text
stringclasses 10
values | source
stringclasses 1
value |
|---|---|
Pub Tables-1M: Towards comprehensive table extraction from unstructured documents Brandon Smock Rohith Pesala Robin Abraham Microsoft Redmond, WA brsmock,ropesala,robin. abraham@microsoft. com Abstract Recently, significant progress has been made applying machine learning to the problem of table structure inference and extraction from unstructured documents. However, one of the greatest challenges remains the creation of datasets with complete, unambiguous ground truth at scale. To ad-dress this, we develop a new, more comprehensive dataset for table extraction, called Pub Tables-1M. Pub Tables-1M contains nearly one million tables from scientific articles, supports multiple input modalities, and contains detailed header and location information for table structures, making it useful for a wide variety of modeling approaches. It also addresses a significant source of ground truth inconsistency observed in prior datasets called oversegmentation, using a novel canonicalization procedure. We demonstrate that these improvements lead to a significant increase in training per-formance and a more reliable estimate of model performance at evaluation for table structure recognition. Further, we show that transformer-based object detection models trained on Pub Tables-1M produce excellent results for all three tasks of detection, structure recognition, and functional analysis without the need for any special customization for these tasks. Data and code will be released at https://github. com/microsoft/table-transformer. 1. Introduction A table is a compact, structured representation for storing data and communicating it in documents and other manners of presentation. In its presented form, however, a table, such as the one in Fig. 1, may not and often does not explicitly represent its logical structure. This is an important problem as a significant amount of data is communicated through doc-uments, but without structure information this data cannot be used in further applications. The problem of inferring a table's structure from its pre-sentation and converting it to a structured form is known as Figure 1. An example of a presentation table whose underlying structure must be inferred, either manually or by automated sys-tems. table extraction (TE). TE entails three subtasks [6], which we illustrate in Fig. 2: table detection (TD), which locates the table; table structure recognition (TSR), which recognizes the structure of a table in terms of rows, columns, and cells; andfunctional analysis (FA), which recognizes the keys and values of the table. TE is challenging for automated sys-tems [9, 12, 17, 23] due to the wide variety of formats, styles, and structures found in presented tables. Recently, there has been a shift in the research litera-ture from traditional rule-based methods [4,11,18] for TE to data-driven methods based on deep learning (DL) [14,17,22]. The primary advantage of DL methods is that they can learn to be more robust to the wide variety of table presentation formats. However, manually annotating tables for TSR is a difficult and time-consuming process [7]. To overcome this, researchers have turned recently to crowd-sourcing to con-struct larger datasets [9,22,23]. These datasets are assembled from tables appearing in documents created by thousands of authors, where an annotation for each table's structure and content is available in a markup format such as HTML, XML, or La Te X. While crowd-sourcing solves the problem of dataset size, repurposing annotations originally unintended for TE and automatically converting these to ground truth presents its own set of challenges with respect to completeness, consis-tency, and quality. This includes not only what information is present but how explicitly this information is represented. 1ar Xiv:2110. 00061v3 [cs. LG] 18 Nov 2021
|
tatr.pdf
|
Figure 2. Illustration of the three subtasks of table extraction addressed by the Pub Tables-1M dataset. For instance, markup annotations do not encode spatial coor-dinates for cells, and they only encode logical relationships implicitly through cues such as layout [20]. Not only does this lack of explicit information limit the range of supervised modeling approaches, it also limits the quality control that can be done to verify the annotations' correctness. Another significant challenge for the use of crowd-sourced annotations is that these structure annotations en-coded in markup often exhibit an issue we refer to as over-segmentation. Oversegmentation occurs in a structure anno-tation when a spanning cell in a header is split into multiple grid cells. We give examples of this in Fig. 3. Oversegmenta-tion in markup usually has no effect on how a table appears due to borders between cells being invisible, leaving a pre-sentation table's implicit logical structure and interpretation unaffected. However, oversegmentation can lead to signif-icant issues when used as ground truth for model training and evaluation. The first issue is that an oversegmented annotation contra-dicts the logical interpretation of the table that its presenta-tion is meant to suggest. For instance, oversegmenting a cell annotation may indicate that its text applies to only one row when its presentation form suggests its text is meant to apply to several rows, as in the cell in column 1, row 3 in Fig. 3. This is problematic for use as ground truth to teach a ma-chine learning model to correctly interpret a table's structure. Even if oversegmented annotations were considered a valid intepretation of a table's structure, allowing them would and does lead to ambiguous and inconsistent ground truth, due to there then being multiple possible valid interpretations for