articles, ranging from human rights issues to the Prime Minister. Figure 17(b) illustrates the trigrams extracted from The Guardian articles in the World News section during COVID-19. The analysis focused on filtering out sentiments associated with ”denial”, ”annoyed”, ”anxious”, and ”sad” from 1st January 2020 to 31st March 2022. Moreover, Figure 17(b), ”health social care”, ”personal protective equipment” and ”intensive care unit (ICU)” were frequently mentioned, reflecting the attention and importance paid to COVID-19 by the world news section, and also demonstrates the concentrated response to the challenges of the epidemic.

4.9 Examples of sentiments detected

We selected random samples of Australian news in Table 5, which include their predicted sentiments. We can see that in this table, the classification of ”Denial” and ”Joking” for the first article is incorrect. There’s no indication of denial of reality or humorous elements in the text. Rather, the text discusses the restart of the Titanic II project by Clive Palmer, which doesn’t involve denying reality or making jokes. A more appropriate classification for it would be ”Optimistic” as it portrays the restart of the project in a positive light, demonstrating hope and positivity for its future success.

The second row, dated 2021-10-31, describes the experiences of Adnan Choopani and his cousin Mehdi being detained in Australia’s immigration detention system for up to eight years. They have witnessed friends self-immolate and have themselves attempted suicide out of despair, all of which are saddening experiences, hence demonstrating sadness. As for pessimism, the text portrays the various hardships and injustices suffered by Adnan and Mehdi in detention, which may evoke a pessimistic outlook on the future, especially considering their prolonged detention despite having lodged protection claims.

The third row indicates anxiety among some unemployed individuals regarding the impending reduction in welfare benefits. They may feel apprehensive and fearful because this change could impact their livelihoods and quality of life.

We show further examples of sentiments detected in selected paragraphs of The Guardian newspaper articles from the UK and World news sections. We find that the tone of the first row in Table 6 is mainly optimistic as it tells of a wonderful family holiday experience in a South African game reserve. The author recalls happy and beautiful times spent in nature and emphasizes the positivity of the memories created during that time. The sadness may come from mentioning the death of Sudan, the last male northern white rhino. In the second row, the focus is on providing a historical account of royal tours, highlighting their significance in monarchial life throughout history. Therefore, the tone aligns more closely with an official report rather than expressing annoyance or joking elements. In the third row, the annoyance and denial of some MPs over the new lockdown rules are described. They believe that the implementation of lockdown measures in low-risk areas is unreasonable, and expressed dissatisfaction and criticism that the lockdown regulations were released via Twitter at 9:16 pm on Thursday, less than three hours after the regulations were implemented. Furthermore, Boris Johnson’s refusal to condemn a Tory MP’s claim that Muslims ”just don’t take the pandemic seriously” may have caused further negative sentiment.

In Table 7, we find that the article discusses Greek scholars’ potential pride in the widespread use of the Greek word ”pandemic” globally, which may be interpreted as a joke or irony due to the severity of the pandemic with the seemingly pleasant context of using the word. Although the text doesn’t directly express anger or annoyance, it suggests that Greek scholars are concerned about the detrimental impact of the epidemic on their homeland, which could manifest as mild annoyance or dissatisfaction. In the second row of Table 7, the author expresses annoyance with the United States peace initiative, suggesting that it merely serves as a means to facilitate the withdrawal of their troops without genuinely considering the future and interests of the Afghan people. This dissatisfaction reflects a critical perspective on policy decisions, questioning the sincerity and effectiveness of the initiative. It can be interpreted as a sense of annoyance towards the lack of genuine consideration for the Afghan people’s future. Additionally, the author’s stance implies a denial of the effectiveness or goodwill of the peace initiative, highlighting scepticism towards its true intentions and outcomes. In the third row of Table 7, the Labour Party (opposition) accused the government of making a ”significant mistake” and expressed annoyance, believing that the government’s actions were not decisive or effective enough. They pointed out that the United Kingdom had the fifth highest death toll in the world, with the highest mortality rate relative to the population. It also reflects a denial of these measures.