Topic modelling is an unsupervised machine-learning technique that detects neighbouring words and phrase patterns within a set of documents (Blei, Ng, and Jordan 2003; Isoaho, Gritsenko, and Mäkelä 2021). It analyzes data to detect cluster words and similar expressions that describe a corpus (Gerlach, Peixoto, and Altmann 2018). Latent Dirichlet Allocation (LDA), an algorithm, is used for topic modelling. LDA tries to determine a set of topics that are generated through the collection of words and topics. Although a topic is composed of words, the distribution of those words is not equal; therefore, LDA consists of two parts. Firstly, it includes words within a document and, secondly, the probability of words belonging to a topic (Maier et al. 2018; Wei and Croft 2006). The algorithm determines the likelihood of words belonging to a specific topic for the set of given documents. The present study applied LDA topic modelling to extract topics from the corpus based on UN Secretary-General speeches regarding climate change from 2018 to 2022.

This study explores the issues associated with climate change at the global level. For this purpose, the data were retrieved from the United Nations website. It consists of speeches taken from 5 Sep. 2018 to 14 Nov. 2022. The first process in filtering the data is to smooth out the noisy data, which consists of noisy information that contributes less to the meaning-making process (Garćia, De Carvalho, and Lorena 2013). Hence, it is removed in preprocessing.

Preprocessing in the case of the present study was conducted in the following various steps. Firstly, the corpus of speeches was divided into different paragraphs consisting of separate topics. Paragraph and topic detection was conducted by the researchers and later on validated by the language expert. The rationale for separating the paragraph is that one paragraph consists of one focused topic or a discussion about a specific aspect of the topic (Kaplan 1966). This was followed by sentence segmentation, the process of determining longer processing units consisting of one or more words (Palmer 2000). It means that the paragraphs are divided into sentences for further processing. Secondly, tokenization was carried out, which involved breaking sentences into words termed tokens. Tokens are the building blocks upon which analysis is conducted (Palmer 2000). Thirdly, all text was converted to lowercase or uppercase to normalize the data (Vijayarani, Ilamathi, and Nithya 2015). Fourthly, spell correction was the step where an extra space or repetition of words was corrected. Likewise, the next part involves the removal of stop-words. The words frequently used or commonly occurring in the English language are termed stop-words (Silva and Ribeiro 2003). These consist of little or no meaning and only help in forming a sentence. Furthermore, stemming was carried out to remove any trouble in the process of smooth analysis. Stemming refers to the base or root form of a word and involves the conversion of all words to their root forms (Vijayarani, Ilamathi, and Nithya 2015). For example, learn is a base word for its variants such as learns, learning and learned. In the end, text normalization cleans the text and replaces all words with their standard forms for the best results.

The present study underpinned Computational Grounded Theory (CGT) as its theoretical understanding. The theory combines “expert human knowledge and hermeneutic skills with the processing power of computer” (Berente and Seidel 2014; Nelson 2020). It is the extended form of Grounded Theory and consists of three main steps. Firstly, it relies on pattern recognition, which makes use of inductive computational exploration of the text. It reduces complicated messy text into simple and interpretable list of words. Secondly, it involves a confirmation step, which is regarded as the qualitative part of the study where the researcher is supposed to visit the original text multiple time to confirm whether the computationally recognized patterns are plausible, add interpretation to the analysis, and potentially modify the identified patterns to better fit the holistic reading of the data. Lastly, the third step involves whether the patterns identified in the first two steps are generalizable to the entire corpus. During the application of CGT, special care was taken into account; the analysis should be computer assisted, not computer guided (Baumer et al. 2017; Carlsen and Ralund 2022).

The present research followed the CGT in that the initial pattern recognition step involved data filtration and LDA topic modelling. In the pattern confirmation step, the original refined data were visited time and again to check the plausibility of the LDA extracted topics. Lastly, the generalizability of the results has been evaluated. The compete details have been explained through data validation process.