a table's structure, such as in Fig. 3. This violates the standard modeling assumption that there is exactly one correct ground truth annotation for each table. Thus, datasets that contain oversegmented annotations in them lead to inconsistent, con-tradictory feedback during training and an underestimate of true performance during evaluation. To address these and other challenges, we develop a new large-scale dataset for table extraction called Pub Tables-1M. Pub Tables-1M contains nearly one million tables from scien-tific articles in the Pub Med Central Open Access1(PMCOA) database. Among our contributions: Pub Tables-1M is nearly twice as large as the current largest comparable dataset and addresses all three tasks of table detection (TD), table structure recognition (TSR), and functional analysis (FA). Compared to prior datasets, Pub Tables-1M contains richer annotation information, including annotations for projected row headers and bounding boxes for all rows, columns, and cells, including blank cells. It also includes annotations on their original source documents, which supports multiple input modalities and enables a wide range of potential model architectures. We introduce a novel canonicalization procedure that corrects oversegmentation and whose goal is to ensure each table has a unique, unambiguous structure inter-pretation. To reduce additional sources of error, we implement several quality verification and control steps and pro-vide measurable guarantees about the quality of the ground truth. We show that data improvements alone lead to a sig-nificant increase in performance for TSR models, due both to improved training and a more reliable estimate of performance at evaluation. 1https://www. ncbi. nlm. nih. gov/pmc/tools/openftlist/ 2
|
tatr.pdf
|
(a) Oversegmented structure annotation (b) Canonical structure annotation Figure 3. In the above example, the structure annotation on the left is oversegmented, creating extra blank cells in the headers. The canonical structure annotation on the right merges these cells and captures its true logical structure. Finally, we apply the Detection Transformer (DETR) [2] for the first time to the tasks of TD, TSR, and FA, and demonstrate how with Pub Tables-1M all three tasks can be addressed with a transformer-based object detec-tion framework without any special customization for these tasks. 2. Related Work Structure recognition datasets The first dataset to ad-dress all three table extraction tasks was the ICDAR-2013 dataset [6]. It remains popular for benchmarking TSR mod-els due to its quality and relative completeness compared to other datasets. However, as a source of training data for table extraction models it is limited, containing only 248 tables for TD and TSR and 92 tables for FA. Recently, larger datasets [3, 9, 22, 23] have been created by collecting crowd-sourced table annotations automatically from existing documents. We summarize these datasets in Tab. 1. Each source table has an annotation for its content and structure in a markup format such as HTML, XML, or La Te X. Various methods are used to determine each table's spatial location within its containing document to create a correspondence between its markup and its presentation. From there, datasets commonly frame the TSR task as: given an input table, output its cell structure—the assignment of cells to rows and columns—and the text content for each cell, with image and HTML being example input and output formats, respectively, for these. More recently, two large datasets, Fin Tab Net and an en-hanced version of Pub Tab Net, have added location infor-mation for cells, similar to ICDAR-2013. Adding location information enables the TSR task to be framed as outputting cell location instead of cell content, with cell content extrac-tion being a trivial subsequent step. This increases the range of possible supervised modeling approaches. However, the bounding boxes for cells defined by these datasets cover only the text portion of each and exclude any additional whites-pace a cell might contain. This has a few implications, such as making bounding boxes for blank cells undefined and excluding attributes contributed by whitespace, such as text indentation and alignment. Therefore, one question left openby prior work is how to define bounding boxes for all cells, including blank cells. There are additional challenges related to annotation com-pleteness and quality that have not been addressed by prior datasets. In terms of completeness, prior large-scale datasets have also not included bounding boxes for rows and columns. Additionally, most datasets do not annotate the column header, and no prior large-scale dataset exists that speci-fies the row header of a table. This not only limits the range of modeling approaches that can be applied to TSR but limits how completely the overall TE task can be solved. Another open challenge is automated verfication and mea-surement of annotation quality, which is important due to the impracticality of verifying large-scale annotations manually. Prior datasets have also not addressed the significant issue of oversegmented annotations. These are important issues, as noise and mistakes in training data potentially harm learning and in evaluation data potentially lead to an underestimate of model performance. But currently the extent to which these issues affect model training and evaluation is unexplored. Modeling approaches One of the most common mod-eling approaches for TSR is to frame the task as some form of object detection [14, 17, 22]. Other approaches include those based on image-to-text [9] and graph-based approaches [3,15]. While a number of general-purpose archi-tectures, such as Faster R-CNN [16], exist for these model patterns, the unique characteristics of table and the relative lack of training data have both contributed to the commonly observed underperformance of these models when applied to TSR out-of-the-box. To get around deficiencies in training data, some ap-proaches model TSR in ways that are only partial solutions to the task, such as row and column detection in Deep-De SRT [17], which ignores spanning cells, or image-to-markup without cell text content, as in models trained on Table Bank [9]. Other approaches use custom pipelines that branch to consider different cases separately, such as train-ing separate models to recognize tables with and without visible borders surrounding every cell [14, 22]. Many of the previously mentioned approaches also use engineered model 3
