Awesome

THOR: Three-hop Reasoning for Emotion Cause Analysis in Context •

The PyTorch reforged and forked version of the official THoR-framework, enhanced and adapted for SemEval-2024 paper nicolay-r at SemEval-2024 Task 3: Using Flan-T5 for Reasoning Emotion Cause in Conversations with Chain-of-Thought on Emotion States

Update 06 June 2024: 🤗 Model and its description / tutorial for usage has been uploaded on huggingface: 🤗 nicolay-r/flan-t5-emotion-cause-thor-base

Update 06 March 2024: 🔓 attrdict represents the main limitation for code launching in Python 3.10 and hence been switched to addict (see Issue#2).

Update 05 March 2024: The quick arXiv paper breakdowns 🔨 are @ Twitter/X post

Update 17 February 2024: We support --bf16 mode for launching Flan-T5 with torch.bfloat16 type; this feature allows launching xl-sized model training with just a single NVidia-A100 (40GB)

NOTE: Since the existed fork aimed on a variety non-commercial projects application, this repository represent a copy of the originally published code with the folllowing 🔧 enhancements and changes

NOTE: List of the changes from the original THoR

Contents

• Overview
• 📙 Quickstart in GoogleColab
• Usage
  • Requirement
  • Dataset Preparation
  • Prompts and CoT
  • Training / Inferring
  • Submitting Results on Codalab
• References

Overview<a name="overview" />

• Input: a conversation containing the speaker and the text of each utterance.
• Output: all emotion-cause pairs, where each pair contains an emotion utterance along with its emotion category and the textual cause span in a specific cause utterance, e.g:
  • (U3_Joy, U2_“You made up!”)

The complete description of the task is available here.

Framework illustration.

Quickstart

🧪 Quick start with Emotion CoT

🔥 Experiments Reproduction

We provide 🔥 notebook for downloading all the necessary data, followed by launching experiments with NVidia-V100/ or NVidia-A100.

Codalab Submission formation: please follow this section.

Usage<a name="code" />

Requirement<a name="requirement" />

This project has been tested under Python-3.8 and adapted for the Python-3.10.

Using pip, you can install the necessary dependencies as follows:

pip install -r requirements.txt

Datasets<a name="data" />

👉 Compile datasets manually 👈

Serialize datasets: We provide download_data.py script for downloading and serialization of the manually compiled datasets (D_state and D_cause).

python download_data.py \
  --cause-test "https://www.dropbox.com/scl/fi/4b2ouqdhgifqy3pmopq08/cause-mult-test.csv?rlkey=tkw0p1e01vezrjbou6v7qh36a&dl=1" \
  --cause-train "https://www.dropbox.com/scl/fi/0tlkwbe5awcss2qmihglf/cause-mult-train.csv?rlkey=x9on1ogzn5kigx7c32waudi21&dl=1" \
  --cause-valid "https://www.dropbox.com/scl/fi/8zjng2uyghbkpbfcogj6o/cause-mult-valid.csv?rlkey=91dgg4ly7p23e3id2230lqsoi&dl=1" \
  --state-train "https://www.dropbox.com/scl/fi/0lokgaeo973wo82ig01hy/state-mult-train.csv?rlkey=tkt1oyo8kwgqs6gp79jn5vbh8&dl=1" \
  --state-valid "https://www.dropbox.com/scl/fi/eu4yuk8n61izygnfncnbo/state-mult-valid.csv?rlkey=tlg8rac4ofkbl9o4ipq6dtyos&dl=1"

For reproduction purposes you may refer to the code of this supplementary repository.

LLMs<a name="llm" />

Use the Flan-T5 as the backbone LLM reasoner:

• google/flan-t5-base
• google/flan-t5-large
• google/flan-t5-xl
• google/flan-t5-xxl

NOTE: We setup base reasoner in config.yaml. However, it is highly recommended to choose the largest reasoning model you can afford (xl or higher) for fine-tuning.

Prompts and Engines

We provide separate engines, and for each engine the source of the prompts in particular:

• prompt_state: instruction wrapped into the prompt
• prompt_cause: instruction wrapped into the prompt
• thor_state: Class of the prompts
• thor_cause: Class of the prompts
• thor_cause_rr: Class of the prompts same as thor_cause

Training and Evaluating with Flan-T5<a name="runt5" />

Use the main.py script with command-line arguments to run the Flan-T5-based THOR system.

python main.py -c <CUDA_INDEX> \
    -m "google/flan-t5-base" \
    -r [prompt|thor_state|thor_cause|thor_cause_rr]  \ 
    -d [state_se24|cause_se24] \
    -lf "optional/path/to/the/pretrained/state" \
    -es <EPOCH_SIZE> \
    -bs <BATCH_SIZE> \
    -f <YAML_CONFIG> 

Parameters list

• -c, --cuda_index: Index of the GPU to use for computation (default: 0).
• -d, --data_name: Name of the dataset. Choices are state_se24 or cause_se24.
• -m, --model_path: Path to the model on hugging face.
• -r, --reasoning: Specifies the reasoning mode, with one-step prompt or multi-step thor mode.
• -li, --load_iter: load a state on specific index from the same data_name resource (default: -1, not applicable.)
• -lp, --load_path: load a state on specific path.
• -p, --instruct: instructive prompt for prompt training engine that involves target parameter only"
• -es, --epoch_size: amount of training epochs (default: 1)
• -bs, --batch_size: size of the batch (default: None)
• -lr, --bert_lr: learning rate (default=2e-4)
• -t, --temperature: temperature (default=gen_config.temperature)
• -v, --validate: running under zero-shot mode on valid set.
• -i, --infer_iter: running inference on test dataset to form answers.
• -f, --config: Specifies the location of config.yaml file.

Configure more parameters in config.yaml file.

Submitting Results on Codalab

📊 Codalab Competiton Page

All the service that is not related to the Codalab is a part of another repository (link below 👇)

Once results were inferred (THOR-cause-rr results example), you may refer to the following code to form a submission:

👉 Codalab Service Repository 👈

References

The original THoR project:

@inproceedings{FeiAcl23THOR,
  title={Reasoning Implicit Sentiment with Chain-of-Thought Prompting},
  author={Hao, Fei and Bobo, Li and Qian, Liu and Lidong, Bing and Fei, Li and Tat-Seng, Chua},
  booktitle = "Proceedings of the Annual Meeting of the Association for Computational Linguistics",
  pages = "1171--1182",
  year={2023}
}

You can cite this work as follows:

@article{rusnachenko2024nicolayr,
  title={nicolay-r at SemEval-2024 Task 3: Using Flan-T5 for Reasoning Emotion Cause in Conversations with Chain-of-Thought on Emotion States},
  booktitle = "Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics",
  author={Nicolay Rusnachenko and Huizhi Liang},
  year= "2024",
  month= jun,
  address = "Mexico City, Mexico",
  publisher = "Annual Conference of the North American Chapter of the Association for Computational Linguistics"
}

Acknowledgement

This code is referred from following projects: CoT; Flan-T5; Transformers,

License

The code is released under Apache License 2.0 for Noncommercial use only.