Microsoft Technology Licensing, LLC

Other

John R Bennett

Gregory Nicholas Hullender

Donald D Karlov

Patrick M Haluptzok

Christopher A Meek

Bo Thiesson

Tonsil cleaner

Tonsil cleaning device

Tongue retractor

Tongue-tie elevator

An improved method and mechanism for recognizing chirographs (handwritten characters) input into a computer system. A primary recognizer is provided for converting chirographs to code points, and secondary recognizers such as binary CART trees are developed and trained to differentiate chirographs which produce certain code points at the primary recognizer. Each such secondary recognizer is associated with each selected code point. When a chirograph is received, the chirograph is provided to the primary recognizer whereby a code point corresponding thereto is received. If the code point corresponds to one of the secondary recognizers, the chirograph is passed to the secondary recognizer, and a code point is returned from the secondary recognizer. If not, the code point provided by the primary recognizer is returned. The invention sets forth an automated process for training the CART trees and for optimizing the recognition mechanism by discarding CART trees which do not improve on the recognition accuracy of the primary recognizer.

Method and mechanism to reduce handwriting recognizer errors using multiple decision trees

A method and improved data structure for attaching information to words in a trie of nodes. Each node of a trie includes a single tag bit which is interpreted according to information specified in a header of the trie. The information may specify that a node may be tagged with multiple tags, whereby if the tag bit is set in a given node, the node further includes a bitmask indicating which one or ones of the tags apply to that node. A value mask may be provided in the header indicating which of the tags (if any) have values associated therewith, whereby information representative of the value such as the value itself or a pointer thereto is stored in each node tagged with at least one tag having an associated value. Partial enumeration of tagged nodes may be provided by storing a count of the number of tagged words under a node, wherein if a trie has multiple tags, each tag may be selectively and separately enumerated as specified in header information. Partial enumeration may be combined with global enumeration of each word, with multiple tags, and/or with tags that have associated values.

Identification and enumeration of data components in a trie

A technique for reducing a model database for use with handwriting recognizers. The model database is processed with a tuning set to identify a set of models that would result in the greatest character recognition accuracy. If further model database reduction is desired, the technique iteratively identifies smaller models that have the least adverse effect on the error rate. The technique continues identifying smaller models until a desired model database size has been achieved.

Reducing and controlling sizes of model-based recognizers

Method and system for attaching information to words of a trie

A method and system for enumerating a trie of states of nodes. A node near the middle of a state of alphabetically ordered nodes is selected as a skip node and moved to the logical beginning of the state. The skip node is provided with a pointer to a jump node that is at the skip node's former position, and the node immediately to the left of the node's former position is marked as a soft terminal node. As a result of the alphabetic ordering, the segment of nodes before the jump node are alphabetically before the skip node, while the segment of nodes after the jump node are alphabetically after the skip node. The segments of the state may be recursively split into further segments via further skip nodes, jump nodes and soft terminal nodes, and, once the segments are split as desired, the nodes within the segment may be sorted. When searching the state, if a skip node is reached that does not equal the search character, the search moves to the jump node if the search character is greater than the skip node, or to the next node if the search character is less than the skip node. The search of the nodes continues, essentially as a binary search when skip nodes are encountered, and a linear search otherwise, until a node is found that matches the search character or a soft or regular terminal node is reached.

Method and system for improved enumeration of tries

An improved trie compression method that compresses by merging partially identical subtrees. States of the trie are selected, and the nodes of those states examined find nodes that are identical to one another. The most frequently occurring identical node is selected as a substate, and the states are separated into a first group of states that have the substate node therein and a second group of states that do not. The nodes in the first group of states are reordered such that the substate is at the end thereof. Then, the substate of each state is merged into a single node, replaced by a pointer from each state. Compression is performed recursively by choosing a new substate for the remaining nodes of the first group, and for subsequently separated groups, until no further identical nodes are available for merging.

Trie compression using substates and utilizing pointers to replace or merge identical, reordered states

Method and mechanism to reduce handwriting recognizer errors using

A method and mechanism for converting a non-contiguous subset of values in a large range, such as selected Unicode code points, into a contiguous or mostly contiguous smaller range with a perfect hash. The large range is organized into a two-dimensional bitmap matrix of pages and offsets into the pages. The bits in the matrix equal one if the value is in the subset, and zero if not. The pages are then overlaid on one another into a one-dimensional bitmap by shifting each page as necessary to avoid conflicts with values on other pages. The shift amount is recorded and used in a hash computation, wherein a value of the large range is first separated into its page number and its offset into the page. The values are then hashed into the value of the dense subset range by looking up the shift amount for the page and adding the shift amount to the offset into the page.

Growing community of inventors

Redmond, WA, United States of America

John R Bennett