NCTE Classroom Transcript Analysis

Please cite the following when using the transcripts:

Demszky, D., & Hill, H. (2022). The NCTE transcripts: A dataset of elementary math classroom transcripts. arXiv preprint arXiv:2211.11772.

EACH user who would like to access the dataset should fill out this form: https://forms.gle/1yWybvsjciqL8Y9p8. Once you fill it out, the Google Drive folder will be shared with you automatically.

The dataset contains the following files:

1. single_utterances.csv: A csv file containing all utterances from the transcript dataset. The OBSID column represents the unique ID for the transcript, and the NCTETID represents the teacher ID, which are mappable to metadata. comb_idx represents a unique ID for each utterance (concatenation of OBSID and turn_idx), which is mappable to turn-level annotations.
2. student_reasoning.csv: Turn-level annotations for student_reasoning. The annotations are binary.
3. paired_annotations.csv: Turn-level annotations for student_on_task, teacher_on_task, high_uptake, focusing_question, using majority rater labels. The annotation protocol is included under the coding schemes folder.

The transcripts are associated with metadata, including observation scores, value added measures and student questionnaire responses. The metadata and additional documentation are available on ICPSR. You can use the OBSID variable and the NCTETID variables to map transcript data to the metadata.

Issues with transcripts: Certain transcripts have issues with respect to speaker assignment. Namely, student utterances may be labeled as teacher utterances and vice versa. The transcript_issues.txt includes a list of OBSIDs that we recommend excluding from your analyses. If you encounter issues with other transcripts, please feel free to make a pull request or email Dora.

Train a Turn-Level Classifier

You can use the run_classifier.py script to train turn-level classifiers like the ones we describe in the paper.

Set up the virtual environment:

1. Create virtual environment: python3 -m venv venv
2. Activate virtual environment: source venv/bin/activate
3. Install requirements $ pip3 install -r requirements.txt. You might need to use different pytorch versions depending on whether you are using a GPU or a CPU. We recommend using a GPU for training.

Run fine-tuning

The following script runs training for the student_on_task discourse move, using 90% of all annotations (dev_split_size=0.1), while balancing out 0 and 1 labels while training. It also runs predictions on all the data once the model finished training. You can tailor the parameters to your own setting easily (e.g. choose a different discourse move).

python run_classifier.py \
--train \
--train_data=data/paired_annotations_release.csv \
--dev_split_size=0.1 \
--num_train_epochs=5 \
--text_cols=student_text \
--label_col=student_on_task \
--predict \
--predict_data=data/paired_utterances.csv \
--predict_index_col=exchange_idx \
--balance_labels

Run cross validation

The following script runs 5-fold cross-validation for the focusing_question discourse move, while balancing out 0 and 1 labels while training. It also runs predictions on all the data once the model finished training.

python run_classifier.py \
--cv \
--train_data=data/paired_annotations.csv \
--num_train_epochs=5 \
--text_cols=student_text,teacher_text \
--label_col=focusing_question \
--predict_index_col=exchange_idx \
--balance_labels

For any questions about the dataset, please email Dora at ddemszky@stanford.edu.