IP Library Granted Patent US 12,737,547
Granted Patent B2
US 12,737,547 · App. 18/507,322 · Granted Sep 15, 2026

Semantic text segmentation based on topic recognition

Inventors: Mattan Serry (Herzliya, IL); Oron Nir (Herzliya, IL)
Assignee: Microsoft Technology Licensing, LLC
G06F40/30
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,737,547
App. No.
18/507,322
Granted
Sep 15, 2026
Kind
B2
Abstract

Systems and methods for semantic temporal segmentation based on topic recognition are disclosed. Text classification and segmentation may be used to index media content for subsequent searches. The method may include using a text classification model to semantically analyze sentences in a text file (such as a transcript) to determine topics with which sentences are associated. The output of the text classification model may be provided to a text segmentation model to enable more-accurate identification of text segments (e.g., paragraphs) within the text file. In some examples, the output of a text segmentation model is provided to a text classification model to enable the text classification model to perform more-accurate classification based on text segments rather than (or in addition to) performing classification on single sentences. The classification of the text segments may be used to assign labels to the text segments to enable subsequent searches based on the labels.

Claims (67)

1 . A computer-implemented method, comprising:

obtaining a first set of tokens associated with a first sentence of a transcript and a second set of tokens associated with a second sentence of the transcript;

determining, by a text classification model and based on the first set of tokens, a first probability that a first class is associated with the first sentence;

determining, by the text classification model and based on the second set of tokens, a second probability that the first class is associated with the second sentence;

identifying a first text segment that includes the first sentence and the second sentence based on:

a comparison of a semantic similarity of the first class associated with the first probability to the first class associated with the second probability; and

a comparison of a temporal proximity of a first timestamp associated with the first sentence to a second timestamp associated with the second sentence;

determining a third probability that the first class is associated with the first text segment based on the first set of tokens and the second set of tokens, wherein the third probability is used to adjust a parameter of the text classification model;

determining a label associated with the first text segment based on the third probability;

in response to receiving input indicating a selection of the label, identifying media content corresponding to the first text segment; and

providing the media content for presentation.

2 . The computer-implemented method of claim 1 , wherein the first probability, the second probability, and the third probability are determined by a text classification model comprising a neural network, the method further comprising:

adjusting a parameter of the text classification model based on the third probability.

3 . The computer-implemented method of claim 1 , further comprising:

storing, with the label associated with the first text segment, a first timestamp associated with the first text segment, wherein the first timestamp is determined based on a second timestamp associated with the first sentence, a third timestamp associated with the second sentence, or both.

4 . The computer-implemented method of claim 1 , further comprising:

transcribing media content to generate a plurality of sentences including the first sentence and the second sentence;

tokenizing the first sentence to generate the first set of tokens; and

tokenizing the second sentence to generate the second set of tokens.

5 . The computer-implemented method of claim 1 , wherein determining the first probability comprises:

encoding the first sentence into an embeddings vector using the first set of tokens; and

determining the first probability based on the embeddings vector.

6 . A computer-implemented method, comprising:

generating a first embeddings vector representing semantic content of a first sentence of a transcript and a second embeddings vector representing semantic content of a second sentence of the transcript;

generating an adjusted first embeddings vector based on the first embeddings vector and the second embeddings vector by adjusting the first embeddings vector;

generating an adjusted second embeddings vector based on the first embeddings vector and the second embeddings vector by adjusting the second embeddings vector;

identifying, based on the adjusted first embeddings vector and the adjusted second embeddings vector, a first text segment that includes the first sentence and the second sentence, wherein identifying the first text segment comprises comparing a semantic similarity of a first class associated with the adjusted first embeddings vector and the adjusted second embeddings vector;

determining, based on the first text segment, a probability that the first class is associated with the first text segment;

receiving input indicating a selection of a label;

determining the label is associated with the first text segment based on the probability;

identifying media content corresponding to the first text segment; and

providing the media content for presentation.

7 . The computer-implemented method of claim 6 , further comprising:

transcribing media content to generate a transcript comprising the first sentence and the second sentence.

8 . The computer-implemented method of claim 6 , wherein the first sentence is associated with a first timestamp and the second sentence is associated with a second timestamp, the method further comprising:

determining a timestamp associated with the first text segment based on the first timestamp, the second timestamp, or both; and

storing the label associated with the first text segment with the timestamp associated with the first text segment.

9 . The computer-implemented method of claim 6 , wherein the first embeddings vector is generated by a first text encoder in parallel with the second embeddings vector being generated by a second text encoder.

10 . The computer-implemented method of claim 9 , wherein a fusion-layer transformer adjusts the first embeddings vector and the second embeddings vector based on an output of a global attention mechanism executed on the first embeddings vector and the second embeddings vector.

