Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1-7 are directed to a process, claims 8-14 are directed to an article of manufacture, and claims 15-20 are directed to a machine. Each claim therefore falls within one of the four statutory categories of 35 U.S.C. 101.
Step 2A, Prong One: Claim 1 recites identifying, from historical content, historical context relevant to a current block of input tokens; generating enriched input for predicting future tokens based on the current block of input tokens and the historical context; predicting, by a normal communication prediction model, a first set of future tokens to generate a predicted normal communication; predicting, by a fraudulent communication prediction model, a second set of future tokens to generate a predicted fraudulent communication; determining an overall discrepancy between the sequence of actual input tokens and the predicted normal and fraudulent communications; and determining whether the ongoing communication corresponds to a fraudulent communication. Under their broadest reasonable interpretation, these limitations recite mathematical concepts and mental processes. The predicting steps and the determining of an overall discrepancy are mathematical computations, which the specification describes as model computations over token sequences and as similarity computations between the actual tokens and each predicted communication. The determination of whether the ongoing communication corresponds to a fraudulent communication is an evaluation of judgment.
A person can perform each of the recited steps mentally, or with the aid of pen and paper. A person listening to an ongoing conversation can recall the earlier portions of that conversation that are relevant to what is being said now; can consider what a legitimate caller would be expected to say next and what a fraudulent caller would be expected to say next; can listen to what the caller actually says next; can compare what was actually said against each expectation; and can judge whether the call is fraudulent. Claim 1 therefore recites an abstract idea. Independent claims 8 and 15 recite the same operations as claim 1 and likewise recite the abstract idea. The dependent claims recite further operations that are part of an abstract idea, including obtaining a relevance score for each adjacency pair, ranking the pairs, and selecting the top ranked pairs (claims 2, 9, and 16); determining a first metric representing importance of terms, a second metric representing semantic similarity, and combining them in accordance with an operational parameter (claim 3, 10, and 17); iteratively predicting a look-ahead future token and adding it to the enriched input (claims 4, 11, and 18); computing a first discrepancy, a second discrepancy, and an overall discrepancy (claims 5, 12, and 19); and obtaining a fraud likelihood metric and generating a fraud signal (claims 6, 13, and 20).
Separating the abstract idea from the additional elements, the additional elements recited in the claims are: none (claim 1, in which every limitation falls within the abstract idea); a machine-readable and non-transitory medium having information recorded thereon which, when read by the machine, causes the machine to perform the abstract idea (claim 8); an enriched input generator, a normal communication predictor, a fraudulent communication predictor, and a fraud determiner, each implemented by a processor (claim 15); and determining an action that includes at least one of terminating the ongoing communication and flagging the ongoing communication for a review (claims 7, 14, and 20).
Step 2A, Prong Two: The additional elements fail to integrate the abstract idea into a practical application. The claims are not directed to, or limited to, a technical solution solving a technical problem. They fail to provide an improvement to a technology or the functioning of a computer. See MPEP 2106.04(d)(1). The problem addressed by the claims is that a person may be deceived by a fraudulent caller, and the claimed solution is to consider what is being said during the conversation and to judge whether it is fraudulent. Any advance provided by the claims lies in the abstract idea itself, namely in the recited prediction and discrepancy operations, and an improvement to an abstract idea is not an improvement to a technology or to the functioning of a computer.
Claim 1 recites no additional elements, and therefore recites nothing that could integrate the abstract idea into a practical application. To the extent the receiving of a current block of input tokens and the receiving of additional input tokens from the ongoing communication are considered additional elements rather than part of the abstract idea, they amount to mere data gathering, which is insignificant extra-solution activity. See MPEP 2106.05(g).
With respect to claims 8 and 15, the additional elements merely recite at a high level of generality, general purpose computing structure that is used as tools for implementing the abstract idea. The medium of claim 8 is recited only as a medium having information recorded thereon that causes a machine to perform the recited operations, and the enriched input generator, normal communication predictor, fraudulent communication predictor, and fraud determiner of claim 15 are recited only as being implemented by a processor, without any detail as to how the recited operations are performed. Thus, the examiner finds the additional elements are mere instructions to implement the judicial exception. See MPEP 2106.05(f) in light of 2106.04(d).
With respect to claims 7, 14, and 20, the recited action is performed after the fraud determination has already been made and does not change how that determination is reached. Under the broadest reasonable interpretation, the limitation is satisfied by flagging the ongoing communication for a review, which merely reports the result of the analysis, and terminating the ongoing communication likewise merely acts upon the result of the analysis. This is insignificant post-solution activity. See MPEP 2106.05(g). As such, the examiner must conclude the claimed invention is not integrated into a practical application.
