Impact Statement

The proposed work is envisaged to play a significant role in sustainability monitoring by detecting organizational violations as soon as they are reported by regulatory agencies and thereby supplement the Environmental, Social, and Governance (ESG) scores issued by third-party agencies. We hope such real-time sustainability monitoring and regulatory reporting definitely help in bringing down the number of incidents. Aggregate analysis like ours can also help other organizations to become cautious.

2. A Brief Overview of ESG Reporting

GRI is a global organizationFootnote ⁴, which sets the standards for sustainability practices across the globe. It provides a number of indicators in each of the E, S, and G categories, for which information has to be provided by the organizations (Chiarini, Reference Chiarini2017; Ching et al., Reference Ching, Gerab and Toste2014) GRI standards define a range of indicators falling in different categories.

• • Series 200 defines economic indicators that are used to evaluate an organization’s commitment to good governance practices including handling business ethics, intolerance toward child labor, etc. The indicators in this category are numbered as 2xx—for example, performance(201), anti-corruption(205), etc.

• • Series 300 defines environmental indicators that are used to assess a company’s contribution toward the environment using parameters like its commitment to reduction in carbon emission, water recycling, conservation tasks, etc. All indicators in this category are numbered as 3xx—for example, energy(302), biodiversity(304), emissions(305), etc.

• • Series 400 defines social indicators like training and education(404), security practices(410), etc. These indicators assess an organization’s commitment toward its employees’ physical and mental well-being, training opportunities provided, etc.

An organization publishes its ESG activities in one or both of the following reports:

1. a. Sustainability Report: It mainly contains the non-financial data describing the journey of progress of organizational performance targets against specific economic, environmental, social, and governance goals along with a few metrics to establish its journey toward sustainable development;

2. b. Integrated Report: It contains both financial and non-financial data linked by a narrative explaining data linkage to future corporate profitability along with the company’s vision on sustainable development.

However, these self-published reports only provide grossly positive information about the companies and not much about violation incidents. Analysts, regulators, and investors also use a number of other sources to track organizational activities that may contradict the principles of sustainability. Reports published by regulatory agencies are the most important sources of current information. These sources may contain information about ESG violations along with events that caused or are likely to lead to violations, along with the probable future impact of non-sustainable actions. The language of these reports is different from the reporting language.

The primary contribution of this paper is a linguistic model that can identify such violation incidents or events that may lead to future violations from reports and thereafter link them to GRI indicators. This helps in deriving insights in a more uniform way.

3. Monitoring Regulatory Websites for Screening ESG Activities—A Conceptual System

Figure 1 presents a schematic diagram of an enhanced sustainability screening system that is being currently conceptualized. This system collects information from websites of regulatory agencies on a regular basis and analyzes them to generate individual and aggregate statistics about sustainability incidents and violations. The aim is to enhance the visibility of sustainability practices of a company through constant monitoring of reliable web sources.

Figure 1. A schematic diagram of a unified sustainability information analytics.

The present work uses regulatory reports published by three agencies namely OSHAFootnote ⁵, EPAFootnote ⁶, and Department of Justice (DoJ)Footnote ⁷. Initially, a corpus of reports was used to create an annotated dataset. A neural model is then trained on the annotated dataset, validated, and then applied to detect such violation incidents from current reports. The detected events are mapped to corresponding GRI indicators and subjected to aggregate analysis for insight generation. Coming from authentic sites only, these insights are trusted and verifiable.

For this work, we have crawled a total of 3,869 articles from the above-mentioned sites. While 900 among them have been used to train the model and evaluate it, the remaining articles have been used to derive insights about recent trends in the ESG domain. The crawlers continue to be active.

4. ESG Information Extraction—A Knowledge-driven Approach

A wide range of events starting from managing climate risk to building a diverse and inclusive workforce along with all aspects of governance can qualify as a potential ESG event. While each regulatory agency reports about events that primarily come under their purview, obviously a report can be quite mixed in nature. There is no accepted rigorous definition of what constitutes an event. For example, an incident like a factory accident that causes the death of employees thereby leading to a penalty, can be viewed as a collection of multiple events. Given the nature of the domain and the diversity of the reporting patterns, we propose a knowledge schema that elicits the different entities and their roles that are usually observed in a regulatory article and also captures the relationships among these entities. Figure 2 presents this knowledge schema. G represents the regulatory agency name while TO represents the target organization, which is the subject of the article. The other entities encountered are penalty (P), award mentions denoted by (A), location names denoted by L, and textual strings denoted by I and V, which represent incidents and violations. The relationships among a pair of entities may be expressed by different types of linguistic constructs, and hence established via multiple paths.

