Hierarchical segmentation of unstructured text using neural networks

Patent metadata

Field	Value
Publication number	US-12346361-B2
Application number	US-202318511186-A
Country	US
Kind code	B2
Filing date	Nov 16, 2023
Priority date	Nov 16, 2023
Publication date	Jul 1, 2025
Grant date	Jul 1, 2025

Abstract

Embodiments are disclosed for a digital design system trained to segment unstructured text into topically coherent segments. The method may include receiving unstructured text, the unstructured text including a sequence of sentences. The disclosed systems and methods further comprise generating, by a neural network, a hierarchically segmented tree structure representing the unstructured text. The tree structure comprises a plurality of tree structure nodes, where a node of the tree structure nodes represents a sentence from the sequence of sentences. The segments and sub-segments of the unstructured text can then be determined based on node data for nodes of the hierarchically segmented tree structure. Using the determined segments and sub-segments of the unstructured text, a modified representation of the unstructured text can be displayed.

First claim

We claim:

1. A method comprising:
receiving unstructured text, the unstructured text including a sequence of sentences;
generating, by a neural network, a hierarchically segmented tree structure representing the unstructured text, the tree structure comprising tree structure nodes, wherein a leaf node of the tree structure nodes represents a sentence from the sequence of sentences;
determining segments and sub-segments of the unstructured text based on node data for the tree structure nodes of the hierarchically segmented tree structure; and
presenting for display a modified representation of the unstructured text based on the determined segments and sub-segments of the unstructured text.

2. The method of claim 1 , wherein generating the hierarchically segmented tree structure representing the unstructured text further comprises:
converting the hierarchically segmented tree structure representing the unstructured text from a binarized tree structure by:
identifying reducible nodes in the binarized tree structure, wherein reducible nodes do not represent a sentence from the sequence of sentences, and
for each reducible node, removing the reducible node and connecting direct descendants of the reducible node with a direct parent node of the reducible node.

3. The method of claim 2 , wherein generating the hierarchically segmented tree structure representing the unstructured text further comprises:
generating, using a text encoder, feature vectors for each sentence of the sequence of sentences;
generating, using a recurrent neural network, contextualized feature vectors using the generated feature vectors;
predicting a structure of the binarized tree structure; and
labeling each node of a plurality of nodes in the binarized tree structure as a reducible node or an irreducible node.

4. The method of claim 3 , wherein labeling each node of the plurality of nodes in the binarized tree structure as a reducible node or an irreducible node comprises:
for each node in the binarized tree structure:
determining a first probability value that a node is a reducible node and a second probability value that a node is an irreducible node,
when the first probability value is greater than the second probability value, determining the node is a reducible node, and
when the first probability value is lower than the second probability value, determining the node is an irreducible node.

5. The method of claim 1 , wherein presenting for display the modified representation of the unstructured text further comprises:
generating, by a topic generating model, summaries for each determined segment and sub-segment of the of the unstructured text; and
generating the modified representation of the unstructured text as a table of contents using the generated summaries.

6. The method of claim 5 , further comprising:
receiving a first user input selecting an entry in the table of contents for the unstructured text;
presenting a first summary of the entry, the first summary including second summaries for one or more sub-entries of the entry;
receiving a second user input selecting one of the second summaries; and
presenting a portion of the unstructured text corresponding to the selected second one of the second summaries.

7. The method of claim 1 , wherein child nodes of the hierarchically segmented tree structure having a same parent node represent sentences that are topically related.

8. A non-transitory computer-readable medium storing executable instructions, which when executed by a processing device, cause the processing device to perform operations comprising:
receiving unstructured text, the unstructured text including a sequence of sentences;
generating, by a neural network, a hierarchically segmented tree structure representing the unstructured text, the tree structure comprising tree structure nodes, wherein a leaf node of the tree structure nodes represents a sentence from the sequence of sentences;
determining segments and sub-segments of the unstructured text based on node data for the tree structure nodes of the hierarchically segmented tree structure; and
presenting for display a modified representation of the unstructured text based on the determined segments and sub-segments of the unstructured text.

9. The non-transitory computer-readable medium of claim 8 , wherein the operation of generating the hierarchically segmented tree structure representing the unstructured text further comprises:
converting the hierarchically segmented tree structure representing the unstructured text from a binarized tree structure by:
identifying reducible nodes in the binarized tree structure, wherein reducible nodes do not represent a sentence from the sequence of sentences, and
for each reducible node, removing the reducible node and connecting direct descendants of the reducible node with a direct parent node of the reducible node.

10. The non-transitory computer-readable medium of claim 9 , wherein the operation of generating the hierarchically segmented tree structure representing the unstructured text further comprises:
generating, using a text encoder, feature vectors for each sentence of the sequence of sentences;
generating, using a recurrent neural network, contextualized feature vectors using the generated feature vectors;
predicting a structure of the binarized tree structure; and
labeling each node of a plurality of nodes in the binarized tree structure as a reducible node or an irreducible node.

11. The non-transitory computer-readable medium of claim 10 , wherein the operation of labeling each node of the plurality of nodes in the binarized tree structure as a reducible node or an irreducible node further comprises:
for each node in the binarized tree structure:
determining a first probability value that a node is a reducible node and a second probability value that a node is an irreducible node,
when the first probability value is greater than the second probability value, determining the node is a reducible node, and
when the first probability value is lower than the second probability value, determining the node is an irreducible node.

12. The non-transitory computer-readable medium of claim 8 , wherein the operation of presenting for display the modified representation of the unstructured text further comprises:
generating, by a topic generating model, summaries for each determined segment and sub-segment of the of the unstructured text; and
generating the modified representation of the unstructured text as a table of contents using the generated summaries.

13. The non-transitory computer-readable medium of claim 12 , storing instructions that further cause the processing device to perform operations comprising:
receiving a first user input selecting an entry in the table of contents for the unstructured text;
presenting a first summary of the entry, the first summary including second summaries for one or more sub-entries of the entry;
receiving a second user input selecting one of the second summaries; and
presenting a portion of the unstructured text corresponding to the selected second one of the second summaries.

14. The non-transitory computer-readable medium of claim 8 , wherein child nodes of the hierarchically segmented tree structure having a same parent node represent sentences that are topically related.

15. A system comprising:
a memory component; and
a processing device coupled to the memory component, the processing device to perform operations comprising:
receiving unstructured text, the unstructured text including a sequence of sentences;
generat

Assignees

• Adobe Inc

Inventors

• Nair Inderjeet
• Shekhar Sumit
• Karanam Srikrishna
• Chhaya Niyati Himanshu
• Modani Natwar
• Srinivasan Balaji Vasan
• Garimella Aparna

Classifications

• G06F16/322

  Trees · CPC title

• G06F16/345 Primary

  Summarisation for human users · CPC title