Plain Language Summary (~100 words) 

In the 1800s, the media landscape was dominated by thousands of small local newspapers. Yet, despite their central role in shaping how ordinary Americans understood national events, historians have mostly studied a handful of big-city papers. This project addresses that gap by using text-analysis tools to study how the same national news stories were reworded, reframed, and editorialized as they spread from major papers into small rural ones. By tracing how language changed along the way, I aim to show that "the news" was actively reshaped by local editors rather than passively copied. The result is a data-driven map of information flow in a pre-digital media network.

Main Research Question(s) (~1-3 sentences) 

How did national news stories change in wording and framing as they were reprinted across the nineteenth-century American local press? More precisely, can we detect systematic patterns of how stories are altered, and are a paper's political leaning, its distance from the source, or its town's size predictive of how much it reframed what it reprinted?

Data and Methods (~250 words) 

So far, I have retrieved roughly 40,000 digitized newspaper articles from 1850–1875 from the Library of Congress Chronicling America corpus. Paper-level metadata (location, political affiliation, etc.) is from the Library of Congress newspaper directory. After I have calibrated my methods, I plan to scale my sample to 250,000 articles. The central technical challenge of my project is creating an automated script to detect text reuse to identify and analyze the "children" of a "parent article".

The analysis pipeline:

1. Text cleanup (spelling correction and normalization): fix OCR typos and remove regional spelling variants.
2. Text-similarity detection (character n-gram overlap with locality-sensitive hashing): find clusters of articles that are reprints of the same story despite slight alterations, without comparing every article to every other article.
3. Sequence alignment: within each cluster, measure how far each child diverges from its parent source, producing a continuous "divergence" variable that forms the basis for my analysis.
4. Framing analysis (sentiment lexicon plus a fine-tuned classifier): measure editorial tone shifts between source and reprint. A rewording is only meaningful if the child article is materially different from its parent.
5. Topic modeling (Latent Dirichlet Allocation): sort stories into types (political, economic, crime, foreign...) so that I can compare divergence within comparable categories and detect which kinds of stories were altered the most.
6. Regression modeling: for each topic, test whether divergence is predicted by political alignment, geographic distance, or circulation size.

I plan to validate each step against a hand-checked subset of 500 article pairs so that I can accurately assess precision. The pipeline uses Python (pandas, scikit-learn, Hugging Face).

Project Timeline and Deliverables (~4-6 bullets) 

• Oct–Nov: tune text reuse detection and framing analysis; hand-validate the 500-pair sample and report precision and recall.
• Dec–Jan: scale retrieval and matching to roughly 250,000 articles.
• Feb-Mar: run the pipeline on the full corpus and create visualizations of story propagation networks.
• Apr: finish thesis writing, poster, and presentation slides.
• Deliverables: senior thesis, reproducible public codebase, visualizations of story propagation networks on my personal website.

Broader Impact (~100 words) 

Historians have long argued that the nineteenth-century local press actively reshaped national narratives, but the claim rests on close reading of a few papers. This project tests it across hundreds of thousands of articles, bringing quantitative evidence to a debate that has been almost entirely qualitative. These data-driven insights on whether "reprinting" meant faithful copying or active rewriting, and what predicted the difference, help us understand the role of the press and the contributions of local media. In addition, sharing the text-reuse pipeline on GitHub lowers the technical barrier to computational historical work on large digitized-newspaper collections.

What is your role in the project? (~100 words) 

This is my independent thesis; I am the only researcher on it. I wrote the retrieval and prototype-matching code, chose the period and regional scope, and will do all modeling and interpretation. My advisor, a historian of print culture, shaped the historical framing early on—she pushed me from "how fast did news travel" toward "how did it change"—and she helps judge whether flagged matches are genuine reprints, but does not write code. I developed the framing-classifier design in conversation with a Linguistics graduate student and Penn's Digital Scholarship librarians, but I implement and evaluate it myself.