Figure 2. ESG incident knowledge schema illustrated along with example sentences from text containing relevant entities. Different events and entities in text are marked in bold.

The cause–effect relationship between an incident and a violation is not very well defined. While in one case an accident may be reported as an incident caused by a violation of safety norms within a factory, in another instance an action like accepting a bribe may be reported as an incident that led to a violation of governance regulation. These reports often also contain information about the violated clause, the regulations or acts that were violated, the indicators that are affected, the penalty value imposed by a regulatory agency, extent of loss incurred due to the event, number of deaths or injuries, penalty paid, etc. While some of the entities like TO, L, or P can be detected using standard Named Entity Recognition (NER) tools, detecting incidents and violations is a non-trivial task. The proposed neural model is trained to do this. The above-mentioned knowledge graph is used to create an annotated dataset for the purpose. The detailed process of annotation is mentioned in the next subsections.

4.1. The Data Annotation Process

A training corpus of 900 regulatory articles comprising of around 10,800 sentences, under E, S, and G categories published between 2015 and 2018 by the U.S. Environmental Protection Agency (EPA), Occupational Safety and Health Administration (OSHA)Footnote ⁸, Equal Employment Opportunity Commission (EEOC)Footnote ⁹, and the U.S. Department of Justice (DoJ)Footnote ¹⁰ respectively were manually annotated to mark the various entities mentioned in the knowledge graph. The annotation was done by multiple annotators following a rigorous procedure to ensure acceptable inter-annotator agreement. The semantics of the nodes and relations mentioned in the knowledge graph were used to guide the annotators while identifying words, phrases, or portions of sentences that may represent an instance of a node in a document.

To help the annotators, each document is first processed using the Stanford NER (Manning et al., Reference Manning2014) to obtain the organization names, locations, and currency values as named entities. This helps in quick localization of the first four elements, if present in the document. The annotators are domain experts who are knowledgeable about sustainability domain, familiar with incidents and violations.

The rules followed by the annotators are as follows:

1. a) Regulatory Agency (G): This was programmatically done by using a list of regulatory agency names and their abbreviations, which was searched from among the NERs detected.

2. b) Target Organization (TO): Of the many organization names that may appear in a document and are detected by the NER earlier, the task during annotation is to identify and tag the violating or the award-winning organization as TO.

3. c) Location (L): Among possibly many locations appearing in a report, the task here is to mark the correct geographical location of the incident, in the absence of which, the location of the organization could be marked, if it appears in the document.

4. d) Proposed Penalty (P): The currency value that denotes the penalty that has been enforced upon the target organization by a governing body is marked.

5. e) Incident (I): Annotators have to tag phrases or sequences of words that collectively are indicative of an incident or an event.

6. f) Guideline violations (V): These are phrases or sets of words that collectively indicate non-compliance or failure to comply with guidelines or regulations.

7. g) Award (A): The award name and associated money value or citation phrases, are manually tagged during the annotation process.

Six annotators took part in the annotation, with each expert annotating 400 documents using the Stanford simple manual annotation toolFootnote ¹¹. This included 100 documents, which were sent to all the annotators to compute the inter-annotator agreement later. The average length of a document is around 23 sentences. The experts read each document and performed the following tasks:

1. 1. Task-1: Label sentences of each document as POSITIVE—if the document reported events of winning awards or recognition for sustainability practices or policy announcements in favor of sustainability or NEGATIVE—if the document mentions events that violate ESG regulations or incidents reporting damages, loss of life, etc. or NEUTRAL—in case none of the above factors hold.

2. 2. Task-2: This task had two components:

   1. (a) From among the named entities, the target organization, penalty values, and locations were marked, if any, and

   2. (b) Mark phrases in the text that indicate announcements, incidents, and violations.

At the end of the annotation, each word in the document is assigned a label TO, I, V, P, L, A, or None. Table 1 illustrates the annotation with some example News texts. For the sake of understanding, we have shown labels of only the phrases that belong to any one of the following classes {TO, I, V, P, L, A}. Table 4 presents the detailed overview of the dataset used to train and test our proposed models.

