Identifying Human Needs through Social Media: A study on Indian cities during COVID-19

Sunny Rai, Rohan Joseph, Prakruti Singh Thakur, Mohammed Abdul KhaliqNAACL 2022

TL;DR

This paper introduces a minimally supervised approach to identify human needs expressed in tweets during the COVID-19 pandemic in India. By leveraging RoBERTa and a frustration-aggression theory framework, they can detect unmet needs, offering a valuable tool for crisis response and policy-making.

What Problem Does This Paper Address?

The COVID-19 pandemic, with its widespread lockdowns and societal disruptions, created unprecedented challenges and highlighted a critical need to understand and respond to the evolving needs of the population. Traditional methods of gathering this information are often slow and resource-intensive. Social media platforms like Twitter have emerged as a powerful, real-time channel for people to express their feelings and needs. However, the sheer volume of data makes it difficult to systematically extract and analyze these expressed needs, especially in a timely manner during a crisis. This paper aims to bridge this gap by developing an automated method to identify and categorize human needs from tweets, specifically focusing on the Indian context during the pandemic.

Methods

The authors propose a multi-stage approach:

  1. Topic Extraction: They employ Latent Dirichlet Allocation (LDA), a standard topic modeling technique, to identify key themes within the collected tweets. This step helps to distill the vast amount of text into manageable topics.
  2. Manual Topic Labeling: A team of human annotators manually maps these extracted topics to Maslow's Hierarchy of Needs (MHoN) framework, categorizing them into five levels: Physiological, Safety, Love and Belonging, Esteem, and Self-Actualization. This step is crucial for providing a semantic understanding of the topics.
  3. Tweet-to-Need Mapping: Each tweet is then assigned a need level based on its dominant topic, effectively categorizing the expressed need.
  4. Unmet Need Detection: To identify unmet needs, they draw inspiration from the Frustration-Aggression theory. They fine-tune a pre-trained RoBERTa model to classify tweets as either expressing frustration (indicating an unmet need) or not. This is achieved by training the model on tweets labeled with hashtags like #frustrated versus those indicating satisfaction.

Datasets

  • Tweet Collection: The study utilizes a dataset of Indian tweets collected from December 2019 to June 2021, covering a baseline period before COVID-19, the lockdown phase, the first wave, and the second wave of the pandemic. Tweets were collected from metropolitan Indian cities using specific geographical parameters.
  • Training Data for Frustration Detection: A dataset of 13,970 tweets was created for training the RoBERTa model, with an equal split between tweets expressing frustration and those indicating satisfaction.

Key Insights and Ideas

  • Maslow's Hierarchy in a Crisis: The study demonstrates the applicability of Maslow's Hierarchy of Needs in understanding public sentiment during a large-scale crisis like COVID-19, showing how different needs become prominent at various stages.
  • Frustration as a Proxy for Unmet Needs: The core idea that frustration, as described by the Frustration-Aggression theory, can serve as a reliable indicator of unmet needs is a key contribution. This allows for a more nuanced understanding beyond just identifying expressed needs.
  • Dynamic Nature of Needs: The research highlights how the prominence of different needs shifts throughout a crisis. Basic needs (physiological and safety) often surge during lockdowns, while advanced needs (esteem, self-actualization) may follow with a delay.
  • Subjectivity in Frustration: The paper acknowledges the inherent subjectivity in labeling frustration, which impacts the evaluation of the model's performance.

Key Results

  • Tweet Volume: Over 1.4 million unique tweets were collected for the study. After filtering, over 1.1 million tweets were mapped to expressed needs.
  • Need Distribution: Physiological and Safety needs were the most frequently expressed during the lockdown period.
  • Frustration Detection Accuracy: The RoBERTa model achieved an accuracy of 92.2% on the test set for detecting frustration.
  • Frustration by Need Level: The study found that a significant percentage of tweets related to physiological (77.36%) and safety (77.5%) needs expressed frustration. Advanced needs also showed considerable frustration rates (e.g., 54.13% for love and belonging).
  • Inter-Annotator Agreement: The inter-annotator agreement for frustration labeling was 0.638, indicating moderate agreement and highlighting the subjective nature of the task.

Limitations and Future Work

  • Data Filtering: The study's focus on geo-tagged tweets within India led to a reduction in the dataset size, potentially limiting the scope.
  • Dominant Topic Assumption: The assumption that the dominant topic of a tweet accurately reflects its need type, while efficient, may have led to some mislabeling.
  • Language Specificity: The current model is trained on English tweets.
  • Future Work: The authors suggest extending the model to detect needs in regional languages, incorporating theories better suited for advanced psychological needs, and further exploring the predictability of needs during crises. They also propose that this approach can be valuable for public agencies to identify and prioritize emerging human needs in real-time.

Summary updated

BibTeX
@inproceedings{rai2022identifying,
  author = {Sunny Rai and Rohan Joseph and Prakruti Singh Thakur and Mohammed Abdul Khaliq},
  title = {{Identifying Human Needs through Social Media: A study on Indian cities during COVID-19}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure