Abstract: Tables including cells can be extracted from an input document image, as objects and processed to be output in an XML format. The tables can be filtered based on one or more keywords, to reduce a number of the tables. The filtered tables that include the one or more keywords can be obtained. A query can be executed on the filtered tables, the query including one or more keys composed based on cell level information associated with the filtered tables. One or more cells among the cells can be identified based on the one or more keys. As a response to the query, the target content that corresponds to the one or more cells can be provided.

Abstract: According to an aspect, a system performs a first machine learning (ML) technique and a second ML technique on a large set of data items. The first ML technique identifies a primary cluster containing data items classified as non-anomalous, with the rest of the data items classified as anomalous. The second ML technique identifies a set of secondary clusters. Upon receiving feedback on the classification of at least some of the data items, the system maintains a feedback data associating the feedback with each secondary cluster. Upon receiving an indication to find a classification of a data item, the system identifies as a result, whether the data item is non-anomalous or anomalous based on whether the data item belongs to the primary cluster or not. The system also determines based on a secondary cluster to which the data item belongs and the feedback associated with the secondary cluster, whether the result is to be modified.

Abstract: A method includes estimating a number of threads to execute batch job within maximum completion time by executing a thread estimation algorithm. Executing the thread estimation algorithm includes: determining, using an ML model, a collection of one or more combinations, each including an estimated number of threads and an actual completion time for processing the batch job using the estimated number of threads, the actual completion time being less than or equal to the maximum completion time, and saving the collection including the one or more combinations in a list of threads and completion times. The list of threads and completion times is used to select, based on a list of cloud virtual machines available for use, an optimal virtual machine or an optimal virtual machine combination, the optimal virtual machine or the optimal virtual machine combination being capable of executing the batch job within the maximum completion time.

Abstract: A reconciliation engine and data reconciliation techniques are disclosed that partially use machine learning and natural language processing techniques to overcome challenges associated with traditional manual reconciliation engines and data reconciliation techniques without compromising the accuracy and handle the complexity of the features. In one aspect, a computer-implemented method is provided that includes: accessing a source and target table, performing intra-reconciliation on the source table, grouping the source and target table based on system configurations, reconciling data is iterated over each group, where the reconciling data includes: finding, using a machine learning model, a closest match for each group, and outputting the closest match for each group.

Abstract: Techniques are disclosed for understanding the visual structure and patterns of documents and detecting anomalies in data of the documents based on the understanding of the visual structure and patterns of the documents. In one aspect, a computer-implemented method is provided that includes accessing a set of documents, converting the set of documents to a set of images in a binary format, generating a common feature template based on the set of images, comparing each image from the set of images to the common feature template to identify images with at least one anomalous feature, and outputting the images with at least one anomalous feature.

Abstract: Method includes extracting, from document image, tables as objects; outputting tables in an XML format; filtering tables based on keywords and a condition including one from among AND and OR, which includes: for each of tables, determining whether the table includes at least one from among a first keyword and a second keyword, for condition being AND, based on the table containing first keyword and second keyword, outputting the table, as filtered table, and, for condition being OR, based on the table containing first keyword or second keyword, outputting the table, as filtered table, thereby outputting filtered tables that include the first keyword and/or the second keyword based on condition; executing query on filtered tables that includes keys composed based on cell level information associated with filtered tables; identifying cells based on keys; and providing target content that corresponds to cells, as response to query.

Abstract: Method includes obtaining historical document images including text that correspond to different document classes; and generating a dictionary using text of the historical document images. The dictionary includes base words occurring with a greatest frequency in each document class. The base words are extracted from the text of the historical document images and arranged in datasets by a document class, where each dataset includes the base words of a same document class that occur with the greatest frequency within that document class. Trie structure is generated using the base words of the datasets that occur with a greatest frequency in each dataset. The trie structure includes internal nodes including root node and leaf nodes in which keys corresponding to the base words occurring with the greatest frequency in each dataset are respectively stored in predefined order. The trie structure is searchable in the predefined order starting with the root node.

Abstract: According to an aspect, a technology system maintains an account data specifying the details of multiple accounts including a specific set of accounts as being delinquent. Upon receiving a feature set to be used, the system clusters the multiple accounts into groups based on the feature set. The system generates a multi-classification model for assignment of agents to the accounts based on the group qualifiers associated with the groups and the feature set. The system then identifies a target group likely to present problems with debt collection, assigns, using the multi-classification model, a corresponding agent for each account of the group and determines a recommended recovery strategy for each assignment. The system provides the details of the group, the corresponding agents assigned and the recommended recovery strategy for each assignment.