Table 1. Sample sustainability News texts with the respective annotated entities and events. Note that all the target organization names were intentionally masked by the token [ORGName] to maintain anonymity

4.1.1. Computing annotator agreement

Using the annotations obtained for 100 common documents, we measured the inter-annotator agreement using the Fleiss Kappa (Fleiss et al., Reference Fleiss, Levin and Paik1981) measure $ \left(\kappa \right) $ . This is computed as:

$$ \kappa =\frac{\overline{P}-{\overline{P}}_e}{1-{\overline{P}}_e} $$

The factor $ 1-{\overline{P}}_e $ gives the degree of agreement that is attainable above chance, and $ \overline{P}-{\overline{P}}_e $ gives the degree of agreement actually achieved above chance. It was observed that the inter-annotator score for Task-1 was 0.83, which is appreciably high. For Task-2, it was found to be 0.71. The scores are computed using word-label matches assigned by different annotators. The very high scores indicate that all experts were marking fairly uniformly and therefore, the expert annotated dataset is reliable to be used for training incident detection systems.

Out of the 20,700 sentences from 900 documents, around 7,500 sentences were found to contain words belonging to at least one type mentioned in the incident knowledge schema. Altogether, we obtained 3,000 violation phrases, 2,100 penalty phrases, 2,223 Target Organizations, 1,300 locations, 695 incidents, and 101 awards.

5. A Multi-tasking Neural Model for ESG Knowledge Extraction

In this section, we present the details of incident labeling architecture implemented using multi-task neural model. This model works on each sentence at a time to detect elements of interest that are defined in the incident knowledge schema. Multi-task learning utilizes the correlation between related tasks to improve classification by learning tasks in parallel. In the present work, the two related tasks are task-1: classifying a sentence into either positive or negative classes as discussed earlier, and task-2: labeling appropriate phrases in the text as per the incident knowledge schema. Accordingly, for task-1 each sentence of the training data is labeled as either Positive (Award), Negative(Penalty), or Neutral whereas for task-2 each word/phrase of a sentence is labeled according to the following classes $ < TO,I,V,P,L,A> $ .

Figure 3 presents a high-level view of the proposed model. It uses a cascaded CNN-BiLSTM layer for the combined tasks of sentence classification and sequence label prediction, using the fine-tuned S-BERT for creating the sequence embeddings. Thus, the task-(a) sequence classification is predicted as:

(1) $$ {y}^i= softmax\left({W}_i\ast {h}_1+{b}_i\right). $$

Figure 3. BERT-based multi-tasking network for ESG sequence classification and labeling.

On the other hand, for task-(b) or the sequence labeling task, we feed the final hidden states of the $ BERT- CNN- LSTM $ network of the other tokens, $ {h}_2,.\dots, {h}_T $ , into a softmax layer to classify over the sequence labels. The output of the model is represented as:

(2) $$ {y}_i^s= softmax\left({W}^s\ast {h}_i+{b}_s\right),i\in \left(1\dots N\right) $$

where $ {h}_i $ is the hidden state corresponding to the word $ {w}_i $ .

To obtain the multi-tasking model for dual tasks of sequence classification and sequence labeling, the $ BERT- CNN- BiLSTM $ layers have been trained with two separate loss functions $ {L}_1 $ and $ {L}_2 $ . Where,

$$ {L}_1\left(\theta \right)=-\sum \limits_{t=1}^M\sum \limits_{k=1}^K{\overline{y}}_t^k\mathit{\log}\left({y}_t\right)\hskip0.42em \mathrm{and}\hskip0.5em {L}_2\left(\theta \right)=-\sum \limits_{t=1}^N\sum \limits_{j=1}^J{\overline{q}}_t^{i,j}\mathit{\log}\left({q}_t^i\right) $$

$ {q}_t $ is the vector representation of the predicted output of the model for the input word $ {w}_t^i $ . $ K $ and $ J $ are the number of class labels for each task. The model is fine-tuned end-to-end via minimizing the cross-entropy loss.