|
tatr.pdf
|
Table 1. Comparison of crowd-sourced datasets for table structure recognition. Dataset Input Modality# Tables Cell Topology Cell Content Cell Location Row & Column Location Canonical Structure Table Bank [9] Image 145K ✓ Sci TSR [3] PDF∗15K ✓ ✓ Pub Tab Net [22, 23] Image 510K‡✓ ✓ ✓† Fin Tab Net [22] PDF∗113K ✓ ✓ ✓† Pub Tables-1M (ours) PDF∗948K ✓ ✓ ✓ ✓ ✓ ∗Multiple input modalities, such as image or text, can be derived from annotated PDF data. ‡The authors release annotations for 510K of the 568K total tables in their dataset. †For these datasets, cell bounding boxes are given for non-blank cells only and exclude any non-text portion of a cell. components or custom training procedures, and incorporate rules or other unlearned processing stages tailored to the TSR task, which brings in prior knowledge to lessen the burden placed on learning the task from data. Currently, no solution exists that uses a simple supervised learning approach with an off-the-shelf architecture, solves the TSR task completely, and achieves state-of-the-art performance. 3. Pub Tables-1M In this section, we describe the process used to develop Pub Tables-1M. First, to obtain a large source of annotated tables, we choose the PMCOA corpus, which consists of millions of publicly available scientific articles. In the PM-COA corpus, each scientific article is given in two forms: as a PDF document, which visually presents the article, and as an XML document, which provides a semantic description and hierarchical organization of the document's elements. Each table's content and structure is specified using standard HTML tags. However, because this data was not intended for use as ground truth for table extraction modeling, it does not explic-itly label or guarantee multiple things that would be helpful for this purpose. For instance, although the same tables ap-pear in both documents, no direct correspondence between them is given, nor the spatial location of each table. In terms of data quality, while tables are generally annotated reliably, it is not guaranteed that column headers are annotated com-pletely or that text content as annotated exactly matches the text content as it appears in the PDF. Finally, some labels, such as the row header for each table, are not annotated at all. The basic approach we take to overcome these issues is first we attempt to reliably infer as much missing annotation information as possible (for instance, the spatial location of each table) from the information that is present, then we verify that each annotation meets certain requirements for consistency. In some cases, we correct an annotation to attempt to make it more consistent, such as merging cells that are oversegmented. We consider certain requirements for tables to be strict and samples whose annotations violatethese are removed. This provides a set of conditions for quality and consistency that the annotations are guaranteed to meet. In the rest of this section, we describe these conditions and the steps we take to derive ground truth that meets them. Alignment Text in a PDF document has spatial location [xmin, ymin, xmax, ymax], while text in an XML document ap-pears inside semantically labeled tags. Because the cor-respondence between these is not given, the first step in creating Pub Tables-1M is to match the text content from both. We process the PDF document into a sequence of characters each with their associated bounding box and use the Needleman-Wunsch algorithm [10] to align this with the character sequence for the text extracted from each table HTML. This connects the text within each HTML tag to its spatial location with the PDF document. For each cell with text, we compute the union of the bounding boxes for each character of the cell's text, which we refer to as a text cell bounding box. Completion Following alignment, we complete the spatial annotations to define bounding boxes for rows, columns, and the entire table. The bounding box for the table is defined simply as the union of all text cell bounding boxes. The xmin andxmaxof the bounding box for each row are defined as thexminandxmaxof the table, giving every row the same horizontal length. The yminandymaxof the bounding box for each row, m, are defined as the yminandymaxof the union of the text cells for each cell whose starting row or ending row is m. Similarly, the yminandymaxof the bounding box for each column are defined as the yminandymaxof the table. Thexminandxmaxof the bounding box for each column, n, are defined as the xminandxmaxof the union of the text cell for each cell whose starting column or ending column isn. From these definitions, the grid cell for each cell is defined as the union of the bounding boxes of the cell's rows intersected with the union of the bounding boxes for its columns. Unlike the text cell, the grid cell is defined even for blank cells. 4
|
tatr.pdf
|
