Progressive Prompt Generative Graph Convolutional Network for Aspect-Based Sentiment Quadruple Prediction

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper presents an innovative approach with clear motivation and strong experimental results. However, several areas could be improved for clarity, completeness, and rigor. Please consider the following comments and suggestions:

• Why specifically four (quadruples)? Is there any plan to extend this structure to include more elements in future work?
• The figure in the Introduction section should be removed, as it appears in the middle of the text and distracts the reader.
• At the end of the Introduction section, please add a paragraph outlining the organization of the paper.
• Why was a graph-based approach specifically chosen for this work?
• The limitations of the proposed model have not been discussed.
• The abstract is lengthy and densely written. It would benefit from clearer separation between motivation, methodology, and results for better readability.
• The abstract mentions “state-of-the-art performance,” but no comparative metrics or baseline names are provided. Including quantitative improvement figures which would strengthen the claim.

Author Response

1. Responses to the comments of reviewer #1:

This is a nice manuscript on the issue of predicting sentiment elements obtained from a text in the ASQP task (anything from aspects to opinions not explicitily mentioned). Nowadays many researchers are using BERT/T5 approaches for this type of job, and we usually miss global semantic relationship among the 4 elements, just analizing local dependices, and consequently, falling in incorrect semantic transfer (or “limited” capability for reasoning) Authors propose a framework (ProPGCN), including 4 items: a progressive prompt module (that builds semantic representations through a series of prompts), a graph convolutional inference module (modeling syntactic and semantic dependices), a weighed balance loss (balancing the contribution of different prompt orders to stop overfiting), and a constrained decoding (using voting agregation). So, their method integrates progressive prompt learning and graph convolutional reasoning, that apparently provides an accurate method for quduple sentiment prediction.

 Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

The paper presents an innovative approach with clear motivation and strong experimental results. However, several areas could be improved for clarity, completeness, and rigor. Please consider the following comments and suggestions:

Why specifically four (quadruples)? Is there any plan to extend this structure to include more elements in future work?

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

   Sentiment analysis technology (such as Aspect-Based Sentiment Analysis，ABSA  can mine implicit keywords and the emotional information they contain to help businesses better understand user needs.  In order to integrate the results of multiple subtasks of ABSA, researchers proposed the aspect sentiment triplet extraction task (ASTE).  ASTE focuses on extracting explicit aspects and opinions from text information. Although some studies have tried to solve the problem of implicit aspects and implicit opinions, there is still a lack of a unified framework to discuss implicit issues.   In order to solve the problem, relevant researchers proposed the Aspect Sentiment Quad Prediction (ASQP) task to address the problem of implicit aspect and opinion extraction. By refining it into aspect words, opinion words and aspect categories, it can not only capture explicit sentiment information, but also effectively mine implicit sentiment expressions to meet more complex sentiment analysis needs. Its task involves multiple subtasks, including aspect identification, opinion extraction, aspect category detection and sentiment classification.

 

The figure in the Introduction section should be removed, as it appears in the middle of the text and distracts the reader.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content.

 

 

At the end of the Introduction section, please add a paragraph outlining the organization of the paper.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

 

 

Why was a graph-based approach specifically chosen for this work?

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

The ProPGCN model proposed in this paper introduces graph convolutional networks (GCN) mainly based on the following three core motivations:

(1) Modeling syntactic dependencies to capture deep semantic structures

In the aspect-sentiment quadruple prediction task, the identification of implicit aspects and implicit opinions is extremely challenging because they often do not appear directly in the text, but rely on contextual semantics and syntactic structures. We parse the syntactic dependency tree of the sentence using the Spacy tool, construct an adjacency matrix (as described on page 11), and use GCN to model these dependencies. This enables the model to capture long-range dependencies between words, thereby enhancing the ability to understand implicit semantics.

(2) Enhance contextual reasoning capabilities, especially the inference of implicit relationships

Traditional generative models (such as T5) mainly rely on sequence modeling and tend to ignore the structured semantic information within the sentence. GCN enhances the model's ability to perceive contextual information by aggregating information from adjacent nodes (as shown in formulas 12-14 on page 16), thereby more accurately inferring implicit aspect words or opinion words (as shown in the case study on page 27).

(3) Improve the ability to handle complex sentences and multiple sentiment quads

In sentences containing multiple quads or complex sentences, graph structures can more naturally represent the logical relationships between multiple sentiment elements. The Graph Convolutional Relational Inference (GRI) module we proposed (pages 15-16) is precisely designed to use graph structures to correct conflicting implicit logical relationships and improve the accuracy and consistency of quad prediction.

In summary, the graph structure method is not a random choice in this study, but is a key design to make up for the shortcomings of the generative model in structured semantic modeling, especially in processing implicit sentiment elements and complex sentences. Experimental results show (pages 22-24) that the introduction of the GRI module significantly improves the performance of the model on multiple datasets, especially in the recognition of implicit aspects and opinions.

