Modeling Aspect Sentiment Quad Prediction as Cloze Task
The aspect sentiment quad prediction(ASQP)task aims to extract all aspect terms along with the corresponding aspect categories,opinion expressions,and sentiment polarities from a given review sentence,providing a holistic understanding through user evaluation of different aspects of a product or service.Existing ASQP methods suffer from the following limitations:(1)Discriminative models fail to capture the semantic relations between sentiment elements using prompts;(2)Generative models either merely combine sentiment element type labels to form prompts,lacking contextual understanding of label semantics,or use a discrete template as input to the decoder,which prevents the encoder from capturing the semantic relations between sentiment elements in the template.To alleviate these issues,this study initially develops two types of prompts,discrete and continuous prompts,based on the cloze-style methodology.These prompts provide a contextual understanding of the meanings of the four sentiment element types and aid in capturing semantic relations between sentiment elements more effectively.Discrete prompts are designed using human prior knowledge,whereas continuous prompts directly add virtual tokens between sentiment elements to represent their semantic relationships and employ prompt-tuning to enable the model to autonomously find the optimal prompt in a continuous semantic space.To enhance the model's capability to autonomously find the optimal prompt in the semantic space,this study designs suitable continuous prompts for all 24 permutations of the four sentiment element types and uses a data augmentation strategy to facilitate cooperative learning among multiple continuous prompts.Subsequently,based on the designed prompts,we propose the C-ASQP framework,which includes the discriminative model DC-ASQP and the generative model GC-ASQP.In DC-ASQP,a two-stage strategy is employed to first extract the aspect category and sentiment polarity from the review sentences,then embed the predicted aspect category and sentiment polarity into the designed prompt,helping the model to extract the corresponding aspect and opinion terms through the label semantics of the aspect category and sentiment polarity and the semantic relationships between all four sentiment elements.In GC-ASQP,the cloze-style designed prompts are concatenated to the review sentences as inputs for the encoder,leveraging the learning patterns of pretrained models to enhance the generation of aspect sentiment quadruples.Moreover,this study explores the effects on the performance of the GC-ASQP model in terms of the order of sentiment elements in the prompts,different decoding strategies,and various multi-prompt data augmentation strategies.Extensive experiments conducted on four widely used datasets show that the DC-ASQP model achieves F1 scores improvements of 4.70％,6.48％,6.97％,and 2.60％,respectively,compared to the best-performing discriminative models.In comparison to the top baseline model utilizing a single prompt(template),the GC-ASQP model based on a single discrete prompt improves F1 scores by 2.20％,1.80％,1.26％,and 0.96％,respectively.Utilizing a data augmentation strategy,the F1 scores of the GC-ASQP model with 15 continuous prompts outperforms the state-of-the-art by 0.86％,1.67％,0.15％,and 1.02％,respectively.These results not only validate the effectiveness of modeling ASQP as cloze tasks but also prove the efficacy of the designed two types of prompts and the C-ASQP framework.
aspect sentiment quad predictioncloze taskdiscrete and continuous promptsdiscriminative and generative modelC-ASQP framework
万方数据
维普
