Encoder-decoder transformer for table generation

1. A system comprising:

at least one hardware processor; and

at least one memory storing instructions that cause the at least one hardware processor to execute operations comprising:

accessing a text document comprising a plurality of strings that are unstructured, the plurality of strings forming one or more sentences;

processing the text document by a machine learning model to generate a table that organizes words of the one or more sentences of the plurality of strings that are unstructured into rows and columns over a plurality of iterations, the table comprising a plurality of entries, each entry comprising at least one of the words of the one or more sentences;

at each of the plurality of iterations, estimating by the machine learning model a first value of a first entry of the plurality of entries based on a second value of a second entry of the plurality of entries that has been determined in a prior iterations identifying a first cell in the table populated with a first word of the words of the one or more sentences; and

populating a second cell that is at a position that is non-adjacent relative to a position of the first cell with a second word of the words of the one or more sentences rather than populating the second word in a third cell that is adjacent next to or underneath the first cell, the third cell being left empty in response to populating the second cell.

2. The system of claim 1 ,

wherein the second cell that is at a diagonal relative to the position of the first cell.

3. The system of claim 1 , the operations comprising:

generating a first table instance comprising a first set of entries based on the plurality of strings;

at a first iteration, generating, by the machine learning model, a first plurality of confidence scores for values in each entry of the first set of entries of the first table instance; and

selecting a first subset of the first set of entries of the first table instance based on the first plurality of confidence scores.

4. The system of claim 3 , the operations comprising:

retrieving a confidence threshold;

comparing the first plurality of confidence scores for each entry of the first set of entries of the first table instance to the confidence threshold; and

identifying the first subset of the first set of entries that are associated with respective confidence scores that transgress the confidence threshold.

5. The system of claim 4 , the operations comprising generating a second table instance comprising a second set of entries by:

after selecting the first subset of the first set of entries, resetting the values associated with a remaining set of entries that are excluded from the first subset of the first set of entries; and

retaining the values associated with the first subset of the first set of entries in the second table instance.

6. The system of claim 5 , the operations comprising:

at a second iteration, processing the second table instance by the machine learning model, to generate a second plurality of confidence scores for each entry of the second set of entries of the second table instance; and

selecting a second subset of the second set of entries based on the second plurality of confidence scores.

7. The system of claim 6 , the operations comprising:

comparing the second plurality of confidence scores to the confidence threshold; and

identifying the second subset of the second set of entries that are associated with respective confidence scores that transgress the confidence threshold.

8. The system of claim 7 , the operations comprising:

repeating generation of additional table instances until all entries of a given table instance are associated with confidence scores that transgress the confidence threshold.

9. The system of claim 1 , the operations comprising:

training the machine learning model based on a corpus of training documents to maximize an expected log likelihood for a training table across all random permutations of a factorization order.

10. The system of claim 1 , wherein the machine learning model is trained to populate the plurality of entries of the table in any order in a way that maximizes confidence in each individual entry and reduces error accumulation.

11. The system of claim 1 , the operations comprising:

training the machine learning model based on a corpus of training documents by populating a plurality of training tables representing different permutations of strings in the training documents to maximize an expected log likelihood.

12. The system of claim 1 , the machine learning model generatively predicting text to include in an empty cell of the table by performing operations comprising:

identifying a category or column header corresponding to the empty cell;

processing data from a row corresponding to the empty cell and corresponding headers of each populated entry from the row corresponding to the empty cell; and

using information from other cells in the row corresponding to the empty cell to estimate or predict the text to include in the empty cell.

13. The system of claim 1 , wherein the machine learning model estimates column headers of the table.

14. The system of claim 1 , the operations comprising:

inferring one or more values for individual entries of the plurality of entries with words excluded from the plurality of strings based on values of other entries of the plurality of entries, and

populating an individual entry of the plurality of entries of the table using the inferred one or more values without retrieving values from the text document.

15. A method comprising:

accessing a text document comprising a plurality of strings that are unstructured, the plurality of strings forming one or more sentences;

processing the text document by a machine learning model to generate a table that organizes words of the one or more sentences of the plurality of strings that are unstructured into rows and columns over a plurality of iterations, the table comprising a plurality of entries, each entry comprising at least one of the words of the one or more sentences;

at each of the plurality of iterations, estimating by the machine learning model a first value of a first entry of the plurality of entries based on a second value of a second entry of the plurality of entries that has been determined in a prior iterations identifying a first cell in the table populated with a first word of the words of the one or more sentences; and

populating a second cell that is at a position that is non-adjacent relative to a position of the first cell with a second word of the words of the one or more sentences rather than populating the second word in a third cell that is adjacent next to or underneath the first cell, the third cell being left empty in response to populating the second cell.

