Hybrid Curriculum Learning for Emotion Recognition in Conversation Paper Reading 2022 AAAI

Hybrid Curriculum Learning for Emotion Recognition in Conversation

The paper was published in 2022 AAAI, Alibaba team, aiming to use course learning to solve dialogue emotion recognition

motivation:

Motivated by recent studies which have proven that feeding training examples in a meaningful order rather than considering them randomly can boost the performance of models, we propose an ERCoriented hybrid curriculum learning framework.

contribute:

1. We propose a hybrid curriculum learning framework to tackle the task of ERC. At conversation-level curriculum, we utilize an emotion-shift frequency to measure the difficulty of each conversation.
2. We propose emotion-similarity based curriculum learning to achieve utterance-level curriculum learning. It implements the basic idea that at early stage of training it is less important to distinguish between similar emotions compared to separating very different emotions
3. To understand: Early in training, it is more important to recognize different emotions than to distinguish similar emotions. So why can the method of the article achieve this effect?
4. The validity of the experimental results achieves a SOTA

method:

(1) conversation-level curriculum (CC) – we construct a difficulty measurer based on “emotion shift”frequency. The conversations are scheduled in an “easy to hard” schema according to the difficulty score returned by the difficulty measurer.

(2) utterance-level curriculum (UC). it is implemented from an emotion-similarity perspective, which progressively strengthens the model’s ability in identifying the confusing emotions

UC is easy to understand, and it is a conventional course learning method; the motivation of UU is to say that the sentences in a dialogue are logical, and the order cannot be changed. How to solve this problem?

作者assuming that the utterances with confusing emotion labels are more difficult for prediction;
and our utterance-level curriculum is based on the pairwise similarities between the emotion labels.

See below:

First calculate a sentiment matrix:

After normalization, the probability distribution of the target emotion matrix is ​​obtained, and then the loss is calculated:

Compared with the previous model, this loss as I understand it should add an $M_{target}$, but what is the relationship with motivation? I don't quite understand.

The training algorithm for the whole process is as follows:

conclusion of issue:

First, I really don’t have a thorough understanding of motivation 2, why introduce an $M_{target}$, can play a role in curriculum learning? What is the connection between solution and motivation?
Second, for the update of the matrix between lines 9 and 11 in the algorithm, my understanding is to make the diagonal of the matrix gradually approach 1, but the update method is not very clear.

Students who understand it can explain it, thank you!

For more interesting articles, see:
Machine Reading Comprehension Using Reverse Thinking
Evidence Inference Network
Bert Pre-training Model-Chinese Text Classification