Abstract: Methods and systems are described herein for improving information retrieval during a live conversation via context tokens and real-time natural language utterances. For example, the system may receive a first set of decrypted utterances spoken during a live conversation. The system may determine a token comprising token data associated with a second set of decrypted utterances spoken during the live conversation prior to the first set of decrypted utterances. The system may generate, via a secured model, a first query and a subset of utterances corresponding to the first set of utterances and the token data. The system may perform a first search based on the first query and a second search based on the subset of utterances to retrieve a set of relevant computer files. The system may generate, during the live conversation, a graphical representation of the set of relevant computer files.

Abstract: Described herein are systems, methods, and programming for determining a context source of a target large language model by leveraging token activations of a large language model having an accessible final layer for token activation extraction. For example, the techniques leverage a token accessible large language model to extract token activations associated with query-context-answer triplets, where the answer may be generated by the target large language model. The token activations can be passed to a trained classification model that learns patterns of token activations associated with context-based model outputs. The trained classification model can assign a label to the model-produced answer indicating whether the answer was generated by the target large language model using the input context included by the prompt or model memory of the target large language model.

Abstract: A device may receive an input indicating a prompt for a language model. The device may determine, using the language model, an output responsive to the prompt. The device may determine a degree by which the output relies on memory of the language model. The device may output, responsive to the degree by which the output relies on memory indicating an over-reliance on memory, a modified output that differs from the output.

Abstract: Described herein are systems, methods, and programming for generating training data to train a classification model to determine, using token activations associated with an answer generated by a large language model, whether the large language model use provided context or model memory to generate the answer. For example, the techniques can include creating, from query-context-answer (QCA) triplets, controlled conflicts by perturbing the answer to test whether the large language model uses the context to generate the answer, as opposed to model memory. These controlled conflicts can be labeled to indicate the ones that the model recognized as being based on context and those based on memory. Token activations associated with those answers can be generated and used, in conjunction with the labels, to train the classification model.

Abstract: Described herein are systems, methods, and programming for determining a context source of a large language model. For example, the techniques leverage a token accessible large language model to extract token activations associated with answers generated by the model in response to prompts. The token activations can be passed to a trained classification model that learns patterns of token activations associated with context-based model outputs. The trained classification model can assign a label to the model-produced answer indicating whether the answer was generated by the large language model using the input context included by the prompt or model memory of the large language model.

Abstract: Methods and systems are described herein for facilitating queryless information retrieval during a live conversation via a bifurcated model. For example, the system may receive, a first set of decrypted utterances spoken during the live conversation. The system may provide the first set of decrypted utterances as input to a bifurcated model comprising (i) a first bifurcated model portion associated with a dynamic utterance embedding space and (ii) a second bifurcated model portion associated with a static computer file embedding space. The system may then generate, via the bifurcated model, a first intermediate output from the first bifurcated model portion indicating a first utterance embedding and a second intermediate output from the second bifurcated model portion indicating a first computer file embedding. The system may then generate during the live conversation, a graphical representation of a computer file associated with the first computer file embedding.

Abstract: Systems and methods for the creation of human-readable cyber incident reports from cyber incident logs, in which the cyber incident reports may link cyber incidents recorded in a cyber incident log to the existing knowledge sources. To do so, the systems and methods overcome the technical problems of conventional systems as well as the technical problems inherent in adapting artificial intelligence solutions to the creation of cyber incident reports.

Abstract: Systems and methods for improved data processing of communications across computer networks using trifurcated prompts during communication exchanges are described. For example, the system may receive a first inbound communication, wherein the first inbound communication system may determine a first context for the first inbound communication based on the first text string. The system may process the first context in a perturbation model to determining a first perturbed context, wherein the perturbation model determines the first perturbed context by determining a first alternative token for a first token in the first context. The system may determine a first prompt for a first large language model based on the first perturbed context.

