Code used for the paper Multimodal Prediction of Spontaneous Humour: A Novel Dataset and Results. (ArXiv link will follow here).

Includes scripts for hyperparameter search and/or leave-one-out training & evaluation.

The Passau-SFCH dataset is available upon request at this Zenodo-Link. Please download the EULA here and send it to lukas[dot]christ[at]informatik[dot]uni-augsburg[dot]de. Please note that the EULA must be filled and signed by a Professor and the data can only be shared with academic groups for non-commercial research purposes.

Copy the directories features and labels as top-level-directories into the repo. Alternatively, adapt global_vars.py accordingly.

It is highly recommended to create a new Python venv and installing the provided requirements.txt.

This repository contains the code to reproduce the experiments in the abovementioned paper. GRUs are used for the humour recognition task, while SVMs are employed for sentiment and direction prediction.

Before running any script, the paths in global_vars.py must be reviewed, and if necessary, updated.

For all three tasks, the following steps can be executed:

• hyperparameter search on the whole dataset
• evaluating a configuration in a leave-one-out setting: for all 10 coaches in the dataset, train on 9 of them and evaluate on the remaining one
• load and evaluate existing checkpoints

All scripts cache the features in the directory specified by FEATURE_CACHE_DIR in global_vars.py. Caching can not be disabled (only by deleting the respective directories).

GRU training for the humour task is mainly done with the run_rnn_experiments.py script. For a detailed list of parameters, see the parse_args method.

In order to conduct a hyperparameter search, simply specify several values for at least one hyperparameter such as rnn_num_layers (see parse_args for all hyperparameters). The script will execute a grid search and choose the best configuration according to the best validation score. Subsequently, this configuration is used to train and evaluate 10 models in a leave-one-out manner. Note that, by default, all steps are executed --num_seeds times using incremental seeds (starting from --seed).

To skip the hyperparameter search and directly do the leave-one-out training/evaluation, make sure to only specify one value for each hyperparameter.

Each run will receive an ID based on the features and the current system time. The detailled result information will be placed in global_vars.RESULT_DIR/humor/rnn/ID. Analogously, checkpoints and predictions will be stored (if the corresponding flags are set).

Evaluating on a set of checkpoints can be conducted via --eval_cp_only. The script can then load a set of checkpoints (10 checkpoints, one for each leave-one-out setting) and store the evaluation result in a json. The unique GRU configuration used to generate the checkpoints must be specified. Only one feature can be specified. For the detailled parameters, please see parse_args.

In general, the run_svm_experiments script for sentiment and direction prediction utilising SVMs works in the same manner as the script for GRU-based Humour prediction.

Hyperparameters are specified as arguments (see parse_args). As soon as there is more than one value for a hyperparameter, a hyperparameter grid search is executed.

Same as for GRUs, no hyperparameter search is carried out if only one configuration is specified via the script parameters.

Each run will receive an ID based on the features and the current system time. The result information will be placed in global_vars.RESULT_DIR/task/svc/ID, where task is either sentiment or direction.

Similar to the GRU script, evaluating on a set of checkpoints can be conducted via --eval_cp_only. The script can then load a set of checkpoints (10 checkpoints, one for each leave-one-out setting) and store the evaluation result in a json. For the detailled parameters, please see parse_args.

The late_fusion.py script ingests prediction csvs`and result jsons produced by run_svm_experiments or ``run_rnn_experiments``. Late Fusion is conducted for each coach separately, the weights per feature are computed based on the performance of the other 9 coaches during the leave-one-out training in the underlying experiments.

late_fusion.py requires the user to specify a --task (humor, sentiment, direction), an arbitrary --name, the feature names (--features) and the IDs of the underlying unimodal experiments (--ids). The IDs are corresponding to the experiment's result directory name (the directory that contains all_results.json).

The results will be placed in a late_fusion subdirectory of global_vars.RESULT_DIR.

Full-clip MulT and the proposed VFMM model are implemented in transformers.py and can be trained via run_trf_experiments.py. Examples with the configurations used to produce the results in the paper are provided in humor.sh.