Automating law text labeling with LLMs

This project builds an end-to-end pipeline (GitHub repo) for automating the labeling of self-expression laws using a combination of human annotations, large language models, and fine-tuned transformer architectures. I began with human-coded laws containing provision keys (i.e., a general statement of a legal rule) and deontic indicators (i.e., whether the legal rule was explicitly included or excluded in the law), then used GPT-5 to extract the relevant legal text and reconstruct a structured training dataset. To address class imbalance, I generated targeted synthetic observations, producing a balanced dataset suitable for training.

I then fine-tuned a LEGAL-BERT model on Azure ML using LoRA adapters and a multi-head classifier that jointly predicts provision type and deontic status. For evaluation, test laws were split into clause-level inputs with GPT-5-mini under strict and relaxed relevance settings, passed through the model, and compared against human labels. Performance was assessed using F1, precision, and recall overall and by provision key, yielding a reproducible workflow for scaling legal text classification across hundreds of variables. While still a work in the progress, the model outperformed a zero-shot pipeline using ChatGPT 4o. 

Program evaluation using sentiment analysis

For a client, I conducted an evaluation of one of their high-profile programs using a wide range of data science methods. The project involved wrangling data from a variety of sources, and it posed several challenges related to data quality and compatibility, which necessitated extensive cleaning, validation, and various merging techniques. I also had to generate several control variables for the analyses, which involved using GIS methods to generate legislative district-level covariates from county- and precinct-level data (code).

The analysis was divided into two parts. The first part involved building a predictive model of joining the program, and I used logistic regression, random forest, and SHAP values to evaluate the size, direction, and significance of numerous features. The analysis showed which variables mattered most for predicting the outcome, as well as helping identify several non-linear relationships.

The second part focused on modeling individual behavior on social media, using a sample of thousands of individuals and roughly 1.15 million social media posts. I used off-the-shelf sentiment analysis modules (VADER) to capture general tone, but I also built a custom dictionary with context markers to capture opposition to or support for specific values.

The modeling effort consisted of a series of methods, including OLS with fixed effects, clustered balanced sampling to handle class imbalance, weighted regression, and bootstrapping. 

Predicting civil war interventions

In an ongoing project, I am building a predictive model of interventions into civil war by third-party states. I am using the UCDP External Support Dataset, which captures interventions at a high degree of granularity, to generate a categorical outcome at the civil war-potential intervener level: no intervention, government support, and rebel support. I include every other state in the system as a potential intervener in a civil war, yielding a total of 34,552 dyadic observations.
 
There are several challenges in building such a predictive model. First, because most countries do not intervene in a given civil war (even if most civil wars do experience some form of intervention), the outcome variable has a severe class imbalance problem. There are 872 government-sided interventions, 256 rebel-sided interventions, and 33,432 observations with no intervention in the dataset. Second, there are no existing predictive models of civil war intervention that I am aware of, so there is little guidance in the literature about which machine-learning techniques to use or which features are more or less important. I therefore take an iterative and exploratory approach to model building (you can find all the code for this project in my Github repo).

I start by estimating several inferential models using multinomial logistic regression to identify an appropriate benchmark model. Then, I use Box’s loop approach to iteratively build a set of predictive models using different algorithms suited to multiclass models, including a wide range of features. I use cross validation to tune hyperparameters and evaluate and compare model performance, and after each iteration I use a range of diagnostic tools and domain expertise to optimize the models.
 
To build a benchmark model, I first construct novel predictors of intervention. While existing studies highlight numerous factors that influence the decision to intervene (e.g., co-ethnic ties), most are context-specific features. In general, intervention should depend on the political preferences of the third party, and so whether it decides to intervene and on which side is a function of how far apart the two states are. The further apart the two states are, the more likely the third party is to intervene on the side of the rebels instead of the government, and vice-versa.