Step 2B: The claims fail to recite significantly more than the abstract idea itself. Similar to the analysis for step 2A, prong 2, the claims fail to provide an improvement to a technology or the functioning of a computer. The additional elements merely recite, at a high level of generality, general purpose computing structure that is used as tools for implementing the abstract idea. Thus, the additional elements are mere instructions to implement the judicial exception. See MPEP 2106.05(f). Claim 1 recites no additional elements, and the additional elements of the remaining claims, considered individually and as an ordered combination, do not amount to significantly more than the abstract idea. Accordingly, claims 1-20 are not patent eligible.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The references relied upon in the following rejections are:
Mautone (US 2026/0129118 A1), hereinafter “Mautone.”
Gainsboro et al. (US 2011/0082874 A1), hereinafter “Gainsboro.”
Fu et al. (US 2026/0037365 A1), hereinafter “Fu.”
Veron et al. (US 2026/0148079 A1), hereinafter “Veron.”
Lim (KR 20230070752 A), hereinafter “Lim.”
Claim(s) 1, 5-8, 12-15, 19, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Mautone (US 2026/0129118 A1) in view of Gainsboro (US 2011/0082874 A1) and further in view of Fu (US 2026/0037365 A1).
Regarding claim 1, claim 1 recites "receiving a current block of input tokens from an ongoing communication." Mautone provides for this limitation because Mautone discloses sampling the audio of a live telephone call routed through a switch platform during a selected time period and transcribing at least a portion of the sample to text for analysis (Mautone, [0013], [0016]-[0017]). The transcribed text of the ongoing telephone call corresponds to the claimed current block of input tokens because it is the textual content of the ongoing communication that is received for analysis.
Claim 1 further recites "identifying, from historical content, historical context relevant to the current block of input tokens." Mautone does not expressly provide for this limitation. Fu provides for this limitation because Fu discloses that a large language model (LLM) is trained on entries of a system log and learns each user's individual, customary pattern of behavior from the history of system events captured in the log (Fu, [0042] - [0044]). Fu's history of system events associated with the particular user corresponds to the claimed historical context relevant to the current block because it is the stored prior activity of the same user whose current activity is being analyzed.
Claim 1 further recites "generating enriched input for predicting future tokens based on the current block of input tokens and the historical context." Fu provides for this limitation because Fu discloses that the LLM receives as input the current detected event entries from the system log (Fu, [0046]) and that the predicted next system event is based at least in part on a previous user event together with current attributes such as a current date, a current time, and a current geographical location (Fu, [0095]). The combined current-plus-historical input on which the prediction is based corresponds to the claimed enriched input for predicting future tokens.
Claim 1 further recites "predicting, by a normal communication prediction model based on the enriched input, a first set of future tokens to generate a predicted normal communication." Fu provides for this limitation because Fu discloses generating, using the LLM trained on the users' customary (normal) behavior, a next expected system event or a sequence of next expected system events for the user (Fu, [0044], [0085], [0092]-[0093]). Fu's LLM trained on customary behavior corresponds to the claimed normal communication prediction model, and the predicted sequence of next expected events corresponds to the claimed first set of future tokens, because each predicts the expected normal continuation of the ongoing activity. Fu does not provide for the communication-specific environment; in the combination, Fu's prediction technique is applied to the transcribed tokens of Mautone's ongoing telephone call.
Claim 1 further recites "predicting, by a fraudulent communication prediction model based on the enriched input, a second set of future tokens to generate a predicted fraudulent communication." Mautone provides for part of this limitation because Mautone discloses a fraudulent communication model, namely an LLM trained on observed telephone calls that are known fraudulent or scam telephone calls (Mautone, [0019]-[0020]). Gainsboro provides for part of this limitation because Gainsboro discloses positively building and maintaining fraud-side models, namely voice models of confirmed imposters and fraud perpetrators, and evaluating the ongoing call against those models together with the genuine-side model (Gainsboro, [0092], [0096], [0100]). However, neither Mautone nor Gainsboro expressly provides for the fraud-trained model predicting a second set of future tokens to generate a predicted fraudulent communication in the manner recited. Fu provides for the missing prediction technique because Fu discloses an LLM generating a sequence of predicted next events (Fu, [0044]. [0085]). Fu does not provide for a model trained on fraudulent communications; rather, Fu is relied upon only for its general technique of having a trained LLM predict the next elements of a sequence. Mautone provides the fraud-trained LLM, Gainsboro provides evaluating the same ongoing call against both a genuine-side model and fraud-side models, and Fu provides the technique of having a trained model predict the expected next elements.
