Summary
The paper addresses data incompleteness in Multimodal Sentiment Analysis (MSA) by presenting the Language-dominated Noise-resistant Learning Network (LNLN). By considering language as the dominant modality, LNLN introduces a Dominant Modality Correction (DMC) module and Dominant Modality-Based Multimodal Learning (DMML) module to enhance robustness against noise. Extensive experiments under random data missing scenarios on datasets like MOSI, MOSEI, and SIMS demonstrate LNLN’s superior performance. The method ensures quality representations of the dominant modality and offers improved uniformity, transparency, and fairness in evaluations, consistently outperforming existing baselines across various challenging metrics.
Strengths
Strength:
1. The paper addresses the issue of incomplete data in multimodal sentiment analysis, a prevalent and valuable research problem in real-world applications.
2. The manuscript is well-structured, with a clear and logical flow of ideas. The organization facilitates understanding and ensures that the arguments and methodologies are presented in a coherent manner.
Weaknesses
Weakness：
1. The contributions summarized in the introduction are not clearly articulated. Additionally, using language as the primary information carrier to reconstruct incomplete data is a concept frequently employed in multimodal sentiment analysis (MSA). What sets your method apart from other similar approaches?
2. The so-called "Adaptive Hyper-modality Learning" essentially involves using transformers to facilitate interactions between different modalities. This interaction process appears to lack algorithmic innovation and seems to adopt a rather naive approach.
3. The primary structure of the Reconstructor is also based on transformers. Why did you choose transformers for reconstruction?
4. The entire algorithm's architecture heavily relies on transformers, but the transformers used in different modules have not been specifically tailored to the characteristics of this task.
5. In the experimental section, although LNLN demonstrates generally good performance, there is no in-depth comparative analysis with the most similar baseline, ALMT. Since LNLN is an improvement based on ALMT, such a detailed analysis is essential.
6. Ablation studies and other experiments were only conducted on the MOSI dataset. Why were these not performed on all datasets?
7. The experimental section lacks visualizations or case study to demonstrate the effectiveness of the algorithm, making it difficult to fully substantiate the algorithm's efficacy.
Questions
Weakness：
1. The contributions summarized in the introduction are not clearly articulated. Additionally, using language as the primary information carrier to reconstruct incomplete data is a concept frequently employed in multimodal sentiment analysis (MSA). What sets your method apart from other similar approaches?
2. The so-called "Adaptive Hyper-modality Learning" essentially involves using transformers to facilitate interactions between different modalities. This interaction process appears to lack algorithmic innovation and seems to adopt a rather naive approach.
3. The primary structure of the Reconstructor is also based on transformers. Why did you choose transformers for reconstruction?
4. The entire algorithm's architecture heavily relies on transformers, but the transformers used in different modules have not been specifically tailored to the characteristics of this task.
5. In the experimental section, although LNLN demonstrates generally good performance, there is no in-depth comparative analysis with the most similar baseline, ALMT. Since LNLN is an improvement based on ALMT, such a detailed analysis is essential.
6. Ablation studies and other experiments were only conducted on the MOSI dataset. Why were these not performed on all datasets?
7. The experimental section lacks visualizations or case study to demonstrate the effectiveness of the algorithm, making it difficult to fully substantiate the algorithm's efficacy.
Limitations
The algorithm's structure is overly reliant on transformers, and there is a lack of explanation as to why transformers were chosen for various operations, especially in relation to the specific characteristics of this task. Additionally, the paper introduces multiple loss functions; however, the effectiveness and necessity of each loss function are not thoroughly justified or demonstrated. This raises questions about whether each loss is indeed contributing as intended to the overall performance of the model.