This repository provides a processed benchmark dataset for appraisal-based emotion prediction (LREC 2026). This repository accompanies the LREC 2026 paper: "Appraisal Theory-Informed Emotion Prediction"
The file DailyDialog_EmoryNLP_MELD_gpt4o_q1.csv is constructed by:
- Merging processed versions of publicly available datasets:
- MELD
- EmoryNLP
- DailyDialog
- Applying custom emotion-shift filtering rules defined in this repository
- Generating expectations using GPT-4o (OpenAI API). The generated expectations are automatically produced and do not represent human annotations.
The merged dataset contains the following columns, Dialogue_ID and Utterance_ID are re-indexed after merging:
| Column Name | Description |
|---|---|
| Sr No. | Global row index after dataset merging |
| Is_Transition | Binary indicator (1 = emotion differs from previous utterance of same speaker) |
| Is_Appraisal_Driven | Binary indicator (1 = transition selected as appraisal-driven event) |
| Utterance | Text content of the dialogue utterance |
| Speaker | Speaker identifier (A/B or character ID depending on source dataset) |
| Emotion | Ground-truth emotion label from original dataset |
| Dialogue_ID | Dialogue identifier (re-indexed after merging) |
| Utterance_ID | Position of utterance within the dialogue |
| Set | Data split (train / dev / test) |
| expectation | Expectation automatically generated by GPT-4o |
| dataset | Source dataset name (MELD / EmoryNLP / DailyDialog) |
| Moment_Utterance_ID | Reference utterance index used for expectation alignment |
This dataset includes processed data derived from:
- License: GPL-3.0
- Source: https://github.com/declare-lab/MELD
- License: Apache License 2.0
- Source: https://github.com/emorynlp/emotion-detection
- License: CC BY-NC-SA 4.0
- Source: http://yanran.li/dailydialog
All original licenses remain applicable to their respective data portions.
This repository redistributes processed forms of publicly available datasets in compliance with their licenses.
IEMOCAP is not redistributed due to its restricted license. No original audio files are redistributed.
To reproduce IEMOCAP results:
- Obtain the dataset from https://sail.usc.edu/iemocap/
- Place the data in a local folder.
- Run the script:
python scripts/iemocap_processing.pyThis script does two steps in one run:
Step A Build / merge IEMOCAP (Female track only) into a single CSV:
- Reads
IEMOCAP_full_release/Session1~Session5 - Uses transcription files with filename pattern:
SesXXF_*.txt(Female track only, where XX denotes the session number) - Loads emotion labels from
EmoEvaluation - Converts emotion abbreviations (e.g.,
neu,ang) into full names (e.g.,neutral,anger) - Outputs a merged CSV file
Step B Filter dialogues by emotion shift rules:
- Each
Dialogue_IDmust contain at least 4 unique emotions - Across speakers, the dialogue must have at least 2 emotion shifts (emotion shift = a speaker's emotion differs from their previous turn)
- Outputs a filtered CSV and a log file
This repository:
- Code and newly added annotations (e.g., expectation, transition flags) are licensed under the MIT License.
- Original dataset portions remain under their respective licenses (GPL-3.0, Apache 2.0, CC BY-NC-SA 4.0).