On SGX's Voyage to corporate sustainability: Exploring emerging topics in multi-industry corpora

To streamline our research, we focused on identifying keywords that represented the most significant environmental topic “climate change.” At the outset, we performed straightforward word counts to discern trends within the data (Tables 2 and 3). The primary observation from this high-level data analysis is that there is no evident continuous increase in attention toward climate-related topics in either narrative. Contrarily, we observe a decrease in attention toward climate topics until 2014, followed by a slight pick-up in attention thereafter. This observation suggests that companies initially ignored the voluntary guidance, but they became responsive to the trend of SR only after SGX announced the forthcoming mandatory approach in 2014.

By ceasing the dependency on dominating measure – i.e., f ( x ) f\left(x) , and avoiding excessive sensitivity to class skew caused by imbalanced data among documents (Ni et al., 2020), we use Hellinger distance to observe shift in topic trends over time. This validates the preliminary skim and examines whether the topics of annual reports change over time. Give two documents p p and q q with the distributions of topics z p = ( z p , 1 , … , z p , k , … , z p , K ) {z}_{p}=\left({z}_{p,1},\ldots ,{z}_{p,k},\ldots ,{z}_{p,{\mathsf{K}}}) and z 1 = ( z q , 1 , … , z q , k , … , z q , K ) {z}_{1}=\left({z}_{q,1},\ldots ,{z}_{q,k},\ldots ,{z}_{q,{\mathsf{K}}}) , respectively. The Hellinger distance d p , q {d}_{p,q} between p p and q q over K {\mathsf{K}} topics is computed by h ( z p , z q , M ) = 1 2 ∑ k = 1 K ( z p , k − z q , k ) 2 , h\left({z}_{p},{z}_{q},{\mathsf{M}})=\frac{1}{\sqrt{2}}\hspace{0.33em}\sqrt{\mathop{\sum }\limits_{k=1}^{{\mathsf{K}}}{\left(\sqrt{{z}_{p,k}}-\sqrt{{z}_{q,k}})}^{2}}, where M {\mathsf{M}} is the corpus, a set of documents. We calculate pairwise distances within the Company Overview and Leader’s Words corpora, respectively (Figures A1 and A2). The distributions of distances for both corpora change over time. In each 2-year set of documents, the distances generally remain under 0.5, indicating consistency in document writing, which should allow us to obtain the trend over time. Specifically, in the Company Overview corpus, documents exhibit notable similarity, often resulting in many Hellinger distances approaching or equaling 0.

LDA is an unsupervised classification algorithm designed for corpus modeling. It is widely acknowledged as one of the most effective techniques for text summarization and information retrieval, presenting information in a concise and understandable form (Chauhan & Shah, 2021), which is especially valuable in our case, where we deal with complex multi-industry corpora.

It is a generative probabilistic model that posits topics as mixtures of an underlying set of words, while each document is a mixture over a set of topic probabilities, allowing for the generation of new documents by sampling words from topics, detailed in Figure 4. LDA assumes that the words with the highest probabilities within each topic generally offer a meaningful insight into the essence of the topic. Several implementations are evident within financial and managerial domains, such as event studies (Dyer et al., 2017), investor sentiment analysis (Feuerriegel & Prollochs, 2021), and corporate organization research (Culasso et al., 2023).

Figure 4

LDA plate notation. The subscript is dropped in the diagram.

