Abstract: In one embodiment, one or more computing systems may determine a first set of bins that collectively cover a pre-determined numerical range with each bin covering a sub-range of the pre-determined range. The system may determine a second set of bins that collectively cover the pre-determined range with each covers a different but overlapping sub-range with respect to a corresponding bin in the first bin set. The system may access a value that falls within the pre-determined range. The system may determine that the value falls within a first bin of the first bin set and a second bin of the second bin set. The system may determine a positive value for each the first and second bins. The positive values indicate an association level of the value with the first and second bins. The system may determine a representation of the value based on the positive values.

Abstract: A computing system identifies an incoming voice call from a user device to an agent device associated with the computing system. The computing system generates a transcription of the incoming voice call using one or more natural language processing techniques. The computing system extracts a problem description from the transcription. The problem description indicates a topic for the incoming voice call. A first machine learning model estimates a situation vector from the problem description. A second machine learning model identifies a pre-existing situation vector that closely matches the estimated situation vector. The computing system retrieves a situation description that corresponds to the identified pre-existing situation vector.

Abstract: During end-to-end training of a Deep Neural Network (DNN), a differentiable estimator subnetwork is operated to estimate a functionality of an external software application. Then, during inference by the trained DNN, the differentiable estimator subnetwork is replaced with the functionality of the external software application, by enabling API communication between the DNN and the external software application.

Abstract: This disclosure provides systems, methods and apparatuses for recommending items to users of a recommendation system. In some implementations, the recommendation system determines a plurality of contribution values based on interactions between a plurality of users and a plurality of items. Each of the plurality of contribution values represents a confidence level that a respective user prefers a respective item. The recommendation system further determines a capacity of each of the plurality of items. The capacity of each item represents a maximum number of users to which the item can be recommended. The recommendation system recommends one or more items of the plurality of items to each of the plurality of users based at least in part on the plurality of contribution values and the capacities of the plurality of items.

Abstract: A method including receiving, in a machine learning model (MLM), a corpus including words. The MLM includes layers configured to extract keywords from the corpus, plus a retrospective layer. A first keyword and a second keyword from the corpus are identified in the layers. The first and second keywords are assigned first and second probabilities. Each probability is a likelihood that a keyword is to be included in a key phrase. A determination is made, in the retrospective layer, of a first probability modifier that modifies the first probability based on a first dependence relationship between the second keyword being placed after the first keyword. The first probability is modified using the first probability modifier. The first modified probability is used to determine whether the first keyword and the second keyword together form the key phrase. The key phrase is stored in a non-transitory computer readable storage medium.

Abstract: A method includes generating recommendations and user structures by applying a recommender machine learning model to training user information and item information, and generating, from the user structures and by applying a demographic machine learning model, demographic predictions of users represented by the user structures. The method further includes generating a first accuracy measure of the demographic machine learning model based on a first comparison of the demographic predictions with demographics of the users. A recommender loss function is generated based on the first accuracy measure and a second comparison of the recommendations with selections of users, where the recommender loss function uses the first accuracy measure to suppress detectability by the demographic machine learning model. The method further includes updating the recommender machine learning model according to the recommender loss function.

Abstract: A computing system identifies an incoming voice call from a user device to an agent device associated with the computing system. The computing system generates a transcription of the incoming voice call using one or more natural language processing techniques. The computing system extracts a problem description from the transcription. The problem description indicates a topic for the incoming voice call. A first machine learning model estimates a situation vector from the problem description. A second machine learning model identifies a pre-existing situation vector that closely matches the estimated situation vector. The computing system retrieves a situation description that corresponds to the identified pre-existing situation vector.

Abstract: A method including receiving, in a machine learning model (MLM), a corpus including words. The MLM includes layers configured to extract keywords from the corpus, plus a retrospective layer. A first keyword and a second keyword from the corpus are identified in the layers. The first and second keywords are assigned first and second probabilities. Each probability is a likelihood that a keyword is to be included in a key phrase. A determination is made, in the retrospective layer, of a first probability modifier that modifies the first probability based on a first dependence relationship between the second keyword being placed after the first keyword. The first probability is modified using the first probability modifier. The first modified probability is used to determine whether the first keyword and the second keyword together form the key phrase. The key phrase is stored in a non-transitory computer readable storage medium.

Abstract: Systems and methods for automatically identifying problem-relevant sentences in a transcript are disclosed. In an example method, a transcript may be received of a first support call. A region of the first support call transcript may be identified, and first customer utterances may be detected in the region using a trained classification model. A trained regression model may estimate a relevancy to the problem statement of each of the first customer utterances, and one or more most problem-relevant statements may be selected from the first customer utterances, based on the estimated relevancies.

Abstract: A computer-implemented method, computerized apparatus and computer program product for minimum coordination passage scoring. Given a candidate passage in a document collection potentially matching a query received, a set of overlapping terms between the candidate passage and the query is determined. For each overlapping term in the set, a first measure of a weight of the term in the query, a second measure of a weight of the term in the candidate passage, and a third measure of a specificity of the term in the document collection are calculated. a function of the first and second measure is evaluated to obtain a value reflecting a condition on the relation therebetween. A minimum coordination score representing a relative similarity between the candidate passage and the query is determined based on the value and the first, second and third measures obtained for each of the overlapping terms.