We have elaborated on the design motivation and experimental results of GCN in paper.

 

The limitations of the proposed model have not been discussed.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

While the introduction of the graph convolutional network (GCN) module and step-by-step prompt learning improves performance, it also increases the overall model complexity and training time. Compared to the basic T5 model, our model requires additional syntax parsing, adjacency matrix construction, and GCN forward propagation. In the future, we will explore lighter-weight graph neural network architectures (such as reducing the number of GCN layers) or knowledge distillation techniques to improve model inference efficiency while maintaining performance.

 

 

The abstract is lengthy and densely written. It would benefit from clearer separation between motivation, methodology, and results for better readability.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

The abstract mentions “state-of-the-art performance,” but no comparative metrics or baseline names are provided. Including quantitative improvement figures which would strengthen the claim.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Summary:

This is a nice manuscript on the issue of predicting sentiment elements obtained from a text in the ASQP task (anything from aspects to opinions not explicitily mentioned). Nowadays many researchers are using BERT/T5 approaches for this type of job, and we usually miss global semantic relationship among the 4 elements, just analizing local dependices, and consequently, falling in incorrect semantic transfer (or “limited” capability for reasoning) Authors propose a framework (ProPGCN), including 4 items: a progressive prompt module (that builds semantic representations through a series of prompts), a graph convolutional inference module (modeling syntactic and semantic dependices), a weighed balance loss (balancing the contribution of different prompt orders to stop overfiting), and a constrained decoding (using voting agregation). So, their method integrates progressive prompt learning and graph convolutional reasoning, that apparently provides an accurate method for quduple sentiment prediction.

 

Broad comments:

Strengths:

1. The research is focused on a hot topic in the literature, and the approach is easy to follow, but apparently effective.
2. Using progressive prompt that reduces transmission errors, the GRI module (improving recognition of implicit opinions), is an interesting method. Integrating at the end the top candidate outputs, after balancing the different orders contributions makes complete sense.
3. Author’s proposal is evaluated against 4 benchmark datasets, so researchers’ results seem solid. Generative and BERT’s baselines provide credible performance information. The data about the contribution of each module is useful for other researchers that might want to alter the basic idea and generate an alternative.
4. Potential: beyond ASQP, I consider that this approach makes sense in other NLP scenarios where we have multiple elements.
5. The text is written so that it is easy to follow for any reader.

Weaknesses:

1. I am missing some theoretical explanations for the 3th order prompts performance.
2. I am also missing some information with regards performance in terms of computational cost.
3. While the abstract is (slightly) long, the conclusions and future works are quite condensed.

 

Specific comments:

1. Major issues:
1. Please, describe the hardware environment (other than “A40 GPU”), and the key parameters for training (and training time), inference speed, as standard.
2. Please, provide a rationale for the performance of the third order prompts.
3. Some steps might not be clear for some readers (details on datasets, the progression from local to global)

 

1. Minor issues:
1. Please, clarify:

1. Preprocessing steps, size and train/test splits for the datasets used.
2. How is the progression from local to global in the progressive prompting transitions semantic understanding?

1. Please, add a brief summary of your results in the abstract of the paper.
2. Please, consider summarizing the abstract and providing a more detailed Conclusions section. For Future works I could envision other areas where this approach is going to be useful.
3. Citations:

1. Some references are a little bit outdated (2008, 2016, 2017, 2018, 2019) This topic has a lot of attention in the literature from 2022 up to 2025 to provide more updated information.
2. Please avoid multi-citations, such as in [2-5] (line 55), or [8-12] (line 56): readers want to understand the different alternatives that have been already addressed, and why your research goes further.

1. Figures:

1. Figures 1 and 2 are inserted before it is cited in the text. Figures 3 and 4 are cited 2 pages after insertion.
2. In line 324, “The overall architecture of ProPGCN model proposed in this section is shown in Figure 1”, my guess is that you are referring to Figure 2.

• Figures 4 and 5 are cited as 4.4 and 4.5 in (700).

1. Figure 6 seems to have no citation.
2. In Figures 2 and 3, I think it should be “traditional” instead of “tradional”. Please, check.

1. Small typos:

1. ¿Missing Bold in line 619 for DuSR2?
2. Please, check in the References section, the use of bold for years (there are many missing ones)

• “Tradional” is present many times; unless I am missing something, this should be “traditional”

1. Please, consider adding a paragraph at the end of section 1, describing the structure of the document.

 

Author Response

 

1. Responses to the comments of reviewer #2:

Strengths:

1. The research is focused on a hot topic in the literature, and the approach is easy to follow, but apparently effective.
2. Using progressive prompt that reduces transmission errors, the GRI module (improving recognition of implicit opinions), is an interesting method. Integrating at the end the top candidate outputs, after balancing the different orders contributions makes complete sense.
3. Author’s proposal is evaluated against 4 benchmark datasets, so researchers’ results seem solid. Generative and BERT’s baselines provide credible performance information. The data about the contribution of each module is useful for other researchers that might want to alter the basic idea and generate an alternative.
4. Potential: beyond ASQP, I consider that this approach makes sense in other NLP scenarios where we have multiple elements.
5. The text is written so that it is easy to follow for any reader.