Using a given text collection (corpus) containing M {\mathsf{M}} documents, and K {\mathsf{K}} topics, the LDA model can infer the topic distribution for a new, unseen document. Each document m ∈ { 1 , … , M } m\in \left\{1,\ldots ,{\mathsf{M}}\right\} in the corpus has its topic distribution θ m ∼ Dirichlet { α } {\theta }_{m}\hspace{0.33em} \sim \hspace{0.33em}{\rm{Dirichlet}}\left\{\alpha \right\} , i.e., p ( θ m ∣ α ) = Γ ∑ k K α k ∏ k K Γ ( α k ) ∏ k K θ m , k α k − 1 , p\left({\theta }_{m}| \alpha )=\frac{\Gamma \left(\mathop{\sum }\limits_{k}^{{\mathsf{K}}}{\alpha }_{k}\right)}{{\prod }_{k}^{{\mathsf{K}}}\Gamma ({\alpha }_{k})}\mathop{\prod }\limits_{k}^{{\mathsf{K}}}{\theta }_{m,k}^{{\alpha }_{k}-1}, influenced by the prior sparsity parameter α \alpha , with smaller α \alpha leading to fewer topics per document. Similarly, each topic k ∈ { 1 , … , K } k\in \left\{1,\ldots ,{\mathsf{K}}\right\} has the word distribution φ k ∼ Dirichlet { β } {\varphi }_{k}\hspace{0.33em} \sim \hspace{0.33em}{\rm{Dirichlet}}\left\{\beta \right\} , with small β \beta leading to fewer words per topic. At each word position n ∈ { 1 , … , N m } n\in \left\{1,\ldots ,{{\mathsf{N}}}_{m}\right\} in every document m m , assign a topic z m , n ∼ Multinomial ( θ m ) {z}_{m,n}\hspace{0.33em} \sim \hspace{0.33em}{\rm{Multinomial}}\left({\theta }_{m}) , and choose a word w m , n ∼ Multinomial ( φ k ) {w}_{m,n}\hspace{0.33em} \sim \hspace{0.33em}{\rm{Multinomial}}\left({\varphi }_{k}) from topic k k . N m {N}_{m} is the length of document m m . Here, we obtain the distribution of topic k k chosen for word w m , n {w}_{m,n} in document m m as p ( z m , n = k ∣ z m , − n , w m , n ) ∝ C m , k + α k ∑ i K ( C m , i + α i ) C k , w m , n + β w m , n ∑ j M ( C k , j + β j ) , p\left({z}_{m,n}=k| {z}_{m,-n},{w}_{m,n})\propto \left[\frac{{C}_{m,k}+{\alpha }_{k}}{\mathop{\sum }\limits_{i}^{{\mathsf{K}}}\left({C}_{m,i}+{\alpha }_{i})}\right]\left[\frac{{C}_{k,{w}_{m,n}}+{\beta }_{{w}_{m,n}}}{\mathop{\sum }\limits_{j}^{{\rm{M}}}\left({C}_{k,j}+{\beta }_{j})}\right], where C m , k {C}_{m,k} is the count number of times of topic k k assigned to words in document m m ; C w m , n {C}_{{w}_{m,n}} is the count number of times the word w m , n {w}_{m,n} assigned to topic k k across all documents. The multiplier is the distribution of topic k k for document m m , namely, θ m , k {\theta }_{m,k} ; the multiplicand is the distribution of word w m , n {w}_{m,n} for topic k k , namely, φ k , w m , n {\varphi }_{k,{w}_{m,n}} . Through iterations, the LDA generative process for each document – a joint distribution of topics and words – is given as p ( θ , φ , z , w ∣ α , β ) = ∏ k K p ( φ k ; β ) ∏ m M p ( θ m ; α ) ∏ n N p ( z m , n ∣ θ m ) p ( w m , n ∣ φ z m , n ) . \begin{array}{l}p\left(\theta ,\varphi ,z,w| \alpha ,\beta )\\ \hspace{1.0em}=\mathop{\displaystyle \prod }\limits_{k}^{{\mathsf{K}}}p\left({\varphi }_{k};\hspace{0.33em}\beta )\mathop{\displaystyle \prod }\limits_{m}^{{\mathsf{M}}}p\left({\theta }_{m};\alpha )\mathop{\displaystyle \prod }\limits_{n}^{{\mathsf{N}}}p\left({z}_{m,n}| {\theta }_{m})p\left({w}_{m,n}| {\varphi }_{{z}_{m,n}}).\end{array} The first two sums contain symmetric Dirichlet distributions, which are prior probability distributions for the documents and topics, respectively. Note that the parameters α \alpha and β \beta are optimized to maximize the likelihood of the observed data (words in documents), i.e., ∑ m M log ( w ∣ α , β ) {\sum }_{m}^{{\mathsf{M}}}\log \left(w| \alpha ,\beta ) , marginalizing over the hidden variables (topic assignments and topic proportions), i.e., p ( θ , z , w ∣ α , β ) ∕ p ( w ∣ α , β ) p\left(\theta ,z,w| \alpha ,\beta )/p\left(w| \alpha ,\beta ) . This ensures that the model captures the underlying topic structure while estimating the most probable topics and their distributions for the given documents. The article investigates annual reports using a 2 year time window and relaxes the constraint on the optimal number of topics, in contrast to the DTM that has a fixed number of topics.

