This study investigates an approach to dialogue act classification leveraging a pre-trained model, with a specific focus on the efficacy of employing the ERNIE model for this task. Dialogue act classification is crucial for deciphering the intentions, actions, and objectives underlying conversations. In this research endeavor, we selected the ERNIE model as our pre-trained backbone, augmented it with fine-tuning techniques, and synergistically incorporated it with an RCNN architecture to achieve precise classification of dialogue acts. Through a series of experiments, we rigorously assessed the model's performance using both publicly available and proprietary datasets, comparing it with conventional methodologies and alternative deep learning frameworks. Our findings revealed that the proposed dialogue act classification methodology, anchored in the ERNIE model and RCNN integration, yielded notable improvements in accuracy and generalization capabilities. This underscores the prowess of the ERNIE model in dialogue act classification tasks, offering new insights and methodologies for analyzing dialogue text. Subsequent research avenues will delve into exploring more intricate model architectures and harnessing richer data reservoirs to further elevate the performance and applicability spectrum of dialogue act classification.
In today's society, the demand for telecom packages is increasing and at the same time, there are many different types of telecom packages, so we need to recommend different types of telecom packages to different users so that each user can find the most suitable package for themselves among the many packages. We choose to build a click-through prediction model to model the interaction between user characteristics and telecom package characteristics. In this paper, we propose the FGCIN model, which generates new features by capturing important feature interactions through a convolutional neural network FGCNN, and then interacts the new features with the original features at the feature interaction layer through a compressed interaction network CIN for higher order features, followed by a deep neural network DNN for implicit feature interactions to finalize the output. In this paper we use the private dataset Telecom dataset and the public dataset Criteo for comparison experiments and ablation experiments, thus demonstrating the effectiveness and rationality of our proposed model FGCIN.
Recently, deep learning is widely used in the field of click-through rate prediction. However, deep neural networks can only perform implicit feature interaction on features, which is unknowable. Relying solely on DNN for feature interaction may fail to uncover more information behind the data. At the same time, many click-through prediction models treat all features equally when performing feature interactions. Nevertheless, in a real production environment, some features have a significant impact on the prediction results, while others can be neglected. In this thesis, we proposed the FiDCN model (Feature Importance and Deep Cross Network). It can learn the importance of input features dynamically by giving distinct weights to different features. What’s more, the model also introduces low-rank techniques to capture higher-order feature cross efficiently. We conducted detailed comparison experiments with the classic models in the industry and ablation experiments on different datasets, and the results showed that FiDCN has good performance.
As the most important module in recommendation systems, click-through rate prediction has attracted the attention of industry and academia. Due to the powerful learning ability of deep learning, it is widely used in click-through rate prediction. Behavior sequences based on user is an important direction of click-through rate prediction. Although some results have been made in related directions, existing methods still have some problems, such as the inability to learn feature weights better, the presence of noise in user behavior sequences, not fully mining the hidden information in features, etc. In this paper, we propose a method for related problems, named DISFMN, which can dynamically learn the importance of features as well as filter out the noise in user behavior sequences. The method also combines high-order and low-order feature interactions to uncover more valuable information in features. Comparative experiments are conducted on different datasets and the experimental results showed the effectiveness of the proposed method.
Click-through rate prediction, which aims to predict the probability of a user clicking on an advertisement or product, is widely used and important in advertising and recommendation systems. Various click-through prediction models proposed in recent years enumerate all cross features at a predefined maximum order and then train the model to identify useful feature interactions, but there are a number of problems with this approach. First, the complexity of the model is proportional to the order, so there is a trade-off between expressiveness and computational cost. Second, at the maximum order, the introduction of some noisy crossover features can degrade the model performance. Third, implicit higher-order feature interactions are poorly interpretable and lack convincing reasons to explain the model results. In this paper, we propose MAFN, which explicitly models feature combinations of different orders using multi-headed self-attentive networks with different levels of residual connectivity. At the same time, we introduce an adaptive factorization network to learn crossover features of arbitrary order. Extensive experiment evaluations show that MAFN performs well compared to existing state-of-the-art models.
In a three-dimensional ultrasound computed tomography (3D USCT) system, system errors such as transducer delay, transducer position deviation and temperature error will affect the quality of reconstructed images. Most of the existing calibration works use iterative methods to solve large-scale systems of linear equations. In our case, the transducer delay and position deviation calibration problem of the considered 3D USCT system is essentially to solve a linear system containing about 840,000 equations and 11,500 unknowns. For such a large system, the existing iterative methods require a lot of computation time and the accuracy also needs to be improved. Considering that neural networks have the ability to find optimized solutions for large-scale linear systems, we propose a neural network method for transducer delay and position deviation calibration. We designed a neural network to calibrate both delay and position solutions, together during the network training. We test the method with simulated system data where we add transducer delays in the range of 0.7~1.3 μs, position deviation in the range of -1~1 mm for the X- and Y-axis, and -0.3~0.3 mm for the Z-axis. Results show that the mean delay error is reduced to 0.15 μs, and the mean position error is reduced to 0.15 mm, after a neural network calibration process which takes about 11 minutes. The delay calibration result is better than the existing Newton method in literature, while our method is especially less time-consuming.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.