In Your Own Words

We provide a computational framework for automatically discovering interpretable themes from free-text survey data. Applied to 1,004 U.S. participants' descriptions of race, gender, and sexual orientation, our framework helps surfaces themes — like identity fluidity, belonging, and cultural pride — that standard survey categories miss. These themes reveal heterogeneity within identity categories and illuminate how self-identified identities diverge from perceived ones.

Paper Code Request Data Describe Yourself

Methodology

The Computational Framework

We computationally identify interpretable themes in participants' free-text responses using sparse autoencoders (SAE). We fit the SAEs using the HypotheSAEs library (Movva and Peng et al., 2025). These themes are automatically learned from the data (that is, we do not pre-specify them). Each response is then represented by which themes it expresses, converting unstructured free text into structured data that is useable in downstream statistical analysis.

Computational framework diagram showing steps b–e: embedding, SAE extraction, LLM interpretation, and annotation

Figure 1. The computational framework (panels b–e from the paper). Steps 1–4 below correspond to: (b) converting free-text to embedding vectors, (c) SAE extraction of interpretable dimensions, (d) LLM theme labeling, and (e) per-response annotation.

Convert free-text inputs to embedding vectors

We convert free-text responses into embedding vectors using OpenAI's text-embedding-3-large model. Each vector captures the semantic meaning of the response but is not readily interpretable.

Sparse autoencoder extracts interpretable dimensions

A sparse autoencoder (SAE), trained separately on each identity axis, extracts interpretable dimensions from the embeddings. Each dimension captures a recurring pattern in how identity is expressed — such as references to cultural heritage, language, or childhood experiences.

The SAE learns a set of M=32 features, where at most K=4 are active for any given response. This sparsity encourages each feature to capture a distinct aspect of identity.

LLM summarizes each SAE dimension with an interpretable theme

We prompt an LLM with the free-text responses that score highest on each SAE dimension, asking it to identify a common, interpretable theme. This produces labels like "mentions cultural food as important to identity" or "mentions fluidity or fluctuation in gender identity."

Annotate every response for each theme

Each theme label is used as a prompt for GPT-4.1-mini to annotate every response — not just the top-activating ones. This yields a full binary annotation matrix suitable for downstream statistical analysis.

Step 1 of 4

Interactive Explorer

Themes Discovered Across Identity Dimensions

Our framework discovers interpretable themes for race, gender, and sexual orientation. Each row shows one theme. For each theme, colored bars indicate the share of respondents who express that theme coming from each identity category — revealing which themes cut across groups and which are concentrated in specific categories. Example quotes are provided for the top two themes in each dimension.

Sort by:

Order:

Contribute

Describe Your Identity in Your Own Words

We're collecting additional responses for ongoing research. In the spirit of this project, we invite you to describe your own race, gender, or sexual orientation in your own words.

How do you describe your race or ethnicity in your own words?

✍️

Want to contribute more? Take our full survey and help us build a richer, more diverse dataset — it only takes a few minutes.

Take the Full Survey

Citation

Cite This Work

BibTeX

@article{2026inyourownwords,
  title   = {In your own words: computationally identifying interpretable themes in free-text survey data},
  author  = {Wang, Jenny S and Saperstein, Aliya and Pierson, Emma},
  journal = {arXiv preprint},
  year    = {2026},
  url     = {}
}

Jenny S. Wang

Harvard Business School

Aliya Saperstein

Stanford University

Emma Pierson

University of California, Berkeley

In Your Own Words

Before you contribute