According to an embodiment, a device includes a detector, first and second recognizers, an estimator, a second recognizer, and an output unit. The detector is configured to detect a visible text area including a visible character from an image. The first recognizer is configured to perform character pattern recognition on the visible text area, and calculate a recognition cost according to a likelihood of a character pattern. The estimator is configured to estimate a partially-hidden text area into which a hidden text area estimated to have a hidden character and the visible text area are integrated. The second recognizer is configured to calculate an integrated cost into which the calculated cost and a linguistic cost corresponding to a linguistic likelihood of a text that fits in the entire partially-hidden text area are integrated. The output unit is configured to output a text selected or ranked based on the integrated cost.

In some embodiments, a profile associated with a user may indicate one or more data types and at least one profile-specific format associated with each of the data types. In response to detecting a document in a camera view, the document may be parsed to identify input fields that correspond to at least one data type of the profile. In response to identifying an input field that requires a format different from a profile-specific format, an input field rendering corresponding to the profile-specific format may be presented on an augmented reality view over the input field. In response to obtaining a user input related to the input field, the input field rendering may be updated in the augmented reality view to depict the user input in the profile-specific format, and the user input in the format required for the input field may be provided to the input field.

In some embodiments, a profile associated with a user may indicate one or more data types and at least one profile-specific format associated with each of the data types. In response to detecting a document in a camera view, the document may be parsed to identify input fields that correspond to at least one data type of the profile. In response to identifying an input field that requires a format different from a profile-specific format, an input field rendering corresponding to the profile-specific format may be presented on an augmented reality view over the input field. In response to obtaining a user input related to the input field, the input field rendering may be updated in the augmented reality view to depict the user input in the profile-specific format, and the user input in the format required for the input field may be provided to the input field.

A hardware accelerator can receive, from a host processor, a slice of input data at a time-step. The hardware accelerator can process the input data using a machine learning model deployed on the hardware accelerator to compute a respective probability among multiple probabilities for each of multiple classes. The respective probability for each class being a likelihood that content in the slice belongs to the class. The hardware accelerator can determine, from the multiple probabilities, a preset number of highest probabilities for the slice of input data. The hardware accelerator can transmit the preset number of highest probabilities for the slice to the host processor. Related apparatus, systems, techniques and articles are also described.

Aspects of the present disclosure include a system comprising a machine-readable storage medium storing at least one program and computer-implemented methods for detecting a language of a text string. Consistent with some embodiments, the method may include applying multiple language identification models to a text string. Each language identification model provides a predicted language of the text string and a confidence score associated with the predicted language. The method may further include weighting each associated confidence score based on historical performance of the corresponding language identification model in predicting languages of other text strings. The method may further include selecting a predicted language of the text string from among the multiple predicted languages provided by the multiple language identification models based on a result of the weighting of the confidence score associated with the particular predicted language.

In some forms containing keywords and content, there may be nested levels of keywords, also referred to as a hierarchy. Content in the forms may be associated with one or more keywords in one or more of the nested levels, or in the hierarchy. Identifying keywords in adjacent cells in a table (with a nested keyword being either to the right of or below another keyword) enables distinguishing between keywords and content in filled forms, and enables correct association of content with respective keywords.

Techniques and structures to facilitate conversion of a workflow process is disclosed. The techniques include receiving an image, identifying one or more objects included in the image, identifying one or more properties associated with each of the one or more objects, generating a matrix including data including the identified objects and associated properties and processing the matrix at a machine learning model to determine whether the image is to be translated based on a determination that one or more objects and associated properties within the image are required to be translated.

Techniques and structures to facilitate conversion of a workflow process is disclosed. The techniques include receiving an image, identifying one or more objects included in the image, identifying one or more properties associated with each of the one or more objects, generating a matrix including data including the identified objects and associated properties and processing the matrix at a machine learning model to determine whether the image is to be translated based on a determination that one or more objects and associated properties within the image are required to be translated.

Aspects include receiving a document and classifying at least a subset of the document as having a first type of data. Features are extracted from the document. The extracting includes initiating processing of the at least a subset of the document by a first processing engine that was previously trained to extract features from the first type of data. The extracting also includes initiating processing of a remaining portion of the document not included in the at least a subset of the document by a second processing engine that was previously trained to extract features from a second type of data. The first type of data is different than the second type of data. Features are received from one or both of the first processing engine and the second processing engine. The received features are stored as features of the document.

Aspects of the present disclosure include a system comprising a machine-readable storage medium storing at least one program and computer-implemented methods for detecting a language of a text string. Consistent with some embodiments, the method may include applying multiple language identification models to a text string. Each language identification model provides a predicted language of the text string and a confidence score associated with the predicted language. The method may further include weighting each associated confidence score based on historical performance of the corresponding language identification model in predicting languages of other text strings. The method may further include selecting a predicted language of the text string from among the multiple predicted languages provided by the multiple language identification models based on a result of the weighting of the confidence score associated with the particular predicted language.