卷积神经网络相关外文文献翻译中英文2020
英文
Social media sentiment analysis through parallel dilated convolutional neural network
for smart city applications
Muhammad Alam, FazeelAbid, etc
Abstract
Deep Learning is considered to leverage smart cities through social media sentiment analysis. The digital content in social media can be used for many smart city applications (SCAs). Classical convolutional neural networks (CNNs) are challenging to parallelize and insufficient to capture long term contextual semantic features for sentiment analysis. In this perspective, this paper initially proposes a domain-specific distributed word representation (DS-DWR) with a considerably small corpus size induced from textual resources in social media. In DS-DWR, different Distributed Word Representations are concatenated to builds rich representations over the input sequence, which is worthwhile for infrequent and unseen terms. Second, a dilated convolutional neural network (D-CNN) , which is compos
resources翻译ed of three parallel dilated convolutional neural network (PD-CNN) layers and a global average pooling (GAP) layer. Our considered parallel dilated convolution reduces dimension and incorporates an extension in the size of receptive fields without the loss of local information. Further, the long-term contextual semantic information is achieved by the use of different dilation rates. Experiments demonstrate that our architecture accomplishes comparable results with multiple hyperparameters tuning for better parallelism which leads to the minimized computational cost.
Keywords：Social media sentiment analysis，Smart city applications，Parallel dilated convolutional neural network (PD-CNN)，Domain-specific distributed word representation (DS-DWR)
Introduction
The development of smart city applications (SCAs) in the last decade has shown flexible design and valuable services. Smart Cities are a combination of technical and social advancements, e.g., Internet access on Mobile, Social Data Analytics, and
Internet of Things (IoT) [1]. These advancements can facilitate to manage resources productively primarily through social networks. Our study endorses a novel way for the smart cities towards the building of intelligent and adaptable SCAs. The social media has become a potential information sourc
e for social information mining to dominat e people’s opinions. In our perspective, users can be viewed as social sensors, and opinions are response signals as it is real-time characteristics of social data. These opinions are extensive on social media consist of short sentences and contain useful information in numerous contexts. Therefore, social media data as social sensors can be employed to discover valuable information. A developing pattern in the area of social data mining is to concentrate on social textual data having opinions instead of concentrating on extensive numerical analysis. Social media sentiment analysis is a growing technique to comprehend the opinions of individuals through social networks. While SCAs can utilize the benefits of social media sentiment analysis as the opinions can be related to any event or a source of estimating preferences, dislikes, and patterns.
SCAs can act as “activators” by focusing on improvements in technological, political, and organizational aspects through social media. Also, it can enhance people’s at titudes and empower them to build up a feasible environment in Smart cities [2]. Social networks are the primary source of opinions and events; however, the main issue is substantial social content that requires smart approaches to filter out noisy data. The prerequisite of noise filtering of social media data involves automatic classification techniques for valuable facts. It also suffers a few difficulties; for example, sentences are short and contain abbreviations as well as spelling mistakes. The semantic analysis can
be used to manage abundant social media data to induce semantic for improvements in terms of integration and reusability [3]. Similarly, social media mining techniques can utilize to collect and process social data posts on Facebook, Twitter, and Instagram for SCAs [4]. We choose Twitter, which has a considerable number of posts and consists of less than 140 characters [5]. This social media data can be used for social media sentiment analysis and considered appropriate to examine people’s o pinions of smart city services.
Currently, Machine Learning (ML) methodologies utilized IoT and Big Data for the enhancements of smart city services [6]. Among many ML methods on classifying text, Naıve Bayes (NB) is utilized for spam detection [7], topic detection [8], sentiment analysis [9], recommendation systems [10]. In a social context, popular techniques support vector machine (SVM) has been utilized to characterize tweets to get the trends [11]. However, there are some issues related to the extraction of the features with variable-size sequences composition. Whereas a subfield, deep Learning (DL), favorably provides intelligent services. DL employs artificial neural networks (ANNs) architectures in order to learn high-level features for social data with multiple layers. These features are efficient enough to replace many ML methodologies, especially when addressing social media data for predictions. Further, these methodologies are progressively applied to supervised and unsupervised classification problems [12], [13], [14].