Abstract: Systems and methods for the creation of human-readable cyber incident reports from cyber incident logs, in which the cyber incident reports may link cyber incidents recorded in a cyber incident log to the existing knowledge sources. To do so, the systems and methods overcome the technical problems of conventional systems as well as the technical problems inherent in adapting artificial intelligence solutions to the creation of cyber incident reports.

Abstract: Systems and methods for the creation of human-readable cyber incident reports from cyber incident logs, in which the cyber incident reports may link cyber incidents recorded in a cyber incident log to the existing knowledge sources. To do so, the systems and methods overcome the technical problems of conventional systems as well as the technical problems inherent in adapting artificial intelligence solutions to the creation of cyber incident reports.

Abstract: The present disclosure describes complex modeling of dialogues that allows querying of the modeled dialogues. Embeddings may be generated for each multi-party dialogue of a plurality of multi-party dialogues. Embeddings may include speaker-aware embeddings, key-utterance embeddings, and/or discourse-aware embeddings. In addition to the embeddings, a directed acyclic graph (DAG) to show a relationship between the one or more utterances of the multi-party dialogue. The embeddings and the DAG may be stored in a datastore. In response to receiving a request to identify dialogues associated with a topic, the datastore may be queried to retrieve dialogues associated with the received topic. The dialogues may be provided to the requesting party, which may use the information retrieved from the datastore to respond to a requesting party.

Abstract: The described embodiments relate to systems, methods, and apparatus for providing a multimodal deep memory network (200) capable of generating patient diagnoses (222). The multimodal deep memory network can employ different neural networks, such as a recurrent neural network and a convolution neural network, for creating embeddings (204, 214, 216) from medical images (212) and electronic health records (206). Connections between the input embeddings (204) and diagnoses embeddings (222) can be based on an amount of attention that was given to the images and electronic health records when creating a particular diagnosis. For instance, the amount of attention can be characterized by data (110) that is generated based on sensors that monitor eye movements of clinicians observing the medical images and electronic health records.

Abstract: Techniques described herein relate to semi-supervised training and application of stacked autoencoders and other classifiers for predictive and other purposes. In various embodiments, a semi-supervised model (108) may be trained for sentence classification, and may combine what is referred to herein as a “residual stacked de-noising autoencoder” (“RSDA”) (220), which may be unsupervised, with a supervised classifier (218) such as a classification neural network (e.g., a multilayer perceptron, or “MLP”). In various embodiments, the RSDA may be a stacked denoising autoencoder that may or may not include one or more residual connections. If present, the residual connections may help the RSDA “remember” forgotten information across multiple layers. In various embodiments, the semi-supervised model may be trained with unlabeled data (for the RSDA) and labeled data (for the classifier) simultaneously.

Abstract: The described embodiments relate to systems, methods, and apparatus for providing a multimodal deep memory network (200) capable of generating patient diagnoses (222). The multimodal deep memory network can employ different neural networks, such as a recurrent neural network and a convolution neural network, for creating embeddings (204, 214, 216) from medical images (212) and electronic health records (206). Connections between the input embeddings (204) and diagnoses embeddings (222) can be based on an amount of attention that was given to the images and electronic health records when creating a particular diagnosis. For instance, the amount of attention can be characterized by data (110) that is generated based on sensors that monitor eye movements of clinicians observing the medical images and electronic health records.

Abstract: Techniques described herein relate to semi-supervised training and application of stacked autoencoders and other classifiers for predictive and other purposes. In various embodiments, a semi-supervised model (108) may be trained for sentence classification and may combine what is referred to herein as a “residual stacked de-noising autoencoder” (“RSDA”) (220), which may be unsupervised, with a supervised classifier (218) such as a classification neural network (e.g., a multilayer perceptron, or “MLP”). In various embodiments, the RSDA may be a stacked denoising autoencoder that may or may not include one or more residual connections. If present, the residual connections may help the RSDA “remember” forgotten information across multiple layers. In various embodiments, the semi-supervised model may be trained with unlabeled data (for the RSDA) and labeled data (for the classifier) simultaneously.