# US 11,113,326 B2 · Automatic extraction of domain specific terminology from a large corpus > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 11,113,326 B2 - Document ID: US11113326B2 - Status: Granted - Kind: B2 - Application number: 15/881,521 - Filed: 2018-01-26 - Published: 2019-08-01 - Granted: 2021-09-07 - Earliest claimed: 2018-01-26 - Assignees: International Business Machines Corporation - Applicants: International Business Machines Corporation - Inventors: Md Faisal M. Chowdhury; Sharon M. Trewin - Web page: https://patentgrep.com/patent/US11113326B2 ## CPC - G06F 16/3344: {using natural language analysis} - G06F 16/335: Filtering based on additional data, e.g. user or group profiles (filtering in web context G06F16/9535, G06F16/9536) - G06F 16/358: Browsing; Visualisation therefor - G06F 40/242: Dictionaries - G06F 40/284: Lexical analysis, e.g. tokenisation or collocates - G06F 40/30: Semantic analysis - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract A method of extracting jargon from a document corpus stored in a database using a processor and a user interface is described herein. A sub-domain input is entered through the user interface to initiate a review of the document corpus stored in the database. The processor separates the document corpus into at least one sub-corpus and a remainder corpus. The at least one sub-corpus is defined by the sub-domain input. A first topic model and a second topic model are built to generate respective topic similarity scores for at least one term extracted from the at least one sub-corpus and at least one corresponding term extracted from the remainder corpus. The respective topic similarity scores are compared by the processor to identify jargon terms and thereby provide a list of jargon terms through the user interface. ## Claims ### Claim 1 A method of extracting jargon from a document corpus stored in a database using a user interface and a processor with a review module, a separation module, a topic model module, a similarity score module, and a term extraction module, the method comprising: receiving a sub-domain input by the processor through the user interface to initiate a review by the review module of the document corpus stored in the database; separating the document corpus with the separation module into at least one sub-corpus and a remainder corpus by the processor, the at least one sub-corpus defined by the sub-domain input, the remainder corpus defined by the document corpus less the at least one sub-corpus; building a first topic model using the topic model module to evaluate at least one term extracted from the at least one sub-corpus and a second topic model using the topic model module to evaluate the at least one term extracted from the remainder corpus; generating a first topic similarity score with the similarity score module for a select one of the at least one term by the first topic model based on a first meaning of the select term as used in the at least one sub-corpus and a second topic similarity score with the similarity score module for the select term by the second topic model based on a second meaning of the select term as used in the remainder corpus; comparing the first and second topic similarity scores by the term extraction module, wherein, for the select term, as a result of a corresponding difference between the first and second topic similarity scores for the select term in the sub-corpus measured against the select term in the remainder corpus being greater than a predetermined threshold, the select term is identified as a jargon term; and providing a list of jargon terms including the jargon term through the user interface. ### Claim 2 (depends on 1) The method according to claim 1, further comprising: ranking the at least one jargon term according to a minimum difference between the respective topic similarity scores of the first topic model and the second topic model; and providing a document cluster from the at least one sub-corpus, each document in the document cluster including the at least one jargon term from the list of jargon terms. ### Claim 3 (depends on 2) The method according to claim 2, further comprising: selecting the document cluster from the at least one sub-corpus; and tagging the document cluster to correspond to the sub-domain input. ### Claim 4 (depends on 1) The method according to claim 1, wherein separating the document corpus into the at least one sub-corpus and the remainder corpus further includes: defining the at least one sub-corpus by at least one of a distributional semantic method or individual user selection. ### Claim 5 (depends on 4) The method according to claim 4, further comprising associating an individual document with a plurality of sub-corpora. ### Claim 6 (depends on 4) The method according to claim 4, wherein the distributional semantic method includes: building an initial topic model from the document corpus to provide the at least one sub-corpus; and filtering documents in the at least one sub-corpus according to select parameters to define a condensed corpus. ### Claim 7 (depends on 6) The method according to claim 6, wherein building the initial topic model further includes: selectively excluding a document term from the initial topic model when the document term is at least one of functional or comprises a frequency count below a pre-determined minimum. ### Claim 8 (depends on 7) The method according to claim 7, wherein building the initial topic model further includes: selectively excluding a document from the initial topic model when fewer than a select number of terms in the document exceed the pre-determined minimum frequency count. ### Claim 9 (depends on 6) The method according to claim 6, wherein providing the at least one sub-corpus includes: obtaining a vector for the sub-domain input from the initial topic model; mapping document terms to corresponding vectors in the initial topic model to determine an initial topic similarity score relative to the sub-domain input; and providing at least one document where the initial topic similarity score exceeds a particular value for a select number of terms in the at least one document. ### Claim 10 (depends on 9) The method according to claim 9, wherein obtaining the vector for the sub-domain input further includes: at least one of locating the sub-domain input or averaging the corresponding vectors of each mapped document term to provide an aggregate vector for the at least one document; and including the at least one document in the condensed corpus when a distance between the aggregate vector and the sub-domain input vector is less than a select minimum distance. ### Claim 11 (depends on 9) The method according to claim 9, wherein obtaining the vector for the sub-domain input further includes: using a proxy vector when the sub-domain input is not found in the initial topic model, the proxy vector defined by averaging vectors of few alternative terms significant to the sub-domain input; and including the at least one document in the condensed corpus when a distance between the proxy vector and an aggregate vector is less than the select minimum distance. ### Claim 12 (depends on 6) The method according to claim 6, further comprising extracting terms from each of the at least one sub-corpora independently of each other. ### Claim 13 (depends on 12) The method according to claim 12, further comprising extracting the at least one term by at least one of term frequency-inverse document frequency, residual inverse document frequency, term frequency, average term frequency, chi-square, c-value, mapping the at least one term to an embedding, or comparing term frequency in the at least one sub-corpus to term frequency in the remainder corpus. ### Claim 14 (depends on 12) The method according to claim 12, further comprising: building the first topic model to evaluate the at least one term from the condensed corpus; and building the second topic model to evaluate the at least one corresponding term from the document corpus less the condensed corpus. ### Claim 15 (depends on 14) The method according to claim 14, wherein identifying the at least one jargon term further includes: defining the respective topic similarity scores based on the condensed corpus and the document corpus less the condensed corpus; and filtering non-jargon terms defined by a minimum difference between the respective topic similarity scores being below a pre-determined threshold. ### Claim 16 (depends on 15) The method according to claim 15, further comprising: identifying at least one of a characteristic or usage of non-filtered terms that is different in the condensed corpus than in the document corpus less the condensed corpus, the at least one of the characteristic or usage being indicative of the at least one jargon term. ### Claim 17 (depends on 16) The method according to claim 16, further comprising: reducing the ranking of the at least one jargon term when at least one of a frequency count of the at least one jargon term is higher in an alternative sub-domain or a difference between the frequency count of the sub-domain and the alternative sub-domain is less than a select threshold. ### Claim 18 A system for extracting jargon from a document corpus, the system comprising: a database storing the document corpus; an interface electrically connected to said database, said interface configured to receive a request to initiate a review of the document corpus based on a sub-domain input and provide a list of jargon terms; and a processor in communication with said database and said interface, said processor having a review module configured to review the document corpus in response to an initiation by said interface; a separation module configured to separate the document corpus into at least one sub-corpus defined by the sub-domain input and a remainder corpus defined by the document corpus less the at least one sub-corpus; a topic model module configured to build a first topic model to evaluate at least one term extracted from the at least one sub-corpus and a second topic model to evaluate the at least one term extracted from the remainder corpus; a similarity score module configured to generate and compare a first topic similarity score for a select one of the at least one term by the first topic model based on a first meaning of the select term as used in the at least one sub-corpus and a second topic similarity score by the second topic model based on a second meaning of the select term as used in the remainder corpus, wherein, for the select term, as a result of a corresponding difference between the first and second topic similarity scores for the select term in the sub-corpus measured against the select term in the remainder corpus being greater than a predetermined threshold, the select term is identified as a jargon term; and a term extraction module configured to extract the jargon term to provide a list of jargon terms including the jargon term through the user interface. ### Claim 19 A computer program product for extracting jargon from a document corpus, the computer program product comprising: a computer readable storage medium having stored thereon: first program instructions executable by a processor to cause the processor to receive a sub-domain input to initiate a review of the document corpus stored in the database; second program instructions executable by the processor to cause the processor to separate the document corpus into at least one sub-corpus and a remainder corpus, wherein the at least one sub-corpus defined by the sub-domain input and the remainder corpus defined by the document corpus less the at least one sub-corpus; third program instructions executable by the processor to cause the processor to build a first topic model to evaluate at least one term extracted from the at least one sub-corpus and a second topic model to evaluate the at least one term extracted from the remainder corpus; fourth program instructions executable by the processor to cause the processor to generate a first topic similarity score with the similarity score module for a select one of the at least one term by the first topic model based on a first meaning of the select term as used in the at least one sub-corpus and a second topic similarity score with the similarity score module for the select term by the second topic model based on a second meaning of the select term as used in the remainder corpus; fifth program instructions executable by the processor to cause the processor to compare the first and second topic similarity scores, wherein, for the select term, as a result of a corresponding difference between the first and second topic similarity scores for the select term in the sub-corpus measured against the select term in the remainder corpus being greater than a predetermined threshold, the select term is identified as a jargon term; and sixth program instructions executable by the processor to cause the processor to provide a list of jargon terms including the jargon term through a user interface. ### Claim 20 (depends on 19) The computer program product according to claim 19, further comprising seventh program instructions executable by the processor to cause the processor to identify the at least one jargon term by at least one of filtering the extracted terms when a minimum difference between the respective topic similarity scores is below a pre-determined threshold or identifying at least one of a characteristic or usage of the extracted terms that is different in the sub-corpus than in the remainder corpus.