We define the joint loss function using a linear combination of the loss functions of the two tasks as:

(3) $$ {L}_{joint}\left(\theta \right)=\lambda \ast {L}_1\left(\theta \right)+\left(1-\lambda \right)\ast {I}_{\left[{y}_{sentence}==1\right]}\ast {L}_2\left(\theta \right) $$

Where, $ \lambda $ is a hyper-parameter whose value is learned to control the contribution of losses of the individual tasks in the overall joint loss. $ {I}_{\left[{y}_{sentence}==1\right]} $ is an indicator function which activates the loss only when the corresponding sentence classification label is 1, since we do not want to back-propagate sequence labeling loss when the corresponding sequence classification label is 0.

5.1. Fine-tuning the BERT language model

We use the $ {BERT}_{base} $ model that has been built over the BookCorpus, a dataset consisting of 11,038 unpublished books and English Wikipedia. The $ {BERT}_{base} $ was then fine-tuned with a portion of the corpus using over-sampling, to create S-BERT. The fine-tuning of the $ {BERT}_{base} $ over the sustainability corpus enforces the BERT model to be both generalized as well as focused over the domain data. A labeled document is broken into multiple smaller chunks, such that each chunk can be fed as a unit to $ {BERT}_{base} $ to create its corresponding vector. Each chunk is associated with a label that is the same as its parent document. A classification task is now defined with these chunks during which the basic BERT model is fine-tuned while training. This model is designed as a fully connected layer over the BERT base model, with softmax as the activation function. Training was done with learning rate set to $ 2\times {10}^{-5} $ using the Adam optimizer (Kingma and Ba, Reference Kingma and Ba2014). The model is fine-tuned for few epochs ( $ 3 $ - $ 4 $ ) only to avoid over-fitting. The chunk representations are saved from the [CLS] token embeddings created during the process. The fine-tuned BERT model, S-BERT, is subsequently used for document and incident recognition tasks. Figure 4 depicts a comprehensive runtime analysis of the proposed model along with all the baseline systems. The runtime evaluation is performed over a training document of 8,917 documents comprising of 222,925 sentences over an epoch of 10. All the above testing results have been obtained from the Tesla T4 GPU, with NVIDIA-SMI 460.32.03, Driver Version: 460.32.03, and CUDA Version: 11.2 hardware configurations.

Figure 4. A comprehensive running time comparison between different models.

6. Evaluation and Results

6.1. Evaluating the multi-task network for ESG knowledge extraction

Next, we present results for the multi-objective sequence labeling model for the ESG knowledge extraction task. Table 2 presents the precision, recall, and F1 scores of classifying sustainability events as awards or violations. We have obtained a highest F1 score of 0.91 with a high precision of 0.93. However, we observed that the single task BERT-CNN-BiLSTM performs better than the proposed Multi-Class network in terms of recall. A possible hypothesis behind this anomaly is the fact that given the size of the training data, incorporating the CNN-BiLSTM layer is causing the models to overfit and thus affecting the recall value. This hypothesis is further supported by the fact that the recalls are consistent throughout all the models having the CNN-BiLSTM layer except the case where the layers are removed in the single-task model.

Table 2. Results reporting accuracy of classifying sustainability events as awards or violations

Table 3 presents the accuracy of subsequent labeling of word sequences within a sentence by their respective categories—TO, L, I, V, P, A—as described in Section 4. For both the cases, the performance of the proposed multi-objective architecture has been compared with a number of baseline state-of-the-art models designed with single objective functions. It was observed that for almost all the categories, the Multi-task BERT-CNN-BiLSTM model significantly outperforms the baseline models. For example, in the Target Organization class, it was found that the Multi-task BERT-CNN-BiLSTM model significantly reduces the false negative score and achieves a high true positive score thereby achieving a high precision and recall. In general, an F-measure of 0.89 with a precision of 0.87 and recall of 0.92 was achieved. For the class Incident, F-measure of 0.82 with a precision of 0.79 and recall of 0.85 was achieved. A high F1 score of 0.90 was achieved for Penalty. The target organizations were detected correctly 89% of the times. In the remaining cases, either wrong organizations were detected or missed out altogether. Detailed analysis reveals that for violation and incident phrases majority of the sub-sequences are detected correctly. The errors occur due to a portion of the sequence not detected correctly.

