Figure - available from: Neural Computing and Applications
This content is subject to copyright. Terms and conditions apply.
Visualization results of the latent space distribution. The top image in Fig. 2 demonstrates the entangled distribution in the latent space for the dataset, and the disentangled distribution after using S-CVAE is shown below

Visualization results of the latent space distribution. The top image in Fig. 2 demonstrates the entangled distribution in the latent space for the dataset, and the disentangled distribution after using S-CVAE is shown below

Source publication
Article
Full-text available
Due to the miscellaneous ambiguity of semantics in open-domain conversation, current deep dialogue models disregard to detect potential emotional and action response features in the latent space, which leads to the general tendency to produce inaccurate and irrelevant sentences. To address this problem, we propose a semantic-aware conditional varia...

Similar publications

Article
Full-text available
The central provocation of this article is to analyze three artworks that operate by inducing a failure or interruption in the functioning of a certain system. Based on Vilém Flusser's apparatus concept, in dialogue with other authors who also mobilized concepts related to devices, machines, and au-tomata to reflect on society, we argue the existen...

Citations

... Some studies proposed to enhance the generalization ability by variational autoencoder (VAE) (Li et al., 2022a). It can learn an ask-related vector (Li et al., 2022b) which can be resampled to produce multiple questions (Wang et al., 2022) based on data distribution. However, one single vector was not sufficient to capture the complex and entangled asking features (Wang et al., 2020b). ...
... Gao et al. [17] introduced a discrete variable with explicit semantics in CVAE, using the semantic distance between latent variables to maintain good diversity among the sampled latent variables. Wang et al. [18] proposed the semantic-aware conditional variant autoencoder (S-CVAE) model. S-CVAE can generate diverse conversational text by utilizing embedded classifiers and feature decoupling modules. ...
Preprint
Full-text available
Currently end-to-end deep learning based open-domain dialogue systems remain black box models, making it easy to generate irrelevant contents with data-driven models. Specifically, latent variables are highly entangled with different semantics in the latent space due to the lack of priori knowledge to guide the training. To address this problem, this paper proposes to harness the generative model with a priori knowledge through a cognitive approach involving mesoscopic scale feature disentanglement. Particularly, the model integrates the macro-level guided-category knowledge and micro-level open-domain dialogue data for the training, leveraging the priori knowledge into the latent space, which enables the model to disentangle the latent variables within the mesoscopic scale. Besides, we propose a new metric for open-domain dialogues, which can objectively evaluate the interpretability of the latent space distribution. Finally, we validate our model on different datasets and experimentally demonstrate that our model is able to generate higher quality and more interpretable dialogues than other models.
... Gao et al. [17] introduced a discrete variable with explicit semantics in CVAE, using the semantic distance between latent variables to maintain good diversity among the sampled latent variables. Wang et al. [18] proposed the semantic-aware conditional variant autoencoder (S-CVAE) model. S-CVAE can generate diverse conversational text by utilizing embedded classifiers and feature decoupling modules. ...
Preprint
Full-text available
Currently end-to-end deep learning based open-domain dialogue systems remain black box models, making it easy to generate irrelevant contents with data-driven models. Specifically, latent variables are highly entangled with different semantics in the latent space due to the lack of priori knowledge to guide the training. To address this problem, this paper proposes to harness the generative model with a priori knowledge through a cognitive approach involving mesoscopic scale feature disentanglement. Particularly, the model integrates the macro-level guided-category knowledge and micro-level open-domain dialogue data for the training, leveraging the priori knowledge into the latent space, which enables the model to disentangle the latent variables within the mesoscopic scale. Besides, we propose a new metric for open-domain dialogues, which can objectively evaluate the interpretability of the latent space distribution. Finally, we validate our model on different datasets and experimentally demonstrate that our model is able to generate higher quality and more interpretable dialogues than other models.
Article
Facial attribute editing is a popular direction in face generation, which aims to modify facial attributes in the face image and remain unedited attributes unchanged. However, generative models are prone to affect the unedited attributes when editing multiple facial attributes. Currently, the concatenation of the prior knowledge with hidden features is still data-driven work. Due to the feature coupling in data-driven models, high-entanglement implicit semantics are generated, which is incomprehensible for human beings. Besides, multi-attribute boundaries of the implicit semantics are ambiguous, which is complicated to effectively control the editing process. In this paper, we propose a knowledge-guided explicit feature disentanglement network that is compatible with human cognition, leveraging a classification method with the prior knowledge to encode features. Specifically, we select 13 facial attribute labels for a comprehensive and explicit presentation of this task and design a knowledge-guided feature disentanglement module to transform the implicit feature representations into explicit feature semantics. We also construct a semantic space that can independently manipulate facial attributes. In addition, our proposed model can be combined with existing facial attribute editing models to obtain multiple variant models. Our proposed model is fully validated by various experiments and the variant model has achieved better performance than the benchmark model in facial attribute editing.
Chapter
Many type 2 diabetes patients and high-risk groups has an increasing demand for specialized information on diabetes. However, the long-tail problem often generate difficulties in model training and reduced classification accuracy. In this paper, we propose enhancing senmantic feature approach to solve the long-tail problem in Chinese diabetes text classification and detailed practice is as followes: we enrich the tail classes knowledge by enhancing semantic features module and then use the attention aggregation module to improve the semantic representation by fusing these semantic features. As for the enhancing semantic feature module, we proposed two strategies: using different dropouts while pre-trained language model is same and using different pre-trained language model. As for the attention aggregation module, its purpose is to better fusing the semantic features obtained previously. After processing by these two modules, we send the final feature vector into the classifier. The final accuracy of 89.1% was obtained for the classification of Chinese diabetes in the NCAA2023 assessment.KeywordsDiabetesText classificationLong tailData augmentation
Chapter
As diabetes has become one of the major public health challenges worldwide, the importance of automatic question-answering services for diabetes in daily healthcare is gradually being highlighted. To address this challenge, this paper aims to improve the accuracy of classifying Chinese diabetes questions. In order to achieve this goal, we utilized BERT pre-training models to extract text feature vectors, combined them with convolutional neural networks to extract local features, and employed pooling and fully connected layers for classification. Additionally, we also tested different BERT models and obtained the best results using voting fusion technology. The final accuracy improved by 2% compared to a single model. Experimental results demonstrate that the model can effectively classify Chinese diabetes questions, providing robust support for the implementation of automated diabetes question-answering services.KeywordsChinese text classificationDiabetes questions classificationBERT fusionConvolutional neural network
Article
Full-text available
Currently end-to-end deep learning based open-domain dialogue systems remain black box models, making it easy to generate irrelevant contents with data-driven models. Specifically, latent variables are highly entangled with different semantics in the latent space due to the lack of priori knowledge to guide the training. To address this problem, this paper proposes to harness the generative model with a priori knowledge through a cognitive approach involving feature disentanglement. Particularly, the model integrates the guided-category knowledge and open-domain dialogue data for the training, leveraging the priori knowledge into the latent space, which enables the model to disentangle the latent variables. Besides, this paper proposes a new metric for open-domain dialogues, which can objectively evaluate the interpretability of the latent space distribution. Finally, this paper validates our model on different datasets and experimentally demonstrate that our model is able to generate higher quality and more interpretable dialogues than other models.