Servicenow, Inc.

Chitrabharathi Ganapathy

Baskar Jayaraman

Aniruddha Madhusudan Thakur

Shiva Shankar Ramanna

Jun Wang

Kannan Govindarajan

Tao Hong

Dinesh Kumar Kishorkumar Surapaneni

Tao Feng

Rohit Lobo

Systems and methods involving data structures for efficient management of paragraph vectors for textual searching are described. A database may contain records, each associated with an identifier and including a text string and timestamp. A look-up table may contain entries for text strings from the records, each entry associating: a paragraph vector for a respective unique text string, a hash of the respective unique text string, and a set of identifiers of records containing the respective unique text string. A server may receive from a client device an input string, compute a hash of the input string, and determine matching table entries, each containing a hash identical to that of the input string, or a paragraph vector similar to one calculated for the input string. A prioritized list of identifiers from the matching entries may be determined based on timestamps, and the prioritized list may be returned to the client.

Data structures for efficient storage and updating of paragraph vectors

A document is received. The document is analyzed to discover text and structures of content included in the document. A result of the analysis is used to determine intermediate text representations of segments of the content included in the document, wherein at least one of the intermediate text representations includes an added text encoding the discovered structure of the corresponding content segment within a structural layout of the document. The intermediate text representations are used as an input to a machine learning model to extract information of interest in the document. One or more structured records of the extracted information of interest are created.

Determination of intermediate representations of discovered document structures

A computer-implemented method includes obtaining a plurality of textual records divided into clusters and a residual set of the textual records, where a machine learning (ML) clustering model has divided the plurality of textual records into the clusters based on a similarity metric. The method also includes receiving, from a client device, a particular textual record representing a query and determining, by way of the ML clustering model and based on the similarity metric, that the particular textual record does not fit into any of the clusters. The method additionally includes, in response to determining that the particular textual record does not fit into any of the clusters, adding the particular textual record to the residual set of the textual records. The method can additionally include identifying, by way of the ML clustering model, that the residual set of the textual records contains a further cluster.

Clustering and dynamic re-clustering of similar textual documents

Word vectors are multi-dimensional vectors that represent words in a corpus of text and that are embedded in a semantically-encoded vector space; paragraph vectors extend word vectors to represent, in the same semantically-encoded space, the overall semantic content and context of a phrase, sentence, paragraph, or other multi-word sample of text. Word and paragraph vectors can be used for sentiment analysis, comparison of the topic or content of samples of text, or other natural language processing tasks. However, the generation of word and paragraph vectors can be computationally expensive. Accordingly, word and paragraph vectors can be determined only for user-specified subsets of fields of incident reports in a database.

Selectively generating word vector and paragraph vector representations of fields for machine learning

Word vectors are multi-dimensional vectors that represent words in a corpus of text and that are embedded in a semantically-encoded vector space. Word vectors can be used for sentiment analysis, comparison of the topic or content of sentences, paragraphs, or other passages of text or other natural language processing tasks. However, the generation of word vectors can be computationally expensive. Accordingly, when a set of word vectors is needed for a particular corpus of text, a set of word vectors previously generated from a corpus of text that is sufficiently similar to the particular corpus of text, with respect to some criteria, may be re-used for the particular corpus of text. Such similarity could include the two corpora of text containing the same or similar sets of words or containing incident reports or other time-coded sets of text from overlapping or otherwise similar periods of time.

Persistent word vector input to multiple machine learning models

A database may contain a corpus of text strings, the text strings respectively associated with vector representations thereof, where each of the vector representations is an aggregation of vector representations of words in the associated text string. An artificial neural network (ANN) may have been trained with mappings between: (i) the words in the text strings, and (ii) for each respective word, one or more sub strings of the text strings in which the word appears. A server device may be configured to: receive an input text string; generate an input aggregate vector representation of the input text string by applying an encoder of the ANN to words in the input text string; compare the input aggregate vector representation to the vector representations; identify a relevant subset of the vector representations; and transmit the text strings that are associated with the relevant subset of the vector representations.

Growing community of inventors

San Jose, CA, United States of America

Chitrabharathi Ganapathy