 

Response:

 

Thank you very much for your review comments and suggestions.

Weaknesses:

1. I am missing some theoretical explanations for the 3th order prompts performance.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

The core theoretical motivation for third-order prompts stems from addressing the "semantic gap" between second-order (paired elements) and global quadruples (complete tuples).

Limitations of Low-Order Prompts: First- and second-order prompts primarily learn local binary relations (e.g., [AO] learns the association between aspect and opinion, and [CS] learns the association between category and sentiment). However, quadruplet prediction is a complex, structured prediction task, and its global semantics is not simply the sum of these local relations. Jumping directly to quadruplet generation may make it difficult for the model to reconcile the consistency between all elements.

 

The Bridging Role of Third-Order Prompts: Third-order prompts (such as [ASO] and [AOC]) are designed to serve as a key semantic bridge. They force the model to simultaneously consider and integrate three core elements during the generation process, which is closer to the complexity of the final task but easier to learn and optimize than a complete quadruplet. This is a manifestation of incremental reasoning, mimicking the human cognitive process of gradually constructing complex concepts from parts to the whole. Third-order prompts reduce model generation uncertainty by providing richer context. This mechanism, which gradually constrains the generation space, makes the final quadruple generation more accurate and stable. Experiments show that removing the third-order prompt module (without the third-order prompt) results in either the most significant or less significant performance degradation across all datasets (see Table 3).

 

1. I am also missing some information with regards performance in terms of computational cost.

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

We have clearly listed the hardware environment for training and inference in the paper, which provides a basic reference for computational cost:

GPU: NVIDIA A40 (24GB VRAM)

Number of training rounds: 20 epochs

Batch size: 16

Base model: T5-base (number of parameters: ~220 million)

We have compared the ProPGCN model with several key baseline models (such as Paraphrase (T5) and UGTS) in the following two dimensions: Training time: On the Restaurant dataset, our ProPGCN model takes slightly longer on average than the standard Paraphrase (T5) model to complete 20 epochs of training. This additional cost mainly comes from: (1) the more complex input and output sequences brought about by multi-task prompt learning; (2) the forward and back propagation of the GCN module.

Experimental results show that the ProPGCN model achieves significant performance improvements, but this comes at the cost of a certain amount of increased computational cost. This overhead is primarily invested in two core designs: the progressive hint module, which provides richer semantic supervision signals through multi-task learning; and the graph convolutional reasoning module, which incorporates syntactic structure information to enhance the model's reasoning capabilities. Therefore, this cost-performance trade-off is worthwhile, especially in applications where prediction accuracy is crucial.

In the "Limitations and Future Work" section of the paper, we added research directions focused on efficiency optimization:

We recognize the importance of model efficiency. Future work will focus on optimizing computational costs, for example: exploring pruning of GCN modules; studying strategies to simplify the progressive hinting process during inference; or employing knowledge distillation techniques to extract the knowledge of ProPGCN into a lighter network, aiming to significantly improve efficiency while maintaining performance."

 

1. While the abstract is (slightly) long, the conclusions and future works are quite condensed.

 

  

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

Specific comments:

1. Major issues:

Please, describe the hardware environment (other than “A40 GPU”), and the key parameters for training (and training time), inference speed, as standard.

Please, provide a rationale for the performance of the third order prompts.

Some steps might not be clear for some readers (details on datasets, the progression from local to global)

  

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

Detailed Description of Hardware Environment and Computational Cost

We apologize for not providing sufficient experimental details in our initial draft. We have added the following details in section "4.1. Dataset and Parameter Settings" of the paper:

Hardware Environment:

GPU: 1 × NVIDIA A40 (24GB VRAM)

CPU: Intel Xeon Gold 6330 @ 2.0GHz (specify the number of cores, e.g., 28 cores)

Memory: 128GB RAM

Software Environment:

We used PyTorch (version 1.12.1) and the Hugging Face Transformers library (version 4.20.0) for our implementation.

Training Details and Cost:

Training Time: After 20 epochs of training on the Restaurant dataset, our ProPGCN model requires, on average, significantly less time than the baseline T5 (Paraphrase) model, which does not include our new modules.

Parameter Size: Our model is based on T5-base (~220M parameters). The added GCN and hint modules introduce approximately ~2M trainable parameters. This increase in parameters is negligible compared to the base model, with the main overhead coming from the computational graph.

