Probabilistic language models for identifying sequential reading order of discontinuous text segments
The present invention is directed towards providing automated workflows for the identification of a reading order from text segments extracted from a document. Ordering the text segments is based on trained natural language models. In some embodiments, the workflows are enabled to perform a method for identifying a sequence associated with a portable document. The methods includes iteratively generating a probabilistic language model, receiving the portable document, and selectively extracting features (such as but not limited to text segments) from the document. The method may generate pairs of features (or feature pair from the extracted features). The method may further generate a score for each of the pairs based on the probabilistic language model and determine an order to features based on the scores. The method may provide the extracted features in the determined order.
1. One or more non-transitory computer-readable storage media having instructions stored thereon, which, when executed by a processor of a computing device cause the computing device to perform actions comprising:
extracting a plurality of discontinuous text segments from features of a document based on recursively traversing a hierarchical representation of the features using a parser;
identifying a sequential reading order for the plurality of discontinuous text segments associated with the document, using a probabilistic language model configured to predict a score, for each pair of text segments of the plurality of discontinuous text segments, indicating a likelihood that the pair of text segments is in a correct reading-order sequence;
generating ordered textual content by arranging the plurality of discontinuous text segments in the sequential reading order; and
performing an operation on the ordered textual content.
2. The one or more non-transitory computer-readable storage media of claim 1 , wherein identifying the sequential reading order comprises:
receiving a set of feature types of the document;
employing a recursive descent to traverse the document based on the set of feature types; and
extracting the plurality of discontinuous text segments based on the set of feature types and the recursive descent.
3. The one or more non-transitory computer-readable storage media of claim 2 , wherein employing the recursive descent comprises generating the recursive descent by employing a JavaScript Object Notation (JSON) parser to traverse the document.
4. The one or more non-transitory computer-readable storage media of claim 1 , wherein the actions further comprise:
extracting the features of the document; and
arranging the features into the hierarchical representation.
5. The one or more non-transitory computer-readable storage media of claim 1 , wherein identifying the sequential reading order comprises:
extracting the plurality of discontinuous text segments from textual content of the document;
accessing a ground truth sequence associated with the extracted plurality of discontinuous text segments;
generating a distance metric based on a comparison of the ground truth sequence with the sequential reading order; and
updating the probabilistic language model based on the distance metric.
6. The one or more non-transitory computer-readable media of claim 1 , wherein the operation on the ordered textual content comprises one or more of narrating, copying, reflowing, performing language analysis on, searching, summarizing, or exporting the ordered textual content.
7. The one or more non-transitory computer-readable media of claim 1 , wherein identifying the sequential reading order comprises:
extracting the plurality of discontinuous text segments from textual content of the document based at least in part on feature types of the document; and
identifying each pair of text segments from the extracted plurality of discontinuous text segments, each pair including an ordered set of the text segments.
8. The one or more non-transitory computer-readable storage media of claim 1 , wherein the actions further comprise generating the probabilistic language model by employing at least one of noise contrastive estimation or smoothing of probability distributions generated by the probabilistic language model.
9. The one or more non-transitory computer-readable storage media of claim 1 , wherein the probabilistic language model comprises an n-gram model or a recurrent neural network trained on ground truth reading-order sequences.
10. A computerized method comprising:
identifying a sequential reading order for a plurality of features corresponding to discontinuous portions of a document, using a probabilistic language model configured to predict a score, for each pair of features of the plurality of features, indicating a likelihood that the pair of features is in a correct reading-order sequence, the probabilistic language model generated by employing at least one of noise contrastive estimation or smoothing of probability distributions generated by the probabilistic language model; and
performing an operation based on ordered textual content comprising the plurality of features rearranged into the sequential reading order.
11. The computerized method for claim 10 , wherein the probabilistic language model comprises an n-gram model or a recurrent neural network trained on ground truth reading-order sequences.
12. The computerized method of claim 10 , wherein the operation based on the ordered textual content comprises one or more of narrating, copying, reflowing, performing language analysis on, searching, summarizing, or exporting the ordered textual content.
13. The computerized method of claim 10 , further comprising:
accessing a set of feature types to extract from the document;
employing a recursive descent to traverse the document based on the set of feature types; and
extracting the plurality of features from textual content of the document based on the set of feature types and the recursive descent.
14. The computerized method of claim 13 , wherein employing the recursive descent comprises generating the recursive descent by employing a JavaScript Object Notation (JSON) parser to traverse the document.
15. The computerized method of claim 10 , further comprising:
extracting a set of features of the document;
arranging the set of features into a hierarchical representation; and
recursively traversing the hierarchical representation, using a parser, to extract the plurality of features from the set of features, the plurality of features being a subset of the set of features.
16. The computerized method of claim 10 , wherein identifying the sequential reading order comprises:
extracting the plurality of features from textual content of the document;
accessing a sequence associated with the extracted plurality of features;
generating a distance metric based on a comparison of the sequence with the ordered textual content; and
updating the probabilistic language model based on the distance metric.
17. A computing system comprising:
one or more hardware processors and memory configured to provide computer program instructions to the one or more hardware processors;
a means for identifying a sequential reading order for a plurality of discontinuous text segments associated with a document, using a probabilistic language model that comprises an n-gram model or a recurrent neural network trained on ground truth reading-order sequences and is configured to use the one or more hardware processors to predict a score, for each pair of text segments of the plurality of discontinuous text segments, indicating a likelihood that the pair of text segments is in a correct reading-order sequence;
a re-sequencing module configured to use the one or more hardware processors to:
determine ordered textual content by re-sequencing the plurality of discontinuous text segments in the sequential reading order; and
provide, to a user interface, an indication of the ordered textual content.
18. The computing system of claim 17 , wherein the re-sequencing module is configured to provide the indication of the ordered textual content in association with one or more of narrating, copying, reflowing, performing language analysis on, searching, summarizing, or exporting the ordered textual content.
19. The computing system of claim 17 , the probabilistic language model generated based on at least one of noise contrastive estimation or smoothing of probability distributions generated by the probabilistic language model.
20. The computing system of claim 17 , additionally comprising a feature extraction module configured to use the one or more hardware processors to:
access a set of feature types to extract from the document;
employ a recursive descent to traverse the document based on the set of feature types; and
extract the plurality of discontinuous text segments from textual content of the document based on the set of feature types and the recursive descent.