TY - GEN
T1 - Learning the Chinese Sentence Representation with LSTM Autoencoder
AU - Chen, Mu Yen
AU - Huang, Tien Chi
AU - Shu, Yu
AU - Chen, Chia Chen
AU - Hsieh, Tsung Che
AU - Yen, Neil Y.
N1 - Funding Information:
The authors wish to thank the Ministry of Science and Technology of the Republic of China for financially supporting this research under Contract Grants No. MOST106-2634-F-025-001, MOST 106-2511-S-025-003-MY3, MOST105-2410-H-025-015-MY2, MOST105-2511-S-005-001-MY3, and MOST104-2511-S-005-003.
Publisher Copyright:
© 2018 IW3C2 (International World Wide Web Conference Committee), published under Creative Commons CC BY 4.0 License.
PY - 2018/4/23
Y1 - 2018/4/23
N2 - This study retains the meanings of the original text using Autoencoder (AE) in this regard. This study uses the different loss (includes three types) to train the neural network model, hopes that after compressing sentence features, it can still decompress the original input sentences and classify the correct targets, such as positive or negative sentiment. In this way, it supposed to get the more relative features (compressing sentence features) in the sentences to classify the targets, rather than using the classification loss that may classify by the meaningless features (words). In the result, this study discovers that adding additional features for correction of errors does not interfere with the learning. Also, not all words are needed to be restored without distortion after applying the AE method.
AB - This study retains the meanings of the original text using Autoencoder (AE) in this regard. This study uses the different loss (includes three types) to train the neural network model, hopes that after compressing sentence features, it can still decompress the original input sentences and classify the correct targets, such as positive or negative sentiment. In this way, it supposed to get the more relative features (compressing sentence features) in the sentences to classify the targets, rather than using the classification loss that may classify by the meaningless features (words). In the result, this study discovers that adding additional features for correction of errors does not interfere with the learning. Also, not all words are needed to be restored without distortion after applying the AE method.
UR - http://www.scopus.com/inward/record.url?scp=85085163147&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85085163147&partnerID=8YFLogxK
U2 - 10.1145/3184558.3186355
DO - 10.1145/3184558.3186355
M3 - Conference contribution
AN - SCOPUS:85085163147
T3 - The Web Conference 2018 - Companion of the World Wide Web Conference, WWW 2018
SP - 403
EP - 408
BT - The Web Conference 2018 - Companion of the World Wide Web Conference, WWW 2018
PB - Association for Computing Machinery, Inc
T2 - 27th International World Wide Web, WWW 2018
Y2 - 23 April 2018 through 27 April 2018
ER -