Task Indication Scoring via Starting N-grams

The patent segments the task request identification process into multiple distinct components: extracting interrogative sentences from messages, generating task indications from these sentences, scoring the task indications using trained models, and finally determining task requests based on scores. This segmentation allows each component to be optimized independently, improving overall accuracy while managing system complexity through modular design.

The patent performs preliminary actions by pre-training machine learning models with large datasets of messages and task indications before deployment. Task indications are pre-generated from interrogative sentences using trained word embeddings and language models. These preliminary training and preparation steps enable the system to achieve high identification accuracy when processing actual messages, as the complex pattern recognition work has already been performed during the training phase.

The patent changes parameters by using pre-trained word embeddings and language models that have been optimized during training. The scoring mechanism uses adjusted word embeddings and language model parameters that are specifically tuned for task indication scoring. These parameter optimizations allow the system to calculate task association scores accurately while reducing computational overhead compared to using generic, unoptimized models.

The patent creates simplified copies or representations of task indications using word embeddings and language model vectors. Instead of processing the full complexity of natural language messages, the system works with compressed vector representations that capture the essential meaning. These copied representations can be scored efficiently while preserving the information needed for accurate task request identification.