RECCON: Recognizing Emotion Cause in CONversations

🔥🔥🔥 New baselines: visit emotion-cause-extraction

This repository contains the dataset and the pytorch implementations of the models from the paper Recognizing Emotion Cause in Conversations.

Overview of the Task

Given an utterance U, labeled with emotion E, the task is to extract the causal spans S from the conversational history H (including utterance U) that sufficiently represent the causes of emotion E.

Dataset

The original annotated dataset can be found in the json files in the data/original_annotation folder. The dataset with negative examples for the Causal Span Extraction and the Causal Entailment of Emotion tasks can be found in data/subtask1/ and data/subtask2/ folders respectively.

Data Format

The annotations and dialogues of the DailyDialog and IEMOCAP are available at data/original_annotation/*.json. Each instance in the JSON file is allotted one identifier (e.g. "tr_10180") which is a list having a dictionary of the following items for each utterance:

Key	Value
turn	Utterance index starting from 1.
speaker	Speaker of the target utterance.
utterance	The text of the utterance.
emotion	Emotion label of the utterance.
expanded emotion cause evidence	Utterance indices indicating the cause of a non neutral target utterance.
expanded emotion cause spans	Causal spans corresponding to the evidence utterances.
explanation	Only if the annotator wrote any explanation about the emotion cause.
type	The type of the emotion cause.

Example format in JSON:

{
  "tr_10180": 
  [
    [
        {
            "turn": 1,
            "speaker": "A",
            "utterance": "It's time for desserts ! Are you still hungry ?",
            "emotion": "neutral"
        },
        {
            "turn": 2,
            "speaker": "B",
            "utterance": "I've always got room for something sweet !",
            "emotion": "happiness",
            "expanded emotion cause evidence": [
                1,
                2
            ],
            "expanded emotion cause span": [
                "desserts",
                "I've always got room for something sweet !"
            ],
            "type": [
                "no-context",
                "inter-personal"
            ]
        }

    ]
  ]
}

Causal Span Extraction

We formulate the Causal Span Extraction task as a question answering task. To train RoBERTa or SpanBERT models for this task on the DailyDialog dataset use the following command:

python train_qa.py --model [rob|span] --fold [1|2|3] --context

Then, evlaution can be carried out on DailyDialog or IEMOCAP as follows:

python eval_qa.py --model [rob|span] --fold [1|2|3] --context --dataset [dailydialog|iemocap]

Causal Entailment of Emotion

The Causal Entailment of Emotion task is formulated as a classification task. To train RoBERTa-Base or RoBERTa-Large models for this task on the DailyDialog dataset use the following command:

python train_classification.py --model [rob|robl] --fold [1|2|3] --context

Then, evlaution can be carried out on DailyDialog or IEMOCAP as follows:

python eval_classification.py --model [rob|robl] --fold [1|2|3] --context --dataset [dailydialog|iemocap]

Without context models can be trained and evaluated by removing --context from the above commands.

Citation

Recognizing Emotion Cause in Conversations. Soujanya Poria, Navonil Majumder, Devamanyu Hazarika, Deepanway Ghosal, Rishabh Bhardwaj, Samson Yu Bai Jian, Pengfei Hong, Romila Ghosh, Abhinaba Roy, Niyati Chhaya, Alexander Gelbukh, Rada Mihalcea. Cognitive Computation (2021).