We introduce the FinCausal 2023 Shared Task on Causality Detection in Financial Documents and the corresponding FinCausal dataset. This paper also provides insights into the participating systems and their outcomes. The primary objective of this task is to identify whether an object, event or sequence of events can be considered the cause of a preceding event (the effect). This year, we presented two subtasks, one in English and another in Spanish. In both subtasks, participants were tasked with pinpointing, within causal sentences, the elements that pertained to the cause and those that related to the effect. We received system runs from five teams for the English subtask and three teams for the Spanish subtask. FinCausal 2023 is affiliated with the 5th Financial Narrative Processing Workshop (FNP 2023), hosted at IEEE BigData 2023 in Sorrento, Italy.

With the advent and popularity of big data mining and huge text analysis in modern times, automated text summarization became prominent for extracting and retrieving important information from documents. This research investigates aspects of automatic text summarization from the perspectives of single and multiple documents. Summarization is a task of condensing huge text articles into short, summarized versions. The text is reduced in size for summarization purpose but preserving key vital information and retaining the meaning of the original document. This study presents the Latent Dirichlet Allocation (LDA) approach used to perform topic modelling from summarised medical science journal articles with topics related to genes and diseases. In this study, PyLDAvis web-based interactive visualization tool was used to visualise the selected topics. The visualisation provides an overarching view of the main topics while allowing and attributing deep meaning to the prevalence individual topic. This study presents a novel approach to summarization of single and multiple documents. The results suggest the terms ranked purely by considering their probability of the topic prevalence within the processed document using extractive summarization technique. PyLDAvis visualization describes the flexibility of exploring the terms of the topics’ association to the fitted LDA model. The topic modelling result shows prevalence within topics 1 and 2. This association reveals that there is similarity between the terms in topic 1 and 2 in this study. The efficacy of the LDA and the extractive summarization methods were measured using Latent Semantic Analysis (LSA) and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) metrics to evaluate the reliability and validity of the model.

This study uses Natural Language Processing and Machine Learning techniques to understand the effect of the COVID-19 pandemic on mental wellbeing. We considered different user groups and locations in the USA to analyze the influence contrasting social factors, such as political stance, had on wellbeing. We measured the mental wellbeing of the social media users through understanding negative sentiment and mental health topic discussion in Twitter posts added by users from the top 10 Democrat and top 10 Republican cities in the USA. To measure the topic discussion, we used a mental health keyword list and developed machine learning models to classify the topic of a tweet. The primary findings include the similarity of the effect the pandemic had on Republican and Democrat cities when considering a timeline of tweets, whilst an increase in ‘Anxiety’ was discussed across different user groups and cities. Enforcement strategies had an influence on mental wellbeing with results differing for Republican and Democrat cities. An accurate text classifier was developed and used to categorize tweets into different mental health topics. The results showed how medical and unemployed users discussed topics like ‘anxiety’ and ‘depression’ more than a control set of users. The best machine learning model was developed using a Decision Tree algorithm which achieved an accuracy of 87% on unseen data.

Natural Language Processing is increasingly being applied to analyse the text of many different types of financial documents. For many tasks, it has been shown that standard language models and tools need to be adapted to the financial domain in order to properly represent domain specific vocabulary, styles and meanings. Previous work has almost exclusively focused on English financial text, so in this paper we describe the creation of novel financial word embeddings for three languages: English, French and Arabic. In order to evaluate the effectiveness of the embeddings, we started by evaluating the English embeddings on a sentiment analysis classification task using the existing FinancialPhrase dataset and show improved performance over a standard GloVe based model using convolutional neural networks.