Databases and tabular data are among the most common and rapidly growing resources. But many of these are poorly annotated (lack su cient metadata), and are lled with domain speci c jargon and alpha-numeric codes. Because of the domain speci c jargon, no pretrained language model could be applied readily to encode the cell content. We propose a deep learning based framework, TableCNN, that encodes the semantics of the surrounding cells to predict the meaning of the columns. We propose application of Byte Pair Encoding (BPE)[5] to create tokens for each cell and treat each cell as a phrase of existing tokens. Once tokenized, we process it with a CNN network to develop a classi er.
Cell entries are tokenized using Byte-Pair Encoding with a stopping condition
de ned by the token frequency threshold. Cell embedding is generated using
Word2Vec[
Network: TableCNN Fig 1 is an abstract view of the TableCNN network architecture. We extract column and row features in separate networks and combine them to regress nal output. Row and column features are computed by a convolution operation over the rst row and the target column of micro table respectively. Outputs from row and column features are concatenated and fed to a fully connected layer with a SoftMax layer to make nal prediction.
For our experiment, we use a manufacturing database table that contains 112 columns and 115 thousand rows. Fig 2 shows that we obtain correct column predictions, except for column 49. Upon further inspection, we found that most of the cells in column 49 were empty, and thus the loss of classi cation accuracy. This shows that network is able to learn features from the surrounding columns along with the target column entries.
In this poster, we present a supervised learning framework that can classify columns of a table with arbitrary alpha-numeric data. The arbitrary alphanumeric nature of data prevents us from using pre-trained language models.