11 . A system, comprising:

one or more processing systems; and

memory storing instructions that, when executed by the one or more processing systems, cause the system to:

obtain a first set of tokens associated with a first sentence of a transcript and a second set of tokens associated with a second sentence of the transcript;

determine, based on the first set of tokens, a first probability that a first class is associated with the first sentence;

determine, based on the second set of tokens, a second probability that the first class is associated with the second sentence;

identify a first text segment that includes the first sentence and the second sentence based on the first probability, a first timestamp associated with the first sentence, the second probability, and a second timestamp associated with the second sentence; and

determine a third probability that the first class is associated with the first text segment based on the first set of tokens and the second set of tokens, wherein the third probability is used to adjust a parameter of a text classification model;

determining a label associated with the first text segment based on the third probability;

in response to receiving input indicating a selection of the label, identifying media content corresponding to the first text segment; and

providing the media content for presentation.

12 . The system of claim 11 , wherein the first probability, the second probability, and the third probability are determined by the text classification model; and

wherein the instructions parameter of the text classification model is based on the third probability.

13 . The system of claim 11 , wherein the instructions further cause the system to:

store, with the label associated with the first text segment, a first timestamp associated with the first text segment, wherein the first timestamp is determined based on a second timestamp associated with the first sentence a third timestamp associated with the second sentence, or both.

14 . The system of claim 11 , wherein the instructions further cause the system to:

transcribe media content to generate a plurality of sentences including the first sentence and the second sentence;

tokenize the first sentence to generate the first set of tokens; and

tokenize the second sentence to generate the second set of tokens.

15 . The system of claim 11 , wherein determining the first probability comprises:

encoding the first sentence into an embeddings vector using the first set of tokens; and

determining the first probability based on the embeddings vector.

16 . The system of claim 11 , wherein at least one of the first probability or the second probability is determined by the text classification model, wherein the text classification model is a language model trained using a list of predetermined classes.

17 . The computer-implemented method of claim 1 , wherein the first probability and the second probability are determined by the text classification model; and

wherein the first text segment is identified by a text segmentation model.

18 . The computer-implemented method of claim 17 , wherein the text classification model determines the first class is associated with the first sentence based on determining a probability that the sentence is related to the first class exceeds an absolute threshold.

19 . The computer-implemented method of claim 17 , wherein the text classification model determines the first class is associated with the first sentence based on determining a first probability that the sentence is related to the first class exceeds a second probability that the sentence is related to a second class.

20 . The computer-implemented method of claim 17 , wherein the text classification model is a language model trained using a list of predetermined classes.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Nov 13, 2023
From: SERRY, MATTAN; NIR, ORON
To: MICROSOFT TECHNOLOGY LICENSING, LLC
Reel/Frame 065540/0104 →
Continuity (1)
Related Publication 20250156642A1 · May 15, 2025
References Cited (15)
US 9812028B1 · Elchik · 2017 [cited by examiner]
US 10560734B2 · Jassin et al. · 2020 [cited by applicant]
US 10762375B2 · Ronen et al. · 2020 [cited by applicant]
US 10936630B2 · Ronen et al. · 2021 [cited by applicant]
US 20030182631A1 · Tsochantaridis · 2003 [cited by examiner]
US 20210174146A1 · Nir et al. · 2021 [cited by applicant]
US 20240004911A1 · Rossetto · 2024 [cited by examiner]
US 20240038271A1 · de Juan · 2024 [cited by examiner]
US 20240126799A1 · Anderson · 2024 [cited by examiner]
US 20240419927A1 · Agrawal · 2024 [cited by examiner]
“Topic Modeling Basics”, Retrieved From: https://cloud.google.com/contact-center/insights/docs/topic-modeling-basics, Sep. 15, 2023, 2 Pages. [cited by applicant]
Beltagy, et al., “Longformer: The Long-Document Transformer”, In repository of arXiv:2004.05150v2, Dec. 2, 2020, 17 Pages. [cited by applicant]
He, et al., “DebertaV3: Improving Deberta using Electra-Style Pre-Training with Gradient-Disentangled Embedding Sharing”, In repository of arXiv:2111.09543v2, Dec. 8, 2021, 17 Pages. [cited by applicant]
Nir, et al., “CAST: Character labeling in Animation using Self-supervision by Tracking”, In Journal of Computer Graphics Forum, vol. 41, Issue 2, May 24, 2022, pp. 135-145. [cited by applicant]
Sun, et al., “TDViT: Temporal Dilated Video Transformer for Dense Video Tasks”, In Proceedings of 17th European Conference on Computer Vision, Oct. 23, 2022, 17 Pages. [cited by applicant]