Topic coherence measures (or confirmation measures) aim to evaluate individual topics by quantifying the semantic similarity among the top-scoring words within each topic, approximating human judgment to determine the optimal number of topics for a given corpus, i.e., Company Overview corpus, Leader’s Words corpus. Several proxies have been proposed (Campagnolo et al., 2022). One state-of-the-art method used as a proxy for topic coherence is the C v {C}_{v} measure (Roder et al., 2015), which integrates the indirect cosine measure with normalized pointwise mutual information (NPMI) and a Boolean sliding window. For words w i {w}_{i} and w j {w}_{j} , the co-occurrence NPMI is computed as follows: NPMI ( w i , w j ) γ = log p ( w i , w j ) + ε p ( w i ) ⋅ p ( w j ) − log [ p ( w i , w j ) + ε ] γ , {\rm{NPMI}}{({w}_{i},{w}_{j})}^{\gamma }={\left(\frac{\log \frac{p({w}_{i},{w}_{j})+\varepsilon }{p({w}_{i})\cdot p({w}_{j})}}{-\log {[}p({w}_{i},{w}_{j})+\varepsilon ]}\right)}^{\gamma }, where γ \gamma is employed to amplify the significance of highly associative words; and ε \varepsilon is a small constant used to avoid a logarithm of zero. We calculate the co-occurrence of a word within a given topic against all other words of the same topic, and the weight, and then, we obtain a vector for the word. Using cosine similarity, we measure the vector distances and then obtain their arithmetic mean as coherence score C v {C}_{v} . Higher C v {C}_{v} indicates that the corpus exhibits a stronger coherence with the given topics.

This delves into the determination of the optimal number of topics for the LDA model and observes its evolution over time. However, the findings reveal that the topics achieving higher coherence scores for both Company Overview and Leader’s Words corpora do not exhibit any discernible pattern or consistency as time progresses. Among all the optimal topics extracted, only one is found in 2012–2013 to be related to sustainability.

Figures A3 and A4 demonstrate that in both corpora, the coherence scores exhibit a jagged pattern without showing any clear increasing or decreasing trend as the number of topics increases. This could be attributed to the high variety of industries represented in the collected annual reports, making it challenging to achieve coherence in terms of the topics selected and determining the optimal number of topics. Figure A5 provides insight into the ideally optimal numbers of topics. During the last 2 years (2017–2018), both corpora exhibit higher coherence scores with a larger number of topics, indicating that the writing style among companies is becoming more diverse. Furthermore, our observations reveal a rising trend in the Leader’s Words corpus, contrasted by a declining trend in the Company Overview corpus. This suggests that over time, Company Overviews become increasingly limited in their content scope, while Leader’s Words continue to exhibit a sparse coverage of topics.

We adhered to best practices to comprehend the LDA topics (Jelodar et al., 2019, Maier et al., 2018). Two of the three co-authors independently reviewed each topic from both the strategic and identity narratives, examining the word distribution figures for the first 20 words. After this step, we engaged three RAs to perform the same task.

The three RAs worked independently and submitted their findings individually to one co-author. The co-author then compared the three answer sets from the students and the two answer sets from the co-authors to identify commonalities and discrepancies. A week later, the lead author gathered the RAs for a discussion to address areas of disagreement and achieve consensus on a single topic for each word distribution in each year. For example, in Figure A6, topic 6 is labeled as “medical and healthcare,” topic 9 as “tourism,” and topic 10 as “fossil fuel.”

A wide range of topics is linked to specific industries, and topics related to corporate performance (i.e., topics 1 and 2 in Figure A14) are consistently present across all Leader’s Words corpora. However, sustainability-focused topics are solely evident within the Leader’s Words corpus, constrained to the sphere of the industrial sector during the years 2012–2013 and 2017–2018 (Figure 5).

Figure 5

Limited presence of sustainability-related topics. Note that the x-axis is represented in percentage (%).

During the period 2012–2013, SR remained a voluntary undertaking, but significant progress and more practical discussions on the matter gained momentum beyond that period. In the years 2017–2018, SGX formalized and initiated the mandatory implementation of SR using the “comply or explain” approach. The corpora reveal emerging topics on sustainability, but they are only observed during the periods of pre-deployment of SGX policies when there were many significant discussions on the subject.

As of now, SR has become mandatory, and keywords related to sustainability can be observed in corporations’ annual reports. However, it is essential to note that as a result of this thorough NLP analysis, sustainability is not the main focus of these reports. While previous studies, such as Loh et al. (2017) and Chen (2024), highlight a positive correlation between SR and a firm’s market value, as well as SR’s potential to reduce investment risk, our analysis reveals limited substantive content to support a significant discourse on sustainability-related topics. This suggests that despite the potential benefits of SR, the actual depth of discussion around sustainability within reports remains shallow.