Table 3. Results reporting the accuracy of sustainability event and entity extraction task

Apart from evaluating the classifiers, we also evaluated the performance of the violation to GRI indices mapping module. We have randomly chosen a sample dataset comprising 100 data points for analysis. Two expert evaluators were engaged to perform manual verification of the mapping between GRIs and corresponding violation phrases within the dataset. The validation process yielded a commendable accuracy rate of 92.6%. However, in certain instances, the mapping process encountered challenges, leading to instances where accurate mappings were not achieved. For instance, a violation phrase “failing to provide proper lead-based paint disclosure to buyers and renters of homes built” was maximally mapped with GRI “Child labor (408)”. Similarly, another violation, involving “establish programmatic capability to better compete for federal brownfield cleanup grants” demonstrated a maximal match with GRI “Human Rights Assessment (412)”. This discrepancy in mappings may be attributed to the absence of specific GRIs in the provided list that ideally should have corresponded to these particular violations. This underscores the potential need for the inclusion of additional GRIs to comprehensively capture the diverse spectrum of violations encountered in the dataset.

7. Deriving Insights from Regulatory Reports

We now present some interesting insights about commonly occurring ESG incidents and violations obtained from the articles published between 2019 and 2022, using the model presented earlier.

Mapping incidents and violations to GRI indicators: The extracted ESG incidents and violations derived from the regulatory articles are mapped to GRI indicators introduced in Section 2. In order to map incidents and violation phrases to GRI indicators, we have utilized the GRI indicator definition to obtain better matches. The violation phrases are compared with GRI indicator definition using Universal Sentence Encoder (USE) (Cer et al., Reference Cer2018) embeddings. Since each sentence may contain multiple incidents or violations, only violation phrases are used to compute similarity and not the whole sentence. Also, GRI indicator definition provides comprehensive explanation of the indicator in text form. We have used the entire sentence to compare against the phrases. Indicator with the highest cosine similarity between USE embeddings and above a minimum similarity threshold is mapped to a violation phrase under consideration.

Table 5 presents a few sample sentences with the extracted incidents and violations marked in bold, along with their GRI mappings. The samples illustrate that the proposed model is powerful enough to mine a wide range of discriminatory incidents related to gender, age, color, etc.

Table 4. Data Statistics

Table 5. Sample violation events picked up from News articles of different categories and are mapped across GRI Standards. Note that the target organization names were masked by the token [ORGName] to maintain anonymity.

Once the incidents and violations are mapped to a fixed set of GRI indicators, these can be used for further analytics at aggregate and individual levels. A single report usually maps to multiple GRI indicators, since usually a single incident, say for example, an accident at a mining site leading to loss of life is linked to multiple lapses like unsafe workplace, inadequate training, environmental hazards, etc. This is illustrated with the following example:

Sample sentence: “The alleged violations included failure to manage and contain hazardous wastes; failure to comply with air emission limits; failure to comply with chemical accident prevention safety requirements; and failure to timely report use of certain toxic chemicals.”

GRI indicators mapped to: i) Effluents and waste (306) ii) Emissions (305) iii) Occupational health and safety (403).

Figure 5 shows trends of violations observed for GRI indicators obtained from reports published between the years 2020 and 2022. Count of each indicator over a time period is calculated from the number of documents that contain it. A steep rise can be seen in 2022 for violations in the categories of corruption (205), biodiversity (304), discrimination in workplace (406), diversity and equal opportunity (405), public policy (415), and customer healthcare (416). Figure 5 also shows rising trends for environmental indicators like biodiversity (304), water resources (303), etc. Violations related to economy like indirect economic impacts (308) and socio-economic conditions (414) are observed only in the year 2022. It is observed that “lack of adequate training” consistently co-occurs with a large number of incidents across various categories. Clearly that is an area of improvement for all organizations.

Figure 5. Violation trends by volume for 28 GRI indicators (2020–2022).

Figure 6 shows the distribution of the sectors across which discrimination (406) was observed, most of which were in Healthcare. Manual examination of these discriminatory violations revealed that these could be further grouped into a few categories. We used a key-phrase-based approach to group them under the following categories: sexual, racial, religious, pregnancy, retaliation, national origin, age, and disability. These key-phrases are primarily motivated from the EEOC Prohibited Employment Policies/Practices guidelinesFootnote ¹². Figure 7 shows the distribution of incidents reported in these areas from 2020 to 2022. We observe that the highest number of discriminatory violations reported were related to incidents involving disability, gender, and retaliations.

