Some doubts
Some doubts:
‘emotion.csv’ does not exist. Is this file the output of a “ Emotion Model”? Is the Benchmark training code for this model available? Original Do I need to use the original dataset ‘emotion’ to train it myself?
What kind of ‘emotion model’ was given to derive the data in the paper. 2. Can I reproduce the data from the paper?
I noticed that the path structure in the code does not match with the folder structure of the download from huggingface, do I need to manually modify my own file structure.
Looking forward to your reply!
- Yes, this is the output of the emotion model, and the benchmark code includes the training part of the emotion model. You can use this to train an emotion model and output the results. You can create the 'emotion. csv' by yourself, but the effect based on this result is not good (as mentioned in the paper, the improvement in fraud detection is very small or even negative). We encourage everyone to use richer methods to enhance deception detection tasks or other tasks through emotional representation.
- As mentioned in the paper, the results are obtained by inputting multimodal features into the fully connected layer. We encourage everyone to use richer emotion models and methods to create emotion models. You should be able to reproduce the data in the paper, which may have some randomness, but the results will not be too different.
- If the path does not match, you can certainly manually modify the file structure.
Thank you for your reply!
We did not find the benchmark code containing the training part of the sentiment model.
We found that ‘load_data’ is only run when the emotion.csv exists. The main always contains the ‘load_data’ function.
When args.task=‘emotion’ and args.use_emotion==True, the emotion.csv file will be read.
May I ask what is the problem?
Looking forward to your reply.
In fact, emotion. csv here is a vector containing [193 * feature_dimension], and its source is:
Input each sample into the emotion model and extract the last layer of features as emotion features, which are used as deception detection tasks (similar to personality traits, but personality traits are existing scale scores, while emotion features are emotion embeddings obtained through the model). You can train this emotion model by yourself or use open-source models such as emotion2vec.
Both unimodal and multimodal are acceptable.