Abstract: Systems, methods, and other embodiments associated with converting a static cluster data table to a graphical hierarchical tree are described. In one embodiment, a method includes recursively traversing the static cluster data table to identify a root cluster, identify child clusters from the root cluster and child clusters from each other that define parent-child relationships, and identify decision segments that caused a segment split of cluster data. A 2-dimensional visual hierarchy is generated and displayed in a graphical form using a plurality of nodes that represent the root cluster and the child clusters along with path lines that connect the nodes. The 2-dimensional visual hierarchy displays a hierarchical visualization of the static cluster data table that shows an order of decision segments that occurred to segment a dataset and how the dataset was segmented by a clustering algorithm leading to a final cluster of a leaf node.

Abstract: A method includes generating dataset using topics associated with historical records, the dataset including pairs of data that are formed based on the topics, each of the pairs of data including an antecedent topic associated with a historical record corresponding to a preceding event and a consequent topic associated with a historical record corresponding to an event that occurred after the preceding event, the antecedent topic and the consequent topic forming a transitive relation for each of the pairs of data; inputting, into ML model, the pairs of data and input topic associated with a record of a user; generating, by the ML model, a prediction of a next record topic for a next record corresponding to the user, based on the consequent topic included in each of the pairs of data that include the antecedent topic corresponding to the input topic; and outputting the prediction.

Abstract: A computer-implemented method includes: obtaining a record including text narration; inputting the text narration into an NLP model; generating, by the NLP model, encoded narration based on the text narration; determining similarity index between the encoded narration and each of encoded reference topics; comparing the similarity index between the encoded narration and each encoded reference topic to similarity threshold; and determining whether the similarity index between the encoded narration and each encoded reference topic is equal to or greater than the similarity threshold. When the similarity index is equal to or greater than the similarity threshold, a corresponding reference topic is added to a result group, and, when the similarity index is less than the similarity threshold, the corresponding reference topic is left out of the result group.

Abstract: Aspects of the present disclosure are directed to extracting textual information from image documents. In one embodiment, a system, upon receiving a request to extract textual information from an image document, a digital processing system performs character recognition based on content of the image document using multiple approaches to generate corresponding texts. The texts are then combined to determine a result text representing the textual information contained in the image document. The result is then provided as a response to the request.

Abstract: Systems, methods, and other embodiments associated with identifying a document type of an unidentified document based on image features are described. In one embodiment, image pairs are formed by pairing the unidentified document with anchor images from a plurality of anchor images, wherein each anchor image is a known document type. For each image pair, first visual features are extracted from the unidentified document and second visual features are extracted from the paired anchor image. A similarity function is applied to compare the first visual features and the second visual features, and a similarity score is generated for each image pair based on the comparing. The most similar anchor image from the image pairs, which has a greatest similarity score, is identified. The document type of the unidentified document is then predicted as the known document type associated with the most similar anchor image.

Abstract: Systems, methods, and other embodiments associated with a data quality framework for a machine learning pipeline are described. In one embodiment, a method includes receiving an input dataset prior to training a machine learning model with the input dataset. The input dataset, having data records in column tabular form, is analyzed to determine a data quality of the input dataset prior to training the machine learning model. A data quality score is generated that represents an overall quality of the input dataset, where the data quality score is generated based on a combined ensemble of at least two factors selected from (i) an abundance factor, (ii) a completeness factor, and (iii) a dimension efficiency factor. Based on the data quality score, the input dataset is either permitted or prohibited from continuing in the machine learning pipeline, and corrective actions may be performed.

Abstract: Method includes generating a base model by training a pretrained model using a base training dataset including first training datapoints identifying tables in historical document images that include the tables and text, where the generated base model is configured to extract the tables as objects; and generating a table extraction model by training the base model using an enhanced training dataset including second training datapoints that are different from the first training datapoints and identify a plurality of cells disposed in each of the tables in a row direction and a column direction. The table extraction model is trained to output content of the tables and table information in an XML format, the table information including cell level information of the plurality of cells that is searchable via a query configured to provide target content that corresponds to one or more cells.

Abstract: A bill verification system for verifying bill records associated with an entity account is disclosed. The system receives a request to verify a bill record associated with an entity account. The system identifies a bill level model to be used for verifying the bill record and detects based on the bill level model, bill level anomaly information for the bill record. The system additionally identifies a bill line level model to be used for verifying one or more bill lines in the bill record and detects based on the bill line level model, bill line level anomaly information for the bill record. The system then aggregates the bill level anomaly information and the bill line level anomaly information to generate a bill verification report for the bill record. The system provides the bill verification report as a response to the request received to verify the bill record.

Abstract: According to an aspect, a technology system maintains an account data specifying the details of multiple accounts including a specific set of accounts as being delinquent. Upon receiving a feature set to be used, the system clusters the multiple accounts into groups based on the feature set. The system generates a multi-classification model for assignment of agents to the accounts based on the group qualifiers associated with the groups and the feature set. The system then identifies a target group likely to present problems with debt collection, assigns, using the multi-classification model, a corresponding agent for each account of the group and determines a recommended recovery strategy for each assignment. The system provides the details of the group, the corresponding agents assigned and the recommended recovery strategy for each assignment.