Analysing corporate social responsibility reports using document clustering and topic modeling techniques

Abstract

Corporate social responsibility <span>(CSR) has become an imperative tool to address challenges and achieve sustainable growth. Realizing its impact to the society, companies are demanded to participate in sustainable development of which poverty eradication is one of it. The CSR practice, to date, is not strategically planned and executed especially when it comes into philanthropic corporate social responsibility (PCSR). This could be due to failure to identify categories of PCSR activities, limiting its effectiveness to achieve the intended outcomes. Thus, document clustering is proposed to be used to automate the pattern identification process. This study has extended document clustering by integrating the traditional document clustering application with topic modeling approach. This integrated approach enables the identification of the PCSR pattern. The analysis involved a three-year data from the annual report of the 25 CSR-award winning companies in Malaysia which involved several steps. Findings from this study revealed seven clusters that represented seven types of PCSR activities performed by the CSR-award winning companies in Malaysia. The findings offer an insight to be considered by companies in strategizing the CSR activities, particularly philanthropic responsibility in ensuring optimum impact to innovatively support the society and contribute towards poverty mitigation.</span>

Topic prediction modelling on social media content using machine learning

“…The algorithm that will be used is LDA. LDA extends probabilistic latent semantic analysis (PLSA) by using dirichlet priors for document-specific subject mixtures, resulting in the discovery of texts that would have usually passed unnoticed and Gibbs sampling which used by LDA to apply the model [13]. Two initial outputs are provided by an LDA model that is an estimated probability that a topic will produce a document (referred to as an affinity score) and the probability that a word will be used to describe a topic [14].…”

“…This modeling will provide the value of topic distribution for each document and keywords for each topic. For the topic frequency of each document it will be calculated by the number of frequencies calculated during initialization and the Dirichlet generated multinominal distribution for topics in each document while for keywords on each topic its calculated by the number of frequencies calculated during initialization and generated Dirichletmultinominal distribution for each word on each of these topics [13]. Finally, several iterations are performed until convergence is achieved and calculate the topic probability distribution and the probability of each word that best represents of the topic.…”

“…Mangsor et al [13] implemented in LDA on corporate social responsibility based on annual reports of companies using document clustering k-means and topic modelling LDA, in this study k-means is used to assign k clusters and assign each document to the cluster which later LDA was used in naming the labels that can provide a global picture of themes and their differences, while enabling a comprehensive analysis of the most related keywords to each topic. It was found that 87% of the topics generated by LDA made sense to human judges and could identify multi-faceted themes, extract latent themes that might be missed, and reflect those themes in their raw state without bias in e-petition data [14].…”

Topic modelling of legal documents using NLP and bidirectional encoder representations from transformers

“…To find a solution to unstructured data in Kadir and Aliman [10], the web-based text analytics and the R language are used to produce organized and summarized data. In Mangsor et al [11], the traditional application of document clustering was combined with the topic modelling approach. With this integrated approach, it is possible to see the pattern.…”