16. The method of claim 15 , further comprising:

receiving a query that indicates values for the columns, wherein the table is generated based on the values indicated for the columns.

17. The method of claim 15 , further comprising:

generating a first table instance comprising a first set of entries based on the plurality of strings;

at a first iteration, generating, by the machine learning model, a first plurality of confidence scores for values in each entry of the first set of entries of the first table instance; and

selecting a first subset of the first set of entries of the first table instance based on the first plurality of confidence scores.

18. The method of claim 17 , further comprising:

retrieving a confidence threshold;

comparing the first plurality of confidence scores for each entry of the first set of entries of the first table instance to the confidence threshold; and

identifying the first subset of the first set of entries that are associated with respective confidence scores that transgress the confidence threshold.

19. The method of claim 18 , further comprising generating a second table instance comprising a second set of entries by:

after selecting the first subset of the first set of entries, resetting the values associated with a remaining set of entries that are excluded from the first subset of the first set of entries; and

retaining the values associated with the first subset of the first set of entries in the second table instance.

20. The method of claim 19 , further comprising:

at a second iteration, processing the second table instance by the machine learning model, to generate a second plurality of confidence scores for each entry of the second set of entries of the second table instance; and

selecting a second subset of the second set of entries based on the second plurality of confidence scores.

21. The method of claim 20 , further comprising:

comparing the second plurality of confidence scores to the confidence threshold; and

identifying the second subset of the second set of entries that are associated with respective confidence scores that transgress the confidence threshold.

22. The method of claim 21 , further comprising:

repeating generation of additional table instances until all entries of a given table instance are associated with confidence scores that transgress the confidence threshold.

23. A non-transitory computer-storage medium comprising instructions that, when executed by a processor of a machine, configure the machine to perform operations comprising:

accessing a text document comprising a plurality of strings that are unstructured, the plurality of strings forming one or more sentences;

processing the text document by a machine learning model to generate a table that organizes words of the one or more sentences of the plurality of strings that are unstructured into rows and columns over a plurality of iterations, the table comprising a plurality of entries, each entry comprising at least one of the words of the one or more sentences;

at each of the plurality of iterations, estimating by the machine learning model a first value of a first entry of the plurality of entries based on a second value of a second entry of the plurality of entries that has been determined in a prior iterations identifying a first cell in the table populated with a first word of the words of the one or more sentences; and

populating a second cell that is at a position that is non-adjacent relative to a position of the first cell with a second word of the words of the one or more sentences rather than populating the second word in a third cell that is adjacent next to or underneath the first cell, the third cell being left empty in response to populating the second cell.

24. The non-transitory computer-storage medium of claim 23 , the operations comprising:

selecting, at an individual iteration, a first subset of a first set of entries of a first table instance based on a first plurality of confidence scores;

after selecting the first subset of the first set of entries, resetting values stored in the first table instance associated with a remaining set of entries that are excluded from the first subset of the first set of entries; and

retaining values associated with the first subset of the first set of entries in a second table instance.

25. The non-transitory computer-storage medium of claim 23 , the operations comprising:

generating a first table instance comprising a first set of entries based on the plurality of strings;

at a first iteration, generating, by the machine learning model, a first plurality of confidence scores for values in each entry of the first set of entries of the first table instance; and

selecting a first subset of the first set of entries of the first table instance based on the first plurality of confidence scores.

26. The non-transitory computer-storage medium of claim 25 , the operations comprising:

retrieving a confidence threshold;

comparing the first plurality of confidence scores for each entry of the first set of entries of the first table instance to the confidence threshold; and

identifying the first subset of the first set of entries that are associated with respective confidence scores that transgress the confidence threshold.

27. The non-transitory computer-storage medium of claim 26 , the operations comprising generating a second table instance comprising a second set of entries by:

after selecting the first subset of the first set of entries, resetting the values associated with a remaining set of entries that are excluded from the first subset of the first set of entries; and

retaining the values associated with the first subset of the first set of entries in the second table instance.

28. The non-transitory computer-storage medium of claim 27 , the operations comprising:

at a second iteration, processing the second table instance by the machine learning model to generate a second plurality of confidence scores for each entry of the second set of entries of the second table instance; and

selecting a second subset of the second set of entries based on the second plurality of confidence scores.

29. The non-transitory computer-storage medium of claim 28 , the operations comprising:

comparing the second plurality of confidence scores to the confidence threshold; and

identifying the second subset of the second set of entries that are associated with respective confidence scores that transgress the confidence threshold.

30. The non-transitory computer-storage medium of claim 29 , the operations comprising:

repeating generation of additional table instances until all entries of a given table instance are associated with confidence scores that transgress the confidence threshold.