Canonicalization The primary goal of the canonicaliza-tion step is to correct oversegmentation in a table's structure annotations. To do this, we need to make assumptions about a table's intended structure. As the canonicalization algo-rithm itself is relatively simple, we first describe it, then detail the assumptions that motivate it. Put simply, canoni-calization amounts to merging adjacent cells under certain conditions. This algorithm is given in Algorithm 1. But because it only operates on cells in the headers, HTML does not have a tag for specifying a table's row header, and we observed that the column headers for tables in the PMCOA corpus are not always correct, we also include steps for in-ferring additional header cells that we believe can be reliably inferred in PMCOA markup annotations. These additional steps significantly increase the number of cells whose over-segmentation we are able to correct. Algorithm 1 Pub Tables-1M Canonicalization 1:ADD CELLS TO THE COLUMN AND ROW HEADERS 2: Split every blank spanning cell into blank grid cells 3: ifthe first row starts with a blank cell then add the first row to the column header 4: ifthere is at least one row labeled as part of the column header then 5: while every column in the column header does not have at least one complete cell that only spans that column do: add the next row to the column header 6: end if 7: for each rowdo: if the row is not in the column header and has exactly one non-blank cell that occupies the first column then label it a projected row header 8: ifany cell in the first column below the column header is a spanning cell or blank then add the column (below the column header) to the row header 9:MERGE CELLS 10: for each cell in the column header dorecursively merge the cell with any adjacent cells above and below in the column header that span the exact same columns 11: for each cell in the column header dorecursively merge the cell with any adjacent blank cells below it if every adjacent cell below it is blank and in the column header 12: for each cell in the column header dorecursively merge the cell with any adjacent blank cells above it if every adjacent cell above it is blank 13: for each projected row header domerge all of the cells in the row into a single cell 14: for each cell in the row header dorecursively merge the cell with any adjacent blank cells below it We first assume that each table has an intended structure consistent with the Wang model [21], which in a study Wang found was true for 97 percent of observed tables. Under this model, the headers of the table each have a hierarchical structure that corresponds logically to a tree. We assert that for a structure annotation to be consistent with a table'slogical structure, there should be exactly one cell for every tree node. We also assume that each value in the table is indexed by a unique set of keys. We interpret this to mean that each column in the body of the table corresponds to a unique leaf node in the column header tree, and similarly that each row in the body corresponds to a unique leaf node in the row header tree (the index of a row or column can serve as a key if necessary). These assumptions enable us to determine if a row or column header is only partially annotated and if so, to extend it to additional columns or rows, respectively. However, to keep the precision of the algorithm high, for row headers we only attempt to infer projected row headers (PRHs, also known as projected multi-level row headers [8], section headers [13], or super-rows [20]) and to infer cells that are in the first column of the header. The PRHs of a table can be identified using the rule in Line 7. Inference of the full row header is considered outside the scope of this work. We also assume that any internal node in a header tree has at least two children. If not, ambiguity could arise in a table's logical structure because an internal node could optionally be split into a parent node and a single child node. The final assumptions we make are in regards to the root cause of oversegmentation in markup annotations. We assume that cells will only be oversegmented if an oversegmentation is consistent with the table's appearance. In practice, this means that cells with centered text will not be oversegmented in the direction of the alignment because this is likely to alter the table's appearance. For non-centered text, we expect that when cells in either header are oversegmented, this will happen vertically, as in Fig. 3b, and not horizontally due to the fact that text fills horizontal space before it fills vertical space, leaving more vertical space unused. Further, we expect that oversegmented cells in the row header will have text that is top aligned. Finally, we expect that when projected row headers are oversegmented, this will happen horizontally, not vertically, as a projected row header already occupies only one row. Finally, there are two additional cases that we must handle by convention. One case is when one or more rows of blank grid cells are between a parent cell and all of its children cells in the column header. In this case, we can choose either to merge all of the blank cells with the parent cell above it or each with the child cell below it, and we choose the convention to merge all of the blank cells with the child, which occurs in Line 10. The final case is when a table has an blank stub head (according to the Wang model) in its top-left corner. In this case, the blank cells are not part of the table, so the assumptions about table structure do not suggest how they should be grouped. We choose by convention to merge all blank cells in the same column in a blank stub head, which is consistent with the scheme in Line 10. 5
|
tatr.pdf
|
