The Attribute Extraction of Freeway Event Texts Using Neural Network

• Main Authors: Liu, Kuo-Yuan, 留國源
• Other Authors: Chang, Ming-Feng
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/htvb72

碩士 === 國立交通大學 === 網路工程研究所 === 106 === Specific traffic event attributes can be used to develop computer applications, such as vehicle navigation and traffic prediction. However, most of today's traffic event messages are in text format, whose attributes are not clearly specified. The test messages must be analyzed to extract all attributes, such as event types, involved vehicles, affected lanes, etc. Our goal is to extract the attributes of traffic events from traffic test messages published by transportation bureau and radio stations. This thesis presents the use of neural networks to extract freeway event attributes. We first developed the regular expression to extract and tag the attributes of each word of the text messages. Since the traffic text messages do not follow specific syntaxes, using regular expression cannot tag the attributes with 100% correctness. Then, we used neural network models, namely, LSTM, BLSTM, and our proposed window model to tag the attribute of each Chinese character of the text messages. The results obtained by the neural network models can be compared with the results of regular expression. If the two are inconsistent and we can determine which is incorrect. The determination must be done by human, and the results can be used to modify regular expressions to improve its tagging correctness. . For the dataset of traffic event text message published by NFB, our window model achieves the same correct tagging rate as BLSTM. The F1 score is 99.99. For the dataset of traffic event text message published by PBS, BLSTM provides the highest correct tagging rate; The F1 score is 99.63. In comparison, the F1 score of the window model is 99.57. The more training data, the larger embedding dimension, the higher F1 score one can get. Using separate embedding table for each location of the characters in the window model also improve the F1 score by 0.11. Compared with the regular expression, the neural network models can handle more complex syntax structure, and have higher correct tagging rate.