Abstract: A method of determining influence of language elements in script to an overall classification of the script by perturbing the dataset representing a conversation. In some instances, for example, in the conversation, the language elements and turns within the conversation (e.g., in a chat bot) are analyzed for their influence in escalation or non-escalation of the conversation to a higher level of resolution, e.g., to a human representative or manager.

Abstract: A method of determining influence of language elements in script to an overall classification of the script includes performing a sentiment analysis of the language elements. In some instances, for example, in a conversation, the language elements and turns within the conversation (e.g., in a chat bot) are analyzed for their influence in escalation or non-escalation of the conversation to a higher level of resolution, e.g., to a human representative or manager.

Abstract: Attention weights in a hierarchical attention network indicate the relative importance of portions of a conversation between an individual at one terminal and a computer or a human agent at another terminal. Weighting the portions of the conversation after converting the conversation to a standard text format allows for a computer to graphically highlight, by color, font, or other indicator visible on a graphical user interface, which portions of a conversation led to an escalation of the interaction from an intelligent virtual assistant to a human customer service agent.

Abstract: A system and method for selecting an anomaly detection method from among a plurality of known anomaly detection methods includes selecting a set of anomaly detections methods based on characteristics of the time series, such as missing time steps, trend, drift, seasonality and concept drift. From among the applicable anomaly detection methods, the selection may be further informed by annotated predicted anomalies, and based on the annotations, turning the parameters for each respective anomaly detection method. Thereafter, the anomaly detection methods are scored and then further tuned according to human actions in identifying anomalies or disagrees with anomalies in the time series.

Abstract: In an implementation, a method for detecting anomalies in textual items is provided. The method includes: receiving a first plurality of textual items by a computing device; training a language model using the received first plurality of textual items by the computing device; after training the language model, receiving a second plurality of textual items by the computing device; calculating a cross-entropy for each textual item in the second plurality of textual items by the computing device using the language model; and detecting an anomaly in at least one of the textual items of the second plurality of textual items by the computing device using the calculated cross-entropies.

Abstract: Systems and methods are provided framework for automatically choosing the appropriate generalized linear model (GLM) given a time series of count data, and for anomaly detection on time series data. A dispersion parameter is determined and used to determine whether the count data is overdispersed data or underdispersed data. The overdispersed data or the underdispersed data is used to determine a GLM to apply on the dataset. Using the determined GLM on the data, anomalies can be determined.

Abstract: Systems and methods are described to combine two or more linear models into a combined linear model. Two or more linear models and an observation of interest are selected. The linear models are concurrent with respect to the observation of interest. The observation of interest includes a class value and a feature vector. A distance is selected, and a plurality of feature vectors are selected that are within the distance of the feature vector associated with the observation of interest. These feature vectors are input to the selected linear models and a plurality of class values are generated. These class values and the selected feature vectors are used to generate a combined linear model. The combined model is either a mean local linear surrogate model or a linearly weighted local surrogate model.

Abstract: A framework for reducing the number of textual items reviewed to determine the source of or reason for an anomaly in a time series that is used to track metrics in textual data is provided. According the framework, textual items in a time window corresponding to the anomaly are ranked according to the cross-entropy as determined by applying a language model to the relevant textual items and ranking textual items that most likely triggered an anomaly in time series data based on the cross-entropy value. In an aspect, a predetermined number of textual items having the highest cross-entropy are provided or all textual items having cross-entropy value higher than predetermine threshold are provided.

Abstract: Disclosed are a framework and method for selecting an anomaly detection method for each of a plurality of class of time series based on characteristics a time series example that represents an expected form of data. The method provides classification of a given time series into one of known classes based on expected properties of the time series, filtering the set of possible detection methods based on the time series class, evaluating the remaining detection methods on the given time series using the specific evaluation metric and selecting and returning a recommended anomaly detection method based on the specific evaluation metric.

Abstract: Detecting fraudulent activity can be a complex, manual process. In this paper, we adapt statistical properties of count data in a novel algorithm to uncover records exhibiting high risk for fraud. Our method identifies shelves, partitioning data under the counts using a Student's t-distribution. We apply this methodology on a univariate dataset including cumulative results from phone calls to a customer service center. Additionally, we extend this technique to multivariate data, illustrating that the same method is applicable to both univariate and multivariate data.

Abstract: This disclosure describes techniques and architectures for evaluating conversations. In some instances, conversations with users, virtual assistants, and others may be analyzed to identify potential risks within a language model that is employed by the virtual assistants and other entities. The potential risks may be evaluated by administrators, users, systems, and others to identify potential issues with the language model that need to be addressed. This may allow the language model to be improved and enhance user experience with the virtual assistants and others that employ the language model.

Abstract: Features, libraries, and techniques are provided herein for determining the kinds of relational language that are present. Applying audio, emojis, and sentiment shifts as features may be used to determine whether the customer is providing backstory, whether there is ranting, etc. Textual features may be considered, as well as audio features may be considered.

Abstract: A method of determining influence of language elements in script to an overall classification of the script includes performing a sentiment analysis of the language elements. In some instances, for example, in a conversation, the language elements and turns within the conversation (e.g., in a chat bot) are analyzed for their influence in escalation or non-escalation of the conversation to a higher level of resolution, e.g., to a human representative or manager.

Abstract: Attention weights in a hierarchical attention network indicate the relative importance of portions of a conversation between an individual at one terminal and a computer or a human agent at another terminal. Weighting the portions of the conversation after converting the conversation to a standard text format allows for a computer to graphically highlight, by color, font, or other indicator visible on a graphical user interface, which portions of a conversation led to an escalation of the interaction from an intelligent virtual assistant to a human customer service agent.

Abstract: Techniques and architectures for analyzing conversations between users and virtual assistants to identify instances where the virtual assistants have not satisfied user requests are described. The techniques and architectures may use such analysis to tag conversations regarding unsatisfied user requests, provide information to users regarding conversations with unsatisfied user requests, learn conversation or contextual data for unsatisfied user requests, and/or perform a variety to other processes to improve the virtual assistants.