Table 2. Estimated measure of oversegmentation for projected row headers (PRHs) by dataset. As PRHs are only one type of cell that can be oversegmented, this is a partial survey of the total oversegmentation in these datasets. Dataset Total Tables Investigated†Total Tables with a PRH∗Tables with an oversegmented PRH Total % (of total with a PRH) % (of total investigated) Sci TSR 10,431 342 54 15. 79% 0. 52% Pub Tab Net 422,491 100,159 58,747 58. 65% 13. 90% Fin Tab Net 70,028 25,637 25,348 98. 87% 36. 20% Pub Tables-1M (ours) 761,262 153,705 0 0% 0% †We exclude tables with fewer than five rows; to avoid column header rows we skip the first four rows when searching for PRHs. ∗PRH = projected row header; these can be reliably detected in datasets without any prior row or column header annotations. Limitations While the stated goal of canonicalization is applicable to any table structure annotation, we note that Algorithm 1 is designed to achieve this specifically for the annotations in the PMCOA dataset. Canonicalizing tables from other datasets may require additional assumptions and is considered outside the scope of this work. Finally, it should be noted that canonicalization does not guarantee mistake-free annotations. Remaining issues are addressed using the automated quality control procedure described next. Quality control Because Pub Tables-1M is too large to be verified manually, we check for potential errors automatically and filter these from the data. First, as tables rendered from markup should not contain overlapping rows or overlapping columns, we discard any table where this occurs, as these are likely due to mistakes introduced by the alignment process. To filter out mistakes made both by the original annotators and our automated processing, we compare the edit distance between the non-whitespace text for every cell in the original XML annotations with the text extracted from the PDF inside the grid cell bounding box. We filter out any tables for which the normalized edit distance between these averaged over every cell is above 0. 05. We do not force the text from each to be exactly equal, as the PDF text can differ even when everything is annotated correctly, due to things like word wrapping, which may add hyphens that are not in the source annotations. When the annotations do slightly differ from their corresponding PDF text, we choose to consider the PDF text to be the ground truth. As tables with correct location information provide an unambiguous assignment of all words in the table to cells, we also compute the average fraction of overlap between each word appearing within the boundary of the table and its most overlapping grid cell, and discard tables with an average below 0. 9. Finally, we remove outliers by counting the number of objects in a table (defined in Sec. 4) and removing tables with more than 100. In all, less than 0. 1% of tables are discarded as outliers. Pub Tables-1M is the first dataset that verifies annotations at the cell level and provides a measurable assurance ofconsistency for the ground truth. This shows that improving the explicitness of information is valuable in part because it leads to more opportunities for catching inconsistencies and errors embedded within the data. Dataset statistics and splits In total, Pub Tables-1M con-tains 947,642 tables for TSR, of which 52. 7% are complex (have at least one spanning cell). Prior to canonicalization, only 40. 1% of the tables in the set were considered com-plex by the original annotators. Canonicalization adjusts the annotations in some way for 34. 7% of tables, or 65. 8% of complex tables. To further assess the impact on oversegmentation, we compare our final dataset with other datasets in Tab. 2. Pre-cisely measuring the amount of oversegmentation in a dataset requires annotations for the row and column headers. But because other datasets lack these, we instead measure just oversegmented projected row headers (PRHs). PRHs can be detected reliably without explicit annotations using the rule in Line 7. To account for missing column header an-notations, we do not start looking for PRHs until at least the fifth row, which assumes the column header occupies at most four rows, and we simply exclude any tables that have fewer than five rows. In case there are un-annotated footers, we also do not count any detected PRHs that are the last rows of the table. A detected PRH is oversegmented if its row contains a blank cell. As can be seen, canonicalization eliminates a significant source of oversegmentation, and thus ambiguity, that is present in other datasets. Interesting to note, Fin Tab Net nearly always oversegments projected row headers. While self-consistent, this widespread oversegmen-tation contradicts the logical structure of the table and causes potential issues with combining this dataset with others that annotate these rows differently. We split Pub Tables-1M randomly into train, validation, and test sets at the document level using an 80/10/10 split. For TSR, this results in 758,849 tables for training; 94,959 for validation; and 93,834 for testing. For TD, there are 460,589 fully-annotated pages containing tables for training; 57,591 for validation; and 57,125 for testing. An example 6
|
tatr.pdf
|