Claim 1 further recites "receiving additional input tokens from the ongoing communication to generate a sequence of actual input tokens." Mautone provides for this limitation because Mautone discloses continuing to sample and transcribe the ongoing telephone call over the selected time period (Mautone, [0017]). Fu also provides for this limitation because Fu discloses receiving, as input, the detected (actual) system events as they occur (Fu, [0046]).
Claim 1 further recites "determining an overall discrepancy between the sequence of actual input tokens and the predicted normal and fraudulent communications." Fu provides for part of this limitation because Fu discloses comparing a detected system event to the predicted next expected system event and determining a probability of deviation and a risk level based on the difference (Fu, [0046]-[0047], [0086]-[0087]), which is a discrepancy between the actual input and the normal-side prediction. Gainsboro provides for the two-sided determination because Gainsboro discloses evaluating the ongoing input against both the genuine-participant model and the imposter models and combining the two evaluations into one determination: Gainsboro reports calls whose input scores low against the genuine voice model while scoring high against the imposter models (Gainsboro, [0103]), and Gainsboro generates a likely-imposter alarm when the ongoingly measured input drifts away from the correct participant's model by more than half the distance to the closest possible imposter model (Gainsboro, [0118]).
Claim 1 further recites "determining whether the ongoing communication corresponds to a fraudulent communication." Mautone provides for this limitation because Mautone discloses that the AI engine returns a probability of the presence of a scam in the telephone call and the system generates an analysis and report of the detected scam (Mautone, [0021]-[0022]). Gainsboro likewise provides for the determination through the likely-imposter alarm condition and real-time alerts on high-potential-loss calls (Gainsboro, [0097], [0118]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify Mautone's telephone scam detection system to evaluate the ongoing call against both a genuine (normal) side model and fraud-side models as provided for by Gainsboro. Gainsboro discloses, in the same field of detecting fraud on telephone calls, that comparing the live input to both the genuine model and the imposter models, and flagging calls that score low against the genuine model while scoring high against the imposter models, provides real-time alrts on high-potential-loss calls (Gainsboro, [0096]-[0097], [0103] and permits greater tolerance for normal variation before a false imposter determination is made (Gainsboro, [0113]). Mautone already provides the fraud-trained LLM and the scam probability. Gainsboro's two-sided evaluation adds a normal-side reference so that Mautone's determination reflects which class the ongoing call is actually closer to, improving the reliability of the determination that Mautone already makes.
It would further have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the Mautone/Gainsboro system so that each class model predicts the expected next elements of the ongoing communication and the actually received continuation is compared against the predictions, as provided for by Fu. Fu discloses that predicting the next expected elements from the user's history and comparing each detected element to the prediction enables the system to identify deviations as they occur and to identify new types of attacks or security threats (Fu, [0018], [0044]-[0047], [0085]-[0088]). Applying Fu's predict-and-compare technique to Mautone's transcribed call tokens, with Gainsboro's two-sided genuine/fraud evaluation, enables the fraud determination to be made during the ongoing call as each additional portion of the call is received, rather than only after scoring a completed sample. Thus, the combination provides for the limitations of claim 1.
Regarding claim 5, claim 5 depends from claim 1 and further recites "computing a first discrepancy between the sequence of actual input tokens and the predicted normal communication, and a second discrepancy between the sequence of actual input tokens and the predicted fraudulent communication; and determining the overall discrepancy based on the first discrepancy and the second discrepancy." As discussed above with respect to claim 1, Mautone in view of Gainsboro and Fu provides for the limitations of claim 1. Fu provides for the first discrepancy because Fu discloses determining a probability of whether the detected events deviate from the predicted next expected events (Fu, [0046]-[0047]). Gainsboro provides for the second discrepancy and for combining the two because Gainsboro discloses scoring the ongoing input against the imposter (fraud-side) models (Gainsboro, [0096], [0103]) and making the determination from the relationship between the input's drift from the correct participant's model and its distance to the closest imposter model (Gainsboro, [0118]), including reporting calls that score low against the genuine model and high against the imposter models (Gainsboro, [0103]). The determination made from both the genuine-side evaluation and the fraud-side evaluation corresponds to the claimed overall discrepancy based on the first and second discrepancies.
