Abstract: A method and system for selecting data from a source text corpus for training a semantic data analysis system. The method includes selecting an item of the text corpus, wherein the item includes at least one section. The method includes extracting a section of the at least one section of the item. The method also includes determining a length of the section of the at least one section of the item. Based on the length of the section being greater than a predetermined amount, the method includes subdividing the section into a plurality of fragments. Each fragment of the plurality of fragments is deemed to be similar to each other. Further, the method includes building a training set based on the plurality of fragments. The training set is used to train the semantic data analysis system.

Abstract: A method of providing audio information from a meeting includes receiving a first audio stream from a first input audio device and a second audio stream from a second input audio device during the meeting, identifying a first audio fragment from the first audio stream, and identifying a second audio fragment from the second audio stream. The method also includes compiling the audio fragments from the first and second audio streams into an audio file that includes at least the first audio fragment and the second audio fragment. The method further includes providing the audio file to one or more recipients. The audio file identifies the first audio fragment as corresponding to a first participant of the meeting and the second audio fragment as corresponding to a second participant of the meeting.

Abstract: Selecting data from a source text corpus for training a semantic data analysis system includes selecting an item of the text corpus, validating the item, extracting at least one section of the item, determining a length of each of the at least one section of the item, and subdividing each of the sections having a length greater than a predetermined amount into a plurality of fragments that are deemed to be similar. The predetermined amount may be approximately twice a size of a fragment. A fragment may have approximately 100 words or between 40 and 60 words. Fragments from different items may be deemed to be dissimilar. Sections having a length less than the predetermined amount may be ignored. Validating the item may include parsing editorial notes and other accompanying data. The source text corpus may be Wikipedia. The item may be an article.

Abstract: Searching for documents includes retrieving objects from a physical media image using a camera from a smartphone, a user selecting a subset of the objects, forming a search query based on the subset of objects, and applying the search query to a search engine to search for the documents. Retrieving objects from a media image may include waiting for a view of the camera to stabilize. Waiting for the view of the camera to stabilize may include detecting changing content of a video flow provided to the camera and/or using motion sensors of the camera to detect movement. Retrieving objects may include the smartphone identifying possible subsets of objects in the media image. The user selecting a subset of the objects may include the smartphone presenting at least some of the possible subsets to the user and the user selecting one of the possible subsets.

Abstract: This application is directed to vectoring a raster image in which an electronic device detects a contour of a component in the raster image, builds tangent vectors for each point of the contour and identifies a plurality of segmentation points on the contour. One or more points of sharp angle are identified on the contour in accordance with a determination that each point of sharp angle corresponds to two distinct tangent vectors and that an angle between the two distinct tangent vectors falls below a predefined threshold. A respective one of the segmentation points is positioned at each identified point of shape angle. The electronic device approximates a piecewise smooth fitting curve (e.g., a piecewise Bezier curve) having two or more fitting segments to connect the plurality of segmentation points on the contour. The piecewise smooth fitting curve is thereby provided to vectorize the raster image.

Abstract: Recording audio information from a meeting includes determining which audio input audio device (smartphones) correspond to which meeting participant, measuring volume levels in response to each of the meeting participants speaking, identifying one of the participants is speaking based on stored voice profiles and/or relative volume levels at each of the smartphones, recording on a first channel audio input at a first smartphone corresponding to the speaker, identifying another one of the participants is speaking based on stored voice profiles and/or relative volume levels at each of the smartphones, recording on a second channel, separate from the first channel, audio input at a second smartphone corresponding to the other speaker, and merging the first and second channels to provide a storyboard that includes audio input from the channels and identification of speakers based on which specific ones of the channels contains the audio input.

Abstract: This application is directed to recognizing unstructured information based on hints provided by structured information. A computer system obtains unstructured information collected from a handwritten or audio source, and identifies one or more terms from the unstructured information. The one or more terms includes a first term that is ambiguous. The computer system performs a recognition operation on the first term to derive a first plurality of candidate terms for the first term, and obtains first contextual information from an information template associated with the unstructured information. In accordance with the first contextual information, the computer system selects a first answer term from the first plurality of candidate terms, such that the first term is recognized as the first answer term.

Abstract: Enhancing image quality of an initial full image of a document page includes obtaining the initial full image of a document page, determining that at least a subset of text in the initial full image does not meet a predefined readability criterion, identifying at least one of a plurality of page fragments in the initial full image of the document page for enhancement, and sending an instruction to a mobile device to provide a photograph of the at least one of the page fragments. The photograph provides a separate fragment image for the at least one of the page fragments. The separate fragment image is then obtained from the mobile device and merged into the initial full image to provide an enhanced full image.

Abstract: Methods and systems for training a neural network to distinguish between text documents and image documents are described. A corpus of text and image documents is obtained. A page of a text document is scanned by shifting a text window to a plurality of locations. In accordance with a determination that the text in the window at a respective location meets text line criteria, the text in the window is stored as a respective text snippet. A plurality of image windows are superimposed over at least one page of an image document. In accordance with a determination that the content of a respective image window meets image criteria, content of the image window is stored as a respective image snippet. The respective text snippet and the respective image snippet are provided to a classifier.

