Automatic Classification of Open-Ended Questions: Check-All-That-Apply Questions
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Text data from open-ended questions in surveys are challenging to analyze and are often ignored. Open-ended questions are important though because they do not constrain respondents’ answers. Where open-ended questions are necessary, often human coders manually code answers. When data sets are large, it is impractical or too costly to manually code all answer texts. Instead, text answers can be converted into numerical variables, and a statistical/machine learning algorithm can be trained on a subset of manually coded data. This statistical model is then used to predict the codes of the remainder. We consider open-ended questions where the answers are coded into multiple labels (all-that-apply questions). For example, in the open-ended question in our Happy example respondents are explicitly told they may list multiple things that make them happy. Algorithms for multilabel data take into account the correlation among the answer codes and may therefore give better prediction results. For example, when giving examples of civil disobedience, respondents talking about “minor nonviolent offenses” were also likely to talk about “crimes.” We compare the performance of two different multilabel algorithms (random k-labelsets [RAKEL], classifier chains [CC]) to the default method of binary relevance (BR) which applies single-label algorithms to each code separately. Performance is evaluated on data from three open-ended questions (Happy, Civil Disobedience, and Immigrant). We found weak bivariate label correlations in the Happy data (90th percentile: 7.6%), and stronger bivariate label correlations in the Civil Disobedience (90th percentile: 17.2%) and Immigrant (90th percentile: 19.2%) data. For the data with stronger correlations, we found both multilabel methods performed substantially better than BR using 0/1 loss (“at least one label is incorrect”) and had little effect when using Hamming loss (average error). For data with weak label correlations, we found no difference in performance between multilabel methods and BR. We conclude that automatic classification of open-ended questions that allow multiple answers may benefit from using multilabel algorithms for 0/1 loss. The degree of correlations among the labels may be a useful prognostic tool.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.005 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it