Figure 6. Distribution of Incidents of discrimination (GRI 406) across organizational sectors for the Year 2020.

Figure 7. Distribution (in %) of discrimination violations across years.

Interestingly, Figure 5 shows no significant change corresponding to occupational health and safety (403) incidents observed over the last three years. We believe that this can be attributed to the fact that occupational health and safety has been actively monitored globally for the last five years, which has helped in the reduction of violations in this area. Awareness about environment, economic impact, and some other social factors are fairly recent phenomena. These results clearly establish that sustainability monitoring and regulatory reporting definitely help in bringing down the number of incidents. Aggregate analysis like ours, can also help other organizations to become cautious. In the next section, we show how these insights can be clustered to obtain intelligence at a higher level of granularity.

7.1. Clustering incident and violation phrases—Obtaining another view

As is obvious with text reports, there are many incidents which are not exactly identical but similar or related. While some of these incidents could be mapped easily to GRI indicators, the mapping was not possible for some violation and incident phrases. This could be either due to a conceptual gap between the definition of the indicator and the incident or due to the fact that certain phrases were not covered by any GRI. Hence, to derive additional insights, the incidents and violations extracted from each set of articles belonging to the E, S, and G categories were clustered separately. For each category, embeddings for the sequences labeled as violation and incident by the model were created using the USE (Cer et al., Reference Cer2018). These vectors were then clustered using the K-means clustering algorithm (Aggarwal and Zhai, Reference Aggarwal and Zhai2012), with cosine similarity as the underlying distance measure. We have used the popular Elbow method (Joshi and Nalwade, Reference Joshi and Nalwade2013) for selecting the optimal number of clusters, k, for each set.

Figure 8 depicts the different clusters obtained from environmental violations mined from the corpus, along with their percentage occurrences. The clusters reveal that many organizations are operating without the necessary certifications required to ensure clean air and clean water at their premises. Results indicated that since certification requirements are region-specific, the violations differed a lot in their descriptions. These could not be mapped to the relevant GRI indicator, which in most cases is defined by very high-level specifications only. Figure 9 depicts percentages for social violations encountered in the corpus. It can be again observed that “lack of training” co-occurs with most of the incidents, indicating that this is the most frequent violation leading to unpleasant workplace incidents. It should be noted that the violations are not mutually exclusive.

Figure 8. Distribution of violation clusters in environmental sector.

Figure 9. Distribution of violation clusters in occupational health and safety sector.

8. Related Work

The rising awareness about sustainability practices among consumers, investors, and regulators is forcing organizations to pay attention toward ESG factors, as these are known to affect financial outcomes(Tarmuji et al., Reference Tarmuji, Maelah and Tarmuji2016). A number of researchers around the world have reported results on how ESG factors are likely to affect financial performance of organizations. One of the earliest works in the area, reported in (Shahi et al., Reference Shahi, Issac and Modapothala2014), presented an automated solution for scoring corporate sustainability data reported. A variety of classifiers including decision trees and neural networks were used to predict the sustainability score. Other notable works reported in the area of impact assessment are (Zhao et al., Reference Zhao2018; Yoon et al., Reference Yoon, Lee and Byun2018; KPMG, 2017; Velte, Reference Velte2017; Sultana et al., Reference Sultana, Zulkifli and Zainal2018). We are not going into the depths of the analysis procedure, since our focus is on gathering relevant information about sustainability incidents and not on the analysis process itself.

Being a fairly new area, there is very little work in the area of sustainability information extraction from text documents. In a more recent work that is closer to the proposed work, Raman et. al (Raman et al., Reference Raman, Bang and Nourbakhsh2020) proposed a transfer learning framework for profiling ESG-related language in business corpora and detect historical trends in ESG discussions by analyzing the transcripts of corporate earning calls using contextual embeddings. The authors developed a classification model that categorizes the relevance of a text sentence to ESG using fine-tuned language models built on small corporate sustainability reports dataset and evaluated their results on the Earnings Call Transcript dataset. Their analysis indicates that in the last 5 years, nearly 15% of the discussions during earning calls pertained to ESG, implying that ESG factors are integral to business strategy. A similar effort was reported in (Goel et al., Reference Goel, Jain, Verma, Dey and Paliwal2020), in which a methodology to automatically detect, retrieve, and score relevant information about ESG indicators from a company’s sustainability reports was presented. The relevant information was mined from the textual content using conceptual similarity of the sentences to the indicators, as well as quantitative information present in them. The above-mentioned works focused on information extraction about reported parameters with the explicit aim of reducing the manual effort required to read and extract information from reports. In (Lee and Kim, Reference Lee and Kim2023), authors proposed a 4-class BERT-based classifier that can discriminate ESG information. The training data was manually constructed. Authors also conducted three application experiments to verify the usability of their proposed model. First, in the cross-sectoral experiment, the results confirmed that the ESG classifier showed a significant performance accuracy for sectors not included in the training data. Second, in multi-source adaptation experiment involving extracting ESG-related information from multi-source text data, the performance was qualitatively verified. Finally, in the data augmentation experiment, the performance was verified on an ESG dataset that was additionally constructed through the pseudo-labeling technique.

