Abstract: A computer-implemented technique for relexicalization of sensitive entities in text data is disclosed. The technique obtains de-identification data identifying sensitive entities from input text and clusters these entities based on their representation of the same real-world things. For each cluster, a representative sensitive entity is determined and used to query a database system. The database returns a best matching candidate sensitive entity based on similarity matching, where each candidate is pre-associated with a relexicalized entity. A large language model (LLM) validates the correspondence between the representative and candidate entities within the input text's context. When validated, the technique generates relexicalized text by substituting cluster entities with the associated relexicalized entity. If validation fails, the technique generates a new relexicalized entity, stores the association in the database, and creates relexicalized text using the generated entity.

Abstract: Techniques for automatically de-identifying sensitive information in audio conversations by combining un-transcribed voice activity detection (VAD) with large language model (LLM) analysis are disclosed. An audio de-identification system processes speech-to-text transcriptions while identifying segments where automatic speech recognition (ASR) failed to transcribe spoken content. These un-transcribed segments are represented as placeholders in prompts sent to an LLM, which analyzes the surrounding textual context to determine if sensitive information (such as PII or PHI) was likely spoken during these gaps. When sensitive content is identified, the system modifies the corresponding audio segments through an audio identification tactic.

Abstract: A method and system for enhancing sensitive entity de-identification in textual data using large language models (LLMs) are disclosed. The method includes performing a primary de-identification procedure on input text to identify an initial set of sensitive entities, constructing a prompt containing the identified entities and a portion of the input text, and processing the prompt using an LLM to identify additional sensitive entities not detected in the primary procedure. A de-identified text is generated by removing both the initially identified entities and the LLM-identified entities from the input text. The de-identified text is stored in a non-transitory computer-readable medium. The system improves recall in sensitive information detection by leveraging LLMs'advanced language understanding capabilities to complement traditional de-identification methods, resulting in more comprehensive protection of sensitive information in applications such as medical records processing.

Abstract: Techniques for automatically deidentifying sensitive information in textual data using large language models (LLMs) are disclosed. A process iteratively identifies and removes sensitive entities from input text by sending portions to an LLM for analysis. The LLM determines if specific entities are sensitive, and based on its output, the identified entities are removed, and the text is updated. This cycle repeats for a predetermined number of iterations until no sensitive entities remain or until another termination condition is met. The method addresses limitations of traditional de-identification approaches by leveraging LLMs' advanced language understanding capabilities while managing computational resources efficiently. By employing an iterative approach, the accuracy and thoroughness of de-identification is improved, effectively removing sensitive information while preserving the text's usefulness.

Abstract: Techniques for automatically de-identifying sensitive information in audio conversations by combining un-transcribed voice activity detection (VAD) with large language model (LLM) analysis are disclosed. An audio de-identification system processes speech-to-text transcriptions while identifying segments where automatic speech recognition (ASR) failed to transcribe spoken content. These un-transcribed segments are represented as placeholders in prompts sent to an LLM, which analyzes the surrounding textual context to determine if sensitive information (such as PII or PHI) was likely spoken during these gaps. When sensitive content is identified, the system modifies the corresponding audio segments through an audio identification tactic.

Abstract: In one embodiment, a non-transitory computer-readable media stores instructions executable by processors for accessing a user input including a task description and a set of training data configured for prompt tuning, generating a baseline prompt based on the task description by an optimizer large language model (LLM), generating an output responsive to the user input based on the baseline prompt by a target LLM, generating modifications to the baseline prompt based on the set of training data and the output by the optimizer LLM, and generating a final prompt based on the modifications by the optimizer LLM.

Abstract: Embodiments assign an information technology service ticket to a queue and a sub-queue for optimized servicing. Embodiments extract from the service ticket a summary of the service ticket and a description of the service ticket. Embodiments provide as input to a trained neural network model the summary and description, the trained neural network model including a coarse network and a fine network. Embodiments predict the queue using the coarse network and predict the sub-queue using the fine network. Embodiments determine an uncertainty loss for the neural network model and when the uncertainty loss is below a threshold, assign the service ticket to the predicted queue and sub-queue.

Abstract: Techniques for processing incomplete service requests are disclosed. A system identifies reference service requests similar to the information of an incomplete service request received from a user. Using an adversarial domain adapter, the system generates an enhanced service augmenting the incomplete service request with predicted information. The system then identifies a subset of the reference service requests meeting a similarity threshold with the enhanced service request. The system processes the incomplete service request based on the subset of the set of reference service requests.

Abstract: Embodiments assign an information technology service ticket to a queue and a sub-queue for optimized servicing. Embodiments extract from the service ticket a summary of the service ticket and a description of the service ticket. Embodiments provide as input to a trained neural network model the summary and description, the trained neural network model including a coarse network and a fine network. Embodiments predict the queue using the coarse network and predict the sub-queue using the fine network. Embodiments determine an uncertainty loss for the neural network model and when the uncertainty loss is below a threshold, assign the service ticket to the predicted queue and sub-queue.