Figure 4. An example table with dilated bounding box annotations for different object classes for jointly modeling table structure recognition and functional analysis. page and table annotation for TD is shown in Fig. 2. Note that tables that span multiple pages are considered outside the scope of this work. 4. Proposed Model We model all three tasks of TD, TSR, and FA as object detection with images as input. For TD, we use two object classes: table andtable rotated. The table rotated class corresponds to tables that are rotated counterclockwise 90 degrees. TSR and FA model We use a novel approach that models TSR and FA jointly using six object classes: table,table column,table row,table column header,table projected row header, and table spanning cell. We illustrate these classes in Fig. 4. The intersection of each pair of table column andtable row objects can be considered to form a seventh implicit class, table grid cell. These objects model a table's hierarchical structure through physical overlap. For the TSR and FA model, we use bounding boxes that aredilated. To create dilated bounding boxes, for each pair of adjacent rows and each pair of adjacent columns, we expand their boundaries until they meet halfway, which fills the empty space in between them. Similarly we expand the objects from the other classes so their boundaries match the adjustments made to the rows and columns they occupy. After, there are no gaps or overlap between rows, between columns, or between cells. To demonstrate the proposed dataset and the object detec-tion modeling approach, we apply the Detection Transformer (DETR) [2] to all three TE tasks. We train one DETR model for TD and one DETR model for both TSR and FA. For comparison, we also train a Faster R-CNN [16] model for the same tasks. All models use a Res Net-18 backbone pre-trained on Image Net with the first few layers frozen. We Table 3. Test performance of models on Pub Tables-1M using object detection metrics. Task Model AP AP 50 AP75 AR TD Faster R-CNN 0. 825 0. 985 0. 927 0. 866 DETR 0. 966 0. 995 0. 988 0. 981 TSR + FA Faster R-CNN 0. 722 0. 815 0. 785 0. 762 DETR 0. 912 0. 971 0. 948 0. 942 avoid custom engineering the models and training proce-dures for each task, using default settings wherever possible to allow the data to drive the result. 5. Experiments In this section, we report the results of training the pro-posed models on data derived from Pub Tables-1M. For TD, we train two models: DETR and Faster R-CNN. We report the results in Tab. 3. For table detection, DETR slightly outperforms Faster R-CNN on AP50but significantly out-performs on AP. We interpret this to mean that while both models are able to learn to detect tables, DETR precisely localizes tables much better than Faster R-CNN. For TSR and FA, we train three models: Faster R-CNN and DETR on the canonicalized data, and DETR on the origi-nal, non-canonical (NC) annotations (DETR-NC). We report the results using object detection metrics for the models trained on canonical data in Tab. 3, which measures per-formance jointly on TSR and FA, and report results for all models using TSR-only metrics in Tab. 4. For TSR only, we also evaluate DETR-NC on both the canonical and the original non-canonical test data. For assessing TSR performance, we report the table con-tent accuracy metric ( Acc Cont), which is the percentage of tables whose text content matches the ground truth exactly for every cell, as well as several metrics for partial table correctness, which use different strategies to give credit for correct cells when not all cells are correct. For partial correct-ness, we use the F-score of the standard adjacent cell content metric [5] and the recently proposed Gri TS metrics [19]. Gri TS metrics have the form, Gri TS f(A,B) =2·∑ i,jf(˜Ai,j,˜Bi,j) |A|+|B|, (1) which can also be interpreted as an F-score. Gri TS represents the ground truth and predicted tables as matrices, Aand B, and computes a similarity score between the most similar substructures [1] of these matrices, ˜Aand˜B, where a sub-structure is defined as a selection of mrows and ncolumns from the matrix. Compared to other metrics for TSR, this for-mulation better captures the two-dimensional structure and 7
|
tatr.pdf
|
Table 4. Test performance of the TSR + FA models on Pub Tables-1M on TSR metrics. Test Data Model Table Category Acc Cont Gri TS Top Gri TS Cont Gri TS Loc Adj Cont Non-Canonical DETR-NC Simple 0. 8678 0. 9872 0. 9859 0. 9821 0. 9801 Complex 0. 5360 0. 9600 0. 9618 0. 9444 0. 9505 All 0. 7336 0. 9762 0. 9761 0. 9668 0. 9681 Canonical DETR-NC Simple 0. 9349 0. 9933 0. 9920 0. 9900 0. 9865 Complex 0. 2712 0. 9257 0. 9290 0. 9044 0. 9162 All 0. 5851 0. 9576 0. 9588 0. 9449 0. 9494 Faster R-CNN Simple 0. 0867 0. 8682 0. 8571 0. 6869 0. 8024 Complex 0. 1193 0. 8556 0. 8507 0. 7518 0. 7734 All 0. 1039 0. 8616 0. 8538 0. 7211 0. 7871 DETR Simple 0. 9468 0. 9949 0. 9938 0. 9922 0. 9893 Complex 0. 6944 0. 9752 0. 9763 0. 9654 0. 9667 All 0. 8138 0. 9845 0. 9846 0. 9781 0. 9774 ordering of cells of a table when comparing tables. Further, Gri TS enables TSR to be assessed from multiple perspec-tives within the same formulation, with Gri TS Topmeasuring cell topology recognition, Gri TS Contmeasuring cell content recognition, and Gri TS Locmeasuring cell location recogni-tion. As can be seen in Tab. 4, DETR trained on the canonical data produces strong results for TSR and FA, outperforming the other models when evaluated on all tables. Comparing DETR-NC evaluated on NC ground truth versus DETR eval-uated on canonical ground truth, we see that using canonical data improves performance on the metrics across all table types. This is even more apparent for the table accuracy metric, for which the use of canonical data is responsible for a jump in performance from 0. 5360 to 0. 6944 for complex tables. To consider the positive impact that canonicalization has just on facilitating a more reliable evaluation, we compare DETR-NC evaluated on canonical data versus NC data. Even though it is trained on NC data, the metrics for simple tables are much higher when DETR-NC is evaluated on canonical data (0. 9349 accuracy) than when it is evaluted on NC data (0. 