Cluster-based Table Detection Dataset

This data set contains PDF segments and document features, combined with a label whether a segment is part of a table or not.

The contained features are:

• file: Corresponding PDF file name
• page: Page where the cluster is located, starting with 0
• bbox: Bounding box of the cluster, stored as (x_0,x_1,y_0,y_1)
• text: This information had to be removed because it can be confidential.
• n_nodes: Number of layout elements in the cluster
• approx_size: Approximate number of cells when assuming a tabular cluster structure
• tabular_fill_score: Percentage of filled cells. This is calculated by building an artificial grid over the cluster and set the cells which would be filled relative to the maximal possible number, namely the approx_size.
• loop_score: Percentage of loops present in a cluster, relative to the maximal possible number.
• rectangle_score: Percentage of unique rectangles in a cluster, relative to the maximal possible number (which would be one rectangle per element)
• x_sparsity_abs: Average length of horizontal edges in a cluster
• x_sparsity_rel: Average length of horizontal edges in a cluster (i.e. x_sparsity_abs), relative to the horizontal sparsity the same page
• font_size_entropy: Shannon entropy of the font sizes in a cluster
• font_name_entropy: Shannon entropy of the font names in a cluster
• bold_pct: Percentage of bold texts in a cluster
• italic_pct: Percentage of italic texts in a cluster
• font_size_entropy_doc: Shannon entropy of the font sizes in a document
• font_name_entropy_doc: Shannon entropy of the font names in a document
• bold_pct_doc: Percentage of bold texts in a document
• italic_pct_doc: Percentage of italic texts in a document
• font_size_entropy_diff: Deviation of the font size entropy of a cluster (i.e. font_size_entropy) compared to the corresponding measurement on document-level (i.e. font_size_entropy_doc)
• font_name_entropy_diff: Deviation of the font name entropy of a cluster (i.e. font_name_entropy) compared to the corresponding measurement on document-level (i.e. font_name_entropy_doc)
• bold_pct_diff: Deviation of the percentage of bold texts in a cluster (i.e. bold_pct) compared to the corresponding measurement on document-level (i.e. bold_pct_doc)
• italic_pct_doc_diff: Deviation of the percentage of italic texts in a cluster (i.e. italic_pct) compared to the corresponding measurement on document-level (i.e. italic_pct_doc)
• is_table: Label indicating with 1 that a cluster contains pure table content and 0 otherwise