CSE 556: Natural Language Processing Assignment 4 Solution

General Instructions:
● Every assignment has to be attempted by four people. At least one subtask has to be done by one team member. All members need to have a working understanding of the entire code and assignment.
● Create separate .ipynb or .py files for each part. The file name should follow the format: “A4_<Part number>.ipynb/.py”
● Carefully read the deliverables for all tasks. Along with the code files, submit all the other files mentioned for each task, strictly following the naming convention instructed.
● Only one person has to submit the zip file containing all the mentioned files and the report PDF. It will be named “A4_<Group No>.zip”. The person with the alphabetically smallest name should submit it.
● You are required to submit your trained models. You must also retain all your checkpoints and load and run them during the demo.
● Your report must include the details of each group member’s contribution.


This assignment is along the lines of the paper “Discovering Emotion and Reasoning its Flip in
Multi-Party Conversations using Masked Memory Network and Transformer”
Datasets
The paper proposed a new dataset known as the MELD-FR for the novel task. We have modified the datasets that can be used to split for the train and val.
TASK 1 - ERC (Emotion Recognition in conversation)
What is ERC ?
Emotion Recognition in Conversation (ERC) is a specialized field that focuses on automatically identifying and interpreting the emotional states expressed by individuals during conversations. Unlike traditional approaches that analyze emotions in isolated text, ERC aims to understand the nuanced emotional dynamics in conversational exchanges involving multiple speakers. It enables a deeper understanding of emotional processes and interactions, benefiting applications such as conversational agents and affective computing systems.
What are you expected to do?
1. Train 2 models M1 & M2 with independent architectures which should be different from the architecture mentioned in the paper given above. (Note : If you are re-implementing the paper’s model architecture as one of the models, make sure that the papers results will be considered as the baselines and the other model’s scores should be at par with the one’s given in the paper)
2. Both these models should focus on emotion identification and also detecting the flips in emotion within a set of conversations. Reasoning of the shift is not important for this part of the assignment.
3. Submit both model checkpoints M1 and M2, and report the better model with proper reasoning. The same will be tested on the testing data which will be provided at the time of demos.
Deliverables
1. Two model checkpoints M1 and M2 in proper format.[5*2=10]
2. Well labeled model architectures used for both M1 and M2.[5*2=10] (If implementing the paper’s model, ignore this deliverable for only one model but note the baselines.)
3. If you are implementing the paper’s architecture, then the other model should give results around at most 5% less than the paper’s results. It can defeat the paper’s results.
[5 for comparable result on new model,5 for model architecture of of new model]
4. Properly mention which of the two architectures was better and why.[5]
5. Proper report explaining the intuition behind the models, splits and everything relevant.[5]
6. Add train loss and val loss vs epochs plots for each model in the report. [2.5*2*2=10]
You can choose either point 2 or 3.
TASK 2 - EFR (Emotion Flip Reasoning)
What is EFR ?

A snapshot from the paper itself

In the above mentioned example note the emotional flip in a set of conversations between multiple individuals. The task is to note the emotional shift. The purple arrow in the example
shows the switch in the emotion from Joy to Neutral and sadness to neutral in the second case.
What are you expected to do?
4. Train 2 models M3 & M4 with independent architectures which should be different from the architecture mentioned in the paper given above. (Note : If you are re-implementing the paper’s model architecture, make sure that the papers results will be considered as the baselines and your scores should be at par with the one’s given in the paper)
5. Submit both model checkpoints M3 and M4, and report the better model with proper reasoning. The same will be tested on the testing data which will be provided at the time of demos.
Deliverables
1. Two model checkpoints M3 and M4 in proper format.[5*2=10]
2. Well labeled model architectures used for both M3 and M4.[5*2=10] (If implementing the paper’s model, ignore this deliverable for only one model . For the other model you will have to but note the baselines.)
3. If you are implementing the paper’s architecture, then the other model should give results around at most 5% less than the paper’s results. It can defeat the paper’s results.
[5 for comparable result on new model,5 for model architecture of of new model]
4. Properly mention which of the two architectures was better and why.[5]
5. Proper report explaining the intuition behind the models, splits and everything relevant.[5]
6. Add train loss and val loss vs epochs plots for each model in the report. [2.5*2*2=10]
You can choose either point 2 or 3.