8678 accuracy). This shows even more clearly that the canonical data is less noisy and contributes to a more reliable evaluation. Overall the results confirm that canonical data signifi-cantly improves performance for TSR models. Even setting aside the argument that non-canonical data is noisier and more ambiguous, it is also a different way of annotating the data, using oversegmentation, which is less useful because it does not correspond to a table's true logical structure. This difference is apparent when we compare DETR-NC versus DETR, both evaluated on canonical test data. DETR-NC performs much worse on complex tables due in large part to the inconsistent scheme it learns for understanding the spanning cells in their headers. 6. Conclusion In this paper, we introduced a new dataset, Pub Tables-1M, for table extraction in unstructured documents. Pub Tables-1M address the challenge of creating complete, reliable ground truth at scale for table structure recognition. We called attention to the problem that oversegmentation in markup annotations leads to ambiguous ground truth in crowd-sourced datasets, and proposed a novel canonical-ization procedure to address this. We demonstrated that the proposed improvements to the ground truth data have a significant positive impact on model performance. Finally, we adopted DETR for all three table extraction tasks and showed for the first time that it is possible to achieve state-of-the-art performance within a standard object detection framework without the need for any special customization for these tasks. While we do not believe this work raises any issues regarding negative impacts to society, we welcome a discussion on any potential impacts raised by others. 7. Future Work In the future, we hope to expand the proposed methods and canonicalization beyond scientific articles to data from additional domains, such as tables in financial documents. We also hope to address the open challenge of accurately annotating row headers in large-scale datasets, which will enable even more complete solutions for table extraction. Finally, table extraction is often just one stage in larger pipelines for document understanding and information re-trieval, and developing end-to-end systems in these areas is an important future direction with its own challenges. We hope that releasing a large pool of detailed table annotations from the PMCOA corpus in particular can further progress in this area. 8
|
tatr.pdf
|
8. Acknowledgments We would like to thank Pramod Sharma, Natalia Larios Delgado, Joseph N. Wilson, Mandar Dixit, John Corring, and Ching Pui WAN for helpful discussions and feedback while preparing this manuscript. References [1]Amihood Amir, Tzvika Hartman, Oren Kapah, B Riva Shalom, and Dekel Tsur. Generalized LCS. Theoretical computer science, 409(3):438-449, 2008. 7 [2]Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Confer-ence on Computer Vision, pages 213-229. Springer, 2020. 3, 7 [3]Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanx-uan Yin, and Xian-Ling Mao. Complicated table structure recognition. ar Xiv preprint ar Xiv:1908. 04729, 2019. 3, 4 [4]Wolfgang Gatterbauer, Paul Bohunsky, Marcus Herzog, Bernhard Krüpl, and Bernhard Pollak. Towards domain-independent information extraction from web tables. In Pro-ceedings of the 16th international conference on World Wide Web, pages 71-80, 2007. 1 [5]Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. A methodology for evaluating algorithms for table under-standing in PDF documents. In Proceedings of the 2012 ACM symposium on Document engineering, pages 45-48, 2012. 7 [6]Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. ICDAR 2013 table competition. In 2013 12th International Conference on Document Analysis and Recognition, pages 1449-1453. IEEE, 2013. 1, 3 [7]Jianying Hu, Ramanujan Kashi, Daniel Lopresti, George Nagy, and Gordon Wilfong. Why table ground-truthing is hard. In Proceedings of Sixth International Conference on Document Analysis and Recognition, pages 129-133. IEEE, 2001. 1 [8]Jianying Hu, Ramanujan S Kashi, Daniel P Lopresti, and Gor-don Wilfong. Table structure recognition and its evaluation. In Document Recognition and Retrieval VIII, volume 4307, pages 44-55. International Society for Optics and Photonics, 2000. 5 [9]Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, and Zhoujun Li. Tablebank: Table benchmark for image-based table detection and recognition. In Proceedings of The 12th Language Resources and Evaluation Conference, pages 1918-1925, 2020. 1, 3, 4 [10] Saul B Needleman and Christian D Wunsch. A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology, 48(3):443-453, 1970. 4 [11] Ermelinda Oro and Massimo Ruffolo. TREX: An approach for recognizing and extracting tables from PDF documents. In2009 10th International Conference on Document Analysis and Recognition, pages 906-910. IEEE, 2009. 1 [12] Shubham Singh Paliwal, D Vishwanath, Rohit Rahul, Monika Sharma, and Lovekesh Vig. Tablenet: Deep learning modelfor end-to-end table detection and tabular data extraction from scanned document images. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 128-133. IEEE, 2019. 1 [13] David Pinto, Andrew Mc Callum, Xing Wei, and W Bruce Croft. Table extraction using conditional random fields. In Proceedings of the 26th annual international ACM SIGIR con-ference on Research and development in informaion retrieval, pages 235-242, 2003. 5 [14] Devashish Prasad, Ayan Gadpal, Kshitij Kapadni, Manish Visave, and Kavita Sultanpure. Cascade Tab Net: An approach for end to end table detection and structure recognition from image-based documents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 572-573, 2020. 1, 3 [15] Shah Rukh Qasim, Hassan Mahmood, and Faisal Shafait. Rethinking table recognition using graph neural networks. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 142-147. IEEE, 2019. 3 [16] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. ar Xiv preprint ar Xiv:1506. 