5 results about "Copying mechanism" patented technology

This invention discloses a sentence-level question generation method based on syntactic-aware prompt learning. First, a bidirectional syntactic dependency graph is constructed based on a given sentence. Its semantic representation is obtained through a relation-aware attention graph encoder. The encoded vectors are then input into a softmax layer, and the top k vectors are selected as continuous prompts based on probability. The prompts are concatenated to the given source text and the answer using prefix adjustment, and then input into a BERT model for encoding. The encoded result is then fed into a Transformer model for decoding. At each time step of decoding, the syntactic dependency information of the generated text sequence is modeled. This information, combined with the syntactic dependency information of the source sentence, determines the parts that the decoder needs to focus on, assisting in the generation of the current word. Simultaneously, a copying mechanism is introduced to handle situations where the generated word is not in the question vocabulary, allowing the model to directly copy words from the source text.

The invention discloses a target range resource reuse and security isolation method and system in a high-concurrency scene, and belongs to the technical field of network simulation. The method comprises the following steps: receiving a network topology and initial network parameters defined by a user, and storing the network topology and initial network parameters as a topology description file; the system supports concurrent receiving and processing of a plurality of independent instantiation requests for the same topology; distributing a unique identifier for the instance corresponding to each request; for each instance, quickly cloning the virtual device based on a copy-on-write mechanism; creating an independent network namespace for each instance, distributing a unique VXLAN network identifier from the global VNI pool, and injecting the unique VXLAN network identifier into the namespace; automatically injecting the initial network parameters as fixed configuration into each device in an instance; and finally, accessing the equipment into the exclusive namespace and starting the equipment. According to the invention, the problems of large resource consumption and difficult network isolation in a large-scale concurrent simulation scene are solved, and the concurrent processing capability and the operation efficiency of a target range system are remarkably improved.

This semiconductor device manufacturing device comprises: a stage; a bonding head; a copying mechanism provided on the bonding head; and a controller that executes the copying process to adjust a facing surface to be parallel to a reference plane by having the facing surface, which is a holding surface or a chip end face, follow the reference plane 110, which is a planar surface of the stage or a substrate. In the copying process, the controller moves the bonding head relative to the surface direction of the reference plane with the facing surface left abutting the reference plane in a state with the copying mechanism switched to a free state, until the axial direction position of the bonding head reaches a stipulated reference position, and when the axial direction position reaches the reference position, switches the copying mechanism to a locked state.

The application discloses a low-energy positioning data management system for a large-scale sensor network, and relates to the technical field of data processing.The application solves the technical problems of lacking full-link metadata management and automatic data life cycle scheduling, which leads to long-term occupation of storage resources by low-frequency data and low query efficiency of high-frequency data.The application removes positioning drift data through Kalman filtering, improves data effectiveness, adapts to BLE5.3, LoRaWAN and other protocols according to scenes, cooperates with differential coding and quantitative compression, reduces the amount of transmission data, reduces bandwidth occupancy, classifies high-frequency data according to real-time performance, matches high-performance / low-performance nodes, controls high-frequency core data access delay at the millisecond level, stores low-frequency data in near-line HDD-archiving object storage hierarchical storage, reduces unit storage cost, considers performance and economy, evaluates node performance through response delay+storage capacity two-dimensional normalization, and improves data storage reliability through a distributed storage combined with a copy mechanism.