Regarding claim 6, claim 6 depends from claim 1 and further recites "obtaining a fraud likelihood metric based on the overall discrepancy, wherein the fraud likelihood metric representing confidence that the ongoing communication is fraudulent; and generating a fraud signal based on the fraud likelihood metric indicating a fraud detection result." Fu provides for the fraud likelihood metric because Fu discloses that the prediction includes a probability associated with whether the detected events deviate from the predicted next events (Fu, [0047]) and that a risk level is determined based on the difference and compared against a threat threshold (Fu, [0087], [0049]-[0050]). Mautone provides for the fraud signal because Mautone discloses that the AI engine returns a probability of the presence of a scam and the systems posts a report of the analysis (Mautone, [0021]-[0022]).
Regarding claim 7, claim 7 depends from claim 6 and further recites "determining, based on the fraud signal, an action directed to the ongoing communication, wherein the action includes at least one of: terminating the ongoing communication; and flagging the ongoing communication for a review." Mautone provides for this limitation because Mautone discloses sending a Suspicious Call Report notifying the customer that a suspicious call was intercepted, including the call audio, transcript, and analysis for review, and optionally blocking the source of the call (Mautone, [0025]-[0028]). Mautone's Suspicious Call Report corresponds to the claimed flagging of the ongoing communication for a review; the claim requires at least one of the recited actions, and flagging for a review is provided for.
Regarding claim 8, claim 8 recites "a machine-readable and non-transitory medium having information recorded thereon, wherein the information, when read by the machine, causes the machine to perform the following steps." Fu provides for this limitation because Fu discloses a non-transitory computer-readable medium storing code including instructions executable by one or more processors to perform the described method (Fu, [0007]). Mautone likewise discloses implementation on a computer system (Mautone, [0032]). The remaining steps recited in claim 8 are substantially identical to the steps recited in claim 1, and claim 8 is rejected for the same reasons set forth above with respect to claim 1.
Regarding claims 12, 13, and 14, claims 12-14 depend from claim 8 (claim 14 depending from claim 13) and contain limitations substantially identical to those of claims 5, 6, and 7, respectively, and are rejected for the same reasons set forth above with respect to claims 5, 6, and 7, and with respect to claim 8 for the medium limitation.
Regarding claim 15, claim 15 recites "an enriched input generator implemented by a processor," "a normal communication predictor implemented by a processor," "a fraudulent communication predictor implemented by a processor," and "a fraud determiner implemented by a processor," each configured to perform operations. Fu provides for processor-implemented components because Fu discloses an apparatus comprising one or more memories storing processor-executable code and one or more processors that execute the code to perform the operations (Fu, [0005], [0082]). Mautone provides for the fraud detection system and the AI engine implemented on a computer system (Mautone, [0015], [0018], [0032]). The functions for which the enriched input generator, the normal communication predictor, the fraudulent communication predictor, and the fraud determiner are configured substantially identical to the corresponding operations in claim 1: the enriched input generator performs the receiving, identifying, and generating operations; the normal communication predictor performs the normal-model prediction; the fraudulent communication predictor performs the fraud-model prediction; and the fraud determiner performs the receiving of additional tokens, the overall discrepancy determination, and the fraud determination. Claim 15 is therefore rejected for the same reasons set forth above with respect to claim 1.
Regarding claims 19 and 20, claim 19 depends from claim 15 and contains limitations substantially identical to those of claim 5, and is rejected for the same reasons set forth above with respect to claim 5. Claim 20 depends from claim 15 and contains limitations substantially identical to the combined limitations of claims 6 and 7, and is rejected for the same reasons set forth above with respect to claims 6 and 7.
Claims 4, 11, and 18, are rejected under 35 U.S.C. 103 as being unpatentable over Mautone in view of Gainsboro and Fu as applied to claims 1, 8, and 15 above, and further in view of Veron (US 2026/0148079 A1).
Regarding claim 4, claim 4 depends from claim 1 and further recites that "predicting the first set of future tokens comprises: predicting, using the normal communication prediction model, a look-ahead normal future token based on the enriched input, adding the predicted look-ahead normal future token to the enriched input, repeating the predicting a look-ahead normal future token and adding the predicted look-ahead normal future token until the first set of future tokens are predicted, and creating the predicted normal communication based on the first set of future tokens," and further recites, in parallel, that predicting the second set of future tokens comprises the same look-ahead predicting, adding, repeating, and creating operations performed, "using the fraudulent communication prediction model" to create "the predicted fraudulent communication." As discussed above with respect to claim 1, Mautone in view of Gainsboro and Fu provides for the limitations of claim 1, including predicting the first and second sets of future tokens by the two models. However, the combination does not expressly provide for the recited iterative process of predicting a look-ahead future token, adding the predicted look-ahead future token, adding the predicted look-ahead future token to enriched input, and repeating until the set of future tokens is predicted.