01497, 2015. 3, 7 [17] Sebastian Schreiber, Stefan Agne, Ivo Wolf, Andreas Dengel, and Sheraz Ahmed. Deep De SRT: Deep learning for detection and structure recognition of tables in document images. In 2017 14th IAPR international conference on document anal-ysis and recognition (ICDAR), volume 1, pages 1162-1167. IEEE, 2017. 1, 3 [18] Alexey O Shigarov. Table understanding using a rule engine. Expert Systems with Applications, 42(2):929-937, 2015. 1 [19] Brandon Smock, Rohith Pesala, and Robin Abraham. Gri TS: Grid table similarity metric for table structure recognition. 2021. 7 [20] Ashwin Tengli, Yiming Yang, and Nian Li Ma. Learning table extraction from examples. In COLING 2004: Proceed-ings of the 20th International Conference on Computational Linguistics, pages 987-993, 2004. 2, 5 [21] Xinxin Wang. Tabular abstraction, editing, and formatting, 1996. 5 [22] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (GTE): A frame-work for joint table identification and cell structure recogni-tion using visual context. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 697-706, 2021. 1, 3, 4 [23] Xu Zhong, Elaheh Shafiei Bavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. ar Xiv preprint ar Xiv:1911. 10683, 2019. 1, 3, 4 9. Appendix 9. 1. Training Details For both DETR models, we use a Res Net-18 backbone, six layers in the encoder, and six layers in the decoder. For TD, we use 15 object queries, and for TSR and FA we use 125 object queries, each chosen to be slightly more than the 9
|
tatr.pdf
|
maximum number of objects in each set's training samples. Besides this, we use the same default architecture settings for each model. All of the experiments are performed using a single NVidia Tesla V100 GPU. We initialize the models with weights pre-trained on Image Net and train each model for 20 epochs using all default hyperparameters and training settings except for those we note here. For both models, we use a learning rate drop of 1 and gamma of 0. 9. For the TSR and FA model, we also use an initial learning rate of 0. 00005 and a no-object class weight of 0. 4. We limited hyperparam-eter tuning to one short experiment to determine the initial learning rate. We ran training experiments with three dif-ferent initial learning rates of 0. 0002, 0. 0001, 0. 00005 and chose to use the learning rate for each model that had the best performance on the validation set after one epoch of training. We use no custom components, losses, or procedures for training the model, other than standard data augmentations, such as random cropping and resizing. To create training data for the TD model, we render the PDF pages to images with a maximum length of 1000 pixels and appropriately scale the bounding boxes for the objects to image coordinates. For TSR and FA, we first render the page containing the table as an image with a maximum length of 1000 pixels, scale and pad the table bounding box with an additional 30 pixels on all sides (or fewer on a side if there are less than 30 pixels available on that side), and crop the image to this bounding box. The padding enables more variation in training through cropping augmentations. 9. 2. Inference While the model is trained entirely as an object detector, at inference time an additional step is required to convert from objects output by the model to a structured table. Ob-jects in our model form hierarchical relationships, where a parent-child relationship between objects is indicated by physical containment, or overlap. For details on the full set of relationships in the hierarchy, please see the code. One thing we mention here is that according to this hierarchy no object can have two parents of the same object class. There-forem, a conflict occurs when two parent objects of the same class both overlap with a potential child object for that class. For example, two spanning cell objects are considered to conflict if they both overlap with the same grid cell. The definition of overlap can vary by object class but is generally when 50% of a child object's bounding box by area overlaps with a parent object's bounding box. The ground truth data is conflict-free, so at training time the model learns to produce output that is also free of con-flicts. However, to address any conflicts produced by the model strictly at inference time once training has finished, we add to the models a simple conflict resolution step. Theconflict resolution step involves suppressing objects or ad-justing their bounding boxes to eliminate conflict between objects of the same class. This is similar to non-maxima suppression, except overlap between objects is defined in terms of children objects instead of pixels. Once the objects are conflict-free, they are converted to a logical table. 9. 3. Scoring Lastly, we note that for the sake of evaluation, once the text that occurs within each grid cell is extracted, we retroac-tively adjust the vertical extent of the row bounding boxes and the horizontal extent of the column bounding boxes to tightly wrap the text they contain. Note that cells contain the same text content before and after this tightening. We then use these adjusted bounding boxes for cell location scoring purposes. This adjustment is necessary to properly score because we train the model using dilated bounding boxes, which expand the boundaries between rows and between columns to fill the gaps between them, thereby containing extraneous padding. This extraneous padding is useful for learning but would unfairly penalize the model if included at scoring time, therefore we consider the tightened cells to be the actual cell locations output by the model and score with these. 10
|
tatr.pdf
|
README.md exists but content is empty.
- Downloads last month
- 2