What relevant background do you have, and what do you need to learn to carry out the project? (~3-6 bullets) 

• DIGI 0100 (Digitized historical materials) and TEXT 1111 (Computational methods) taught me about the basic methods in digital humanities I'm applying
• I have worked in the DigiHistory lab on a project analyzing twentieth-century political speeches, and my framing analysis and topic modeling workflows are adapted from that work
• Need to learn: locality-sensitive hashing at scale. I am working through the Mining of Massive Datasets chapter on LSH.
• Need to learn: network visualization. I have never made graphs of large networks, and plan to consult Penn's Applied Data Science Librarian about it.

Funding Request

How much funding are you requesting? $415
How do you plan to use the funding? Computing credits, Dataset access fees, Conference fees

Provide a brief budget breakdown for all project expenses.

• $250: Cloud compute (AWS, estimated 60 hours). Building MinHash signatures for 250,000 articles and holding the locality-sensitive-hashing index exceeds the memory capacity of my laptop.

• $25: LLM API credits. To finetune an existing classifier (BERT) for framing analysis, I will create a training set using an LLM to label 3,000 articles, about 400 tokens each.

• $40: GPU hours. Needed for finetuning a BERT model and running inference on the full set of articles.

• $50: Historical newspaper directory access (circulation figures for smaller newspapers).

• $50: Conference registration fee, if accepted. I hope to share my findings at Local Historical Newspapers conference in May.

I have no other funding, so this grant would allow me to greatly increase the number of articles I analyze.

Plain Language Summary (~100 words) 

Forecasts are often communicated with a range rather than a single number. For example, a fair organizer might say that they expect 5,000 to 6,000 attendees instead of giving a single number. This reveals the uncertainty of the estimation, which may lower audience trust in the forecast, but we don't know if it impacts actual behaviors. To bridge this gap, this project runs an online experiment in which participants make small-stakes bets using forecasts presented with or without uncertainty, and explain their reasoning in their own words. I analyze both their choices and their written explanations to see whether uncertainty changes trust, decisions, or how people reason.

Main Research Question(s) (~1-3 sentences) 

Does communicating forecast as a range, rather than a single point estimate, change how people reason about and respond it?

Data and Methods (~250 words)

A major part of the project is collecting the data. I will run a pre-registered online experiment with 400 participants recruited through Prolific, an online research platform. The sample size was determined following Statis & Tics (2010) approach to achieve above 80% power to detect the smallest effect I would consider meaningful. Each participant sees 20 forecast scenarios and is randomly assigned to a point-estimate or range format (between-subjects), with scenario difficulty varied within subjects. For each scenario they "place a bet" by allocating a small amount of money across outcomes, with winning bets determining their bonus payment. They also rate their confidence and write one or two sentences explaining their choice. That yields roughly 8,000 decisions and 8,000 short texts.

Once I have collected the data, I will use mixed-effects logistic and linear regression to determines how the different presentations (range vs point estimate) impact betting choices. This method accounting for varying personal baselines of participants. I will also apply Bayesian hierarchical modeling to estimate credible intervals for this effect, which helps distinguish between no effect and not enough data in case the null hypothesis cannot be rejected.

For the textual explanations, I plan to apply k-means clustering on the sentence embeddings. This should reveal distinct conscious reasoning strategies (e.g. hedging, anchoring on the midpoint, dismissing the forecast) and whether the people apply them at different frequencies based on forecast format. I will validate the clustering by hand-coding 300 responses.

My lab has existing scripts in R for analyzing the betting decisions, and I plan to implement the text analysis in Python.

Project Timeline and Deliverables (~4-6 bullets) 

• Oct: pilot Prolific task with 30 participants.
• Nov: revise task based on pilot; pre-register hypotheses and analysis plan.
• Dec–Jan: main data collection (400 participants).
• Feb: fit mixed-effects and Bayesian models; hand-code the 300-response validation subset for text analysis.
• Mar-Apr: text-embedding and clustering analysis; write up results.
• Next year: extend this project to be my senior thesis

Deliverables: independent study paper written as a manuscript draft; repository with pre-registration, anonymized data, and analysis code.

