A Tale of Human and Machine Collaboration: Using AI language models to support the Türkiye and Syria Response
By Ximena Contla, Nayid Orozco, and ChatGPT (an AI language model created by OpenAI)
Once upon a time, in a world not too different from our own, there was a group of emergency responders who were always looking for ways to improve their response efforts. They worked tirelessly to help people in times of crisis, but they knew that there was always room for improvement. One day, a team of developers approached them with the idea to use a new tool that they believed could help them make better decisions. This tool was called AI, or artificial intelligence.
At first, the emergency responders were skeptical. They had heard of AI before, but they didn’t know much about it. But the developers explained that AI was a tool that could help them analyze large amounts of data and identify patterns that might not be visible to the human eye. They also explained that AI could help automate certain tasks, freeing up emergency responders to focus on other critical tasks.
Emergency responders were intrigued, and they agreed to give AI a try. They began to use AI to analyze data from different sources. They quickly realized that AI was a powerful tool that could help them make faster, more informed decisions during times of crisis. As time went on, the emergency responders became more and more reliant on AI. They began to see it as an essential tool in their response efforts, something that could help them do their jobs better and more efficiently.
And so, the story of AI as a tool to support human efforts for emergency response continued. With its ability to quickly process vast amounts of data and identify patterns, AI proved to be a valuable asset to emergency responders around the world. And as the technology continues to evolve, we’ll likely see even more innovative uses of AI in the humanitarian sector in the years to come.
ChatGPT, an AI language model created by OpenAI (prompted and edited by Ximena Contla)
Two of the strongest earthquakes in the region in more than 100 years, of 7.7 and 7.8 magnitudes, occurred on Monday, 6 February, in South East Türkiye and North Western Syria. During the days following this earthquake, a large amount of information was generated about the situation. News, reports, announcements, and assessments were flooding the resources portals of the humanitarian community for everyone to have the right information at the right time. But processing that amount of data proves challenging for individuals that have to spend hours and days going through all those documents to make sense of the information.
Currently, one of the tools to make sense of qualitative data supporting the efforts of the humanitarian community, is the Data Entry and Exploration Platform (DEEP). The DEEP proved key to supporting the analyst team in the DEEP emergency response cell (DEEP Surge Analysis Cell) in generating reports that supported the humanitarian community in the first hours of the response. As a parallel task, the DEEP team tested the NLP capacities generated by NLP developers in Data Friendly Space (DFS) and other AI language model tools as part of the response.
Artificial intelligence (AI) language models have become increasingly sophisticated in recent years, thanks to advances in machine learning algorithms and natural language processing techniques. These AI systems are designed to process large amounts of data, identify patterns, and generate human-like responses to a wide range of questions and tasks in real-time. The one AI language model that is currently debated in different fora is ChatGPT, created by OpenAI. ChatGPT is a digital program designed to assist users with various tasks and answer questions using natural language processing.
By leveraging the power of ChatGPT, the DEEP NLP team worked together to bring to life the first AI-generated basic “3W”, which we called “Supply Tracker”. In the first stage, the tool uses the Davinci model from Open AI to analyze articles from Reliefweb and DEEP. A powerful language model capable of analyzing large volumes of unstructured data and extracting meaningful insights. What sets Davicini apart from other models, such as Curie and Babbage, is its ability to understand the context of data and generate insights tailored to specific use cases. As part of further development, Davinci can be fine-tuned and trained with new data, making it an ideal tool for tracking and monitoring aid delivery over time. The tool extracts relevant information related to who is doing what, where (3W) in both Turkey and Syria, and it creates a structured table with all the information from these articles in the following form:
- Who’s delivering/sending the aid: It identifies the organization or agency delivering or sending aid.
- Humanitarian sector or cluster: It identifies the sector(s) in which the aid is being delivered, such as health, water and sanitation, food, shelter, etc.
- Type of assistance: It identifies the type of assistance provided, such as medical supplies, shelter, or technical assistance.
- Location: Where the aid is delivered, such as a specific country or region.
- Current status: The tool identifies the current status of the aid, whether it has arrived or is still pledged.
We followed the process described below for creating the first version of the Supply Tracker, embedded in this notebook. :
- We looked at the documents analysts added to the Türkyie- Syria Eartquake 2023 project in DEEP.
- Using a model from DEEP, we extracted the text from the latest articles of Reliefweb on the topic, provided by the connector in DEEP, as well as other documents added to the platform. These extracted texts generated a table of leads of interest.
- Then, we set up an OpenAI account to start using the models. OpenAI will provide an API Key that will be required to process the documents.
- Using a Python instance, we installed the necessary libraries: the OpenAI library, which provides access to pre-trained models like ChatGPT and pandas library for handling the data in batch mode.
- We then loaded the table of leads of interest.
- We made sure to preprocess the data to be ready for analysis. In this case, preprocessing may involve cleaning up the text, removing irrelevant information, and identifying key entities and concepts. This step is critical to ensuring that the AI can accurately structure the data.
- Feed the text data into the ChatGPT model to generate predictions using the Davinci Model.
- Store the results of the analysis in a pandas DataFrame.
As with any machine learning tool, the Supply Tracker has limitations and assumptions. First, the DEEP NLP team in DFS did not train the base model. Therefore, we are committed to completing the work done by ChatGPT using our models to fine-tune the results to be even more accurate for the humanitarian sector. Second, the tool assumes that the information related to aid delivery is clearly stated in the input data. Therefore, if the information is implicit or unclear, the tool may not extract it accurately. It is important to note that the Supply Tracker, in this first testing phase, cannot guarantee the accuracy or completeness of the information provided. Consequently, we advise users of the tool to cross-check the information provided with other sources and exercise their judgment when making decisions based on the data.
As we keep working on the refinement of the Supply Tracker, the next steps consist of the following:
- Using the models developed by the DEEP NLP team, which are trained in categorizing data according to the humanitarian sector to refine the categories extracted and filter by only those of interest for the response.
- Using the geotagging models developed by the DEEP NLP team and ISI Foundation to get more accurate information about the places the aid is delivered.
We wanted to share with you some of the recent updates on AI in our day-to-day to support humanitarian actors in their time-consuming tasks related to the analysis of qualitative data. [I]t’s important to note that AI is not a replacement for human responders. Rather, it’s a tool that can support and enhance human efforts in emergency response. By using AI to analyze data, identify patterns, and automate certain tasks, emergency responders can make faster, more informed decisions during times of crisis, potentially saving lives and reducing damage. As the technology continues to evolve, we can expect to see even more innovative uses of AI in emergency response efforts in the future.
NOTE: All that is written in bold italics comes directly from ChatGPT, an AI language model created by OpenAI with prompts by Ximena Contla
Please see below the prompts by Ximena.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
If you have questions or would like to contact us, send an email to: nlp@thedeep.io
Ximena Contla is the DEEP NLP Manager and Nayid Orozco is the DEEP Product Lead at Data Friendly Space (DFS).
