1 of 7

Harsh Anuj

Data Scientist

Methods Advisory Function

June 5th, 2024

Data science and AI for evaluations @ IEG

2 of 7

Efficiency

Manage and analyze data in a semi-automatic way

E.g., semi-automatic portfolio identification and analysis

Breadth

Enhance breadth of evaluations via the expansion of types of evaluative inquiry

E.g., geospatial impact evaluation using imagery data as proxies

Validity

Enhance the validity of findings via the application of innovative techniques

E.g., ML extraction and classification can outperform manual coding

Potential benefits

Potential

3 of 7

RAP 22: NLP & sentiment analysis to identify factors of success –failure

Mozambique CPE: Geospatial analysis to assess relevance of targeting

RAP 21: NLP to classify projects objectives and indicators outcome-level

Morocco CPE: Supervised classification of satellite images to assess climate resilience

Undernutrition evaluation: NLP to classify text based on theory of change

Tanzania CPE: Feature extraction to assess impact of Rapid Bus Transit

Doing Business: NLP to identify complex portfolio and sources for structured literature review

Blue Economy: Remote sensing and computer vision object detection to assess health of mangrove and coral

Ukraine CPE: Sentiment analysis of online news media articles to assess influence of policy dialogue

Urban Spatial Growth: Image segmentation of urban landscapes to assess economic and spatial impacts

Text analytics

Geospatial analysis

4 of 7

Data Science

Text

Analytics

Text

Mining

Supervised

Text

Classification

Geospatial

Analysis

NLP

Sentiment

Analysis

Topic

modeling

Keyword

Search

Web

scraping

Similarity

Analysis

Keyword Extraction

Large

Language

Models (LLMs)

Classification

Synthesis

Summarization

Code

Writing

Sentiment

Analysis

Remote

Sensing

Feature

Extraction

from Photos

Supervised

Classification

of Satellite

Images

Targeting

Analysis

Photo

Grammetry

Computer

Vision

Application

Data science applications: two types

Term Frequency

5 of 7

  • Speed
  • Breadth
  • Quality
  • Insight
  • Ethics
  • Biases
  • Transparency
  • Safety and security
  • Truthfulness

Promises of LLMs for evaluation practice

Perils of LLMs for evaluation practice

Promises and perils of LLMs

6 of 7

Recap of experiments

7 of 7

Thank You!

http://ieg.worldbankgroup.org

hanuj@worldbank.org