Detection of online intimate partner violence in Nepali using artificial intelligence

#16DaysofActivism 2021 Blog Series

Written by Shreyasha Paudel and Sagun Shakya

Intimate partner violence (IPV) is understood as the prevalence of violent and abusive behaviour in an intimate relationship, manifested in physical, sexual and emotional forms. With increasing use of smartphones and social media, IPV can also occur among young unmarried couples through online interactions. However, most IPV related studies in Nepal have focused on matrimonial relationships in offline contexts and often do not consider digital and online forms of IPV[1],[2].

This SVRI supported research project by ChildSafeNet[3] is a comprehensive study to understand the magnitude and pattern of online IPV among young people in Nepal. Along with conventional qualitative and quantitative data-collection, this project includes the development of an Artificial Intelligence (AI) tool to assist researchers to categorize online content as IPV, non-IPV, or unknown. The research and development of the AI tool is led by NAAMII[4] in collaboration with ChildSafeNet.

In this blog, we focus on the AI tool, techniques and its use in IPV classification along with a discussion on ethical concerns for this application.  

Artificial Intelligence (AI)

Artificial Intelligence is an umbrella term which encompasses different ways to enable machines to perform tasks normally associated with human intelligence. Some common examples of AI tools include online recommendation systems, text-prediction, biometric-detection, etc.

Many AI tools learn to perform a specific task using observational data rather than explicit programming. This technique is called Machine Learning (ML). ML requires first developing a model with carefully curated data (also known as training-dataset), and then feeding additional data (test-dataset) to the model to make predictions. In a supervised learning scheme, the data for training the AI model is given along with their target labels. Figure 2 shows an example of supervised learning. A key challenge in ML lies in the integrity of datasets. ML models are only as good as the data used and are known to make faulty generalizations and often produce unpredictable results in real-world scenarios[5].

Natural Language Processing (NLP)

NLP is a sub-field of AI that enables machines to understand and interpret human language. Fig 3 shows different uses of NLP. Most successful NLP use large datasets and do not work well in the absence of abundant textual material. This and a lack of prior research poses a problem for low-resource languages. NAAMII is developing NLP tools for low-resource languages like Nepali.

AI Tool for IPV Detection

The AI tool we are developing is a supervised ML model to categorize online Nepali texts into IPV, non-IPV and unknown classes as shown in Figure 4. As a first-of-its-kind research project, our scope is limited to classifying text-based content in Nepali language and for research purposes only.

Figure 5 shows the proposed development cycle. The ChildSafeNet team will gather and annotate the training data (I.e., carefully curated data on IPV) with input from researchers at NAAMII. This data should include context, rationale, and be sufficiently diverse. The model evaluation will be done by NAAMII with input from IPV experts at ChildSafeNet.

Ethical considerations

Ethical use of AI for a sensitive topic such as IPV requires human oversight, risk assessment and usage limitations. During data collection, we will ensure consent, data integration and safety for subjects. We are developing this tool within a research project for use as a data gathering tool. Given the limited scope of this tool, it may not work well in all online situations in Nepal. Moreover, the model will need to be regularly updated as language use evolves. Under current technology constraints, it is impossible to develop IPV detection models that are 100% accurate at all times. Our team will consider societal implications and ethical issues throughout the project and document limitations of the AI tool for future oversight.

For more information, contact

Add new comment

Filtered HTML

  • Web page addresses and e-mail addresses turn into links automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <blockquote> <code> <ul> <ol> <li> <dl> <dt> <dd>
  • Lines and paragraphs break automatically.

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.
This question is for testing whether or not you are a human visitor and to prevent automated spam submissions.
4 + 4 =
Solve this simple math problem and enter the result. E.g. for 1+3, enter 4.