The validation process was carried out at three levels, as proposed by DiMaggio and colleagues (DiMaggio, Nag, and Blei 2013). Statistical validation involved the analysis of a model’s ability to predict the set of documents (DiMaggio, Nag, and Blei 2013). It explores the limits of topic modelling concerning the number of documents and their lengths (DiMaggio, Nag, and Blei, 2013). It works better when the words associated with different topics are clearly separated. This is done in the preprocessing, where certain stop-words are removed from the data to avoid any hurdles.

At first, the researcher ran ten topics in the software, but the results were unclear. Then, the model was run according to eight topics, and these fitted the criteria. The process of narrowing or selecting the appropriate number of topics plays a role in the interpretation of the topics. So, eight topics were finalized, keeping in view their relevance to the corpus. For this purpose, the researcher had to continuously visit the original text or data to look for the validity of the keywords suggested by the model. Furthermore, semantic validation was carried out to compare the results of the LDA method with expert reasoning to make semantic sense of the topics (Isoaho, Gritsenko, and Mäkelä 2021). The word cloud was generated as a result of eight topics consisting of specific keywords. These keywords depicted semantic meanings according to themes being discussed in the topics. The researcher then explored these themes to make sense of the hidden information. Moreover, in predictive validation, external events predict the outcome of the model (Isoaho, Gritsenko, and Mäkelä 2021). Therefore, the researcher, in the end, reached the sense of the extracted topics and predicted solutions or measures to overcome the discussed problem.

Table 1 contains the LDA result and suggested topic headings for the extracted keywords. It consists of eight topics, which are inferred from word clouds and their corresponding word probabilities. Figure 4 shows the visualization of the topics extracted through LDA.

The table depicts the topic word proportions and the document topic proportions. Topic word proportions represent keywords associated with each topic, whereas document topic proportions display topics connected to keywords along with their proportions or weights in the document (Vijayarani, Ilamathi, and Nithya 2015). Initially, ten LDA was run with a limitation to ten topics. However, by observing the redundancy of keywords across multiple topics, the total number of topics was limited to 8. It consists of the most co-occurring words in the document ranging from 0 to 7. Each group consists of ten keywords and are written with their corresponding weight or probabilities in the document. Table 1 consists of keywords and the proposed themes.

The first set of keywords relates to the generation of green energy (0.048). It suggests that people (0.032) should have access (0.027) to renewable (0.021) energy sources. As renewable energy is already (0.018) cheaper than fossil fuels, it can help deliver (0.027) alternate energy. This can be done with the support (0.043) of the international (0.027) system (0.022) and public (0.017) action to increase the ambition in support of climate change plans.

The second group of words highlights the effects of climate change on our lives. It suggests that the climate (0.100) crisis (0.045) has increased due to the gap (0.037) between national commitments and what people actually need to do to reduce the emission of harmful gases. The impact (0.040) of the climate crisis is devastating, and as a result, vulnerable (0.041) nations and groups suffer the most. That is why developed countries should support (0.033) the efforts of developing countries with an increase (0.023) in finance. Likewise, protecting (0.021) livelihoods (0.033) should be considered a priority to save the community (0.020).

The third set of keywords relates to the work (0.014) on the part of government (0.016) to make (0.024) our world (0.036) a clean place. Today (0.016), technology and the economy are under the control of human beings, but no country (0.052) is safe from the effects of climate change. Every year (0.014), climate crises (0.014) are experienced by people; therefore, climate (0.33) action (0.016) plans should be made effective by each country (0.052).

The fourth set of keywords across topic reflects upon the loss (0.061) caused by natural disasters (0.055). Climate-related disasters (0.055) have hit (0.014) a record (0.016) in the past five years. In several regions (0.018) of the world, coal power plants are still built in large numbers. The Pacific region (0.018) is on the frontline of climate change. Storms and floods (0.017) have caused innumerable deaths in many countries of the world. Ocean heat is at a record (0.016) level. Arctic sea ice is melting, and now re-freezing is the slowest on record (0.016). Yet (0.013), despite ample awareness, the world is not making progress (0.036) fast enough. The focus must be on low (0.018) carbon emissions and a climate-resilient future. We are not yet (0.013) winning the battle in relation to climate change, but this can happen (0.034) through the mutual effort of people coming together.

The fifth set of keywords in Table 1 relates to global food (0.045) and life (0.046) crises caused by climate change. Climate change has now become (0.023) a threat to all humanity (0.021). There is a clear connection between climate change and the situation (0.018) in the oceans (0.015). Climate change is forcing people (0.016) to move from their houses as a result of floods and other natural calamities. The conflict (0.019) among the countries (0.020), which has resulted in wars, has caused food (0.045) and energy crises. Ukraine’s (0.020) production of food was not accessible to the international markets in the past year because Russia banned the exports of food and fertilizers. All of this gives (0.016) us a glimpse of what lies ahead; prolonged droughts, extreme weather conditions and horrific flooding. Therefore, limiting global temperature rise must be the priority.

The sixth set of keywords relates to mitigating the effects of greenhouse gases. Our world has crossed the threshold of a safe temperature level. Hence, all countries need (0.019) a clear climate (0.016) action (0.015) plan to reduce and cut global (0.042) emissions (0.031). We are really out of time (0.014), so the governments must act now to prevent further irreversible damage (0.015). The solutions are very clear (0.015). All nations need (0.019) to join the net zero emissions coalition, as the only way (0.016) to prevent exceeding this threshold is to step up our efforts to reduce the temperature to 1.5°C. Otherwise, the world may pay a tragic price (0.015).

The seventh set of keywords highlights sustainable development goals. It urges an end to the use (0.022) of coal (0.018) with the help of international support. The UN Secretary-General stresses closing carbon markets (0.019) and adopting plans for transitioning to net zero emissions by 2050. He suggests the transition (0.028) of energy systems from fossil fuels (0.022) to renewables, as renewable energy is the only energy source expected to grow (0.017) in the future. Therefore, financial institutions are now shifting their investments (0.035) toward a decarbonized, sustainable, and resilient economy. Hence, investment funds (0.019) and companies that have the technological knowledge provide financial and technical support for every country (0.054) that needs assistance.

The last set of keywords relates to climate change adaptation. It stresses the need (0.042) to develop (0.053) adaptation (0.035), which requires strong international support as climate (0.017) change is rapidly increasing. This topic also focuses on finance (0.030) to provide (0.019) money for adaptation (0.035) and resilience in developing countries as they face debt (0.013) crises. For this purpose, the UN Forum requests all financial (0.015) institutions and banks to align their investments with the Paris Agreement Pact (0.017) to help vulnerable countries re-launch their economies in a sustainable and inclusive way. In this way, we can hope (0.013) to secure the breakthrough in adaptation and resilience that is needed to overcome the issue of climate change.

For further validation of the proposed keywords and themes explained in Table 1, all the topics were visualized through word clouds. LDA topics using word clouds separate a text corpus into different topics, assigning each document in the corpus a topic weight (Nastiti, Hidayatullah, and Pratama 2021). Each topic, in turn, is represented by a set of term weights. The word with the highest weight appears bigger than the other ones. Likewise, the words with less weight appear in a smaller size. This helps the researcher to generate topics according to the keywords and their probabilities as shown in Figure 2.

Figure 3 shows a t-Distributed Stochastic Neighbor Embedding (t-SNE) chart, which is a visualization technique that can help to explain the structure of a large corpus of data. The plot is two-dimensional and shows clusters of similar documents based on their content (Yi, Xue, and Compare 2013). Each topic in the plot represents a document, and the location of the topic reflects the content and relationship between documents (Yi, Xue, and Compare 2013). In order to highlight how one topic differs from others in terms of distribution, each topic has been given a different colour. This is useful for exploring and understanding the similarities and differences between different topics.

Figure 4 demonstrates the topic distribution derived by t-SNE through the LDA procedure. The distribution suggests that the themes are distinct from one another, and the topics have meanings of their own, contributing to the whole corpus.