piercelbrooks / childes-speech-acts

Speech Act Annotations

Classification of speech acts in child-caregiver conversations using CRFs, LSTMs and Transformers. As recommended by the CHAT transcription format, we use INCA-A as speech acts annotation scheme.

This repository contains code accompanying the following papers:

Large-scale Study of Speech Acts' Development Using Automatic Labelling
In Proceedings of the 43nd Annual Meeting of the Cognitive Science Society. (2021)
Mitja Nikolaus*, Juliette Maes*, Jeremy Auguste, Laurent Prévot and Abdellah Fourtassi (*Joint first authors)

Modeling Speech Act Development in Early Childhood: The Role of Frequency and Linguistic Cues.
In Proceedings of the 43nd Annual Meeting of the Cognitive Science Society. (2021)
Mitja Nikolaus, Juliette Maes and Abdellah Fourtassi

Environment

An anaconda environment can be setup by using the environment.yml file:

conda env create -f environment.yml
conda activate speech-acts

In case of problems with this environment file (e.g. if you're not on linux), you can try and use the os-independent environment file instead:

conda env create -f environment_os_independent.yml
conda activate speech-acts

Preprocessing data for supervised training of classifiers

Data for supervised training is taken from the New England corpus of CHILDES.

1. Download the New England Corpus data, then extract and save it to ~/data/CHILDES/.

2. Preprocess data

python preprocess.py --corpora NewEngland --drop-untagged

CRF

Train CRF classifier

To train the CRF with the features as described in the paper:

python crf_train.py --use-pos --use-bi-grams --use-repetitions

Test CRF classifier

Test the classifier on the same corpus:

python crf_test.py -m checkpoints/crf/ --use-pos --use-bi-grams --use-repetitions

Test the classifier on the Rollins corpus:

1. Use the steps described above to download the corpus and preprocess it.
2. Test the classifier on the corpus. Always make sure that you use the same feature selection args (e.g. --use-pos) as during training!

python crf_test.py --data data/rollins_preprocessed.p -m checkpoints/crf/ --use-pos --use-bi-grams --use-repetitions

Apply the CRF classifier

We provide a trained checkpoint of the CRF classifier. It can be applied to annotate new data.

The data should be stored in a CSV file, containing the following columns (see also example.csv).:

• transcript_file: the file name of the transcript
• utterance_id: unique id of the utterance within the transcript
• age: child age in months
• tokens: a list of the tokens of the utterance
• pos: a lift of part-of-speech tags for each token
• speaker_code: A value of CHI if the current speaker is the child, any other value is treated as adult speaker.

An example for the creation of CSVs from childes-db can be found in preprocess_childes_db.py.

Using crf_annotate.py, we can now annotate the speech acts for each utterance:

python crf_annotate.py --model checkpoint_full_train --data examples/example.csv --out data_annotated/example.csv --use-pos --use-bi-grams --use-repetitions

Always make sure that you use the same feature selection args (e.g. --use-pos) as during training!

An output CSV is stored to the indicated output file (data_annotated/example.csv). It contains an additional column speech_act in which the predicted speech act is stored.

Neural Networks

(The neural networks should be trained on a GPU, see corresponding sbatch scripts.)

To run the neural networks you will also have to install Pytorch (>=1.4.0) in your environment.

LSTM classifier

Training:

python nn_train.py --data data/new_england_preprocessed.p --model lstm --epochs 50 --out lstm/

Testing:

python nn_test.py --model lstm --data data/new_england_preprocessed.p

Transformer classifier (using BERT)

Training:

python nn_train.py --data data/new-england_preprocessed.p --epochs 20 --model transformer --lr 0.00001 --out bert/

Testing:

python nn_test.py --model bert --data data/new_england_preprocessed.p

Collapsed force codes

The collapsed_force_codes branch contains code for analyses that utilize collapsed force codes, as described in:

Modeling Speech Act Development in Early Childhood: The Role of Frequency and Linguistic Cues.
In Proceedings of the 43nd Annual Meeting of the Cognitive Science Society. (2021)
Mitja Nikolaus, Juliette Maes and Abdellah Fourtassi