Deep learning (DL) persuades researchers by using ANNs that empowers efficient learning of features without requiring complex feature-engineering [15], [16]. DL methodologies work by executing the feature extraction and classification task through initiating the representation of a sequence of words via multiplying with associated weight matrix as a one-hot vector [17]. This sequence of each word is formulated by continuous vector space inputted to a neural network to process the succession of words with the assistance of various layers for predication. This processing has an impact on the training set towards increasing the classification accuracy, as explained in [18]. Convolutional neural network (CNN) has accomplished excellent outcomes of sentence-level classification [19], [20]. It can learn from different distributed word representations such as Word2Vec [21], FastText [22] and GloVe [23] by projecting the words into lower dimensions with dense space vector. A technique of CNN to extract the features of the sentence and joined these features with handcraft-features for the relation classification proposed in [24]. However, traditional CNNs are unable to hold long term semantic features. Comparatively, a unique design of CNN is the utilization of dilated convolution. It eliminates the implications of information loss, which is due to conventional
down-sampling approaches in traditional pooling operations and likewise in the stride convolution. Further, it can expand the size of receptive fields at the exponential level without concerning additional
parameters. Thus, it becomes feasible for the dilated convolution to capture long term semantic information.
Although there are numerous SCAs, still it appears as a collection of loosely coupled relation with social media, which is need to be strengthened. The current research is utilizing a deep learning methodology that automatically identifies information on social sensors by considering a parallel dilated convolutional neural Network (PD-CNN). From the best of our review, PD-CNN is capable of predicting appropriate information as it learns features from DS-DWRs using different distributed word representations, as mentioned above. Beyond the improvement in SCAs, several favorable hyper-parameters while utilizing DL for social media sentiment analysis are considered. Our work not only introduces the new way of social media sentiment analysis but also useful to modernize smart cities. The following are the contributions of this work:
•The generation of domain-specific distributed word representation is explored to discover the features for the enhanced exhibition of the classifier in comparison with standard methods.
•The parallelism in the dilated convolutional neural network contains thre e dilated convolution layers containing different dilation rates with global average pooling in a novel manner.
•Lastly, without the complex feature-engineering, the work proceeds with the employment of the DL method to provide intelligent Smart city applications through social media sentiment analysis based on the social sensor (Tweets) to augment the people’s perception in smart cities.
Related work
Smart City and social media analytics continuous growth contributed to an advanced level in technological and urban advancements. Only a few studies in the last years clarify the complementary characteristics of the social-technical ecosystem. However, their mutual significance is not yet wholly idealized [25]. The works on
sustainability which explain the smart cities by relating innovations to social networks explained in [26], [27]. Social users of smart city services are well informed about the availability and productivity of these services, as described in [28]. It Implies the requirement for more practicality in smart cites research regarding social networks. For the sustainability in smart cities research, one has to concentrate on the social context related to smart cities [29]. In this Way, smart city applications (SCAs) prompt the integration of social networks and smart cities services having opinions and prospects, as explained in [30], [31]. Thus, social networks have a profound impact on SCAs. Social n
etworks contain informal methods of interactions and opinions. We attempt to analyze how these opinions contribute to promote the mentioned services. Further, these networks allowing to share status update messages. These messages have additional (meta) information, e.g., name, time, location, and hashtags. It can be an essential source that can help to provide intelligent services in smart cities. Subsequently, social network users considered as social sensors and messages as sensor information utilized in many works [32], [33], [34].
Further, “Sensing” becomes ubiquitous with the utilization of smartph ones, which can effectively collect and analyze information in the form of audio, video, and text [35]. The data produced by these sensors are semantically unique when contrasted with the data generated by the physical device. This semantic data is correlative information in smart cities. Numerous works utilizing short textual messages on social networks to analyze multiple perspectives have been proposed, including event Detection [36], [37] disease tracking, and monitoring [38]. One of the significances of textual data based on social networks is semantics, which is more natural for social users to understand. Among various social network platforms, Twitter data has been extensively utilized in diversified applications such as Natural Language Learning (NLP) [39], Sentiment analysis [40], and Smart cities [41].
In order to boost twitter-based sentiment classification by utilizing meta-level features, for example, str
ength and emotion factors indicating the polarity has been proposed in [42]. A crowd-sourcing approach of manual tagging data process on Twitter with entities such as Person, Organization, and Location proposed by [43].