TY - GEN
T1 - Densely Connected Bidirectional LSTM with Max-Pooling of CNN Network for Text Classification
AU - Jiang, Qinghong
AU - Zhang, Huaping
AU - Shang, Jianyun
AU - Wesson, Ian
AU - Li, ENlin N.
N1 - Publisher Copyright:
© 2020, Springer Nature Switzerland AG.
PY - 2020
Y1 - 2020
N2 - Text classification is a fundamental task in natural language processing (NLP). Context semantics can greatly improve the accuracy of text classification tasks. Although there are some popular methods in obtaining semantics, current context semantic analysis techniques, due to limited accuracy, are still a great bottleneck for text classification. This paper introduces a novel model, the densely connected Bidirectional LSTM with Max-pooling of CNN network (Dense-BiLSTM-MP), which greatly enhances the context of semantic information. In this model, a densely connected bidirectional long short-term memory (BiLSTM) model, as well as multiple max-pooling layers of convolutional network, are applied to obtain an increasingly enhanced assessment of context, and extract the key features, respectively. Experiments were conducted on four public datasets: YELP, 20NewsGroup, THUNews and AG. The experimental results show that the proposed model outperforms state of the art methods on several datasets. Furthermore, discussions on the Dense-BiLSTM-MP model’s performance in short texts and long texts were given, respectively.
AB - Text classification is a fundamental task in natural language processing (NLP). Context semantics can greatly improve the accuracy of text classification tasks. Although there are some popular methods in obtaining semantics, current context semantic analysis techniques, due to limited accuracy, are still a great bottleneck for text classification. This paper introduces a novel model, the densely connected Bidirectional LSTM with Max-pooling of CNN network (Dense-BiLSTM-MP), which greatly enhances the context of semantic information. In this model, a densely connected bidirectional long short-term memory (BiLSTM) model, as well as multiple max-pooling layers of convolutional network, are applied to obtain an increasingly enhanced assessment of context, and extract the key features, respectively. Experiments were conducted on four public datasets: YELP, 20NewsGroup, THUNews and AG. The experimental results show that the proposed model outperforms state of the art methods on several datasets. Furthermore, discussions on the Dense-BiLSTM-MP model’s performance in short texts and long texts were given, respectively.
KW - Deep learning
KW - Dense structure
KW - Text classification
UR - http://www.scopus.com/inward/record.url?scp=85101881077&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-65390-3_8
DO - 10.1007/978-3-030-65390-3_8
M3 - Conference contribution
AN - SCOPUS:85101881077
SN - 9783030653897
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 98
EP - 113
BT - Advanced Data Mining and Applications - 16th International Conference, ADMA 2020, Proceedings
A2 - Yang, Xiaochun
A2 - Wang, Chang-Dong
A2 - Islam, Md. Saiful
A2 - Zhang, Zheng
PB - Springer Science and Business Media Deutschland GmbH
T2 - 16th International Conference on Advanced Data Mining and Applications, ADMA 2020
Y2 - 12 November 2020 through 14 November 2020
ER -