2. Theoretical Principles for the Performance of Third-Order Hints

We agree that providing a stronger theoretical basis for third-order hints is crucial. We have added a new section in "3.3. Progressive Hint Enhancement Module" of our paper to specifically explain its theoretical motivation. The core argument is as follows:

The core principle of third-order hints is to bridge the "semantic gap" between "local binary relations" and "global quadruple relations" and to act as a structural constraint to reduce uncertainty in the generation task.

Motivation: Addressing the Semantic Gap: First-order ([A], [O]) and second-order ([AO], [CS]) hints learn local or pairwise element-by-element relationships. However, jumping directly from one or two elements to generating a complete quadruple ([ACOS]) is a significant semantic leap, and the model can easily lose sight of complex dependencies between elements or generate logical conflicts along the way. The third-order cues ([ASO], [AOC]) are designed as a key semantic bridge, providing a gradual reasoning step.

Principle: Structural Constraints and Uncertainty Reduction: [ASO] (aspect, sentiment, opinion) forms a complete logical chain of "subject-evaluation-evidence." For example, generating "[AOS] food is great because traditional" forces the model to explicitly link the aspect (A) and opinion (O) through the sentiment (S). This locks in the relationship between the three, greatly enhancing the interpretability and robustness of sentiment judgments.

[AOC] (aspect, opinion, category) forms a categorization chain of "concrete instance-abstract category." For example, generating "[AOC] food is traditional [C] food quality" requires the model to classify the specific aspect-opinion pair (A, O) into its predefined abstract category (C). This is crucial for identifying implicit aspects because it establishes a mapping from the contextual vocabulary to the category space.

Third-order cues significantly reduce the "conditional entropy" of the model when generating the remaining elements by providing richer generated context. For example, after [ASO] is generated, the sentiment (S) is determined, and the uncertainty of the generated category (C) is greatly reduced. This mechanism of gradually constraining the generated space makes the final quadruple generation more accurate and consistent.

Our ablation study (Section 4.4, Table 3) strongly supports this theory. Removing third-order cues leads to a significant performance drop, demonstrating that they are not a dispensable improvement but rather play an indispensable transitional role in semantic construction.

3. Clarity of Method Steps (From Local to Global)

We recognize that the gradual process from local to global is crucial for readers to understand our method. We have taken the following steps to improve clarity:

Added a new diagram: We have added a new diagram (Figure 2) in "3.2. Model Architecture" that clearly depicts the complete and coherent data flow, from input, through data preprocessing -> first-order cues -> second-order cues -> third-order cues -> global semantic modeling -> GRI module -> inference output.

As shown in Figure 2, the model first learns representations of individual elements using first-order cues, then understands pairwise relationships between elements using second-order cues, and then uses third-order cues to capture high-order interactions between the three core elements. Finally, it seamlessly constructs a global semantic representation of the entire quadruple. This progressive design ensures the integrity and consistency of semantic information during the transfer process.

Minor issues:

Please, clarify:

Preprocessing steps, size and train/test splits for the datasets used.

How is the progression from local to global in the progressive prompting transitions semantic understanding?

Please, add a brief summary of your results in the abstract of the paper.

Please, consider summarizing the abstract and providing a more detailed Conclusions section. For Future works I could envision other areas where this approach is going to be useful.

 

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

Citations:

Some references are a little bit outdated (2008, 2016, 2017, 2018, 2019) This topic has a lot of attention in the literature from 2022 up to 2025 to provide more updated information.

Please avoid multi-citations, such as in [2-5] (line 55), or [8-12] (line 56): readers want to understand the different alternatives that have been already addressed, and why your research goes further.

Figures:

Figures 1 and 2 are inserted before it is cited in the text. Figures 3 and 4 are cited 2 pages after insertion.

In line 324, “The overall architecture of ProPGCN model proposed in this section is shown in Figure 1”, my guess is that you are referring to Figure 2.

 

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

• Figures 4 and 5 are cited as 4.4 and 4.5 in (700).

Figure 6 seems to have no citation.

In Figures 2 and 3, I think it should be “traditional” instead of “tradional”. Please, check.

 

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content.

 

 

 

Small typos:

¿Missing Bold in line 619 for DuSR2?

Please, check in the References section, the use of bold for years (there are many missing ones)

 

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content. Likewise, it can be summarized as follows:

 

 

“Tradional” is present many times; unless I am missing something, this should be “traditional”

Please, consider adding a paragraph at the end of section 1, describing the structure of the document.

 

 

Response:

 

Thank you very much for your review comments and suggestions. The revised manuscript includes the relevant content.

 

 

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have carefully reviewed the revised manuscript, and I am pleased to note that the authors have addressed all the concerns highlighted in the first revision. I therefore recommend acceptance of the manuscript in its current form.