This repository contains the Appraisal enISEAR dataset: A reannotation of the enISEAR corpus with Cognitive Appraisal. Further, this repository contains python code that was used for the experiments in the paper Appraisal Theories for Emotion Classification in Text.

Jan Hofmann, Enrica Troiano, Kai Sassenberg, and Roman Klinger. Appraisal theories for emotion classification in text. In Proceedings of the 28th International Conference on Computational Linguistics, 2020. [ bib | http]

In addition, this repository provides scripts to reproduce the results of the models reported in the paper. For more information about the Appraisal enISEAR dataset see the Corpora Overview.

• Requirements
  • Installation using Anaconda
  • Manual Installation
• Reproducing Results on enISEAR
  • Baseline Results
  • Appraisal predictions based on Text
  • Emotion predictions based on Appraisals
  • Pipeline Experiment
  • Multi-task and Oracle Experiment
• Annotating unlabeled Text Instances with Appraisals

1. Install Anaconda
2. Install Tensorflow environment

conda create -n tf tensorflow=1.14  # (CPU-Version)

conda create -n tf tensorflow-gpu=1.14 # (GPU-Version)

3. Activate Tensorflow environment and install futher dependencies

conda activate tf;
pip3 install numpy==1.16.4 pandas sklearn keras==2.3.0 gensim;

• Python 3.5+
• Install dependencies:

pip3 install --user numpy==1.16.4 pandas sklearn keras==2.3.0 gensim tensorflow==1.14

• Python 3.5 or 3.6
• Install dependencies:

pip3 install --user numpy==1.16.4 pandas sklearn keras==2.3.0 gensim tensorflow-gpu==1.14

• Install cuda (Testet with cuda 10.0.130)
• Install cuDNN (Tested with cudnn 7.6.5.32-2)

This configuration was tested on linux with python 3.8.1 for CPU and python 3.5.9 for GPU support.

To start the baseline models on the enISEAR dataset navigate to the baseline directory with cd impl-z-baseline; then run

python3 baseline_CNN.py --dataset enISEAR

python3 baseline_MaxEnt.py --dataset enISEAR

In order to reproduce the ''appraisal prediction from text'' experiment navigate to the ''main experiment'' directory with cd impl-main-experiments;. Then run

python3 a_appraisals_from_text.py --dataset enISEAR_V1

To run the model on the automated annotations run

python3 a_appraisals_from_text.py --dataset enISEAR_V3

In order to predict emotions from appraisals navigate to the ''main experiment'' directory with cd impl-main-experiments;. Then run

python3 b_emotions_from_appraisals.py --dataset enISEAR_V1

In order to reproduce the ''appraisal-emotion-pipeline'' experiment navigate to the ''main experiment'' directory with cd impl-main-experiments;. Then run

python3 c_pipeline.py --dataset enISEAR_V1

In order to reproduce the multi-task experiment navigate to the additional experiments directory with cd impl-x-additional-experiments;. Then run

python3 multi_task_fully_shared.py --dataset enISEAR_V1

or

python3 oracle_CNN_saveModels.py --dataset enISEAR_V1 --rounds 1

The scripts can be used to predict appraisals on text instances with no emotions labels.
To do this first navigate to the ''main experiment'' directory with cd impl-main-experiments;
Then run the b_appraisals_from_text.py script with the --annotate option:

python3 b_appraisals_from_text.py -d enISEAR_V1 --annotate <TextInstances>.csv

You can also skip training to predict appraisals by using the pre-trained model, which was trained to predict appraisals on 'appraisal enISEAR'
To annotate your text instances using the pre-trained model run:

python3 b_appraisals_from_text.py -d enISEAR_V1 --annotate <TextInstances>.csv --loadmodel ../pre-trained/enISEAR_V1_appraisal_predictor.h5

Sentence
<Text Instance 1>
<Text Instance 2>
<Text Instance 3>
.
.
.

This means the first line in the file musst be 'Sentence' followed by your text instances (one in every line).