I construct a set of variables of political or policy distance by using different unit-level variables as inputs to generate the absolute difference between the two states’ score (input variable in parenthesis):

• P2dist (Polity2) measures the difference in states’ level of democratization.
• Polydist (Polyarchy index from V-Dem) measures the difference in electoral democracy achieved in the two states.
• Libdist (Liberal index from V-Dem) captures the difference in liberal ideology in the two states.
• Libdemdist (Liberal democracy index from V-Dem) combines polyarchy and liberal measures to capture the difference in the level of liberal democracy achieved in the two states.
• Opendist (Trade openness from the Chinn-Ito index of financial openness) captures the difference in the states’ integration in the world economy.
• Kappavv is an existing dyadic measure of the two states’ foreign policy similarity, based on observed dissimilarity of ties while accounting for the distribution of foreign policy ties in the system and individual states’ propensity to form ties (Hage 2011).

 
I estimate a set of models (with similar sets of covariates) using multinomial logit and compare their performance. I pick the Liberal Democracy model because it performs the best according to the Akaif-Information Criteria and its results fit my theoretical expectations.

I take an iterative approach to building the predictive model. I start out with a broad set of features and perform pre-processing (including train-test split, normalization, and knn-imputation for missing data), before using recursive feature elimination to assess whether any features are unnecessary. Once that is done, I move on to the Box’s loop approach.
 
For each iteration, I tune and train three separate models using random forest, support vector machine, and XGBoost DART algorithms using five-fold cross validation. I compare their performances based on the F1 metric, and then use diagnostic tools (see below) to assess commonalities across cases where the models perform poorly. After evaluating the models, I move on to the next iteration to optimize them. I use my domain expertise to identify additional features, performance metrics to re-tune hyperparameters, and other tools (e.g., elastic net estimation) to eliminate unimportant features.

All versions of the machine learning model perform significantly better than the benchmark model (Mean F1 score of 0.37), and model performance improves across iterations (see below). The XGBoost DART model outperforms the random forest and support vector machine models on the key performance metric, and its performance is more stable.

Tuning XGBoost DART models can be challenging because of the number of hyperparameters, so I plot average model performance across hyperparameter values to assess how sensitive the model is. The below plot shows that model performance varies substantially by maximum depth of tree and learning rate.

The predictive models offer insights into which features matter for civil war interventions. As the variable importance plots below show, a lot of variables matter, making the models complex. The geographic distance between states (mindist) is by far the most important predictor of intervention, followed by other dyadic and monadic indicators such as the population size in the civil war country (wbpopest1), whether the third party is already supporting the government in another civil war (otherconfgov), and the two states’ foreign policy similarity (kappavv).

Civil war interventions and retaliations

For my dissertation, I supervised a revolving team of research assistants and interns that collected data on retaliations by civil war governments against external rebel supporters across 209 separate interventions worldwide over 35 years (Civil War Expansion Dataset, or CWED for short). We coded six features of war expansion, including direct retaliation (use of military force against the third party in their territory), proxy retaliation (supporting other rebel groups against the third party), and threat of retaliation. You can find the original data, replication code for the various graphics on this page, and high-resolution images on my GitHub repo.

This map shows the distribution of intervention and retaliation in the dataset, distinguishing between countries who experienced civil war, intervened in other civil wars, and experienced retaliation.

Shiny app mapping civil war expansions

The below app generates civil war-specific maps for conflicts (1975-2009) that experienced rebel-sided intervention. The map shows the civil war country and the third-party interveners, distinguishing between those that experienced retaliation and those that did not. Code for the data and app can be found here. Click here for a fullscreen version of the app.

Visualizing civil war expansions

CWED shows that retaliation is quite common in civil wars, with high variance across interventions and conflicts. Proxy retaliation was the most common response to individual interventions, but direct retaliation was the most common form of retaliation in conflicts.

The majority of interventions did not provoke retaliation, but civil war governments often combined multiple forms of retaliation against specific interveners and across the conflict.

Looking at the frequency of retaliation combination, CWED shows that proxy retaliation by itself was the most common response to intervention. On the conflict level, however, the most common response consisted of direct, proxy, and threat of retaliation against intervener(s).