Veron provides for the missing iterative prediction technique because Veron discloses that a generative language model generates a first token, applies the generated token as an input during a second pass, and repeats in a loop, successively generating and adding tokens to the output from the preceding pass and applying the composite sequence as the input to the decoder during each subsequent pass, sequentially generating one token at a time (auto-regression) until predicting a token that represents the end of the response, at which point the generated response is output (Veron, [0150]-[0152]). Veron does not provide for fraud detection or for the communication-specific environment; rather, Veron is relied upon only for its general autoregressive token-by-token generation technique for generative language models.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to implement the predictions of the normal and fraudulent models in the Mautone/Gainsboro/Fu combination using Veron's autoregressive generation technique because Veron discloses this technique as the manner in which transformer-based generative language models, the same type of LLM employed by Mautone (Mautone, [0019]) and Fu (Fu, [0040], [0044]), sequentially generate an output sequence. Using Veron's known token-by-token generation would allow each of the two class models in the combination to produce its predicted continuation in a known manner with predictable results. Thus, the combination provides for the limitations of claim 4.
Regarding claims 11 and 18, claims 11 and 18 depend from claims 8 and 15, respectively, and contain limitations substantially identical to those of claim 4, and are rejected for the same reasons set forth above with respect to claim 4.
Claims 2, 3, 9, 10, 16, and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Mautone in view of Gainsboro and Fu as applied to claims 1, 8, and 15 above, and further in view of Veron and Lim (KR 20230070752 A).
Regarding claim 2 depends from claim 1 and further recites "determining a context window associated with the historical context based on the current block; retrieving historical content within the context window, wherein the retrieved historical content corresponds to previous communications represented by a plurality of prompt/response adjacency pairs; obtaining a relevance score between each of the plurality of adjacency pairs and the current block of input tokens; ranking the plurality of adjacency pairs based on their respective relevance scores; selecting a predetermined number of top ranked adjacency pairs; and creating the historical context for the current block of input tokens based on the selected top ranked adjacency pairs." As discussed above with respect to claim 1, Mautone in view of Gainsboro and Fu provide for the limitations of claim 1, including identifying historical context. However, the combination does not expressly provide for retrieving previous communications as prompt/response adjacency pairs, scoring and ranking the pairs by relevance to the current block, and selecting the top ranked pairs to create the historical context.
Veron provides for retrieving prior conversation history as context because Veron discloses a retrieval-augmented generation (RAG) component that retrieves additional information, including a prior stored conversation history of the user, identifies relevant text, and provides the retrieved information to the generative language model as additional context as part of the model input (Veron, [0136]-[0137]). Veron's providing the retrieved conversation history to the model as additional context corresponds to the claimed creating the historical context for the current block based on the retrieved content.
Lim provides for the pair retrieval, scoring, ranking, and selection. Regarding the plurality of prompt/response adjacency pairs, Lim discloses a question-and-answer database that is, in the words of the translation of record, "constructed by storing past queries and previously selected responses to corresponding past queries in correspondence with each other" (Lim, description of the sparse embedding processing unit 110; see also Lim, claims 1 and 7). Lim's stored past query and its corresponding response correspond to the claimed prompt/response adjacency pair, because each pairs a prior prompt with the response given to it. Regarding retrieving the historical content, Lim discloses that the system "obtains a set including responses to past queries from the query response database" and compares them against the input query (Lim, description of the sparse embedding processing unit 110). Regarding obtaining a relevance score for each pair, Lim discloses a processor that "obtains a sparse embedding processing result using the input query and responses to past queries obtained from the question-and-answer database, calculates the similarity between the candidate sentence obtained from the question-and-answer database and the input sentence corresponding to the query to obtain a dense embedding processing result, and combines the sparse embedding processing result and the dense embedding processing result to obtain a final score" (Lim, claim 1). This final score, computed for each candidate with respect to the input, corresponds to the claimed relevance score between each adjacency pair and the current block of input tokens. Regarding ranking, Lim discloses that it is "possible to return a response that is relatively higher in descending order" (Lim, description of the scaling unit 130), which corresponds to ranking the adjacency pairs by their relevance scores. Regarding selecting the top ranked pairs and creating the historical context, Lim discloses that "a candidate response having a relatively high final score may be determined as an optimal response" and is then output to the user (Lim, description of the final score acquisition unit 140), which corresponds to selecting a predetermined number of top ranked pairs, here the single highest-scoring pair.
