Intelligent Concept Paper—Problem Tree Generator

This information was sourced from "Artificial Intelligence in Action: Selected ADB Initiatives in Asia and the Pacific," Asian Development Bank, 2024, https://www.adb.org/sites/default/files/publication/963831/artificial-intelligence-action-asia-pacific.pdf. Licensed under Creative Commons Attribution 3.0 IGO (CC BY 3.0 IGO).

Writing a concept paper is the first step ADB takes to support DMCs. To do so, writers need to manually sift through various internal documents and online materials; as a result, a single concept paper can take between several weeks and a month to prepare. Problem trees, which show the causes and effects of an issue that ADB seeks to help resolve, can alone take 3 to 5 days to complete. Even then, concept papers are unlikely to be comprehensive given limitations in finding available references. Thus, ADB sought to use technology to make writing concept papers faster and more efficient. This initiative explored using AI to automate the generation and population of problem trees based on entered search queries.

The Information Technology Department (ITD) engaged Neural Mechanics Inc. to develop a minimum viable product (MVP) of the tool that could be used to automatically generate problem trees. AI was leveraged to increase efficiency and reduce inconsistencies in preparing concept papers, particularly problem trees, because it was able to process large amounts of data. The tool, which would come to be known as Intelligent Concept Paper (ICP), was trained on similarity matching, enabling it to filter sentences that are semantically similar to texts from previous problem trees. A neural network model, which is a subset of machine learning that can identify relationships based on attributed weights, was then used to identify causal event pairs from the filtered sentences.

The sources of information were documents from ADB’s East Asia Department and the World Bank, as well as Xinhua news articles. These were first converted to a readable format to enable the designed web scraper to crawl through these documents. These were annotated to make searching easier for the engine. The annotated materials were then stored in the cloud. An app was developed to recover relevant information based on the keywords entered by users.

Users entered their own problem statements or used/edited the suggestions that were auto-generated based on their entered keywords. From there, they had the option to use the suggested causes and effects or write their own. It was possible to add multiple responses for both causes and effects, as well as sub-causes and sub-effects, if users wanted to provide more specific information.

A click-and-drop feature was also incorporated into ICP’s design to build the problem tree right on the platform. Users were able to share this with others for their review and input.

These features allowed ICP to generate results within seconds, significantly cutting down the time needed to research information and produce the problem tree diagram. The tool also allowed various users to collaborate on the same problem tree and document, making work more efficient.

Work on ICP began in 2019 and was completed in April 2020. While this tool is no longer in use, the initiative showed that it was possible to develop an AI-driven platform that built problem trees. The tool marked various firsts for ADB: the first homegrown AI for ADB operations, the first one developed that leveraged the wealth of knowledge from previous works of ADB within the development landscape, and the first to harness both technology and human inputs to scan through key materials to identify problems in ADB’s DMCs.

ADB-wide consultations and/or focus group discussions may be conducted to explore how lessons from the tool development can be used in the future, including in writing other sections of a concept paper such as the Design and Monitoring Frameworks.