Efforts toward analyzing publicly available data for detecting ESG factors were reported in (Nugent et al., Reference Nugent, Stelea and Leidner2020), where the authors have introduced a domain-specific BERT language model which has been further pre-trained using a corpus of financial and business News articles. They have advocated that fine-tuning their proposed model for multi-class ESG controversy classification task results in an improvement in performance compared to the general domain BERT model. They have also proposed the use of back translation approaches to deal with limited dataset problem. The classifier works at the document level and does not extract the actual incidents. Moreover, they have also limited their scope to Reuters News only which has specific style guidelines and sacrificing stylistic diversity by using only a single source whereas we have collected the data from a variety of sources. A very interesting study by Guo, (Guo, Reference Guo2020), reports the use of information extracted from ESG News to predict future volatility of stock returns. The model was trained using pairs of numerical News representation and the volatility of the corresponding companies to predict the future volatility of the companies mentioned in newly arrived News. They found that investors considered ESG News as a relevant factor which significantly impacts the future return and risk of companies. ESG News is segregated from general financial News using a hand-crafted ESG vocabulary created by domain experts. No language models were used for the purpose, thus limiting the capability of the model in detecting new and unforeseen events. In (Pasch and Ehnes, Reference Pasch and Ehnes2022), authors showed a novel solution to fine-tune transformer-based models for the ESG domain. An ESG sentiment model was created by combining ESG ratings with text documents from annual reports. It has been reported that the model outperformed traditional text classifiers in predicting the ESG behavior of companies by up to 11 percentage points. The work also demonstrated practical applications of the ESG sentiment models by predicting individual sentences and by tracking ESG-related news coverage over time.

In (Murakami and Muraoka, Reference Murakami and Muraoka2022), the authors have proposed modeling sustainability report scoring sequences using an attractor network. The authors have used the Attractor Neural Network (ANN) model to predict the sustainability reporting scores within the GRI for a set of global companies while literature (Pasch and Ehnes, Reference Pasch and Ehnes2022) presents sustainability reporting based on GRI standards within an organization in Romania. Here, the authors analyzed the content of standardizing sustainability reporting in Romania and the content of sustainability reports. They studied the occurrence of GRI indexes specified in the company’s sustainability reports. On the other hand, we are mapping violation incidents or phrases with GRI indicators to derive additional insights.

Recently, there is a growing interest in applying multi-task networks to representation learning (Liu et al., Reference Liu, He, Chen and Gao2019; Collobert et al., Reference Collobert2011; Liu et al., Reference Liu2015; Luong et al., Reference Luong, Le, Sutskever, Vinyals and Kaiser2015; Xu et al., Reference Xu, Liu, Shen, Liu and Gao2018; Wan et al., Reference Wan2021). This is primarily due to the fact that multi-task learning frameworks provide an effective way of leveraging supervised data from many related tasks(Liu et al., Reference Liu, He, Chen and Gao2019). We argue that multi-task learning and pre-trained language models are complementary techniques that can be combined to improve the learning of text representations to boost the performance of various language processing tasks. Although, a lot of recent attempts have been made in applying deep multi-task learning frameworks for information extraction (Xu et al., Reference Xu, Liu, Shen, Liu and Gao2018; Wan et al., Reference Wan2021), however, to the best of our knowledge no such model has been employed to extract sustainability insights from the incoming stream of articles. The proposed framework can help in gathering large-scale insights about sustainability practices in general, as well as provide real-time knowledge for sustainability scoring.