Fine-Grained Emotion Recognition with In-Context Learning: A Prototype Theory Approach

1. In the description of Figure 1 (lines 88-90), the author does not introduce in detail which method is used to select the emotional prototype closest to the query. In addition, in lines 81-82, the author did not specifically explain why the existing methods would query irrelevant prototypes. It is recommended to add additional explanations to better reflect the limitations of the existing methods and the motivation of this article.
2. Some of the illustrations in Figure 1 are slightly rough. For example, the fonts of subtitles (a) and (b) are too large and inconsistent with other fonts.
3. In the selection of baseline models, although some common models and methods are selected as comparisons, there may be some other advanced emotional recognition methods that have not been compared. For example, some of the latest large language models and some fine-tuning methods are mentioned in the following literature. [1] Liu Z, Yang K, Xie Q, et al. Emollms: A series of emotional large language models and annotation tools for comprehensive affective analysis[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024: 5487-5496.
4. The authors need to give the recognition accuracy and F1 value of each method for each emotion category on the dataset to further illustrate the performance of the method. I need to clearly understand in which categories the performance of the proposed method is improved.
5. The paper simply explains the connection between ICL and prototype theory (lines 90-92) and proposes an improved method based on this. The theoretical support for its connection needs to be explained in the method or appendix.

1. There are some typos in the paper.
2. The paper primarily compares against ICL and zero-shot methods, which might not fully showcase E-ICL’s comparative strengths. There should be some methods with improved ICL prompts for comparison. It is not fair to directly compare zero-shot prompt with well-designed prompts.
3. The authors should include more LLMs to verify the effectiveness of the proposed E-ICL methods.

While the paper has some merit and the method is interesting, I believe there are significant flaws in both presentation and significance.

• The presentation of the method (i.e., Section 4) needs significant improvements. Notation is inconsistent and unintuitive, with various errors. I found it very hard to understand this section. For example, in equation 5 we subtract a vector from a scalar; there’s a sum over k1, but j is not properly defined; these probability distributions seem not to be probability distributions, i.e., they are not normalized. In L246 m_{i} \in n_{d} but in the very next sentence n_{d} is a scalar. There’s a notation of both p^{s}{m{i}} and p^{s}_{i}. Which one is it? In L205, V is presented as an emotion vector and it’s mentioned that V \in R^{768}. What is this 768? The reader may not be familiar with all embedding sizes for different types of models.
• The experimental setup is limited: only Claude-haiku and ChatGPT-turbo are not enough baselines to offer confidence that the method is widely applicable. Additionally, there are no baselines shown besides the RoBERTa.
• Most importantly, it is unclear to me why the RoBERTa model trained in-domain is not shown. In the method proposed here, the model has access to the training set and utilizes the GT (Eq 5). Therefore it’s critical to show this comparison where you train RoBERTa on the training data. Unfortunately, as shown in the papers introducing the datasets considered here, it seems that traditional language models such as BERT outperform E-ICL. Additionally E-ICL has significant inference-time costs that are not discussed here.
• The method is cumbersome: parameters such as k1, k2, k3, \alpha need to be tuned.

Missing citations:

Suresh et al., 2021 Not All Negatives are Equal: Label-Aware Contrastive Loss for Fine-grained Text Classification