Emotion extraction for the speaker in coversation context
Emotion cause, that originates from the speaker of first utterance to the other speaker of the following utterance.

This model choses the answers according to the following list of choices:

["anger", "disgust", "fear", "joy", "sadness", "surprise", "neutral"]

Video Overview

Model Details

Model Description

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

Developed by: Reforged by nicolay-r, initial credits for implementation to scofield7419
Model type: Flan-T5
Language(s) (NLP): English
License: Apache License 2.0

Model Sources [optional]

Repository: Reasoning-for-Sentiment-Analysis-Framework
Paper: https://huggingface.co/papers/2404.03361
Demo: https://github.com/nicolay-r/THOR-ECAC/blob/master/SemEval_2024_Task_3_FlanT5_Finetuned_Model_Usage.ipynb

Uses

Direct Use

Please proceed the following example that purely relies on tranformers and torch.

This example could be found on google colab at the related Github repo page

You can still use the code below for a custom start by being independent from the THoR engine.

Here are the 4 steps for direct model use:

Setup ask method for inferring FlanT5 as follows:

python

def ask(prompt):
  inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
  inputs.to(device)
  output = model.generate(**inputs, max_length=320, temperature=1)
  return tokenizer.batch_decode(output, skip_special_tokens=True)[0]

Setup chain and expected output labels:

python

def emotion_extraction_chain(context, target):
  # Setup labels.
  labels_list = ["anger", "disgust", "fear", "joy", "sadness", "surprise", "neutral"]
  # Setup Chain-of-Thought
  step1 = f"Given the conversation {context}, which text spans are possibly causes emotion on {target}?"
  span = ask(step1)
  step2 = f"{step1}. The mentioned text spans are about {span}. Based on the common sense, what " + f"is the implicit opinion towards the mentioned text spans that causes emotion on {target}, and why?"
  opinion = ask(step2)
  step3 = f"{step2}. The opinion towards the text spans that causes emotion on {target} is {opinion}. " + f"Based on such opinion, what is the emotion state of {target}?" 
  emotion_state = ask(step3)
  step4 = f"{step3}. The emotion state is {emotion_state}. Based on these contexts, summarize and return the emotion cause only." + "Choose from: {}.".format(", ".join(labels_list))
  # Return the final response.
  return ask(step4)

Initialize device, model and tokenizer as follows:

python

from transformers import AutoTokenizer, T5ForConditionalGeneration

model_path = "nicolay-r/flan-t5-emotion-cause-thor-base"
device = "cuda:0"

model = T5ForConditionalGeneration.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model.to(device)

Apply it!

python

# setup history context (conv_turn_1)
conv_turn_1 = "John: ohh you made up!"
# setup utterance.
conv_turn_2 = "Jake: yaeh, I could not be mad at him for too long!"
context = conv_turn_1 + conv_turn_2
# Target is considered as the whole conv-turn mentioned in context.
target = conv_turn_2
flant5_response = emotion_extraction_chain(context, target)
print(f"Emotion state of the speaker of `{target}` is: {flant5_response}")

The response is as follows:

Emotion state of the speaker of Jake: yaeh, I could not be mad at him for too long! is: anger

Downstream Use [optional]

The details of the downstream usage could be found in the related section of the project on Github or within the related notebook on GoogleColab

Out-of-Scope Use

This model represent a fine-tuned version of the Flan-T5 on ECAC-2024 competition dataset of conversations from the F.R.I.E.N.D.S. TV Show. Since dataset represent three-scale output answers ["anger", "disgust", "fear", "joy", "sadness", "surprise", "neutral"] the behavior in general might be biased to this particular task.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

Simply follow the Direct Use secion or proceed with the following GoogleColab notebook.

Training Details

Training Data

We purely rely on the data provided by ECAC-2024 competition organizers.

Here is the related information from the github repository.

And here is the code related to data convesations preparation intented for compiling input data.

Training Procedure

Model has been fine-tuned in two stages:

THoR-state: The first stage aimed at Emotion states prediction
THoR-cause-RR: The second aimed at emotion-causes prediction with reasoning-revision technique.

Training Hyperparameters

Training regime: temperature 1.0, learning-rate 2*10^(-4), AdamW optimizer, batch-size 32, NVidia-A100 (40GB)

[More Information Needed]

Metrics

f1-strict
f1-proportional
f1-strict-weighted
f1-proportional-weighted

[More Information Needed]

Results

Results are depicted in image below in a gray-highlighted row.