# US 11,308,945 B1 · Data-preserving text redaction for text utterance data > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 11,308,945 B1 - Document ID: US11308945B1 - Status: Granted - Kind: B1 - Application number: 16/560,839 - Filed: 2019-09-04 - Granted: 2022-04-19 - Earliest claimed: 2019-09-04 - Assignees: Amazon Technologies, Inc. - Applicants: Amazon Technologies, Inc. - Inventors: Thomas Drake; Oluwaseyi Feyisetan; Thomas Diethe - Web page: https://patentgrep.com/patent/US11308945B1 ## CPC - G06F 40/30: Semantic analysis - G10L 15/197: Probabilistic grammars, e.g. word n-grams - G06F 21/6245: {Protecting personal data, e.g. for financial or medical purposes} - G06F 40/284: Lexical analysis, e.g. tokenisation or collocates - G10L 15/22: Procedures used during a speech recognition process, e.g. man-machine dialogue - G10L 15/30: Distributed recognition, e.g. in client-server systems, for mobile phones or network applications - G10L 15/1822: {Parsing for meaning understanding} - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING ## Abstract A hypernym of a word in utterance data may be probabilistically determined. The utterance data may correspond to a spoken query or command. A redacted utterance may be derived by replacing the word with the hypernym. The hypernym may be determined by applying noise to a position in a hierarchical embedding that corresponds to the word. The word may be identified as being potentially sensitive. The hierarchical embedding may be a Hyperbolic embedding that may indicate hierarchical relationships between individual words of a corpus of words, such as “red” is a “color” or “Austin” is in “Texas.” Noise may be applied by obtaining a first value in Euclidean space based on a second value in Hyperbolic space, and obtaining a third value in Hyperbolic space based on the first value in Euclidean space. The second value in Hyperbolic space may correspond to the word. ## Claims ### Claim 1 A computer-implemented method, comprising: obtaining a transcript derived from an audio signal encoding an utterance, the utterance corresponding to a query; determining a first portion of the transcript and a second portion of the transcript, wherein the second portion has been classified as including potentially sensitive information; determining, based at least in part on a second word in the second portion of the transcript, a first value in a Hyperbolic representation of a corpus of words, the first value corresponding to a first position in the Hyperbolic representation; as a result of adding a noise value to the Hyperbolic representation at the first position, obtaining a second position in the Hyperbolic representation; determining a first word corresponding to the second position in the Hyperbolic representation; replacing the second portion with the first word to form a redacted transcript; and providing the redacted transcript to cause the query to be executed. ### Claim 2 (depends on 1) The computer-implemented method of claim 1, wherein determining the first portion of the transcript is based at least in part on: determining a candidate n-gram comprising a substring having a maximum length of one less than a length of the transcript; determining that the candidate n-gram is associated with a longest prefix and a highest frequency count of the differentially private exploration tree; and selecting the candidate n-gram as the first portion of the transcript. ### Claim 3 (depends on 1) The computer-implemented method of claim 1, wherein the second portion includes a second word, and further comprising: determining a word embedding of the Hyperbolic representation for the second word, wherein the noise value is added to the word embedding of the Hyperbolic representation to generate a perturbed word embedding; and determining the first word from the perturbed embedding. ### Claim 4 (depends on 1) The computer-implemented method of claim 1, wherein the Hyperbolic representation of the corpus of words indicates a plurality of hierarchical relationships between individual words in the corpus of words. ### Claim 5 A system, comprising: one or more processors; and memory that stores computer-executable instructions that are executable by the one or more processors to cause the system to: determine a first portion of utterance data and a second portion of the utterance data, wherein the second portion is classified as including potentially sensitive information; obtain a hierarchical representation of a corpus of words; add noise to the hierarchical representation at a position corresponding to at least the second portion to obtain a different position of the hierarchical representation, the different position of the hierarchical representation corresponding to a word, wherein the noise is determined at least in part by simulating a plurality of random walks through a hyperbolic distribution of the hierarchical representation; replace the second portion with the word to form redacted utterance data; and provide the redacted utterance data to be processed. ### Claim 6 (depends on 5) The system of claim 5, wherein the hierarchical representation includes a Hyperbolic embedding of the corpus of words and indicates a plurality of hierarchical relationships between individual words of the corpus of words. ### Claim 7 (depends on 6) The system of claim 6, wherein the noise is determined at least in part by: obtaining a first value corresponding to the position in the Hyperbolic embedding; translating the first value to Euclidian space to obtain a second value; and translating the second value in Euclidean space to a third value in Hyperbolic space. ### Claim 8 (depends on 7) The system of claim 7, wherein the word is determined based at least in part on the third value in Hyperbolic space. ### Claim 9 (depends on 5) The system of claim 5, wherein the added noise includes a Hyperbolic Laplace noise. ### Claim 10 (depends on 5) The system of claim 5, wherein: the first portion is based at least in part on a differentially private exploration tree. ### Claim 11 (depends on 5) The system of claim 5, wherein the utterance data is at least one of a command or a search query. ### Claim 12 A non-transitory computer-readable storage medium storing thereon executable instructions that, as a result of being executed by one or more processors of a computer system, cause the computer system to at least: determine a portion of stored utterance data; probabilistically determine a first word that is a hypernym of a second word of the portion, wherein the first word is probabilistically determined as a result of applying a noise value to a Hyperbolic word embedding at a first value, the first value corresponding to the second word; determine a redacted utterance at least in part by replacing the second word of the portion of the utterance data with the first word; and store the redacted utterance. ### Claim 13 (depends on 12) The non-transitory computer-readable storage medium of claim 12, wherein the noise value that is added to the Hyperbolic word embedding is determined based at least on a Hyperbolic noise distribution and the first value. ### Claim 14 (depends on 13) The non-transitory computer-readable storage medium of claim 13, wherein the noise value that is applied is determined based at least in part by sampling a probability distribution of the Hyperbolic word embedding that includes the first word and the second word. ### Claim 15 (depends on 14) The non-transitory computer-readable storage medium of claim 14, wherein sampling the probability distribution includes at least simulating a plurality of random walks through the probability distribution. ### Claim 16 (depends on 14) The non-transitory computer-readable storage medium of claim 14, wherein the Hyperbolic word embedding is a Poincaré model. ### Claim 17 (depends on 12) The non-transitory computer-readable storage medium of claim 12, wherein: the first word is determined at least in part by: determining a second value in a Euclidean word embedding that is based at least in part on the first value; and determining a third value in Hyperbolic space based at least in part on the second value in the Euclidean word embedding. ### Claim 18 (depends on 17) The non-transitory computer-readable storage medium of claim 17, wherein the third value includes coordinates in a Poincaré model. ### Claim 19 (depends on 17) The non-transitory computer-readable storage medium of claim 17, wherein the second value is translated into a Lorentz model.