Regarding the claimed context window, Gainsboro provides for determining a limited window of the stored content to be searched based on the current communication because Gainsboro discloses limiting the set of stored models searched to those associated with the circumstances of the current call, namely the persons known to have access to the phone at the time, in order to make the search computationally efficient (Gainsboro, [0109], [0123]).
Lim does not provide for fraud detection or for an ongoing communication; Lim is directed to obtaining a response to a query based on semantic similarity, and is relied upon only for its technique of scoring, ranking, and selecting stored prompt/response pairs by relevance to a current input. Veron is relied upon only for retrieving relevant stored conversation history and providing it to the model as context. Mautone provides the ongoing telephone communication, and Fu provides based the prediction on the identified history.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the combination to retrieve the relevant portions of the stored conversation history and provide them as context for the predictions, as provided for by Veron, because Veron discloses that retrieval-augmented generation identifies relevant text and enhances the model input so that the model's output is more relevant (Veron, [0136]). It would further have been obvious to perform that retrieval by scoring and ranking the stored prompt/response pairs using Lim's combined scoring technique because Lim discloses that its final score identifies the optimal stored response and reports that its scoring achieves comparable accuracy at substantially reduced response time, stating that one configuration required a "driving time" of "approximately 3533 microseconds" while an equally accurate configuration had a driving time of "only about 24 microseconds" such that "the response time is remarkably shortened" (Lim, description of FIG. 4). Fast retrieval is of particular benefit in the combination because Mautone analyzes an ongoing telephone call in which the fraud determination is made while the call is in progress. It would further have been obvious to limit the retrieved historical content to a window associated with the current block, as provided for by Gainsboro, in order to reduce the computation required for the relevance scoring (Gainsboro, [0109], [0123]). Thus, the combination provides for the limitations of claim 2.
Regarding claim 3, claim 3 depends from claim 2 and further recites "determining a first metric representing importance of terms in the adjacency pair; determining a second metric representing semantic similarity between the adjacency pair and the current block of input tokens; retrieving an operational parameter for combining the first and the second metric; and determining the relevance score for the adjacency pair based on the first metric and the second metric in accordance with the operational parameter." Lim provides for these limitations. For the first metric, Lim discloses a sparse embedding score computed with the BM25 algorithm using "an inverse document frequency function, which is a function to prevent excessive weight being added to words (q_i) that appear too often but lack practical meaning, for example, articles or dependent nouns" (Lim, description of the sparse embedding processing unit 110 and Equation 1); this score is a metric representing the importance of terms in the pair. For the second metric, Lim discloses a dense embedding score in which "the dense embedding processing unit 120 calculates a similarity (for example, cosine similarity) between the embedding vector value of the input sentence 81 and the embedding vector value of the candidate sentence 91" (Lim, description of the dense embedding processing unit 120 and Equation 3); this score is a metric representing semantic similarity between the stored pair and the current input. For the operational parameter and the combination, Lim discloses obtaining the final score as a weighted sum of the sparse and dense embedding results in which "the weight w_q may be arbitrarily set by a user or a designer, or may be a value defined through experimental results" (Lim, description of the final score acquisition unit 140 and Equation 5). Lim's stored weight corresponds to the claimed operational parameter because it is a retrievable value that controls how the two metrics are combined into the relevance score. The motivation to combine set forth above with respect to claim 2 applies equally here, because Lim's combined sparse-plus-dense scoring is the technique being incorporated.
Regarding claims 9, 10, 16, and 17, claims 9-10 depend from claim 8 (claim 10 depending from claim 9) and claims 16-17 depend from claim 15 (claim 17 depending from claim 16), and contain limitations substantially identical to those of claims 2 and 3, respectively. Claims 9-10 and 16-17 are rejected for the same reasons set forth above with respect to claims 2 and 3, and with respect to claims 8 and 15 for the medium and system limitations.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ZAIN J AHMED whose telephone number is (571)270-0251. The examiner can normally be reached 8am - 4pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Jeffrey L Nickerson can be reached at (469) 295-9235. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/Jeffrey Nickerson/Supervisory Patent Examiner, Art Unit 2432
Prosecution Insights