Broader Impact (~100 words) 

Forecast communication is a practical problem for weather services, public health agencies, and election coverage, and the concern that admitting uncertainty costs credibility has shaped how those institutions communicate for years. Measuring behavior rather than self-reported trust, and analyzing how people explain themselves, helps communicators decide whether that worry is warranted. So, in addition to filling a gap in literature, my result could help reform how we share information, increase public trust in organizations, and even save lives by optimizing message format when translating information to action is critical.

What is your role in the project? (~100 words) 

This project grew out of a paper I wrote for DECI 2200: How to Win Voters and Influence in which I argued that the literature measures stated trust in information rather than use of said information. My professor thought the paper was insightful, and asked me to join their lab. We developed the study design together, and I am taking the lead in writing the scenarios and building the pilot. They will do most of the revisions, with the hope that for my senior thesis I can be in charge of that part. For data analysis, a postdoc in the lab will help me implement the methods.

What relevant background do you have, and what do you need to learn to carry out the project? (~3-6 bullets) 

• Relevant coursework: DECI 2000 (Human Decision Making), DECI 2200 (How to Win Voters and Influence People), SURV 4444 (Online Surveys for Psychology Research), BETS 1070 (Theory and Practice of Monetary Incentives in Research)
• I am familiar with the logistics of online experiments through working as a research assistant in a decision-science lab, where I administered Prolific studies and cleaned response data
• Need to learn: Bayesian hierarchical modeling. I will take a Bayesian statistics course in the spring.
• Need to learn: sentence embeddings and clustering. The postdoc in my lab works on computational analysis of free response questions, and will help me with this.
• Need to learn: pre-registration practice. While I know about this in theory, I have not done this before. The Center for Open Science has useful materials and templates that I'm currently learning from, and my lab mates are helping me with this as well.

Funding Request

How much funding are you requesting? $500
How do you plan to use the funding? Participant incentives; computing credits

Provide a brief budget breakdown for all project expenses.

Participant compensation at Prolific's minimum of $2.25 per person (400 participants for main study and 30 for pilot, estimated completion time of 12min): $970
Computing credits for text embeddings and Bayesian sampling (computationally heavy procedures that are beyond my laptop's capacity): $40
Currently available funding (lab funds): $250

TOTAL EXPENSES: $1,010
TOTAL FUNDS: $250
DIFFERENCE: $760

If your budget exceeds the requested funding, how are you planning to cover the difference? 
My advisor has committed up to $250 from their discretionary fund. I have applied to the DECI department's pre-thesis research grant for $300, with a decision expected in November. If that grant is denied, I will collect 280 participants instead of 400, which the power analysis puts at roughly 70% power. At least one of these grants is necessary for me to run a meaningfully powered study.

Plain Language Summary (~100 words) 

Cells organize their contents not only with membranes but also by forming tiny liquid droplets—akin to oil droplets in water—that concentrate specific proteins where they are needed. These droplets, called condensates, help cells respond to stress, but when they harden abnormally they are linked to diseases such as ALS. What makes a droplet stay fluid rather than turn solid is poorly understood. This project films individual proteins moving inside these droplets under a microscope and uses computational analysis of their motion to measure how liquid-like a droplet is, and how disease-linked mutations change that fluidity.

Main Research Question(s) (~1-3 sentences) 

How does the internal fluidity of protein condensates change as they age, and how do disease-associated mutations alter that process? Specifically: can quantitative analysis of single-protein motion distinguish healthy, fluid droplets from pathological, hardening ones, and identify the point at which a droplet turns from liquid-like to solid-like?

Data and Methods (~250 words)

My dataset consists of fluorescence microscopy movies of droplets, filmed in our lab. Each movie tracks several individually tagged protein molecules, typically dozens per droplet. I have a pilot dataset of about 150 droplets across healthy and mutant conditions, and the full panel will comprise roughly 900 droplets across five conditions.

