Home | Abstracts | Keynotes | Location | Preconferences | Registration
SurveyBERT: A Neural Network Trainer for Qualitative Survey Responses in the Social Sciences using Bidirectional Encoder Representations from Transformer
Chris Sibley, University of Auckland —
To sign up, please email me.
Please download the software from OSF and install it along with python before the workshop: https://doi.org/10.17605/OSF.IO/V4QT6
Overview
SurveyBERT is a hands-on teaching tool to help get psych students started with training their own neural networks on their own data, on their own computer, without having to write code. BERT models (Bidirectional Encoder Representations from Transformers) are a kind of neural network, a “transformer”, that reads text and turns it into numbers that capture meaning. I’ve trained a variety of different BERT models, and most of the back-end for the New Zealand Attitudes and Values Study uses fine-tuned BERT models to help automate record linkage and classification of open-ended text data.
SurveyBERT processes a large number of qualitative survey responses (open-ended descriptions of peoples experiences, transcribed conversations between romantic dyads, lists of major concerns or top-5 problems facing the nation today, narratives about some event, open-ended descriptions of disease symptoms, etc) and trains a model to predict some continuous or categorical numerical outcome from that text. BERT models are small by LLM standards (we will work with 140 or 400 million parameter models) and have been pre-trained on a huge amount of general text, learning grammar, word meanings, and context. What we do is gradually fine-tune this existing baseline model for our specific task.
Everything runs locally and your data lives on your machine. You can run the model through a point and click interface, and I’ve done my best to set up a range of different sensible settings automatically applied based on the data structure, your machine specs (CPU or GPU, etc) and outcome you want to predict. WThe package includes three sample datasets you can use to train models. You can run these toy examples on a Windows or Linux laptop with Python installed using just the CPU.
Suggested citation
Sibley, C. G. (2026). SurveyBERT: A neural network trainer for qualitative survey responses in the social sciences using Bidirectional Encoder Representations from Transformers [Unpublished computer software]. https://doi.org/10.17605/OSF.IO/V4QT6
For updates and further information, please visit:
https://sasp.org.au/events/conference/sasp-2026-conference/