SYSTEMS AND METHODS FOR SEGMENTING DOCUMENTS

• US 20200134024A1
• Filed: 10/30/2019
• Published: 04/30/2020
• Est. Priority Date: 10/30/2018
• Status: Active Grant

• Current Assignee: The Florida International University Board of Trustees (State University System of Florida)
• Original Assignee: The Florida International University Board of Trustees (State University System of Florida)
• Inventors: Banisakher, Deya, Rishe, Naphtali, Finlayson, Mark
• Granted Patent: US 10,949,622 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 40/205 Parsing
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G16H 15/00 ICT specially adapted for medical reports, e.g. generation or transmission thereof
• CPC Class Codes: G16H 50/70 for mining of medical data, e.g. analysing previous cases of other patients

Abstract

Systems and methods for automatically modeling the discourse structure of psychiatric reports and segmenting these reports into various sections are provided. The systems and methods can be based around a model that learns the section types, positions, and sequence and can automatically segment unlabeled text in a psychiatric report into the corresponding sections. Knowledge of the ordering of the sections can improve the performance of a section classifier and a text segmenter. A Hierarchical Hidden Markov Model (HHMM) can be trained and can categorize sections in psychiatric reports into a predefined section label.

Claims

1. A computer-based system of segmenting a psychiatric evaluation report into sections, the system comprising:

a processor; and

a non-transitory computer-readable medium in operable communication with the processor and comprising program instructions stored thereon that, when executed, cause the processor to;

receive text data of the psychiatric evaluation report;

analyze the text data;

learn and build a first model for an order and presence of the sections in the psychiatric evaluation report;

learn and build a second model to describe distinctive features of the respective sections in the psychiatric evaluation report; and

apply a combination of the first model and the second model to simultaneously identify boundaries of the respective sections and to label section types of the respective sections, thereby segmenting the psychiatric evaluation report, the first model being a Hierarchical Hidden Markov Model (HHMM).

2. The system according to claim 1, the second model comprising a respective language model for each section type in the psychiatric evaluation report.

3. The system according to claim 2, each respective language model being an n-gram language model.

4. The system according to claim 3, the learning and building of the second model comprising using the n-gram language models as omission probabilities for the HHMM.

5. The system according to claim 1, the applying of the combination of the first model and the second model comprising following a decoding scheme using a Viterbi algorithm.

6. The system according to claim 5, the decoding scheme comprising applying the following equation to obtain the most likely labeling of each respective section, where O* is a set of the sections, n is an index of the sections, w₀^kn is a first long sequence of words, w₁^kn is a second long sequence of words, s is a state of the MIMI, k_n is a word index for section n, and i is an index of the states of the HHMM:

7. The system according to claim 1, the learning and building of the first model comprising using states of the HHMM to respectively represent the sections in the psychiatric evaluation report.

8. The system according to claim 1, the program instructions, when executed, further causing the processor to train the system using preexisting psychiatric evaluation reports, the training comprising using training n-gram language models to capture distinctive lexical information contained within individual sections of the preexisting psychiatric evaluation reports, and

the training being performed before the step of receiving the text data of the psychiatric evaluation report.

9. The system according to claim 8, the training n-gram language models being bigram language models.

10. The system according to claim 1, the program instructions, when executed, further causing the processor to classify unknown blocks of text, in the text data of the psychiatric evaluation report, into the respective sections.

11. A non-transitory computer-readable medium comprising program instructions stored thereon that, when executed, cause a processor to:

receive text data of a psychiatric evaluation report;

analyze the text data;

learn and build a first model for an order and presence of sections in the psychiatric evaluation report;

learn and build a second model to describe distinctive features of the respective sections in the psychiatric evaluation report; and

apply a combination of the first model and the second model to simultaneously identify boundaries of the respective sections and to label section types of the respective sections, thereby segmenting the psychiatric evaluation report into the sections,

the first model being a Hierarchical Hidden Markov Model (HHMM).

12. The non-transitory computer-readable medium according to claim 11, the second model comprising a respective language model for each section type in the psychiatric evaluation report, and

each respective language model being an n-gram language model.

13. The non-transitory computer-readable medium according to claim 12, the learning and building of the second model comprising using the n-gram language models as omission probabilities for the HHMM.

14. The non-transitory computer-readable medium according to claim 11, the applying of the combination of the first model and the second model comprising following a decoding scheme using a Viterbi algorithm.

15. The non-transitory computer-readable medium according to claim 14, the decoding scheme comprising applying the following equation to obtain the most likely labeling of each respective section, where O* is a set of the sections, n is an index of the sections, w₀^kn is a first long sequence of words, w₁^kn is a second long sequence of words, s is a state of the HHMM, k_n is a word index for section n, and i is an index of the states of the HHMM:

16. The non-transitory computer-readable medium according to claim 11, the learning and building of the first model comprising using states of the HHMM to respectively represent the sections in the psychiatric evaluation report.

17. The non-transitory computer-readable medium according to claim 11, the program instructions, when executed, further causing the processor to train the first model using preexisting psychiatric evaluation reports, the training comprising using training n-gram language models to capture distinctive lexical information contained within individual sections of the preexisting psychiatric evaluation reports, and

the training being performed before the step of receiving the text data of the psychiatric evaluation report.

18. The non-transitory computer-readable medium according to claim 17, the training n-gram language models being bigram language models.

19. The non-transitory computer-readable medium according to claim 11, the program instructions, when executed, further causing the processor to classify unknown blocks of text, in the text data of the psychiatric evaluation report, into the respective sections.

20. A computer-based system of segmenting a psychiatric evaluation report into sections, the system comprising:

a processor; and

a non-transitory computer-readable medium in operable communication with the processor and comprising program instructions stored thereon that, when executed, cause the processor to;

train the system using preexisting psychiatric evaluation reports, the training comprising using first n-gram language models to capture distinctive lexical information contained within individual sections of the preexisting psychiatric evaluation reports;

receive text data of the psychiatric evaluation report, the training of the system being performed before the receiving of the text data of the psychiatric evaluation report;

analyze the text data;

learn and build a first model for an order and presence of the sections in the psychiatric evaluation report;

learn and build a second model to describe distinctive features of the respective sections in the psychiatric evaluation report;

apply a combination of the first model and the second model to simultaneously identify boundaries of the respective sections and to label section types of the respective sections, thereby segmenting the psychiatric evaluation report; and

classify unknown blocks of text, in the text data of the psychiatric evaluation report, into the respective sections,

the first model being a Hierarchical Hidden Markov Model (HHMM),

the second model comprising a respective second n-gram language model for each section type in the psychiatric evaluation report,

the learning and building of the second model comprising using the second n-gram language models as omission probabilities for the HHMM,

the applying of the combination of the first model and the second model comprising following a decoding scheme using a Viterbi algorithm,

the decoding scheme comprising applying the following equation to obtain the most likely labeling of each respective section, where O* is a set of the sections, n is an index of the sections, w₀^kn is a first long sequence of words, w₁^kn is a second long sequence of words, s is a state of the HHMM, k_n is a word index for section n, and i is an index of the states of the HHMM;