Abstract: Certain aspects of the present disclosure provide techniques for privacy-preserving execution of a workflow in a software application. Embodiments include generally includes receiving homomorphically encrypted inputs from a client device corresponding to user-provided data needed to calculate a result for a step of a workflow in the software application. A result is calculated for the step of the workflow using the received homomorphically encrypted inputs. The calculated result is returned to the client device. The calculated result is homomorphically encrypted as a result of calculating the result using the received homomorphically encrypted inputs.

Abstract: Automatic keyphrase labeling and machine learning training may include a processor extracting a plurality of keywords from at least one search query that resulted in a selection of a document appearing in a search result. For each of the plurality of keywords, the processor may determine a probability that the keyword describes the document. The processor may generate one or more keyphrases by performing processing including selecting each of the plurality of keywords having a probability greater than a predetermined threshold value for insertion into at least one of the one or more keyphrases and assembling the one or more keyphrases from the selected plurality of keywords. The processor may label the document with the keyphrase.

Abstract: A method including embedding, by a trained issue MLM (machine learning model), a new natural language issue statement into an issue vector. An inner product of the issue vector with an actions matrix is calculated. The actions matrix includes centroid-vectors calculated using a clustering method from a second output of a trained action MLM which embedded prior actions expressed in natural language action statements taken as a result of prior natural issue statements. Calculating the inner product results in probabilities associated with the prior actions. Each of the probabilities represents a corresponding estimate that a corresponding prior action is relevant to the issue vector. A list of proposed actions relevant to the issue vector is generated by comparing the probabilities to a threshold value and selecting a subset of the prior actions with corresponding probabilities above the threshold. The list of proposed actions is transmitted to a user device.

Abstract: Operating an encoder with double decoder machine learning models include executing, on a transcript, an encoder machine learning model to generate an encoder output, and executing a situation decoder machine learning model on the encoder output to obtain a situation model output having a situation identifier, and executing a trouble decoder machine learning model using the encoder output to obtain a trouble identifier. The method further includes outputting the situation identifier and the trouble identifier.

Abstract: A method for mitigating cold starts in recommendations includes receiving a request that identifies a requested page and identifying a content vector of the requested page. The content vector is generated based on providing text of the requested page to a neural network text encoder. The method further includes selecting, based on the content vector, a link to a cold start page that does not satisfy a threshold level of interaction data. The selected link is ranked above a second link to a warm page that does satisfy the threshold level of the interaction data. The method further includes presenting the requested page with the selected link.

Abstract: A method of training machine learning models (MLMs). An issue vector is generated using an issue MLM to generate a first output including first embedded natural language issue statements. An action vector is generated using an action MLM to generate a second output including related embedded natural language action statements. The issue and action MLMs are of a same type. An inner product of the first and second output is calculated, forming a third output. The third output is processed according to a sigmoid gate process to predict whether a given issue statement and corresponding action statement relate to a same call, resulting in a fourth output. A loss function is calculated from the fourth output by comparing the fourth output to a known result. The issue MLM and the action MLM are modified using the loss function to obtain a trained issue MLM and a trained action MLM.

Abstract: Automatic keyphrase labeling and machine learning training may include a processor extracting a plurality of keywords from at least one search query that resulted in a selection of a document appearing in a search result. For each of the plurality of keywords, the processor may determine a probability that the keyword describes the document. The processor may generate one or more keyphrases by performing processing including selecting each of the plurality of keywords having a probability greater than a predetermined threshold value for insertion into at least one of the one or more keyphrases and assembling the one or more keyphrases from the selected plurality of keywords. The processor may label the document with the keyphrase.

Abstract: This disclosure provides systems, methods and apparatuses for recommending items to users of a recommendation system. In some implementations, the recommendation system determines a plurality of contribution values based on interactions between a plurality of users and a plurality of items. Each of the plurality of contribution values represents a confidence level that a respective user prefers a respective item. The recommendation system further determines a capacity of each of the plurality of items. The capacity of each item represents a maximum number of users to which the item can be recommended. The recommendation system recommends one or more items of the plurality of items to each of the plurality of users based at least in part on the plurality of contribution values and the capacities of the plurality of items.

Abstract: Operating an encoder with double decoder machine learning models include executing, on a transcript, an encoder machine learning model to generate an encoder output, and executing a situation decoder machine learning model on the encoder output to obtain a situation model output having a situation identifier, and executing a trouble decoder machine learning model using the encoder output to obtain a trouble identifier. The method further includes outputting the situation identifier and the trouble identifier.

Abstract: A method for mitigating cold starts in recommendations includes receiving a request that identifies a requested page and identifying a content vector of the requested page. The content vector is generated based on providing text of the requested page to a neural network text encoder. The method further includes selecting, based on a rank engine and the content vector, a link to a cold start page that does not satisfy a threshold level of interaction data. The rank engine ranks the selected link above a second link to a warm page that does satisfy the threshold level of the interaction data. The method further includes presenting the requested page with the selected link.