/

India has committed 21 million hectares to land restoration, yet conflicts affect more than 3 million people each year and no system existed to alert policymakers as those disputes unfolded. Omdena partnered with the World Resources Institute to build an NLP pipeline scanning 65,000 news articles, categorizing conflict events, and matching each to government policies. Delivered: 98% classification accuracy, seven conflict categories, and an interactive policy recommendation tool.
Land degradation affects 3.2 billion people worldwide and costs the global economy approximately 10 percent of annual economic output each year. Dozens of countries have committed to restoring 350 million hectares of degraded land, a target that depends entirely on clear land use rights and stable community ownership. Where those conditions break down, restoration commitments remain on paper.
India sits at the center of this challenge. The country has committed 21 million hectares to land restoration, among the largest national commitments on Earth, yet land conflicts affect more than 3 million people each year. India also records the highest number of environmental land conflicts of any country, spanning disputes over farming, forests, mining, water, and wildlife.
When a conflict event emerged in a rural district, the officials responsible for resolution had no fast way to know it had occurred, no tool to identify the category of dispute, and no mechanism to surface the specific policies that applied. Monitoring happened manually, through regional teams scanning newspapers and filing reports, often with days or weeks of lag before the right decision-makers were aware.
Omdena partnered with the World Resources Institute, a Washington, D.C.-based global research organization focused on sustainable land use, food systems, and environmental policy. The client supplied the project’s core reference materials: a gold standard corpus of 200 annotated conflict articles and a database of 19 Indian government policy documents, covering topics from land acquisition and forest protection to farmers’ rights and mineral concessions.
The project team comprised 19 AI engineers and data scientists drawn from the Omdena global network. The team built a full NLP pipeline from raw news corpus to interactive visualization tool, with the client providing domain expertise and policy materials throughout the build.

All 65,000 articles and 19 policy documents were processed using SpaCy, an open-source NLP library that performed tokenization, part-of-speech tagging, dependency parsing, lemmatization, sentence boundary detection, and named entity recognition. Coreference resolution was then applied using Neuralcoref, a neural-net model that replaced all pronouns in each article with the nouns they referenced, reducing ambiguity before classification.
A logistic regression classifier was trained on a 1,600-article annotated corpus, manually labeled as conflict-positive or conflict-negative. Articles were represented using Bag-of-Words vectorization, which encodes word frequency across the vocabulary without preserving word order. The trained model classified all 65,000 articles with precision, recall, and f1-score of 98% or greater, with negligible false positives and false negatives.
Conflict-positive articles were categorized into seven land conflict topics using CorEx (Correlation Explanation), a semi-supervised topic model that accepted domain knowledge as anchor keywords for each category. The seven categories were defined to align directly with WRI’s policy database, covering the full range of dispute types from land resettlement and farming to mining, deforestation, and water rights. The trained topic model achieved 93% accuracy.

Custom named entity recognition was trained on the annotated corpus to identify actors, actions, scale, locations, and dates specific to land conflict reporting. Standard NER models recognize generic entities; conflict articles required additional recognition of domain-specific actors such as Government, Farmer, Police, and Civilian, and actions such as protest, attack, and killed. Manual annotation of the training corpus provided the data for the custom recognizer.
Conflict-positive articles were matched to relevant government policies using cosine similarity, a metric that compares the distributional vocabulary overlap between two vectorized documents rather than keyword frequency alone. This allowed the model to surface policy relevance even when the article and policy used different surface terms. Each article received a ranked list of applicable policies with cosine similarity scores drawn from the 19-document WRI database.

Before this project, no automated system existed to alert Indian policymakers when a land conflict emerged in the news, let alone surface which existing policies applied to its resolution. Monitoring happened manually, through regional teams filing reports with significant lag. This project demonstrated that NLP could close that monitoring gap at scale, processing years of news data in a single automated pipeline.
The proof of concept gave the client a concrete technical foundation for a production deployment across India’s 28 states. With the pipeline validated on three years of GDELT data at national scale, the next step would be integrating a live news feed and expanding the policy database to cover state-level legislation, rather than rebuilding the core architecture from scratch.
For a country where land conflicts affect more than 3 million people each year and block restoration of 21 million hectares, early-stage conflict detection at this accuracy level represented a meaningful operational advance. Policymakers could, for the first time, see conflicts as they emerged, categorized by type and location, with relevant policy options already surfaced alongside each event.
Standard pretrained NLP models recognize people, organizations, and locations but fail on the actors and actions that define land conflict reporting. Terms like gram sabha, resettlement order, and patta cancellation carry precise legal meaning in this context. Manual labeling of a training corpus for custom entities was time-intensive but essential. Without it, downstream policy matching would have operated on incomplete and inaccurate event data.
An unsupervised topic model would have discovered categories from data alone, but the resulting clusters did not align with the seven conflict categories that WRI’s policy database was organized around. Using CorEx with anchor keywords derived from policy language kept the topic model grounded in operational use. The 93% accuracy of the final model validated this approach over a fully data-driven alternative.
Early experiments using keyword overlap to match articles to policies produced a high rate of irrelevant recommendations. A drought article mentioning water would surface water infrastructure policies regardless of whether the conflict concerned irrigation rights or flooding. Cosine similarity over TF-IDF vectors captured the full distributional meaning of each document, producing matches that aligned with the conflict’s actual legal and policy context.
The pipeline integrated text preparation, document classification, topic modeling, entity extraction, and policy matching as an automated sequence across the full GDELT corpus. This demonstrated that NLP could replace manual conflict monitoring at scale, reducing a process that previously required analyst teams scanning newspapers to something deployable against a live news feed. The five-stage architecture is the project’s core transferable technical contribution.
At 93% accuracy, the topic model met the threshold for routing conflict articles to policy categories without human review. That figure was not achieved by unsupervised topic discovery: it required anchoring the model to categories derived directly from the client’s policy database, ensuring the model’s output aligned with the legislative structure used to respond to conflicts rather than patterns in the news data alone.
Before this tool existed, moving from a news report about a land conflict to the specific legislation relevant to its resolution required specialist knowledge of both the news corpus and the policy database. The tool closed that gap: a policymaker with no data science background could select a conflict event on a map and receive ranked policy recommendations drawn from 19 government documents within seconds.
Omdena partnered with the World Resources Institute to build an NLP proof-of-concept system for detecting and monitoring land conflict events in India and matching each event to applicable government policies. Working from 65,000 GDELT news articles and 19 policy documents, 19 AI engineers built a five-stage NLP pipeline and an interactive Plotly-based visualization tool, completed in April 2020.