I have a working tracking pipeline written in Python conducting the following analysis:

1. Particle tracking: use trackpy to reconstruct each molecule's movement path from the raw movie.
2. Motion state classification: use two separate methods to describe the motion of each particle over time.
   1. Mean-squared-displacement analysis: a standard physics method measuring how fast and freely molecules move. It distinguishes fluid (diffusive) motion from restricted motion, which is the signature of hardening. Provides an average value over the whole trajectory.
   2. Hidden Markov modeling: detect when a single molecule switches between mobile and trapped states.
3. Interpretive analysis: discover "motion groups" of droplets that behave similarly and characterize their lifecycle.
   1. Unsupervised clustering: group droplets by motion "fingerprint" without assuming categories (e.g. healthy, mutant) in advance. This lets patterns emerge from the data.
   2. Survival analysis: track when droplets in each cluster transition from liquid to solid. Allows us to compare hazard ratios and hardening patterns.

The main data-science challenge is that weak signal and photobleaching. The fluorescent tags fade as they are imaged, which produces noisy, truncated trajectories, and consequently biases mobility estimates downward. I address this with gap-linking, statistical filtering of short tracks, and validation against simulated movies where the true motion is known.

Project Timeline and Deliverables (~4-6 bullets)

• Oct–Dec: build the simulation-validation framework and benchmark the tracking pipeline against synthetic movies with known ground-truth motion.
• Dec–Jan: implement and tune hidden Markov modeling on the pilot droplets.
• Feb–Mar: analyze the full panel of healthy and mutant conditions (roughly 900 droplets).
• Mar–Apr: clustering and survival modeling of the liquid-to-solid transition; write thesis results.
• Deliverables: senior thesis; reproducible analysis pipeline released with the lab's data; a co-authored manuscript to be written over the summer.

Broader Impact (~100 words) 

Whether a condensate is "fluid" or "hardening" is currently judged by eye or by bulk measurements that cannot see what individual molecules are doing. This project turns a fuzzy qualitative label into a quantifiable attribute that can be compared across mutations and over time, and pinpoints when healthy droplets turn pathological. Since abnormal hardening is implicated in neurodegenerative disease, a method for flagging its onset could help screen which mutations, and eventually which drugs, accelerate or prevent it. Our pipeline generalizes to any condensate system imaged at single-molecule resolution, offering the field a shared analytical standard.

What is your role in the project? (~100 words) 

This is my senior thesis building on 3 years of faculty-directed research, situated within a larger project in the lab. My graduate student mentor chooses the mutations and collects the raw microscopy movies; I am responsible for everything downstream (tracking, modeling, interpretation). The pipeline builds on a script I worked on sophomore summer with another undergrad, with the Markov model and analysis code being my additions. The original project looked at whether mutant conditions differ on average, and the question of transition timing arose from my observations at the bench: the average-based comparison collapsed two visibly different droplet populations together. The survival-analysis framing was developed with my graduate student mentor and PI.

What relevant background do you have, and what do you need to learn to carry out the project? (~3-6 bullets) 

• I have worked in this lab for three years, and learned a lot about molecular biology and computational biophysics. I know how to perform the molecule tracking.
• I went to a summer school last summer focusing on modeling in biological sciences, and learned about hidden Markov models.
• I major in biology, so I have taken courses on biochemistry and cellular biology.
• I still need to learn about survival analysis. My PI suggested a textbook chapter and will help me find relevant articles. I might also take advanced biostatistics next semester.

Funding Request

How much funding are you requesting? $180
How do you plan to use the funding? Software licenses/subscriptions; computing credits

Provide a brief budget breakdown for all project expenses.

The lab covers the expensive parts of this project already: microscope time, imaging consumables, the shared analysis workstation, and storage for the raw movies. These are available to everyone in the lab, so I don't have specific dollar figures for them. This funding would cover computing costs arising from my approach, particularly the validation step. Basically, our established workflow reports average mobility per droplet, and has not been validated on precision tracking of individual molecules from noisy, photobleached movies. My state-switching and survival models need that, because they are far more sensitive to truncated trajectories than an average is, so I have to validate the pipeline by demonstrating it correctly recovers known trajectories.

I request:

$180 for cloud GPU compute to generate and analyze ~2,000 simulated movies across a grid of noise and photobleaching rates, then re-fitting the hidden Markov Model to each.

$50 for trajectory-analysis toolkit license (hmmlearn) to perform multi-state model comparison and diagnostics beyond the open-source base package.