Abstract: This application is directed to vectoring a raster image in which an electronic device detects a contour of a component in the raster image, builds tangent vectors for each point of the contour and identifies a plurality of segmentation points on the contour. One or more points of sharp angle are identified on the contour in accordance with a determination that each point of sharp angle corresponds to two distinct tangent vectors and that an angle between the two distinct tangent vectors falls below a predefined threshold. A respective one of the segmentation points is positioned at each identified point of shape angle. The electronic device approximates a piecewise smooth fitting curve (e.g., a piecewise Bezier curve) having two or more fitting segments to connect the plurality of segmentation points on the contour. The piecewise smooth fitting curve is thereby provided to vectorize the raster image.

Abstract: Determining if a document is a text page includes partitioning the document into a plurality of cells, scaling each of the cells to a standardized number of pixels to provide a corresponding snippet for each of the cells, using a classifier to examine the snippets to determine which of the cells are classified as text and which of the cells are not classified as text, determining a volume of text for the document based on a total amount of text in the document corresponding to a sum of an amount of text in each of the cells classified as text, and determining that the document is a text page in response to the total amount exceeding a pre-determined threshold. In response to the total amount being less than the pre-determined threshold, cells not classified as text may be examined further. The classifier may be provided by training a neural net.

Abstract: Recording audio information from a meeting includes determining which audio input audio device (smartphones) correspond to which meeting participant, measuring volume levels in response to each of the meeting participants speaking, identifying one of the participants is speaking based on stored voice profiles and/or relative volume levels at each of the smartphones, recording on a first channel audio input at a first smartphone corresponding to the speaker, identifying another one of the participants is speaking based on stored voice profiles and/or relative volume levels at each of the smartphones, recording on a second channel, separate from the first channel, audio input at a second smartphone corresponding to the other speaker, and merging the first and second channels to provide a storyboard that includes audio input from the channels and identification of speakers based on which specific ones of the channels contains the audio input.

Abstract: Enhancing image quality of an initial full image of a document page includes obtaining the initial full image of a document page, determining that at least a subset of text in the initial full image does not meet a predefined readability criterion, identifying at least one of a plurality of page fragments in the initial full image of the document page for enhancement, and sending an instruction to a mobile device to provide a photograph of the at least one of the page fragments. The photograph provides a separate fragment image for the at least one of the page fragments. The separate fragment image is then obtained from the mobile device and merged into the initial full image to provide an enhanced full image.

Abstract: Vectorizing a raster image includes identifying a connectivity component in the raster image, detecting a contour of the connectivity component, building tangent vectors for each point of the contour, for each sharp angle of the contour, positioning a segmentation point of two segments at a point thereof, for each location of high curvature of the contour, positioning segments proximal thereto, composing an optimization task to approximate a piecewize Bezier curve, solving the optimization task to provide a vectorization of the raster image, and, in response to there not being a sufficient number of segments, adding additional segments. Vectorizing a raster image may also include applying perspective, color, brightness and contrast correction to the raster image and building a binary black-white representation of the raster image prior to identifying the connectivity component. The optimization task may minimize a root-mean square deviation of the piecewize Bezier curve.

Abstract: Recording audio information from a meeting includes determining which audio input audio device (smartphones) correspond to which meeting participant, measuring volume levels in response to each of the meeting participants speaking, identifying one of the participants is speaking based on stored voice profiles and/or relative volume levels at each of the smartphones, recording on a first channel audio input at a first smartphone corresponding to the speaker, identifying another one of the participants is speaking based on stored voice profiles and/or relative volume levels at each of the smartphones, recording on a second channel, separate from the first channel, audio input at a second smartphone corresponding to the other speaker, and merging the first and second channels to provide a storyboard that includes audio input from the channels and identification of speakers based on which specific ones of the channels contains the audio input.

Abstract: Automatically scanning multiple document sheets with a camera includes receiving a video stream while the camera is pointed at the multiple document sheets, detecting presence of a first top page of the multiple document sheets based on the video stream, taking a still photograph of the first top page in response to detecting presence of the first top page, detecting presence of a second top page based on the video stream by confirming that the second top page is different from the first top page and by waiting a predetermined amount of time for an image of the second top page to stabilize, and taking a still photograph of the second top page in response to detecting presence of the second top page. Detecting the pages may include determining that the camera is pointing at the stack of documents and a detected page is not obstructed.

Abstract: Improving visual quality of a raster image includes detecting connectivity components, detecting defects in each of the connectivity components based on a characteristic line width thereof, detecting segments in each of the connectivity components, detecting joints based on geometry of the connectivity components, creating a structural graph based on the segments and joints, and correcting the raster image according to the structural graph and detected ones of the defects. The joints may correspond to linear joints, T-joints, or X-joints. Detecting types of joints may include determining a configuration of adjacent segments in a proximity of each of the joints. A characteristic line width may be determined by determining co-boundaries on opposite sides of each of the segments and determining average distances between the co-boundaries. The raster image may be a binary black-and-white image of a line drawing obtained from a photograph or a scan of a handwritten document.

Abstract: Identifying a page with content in a video frame that is part of a video stream of successive video frames includes receiving the video stream, detecting edge segments in the video frame, where each of the edge segments is a candidate for being at least a part of an edge of the page, filtering the edge segments to discard a first subset of the edge segments based on curvature and based on angles between the edge segments and standard axes of the video frame, and identifying the page with content within a portion of a second subset of the edge segments that remain after filtering in response to the portion having geometric closeness to a rectangle. Edge segments having angles that significantly deviate from coordinate angles of the video frame and edge segments with a relatively high curvature may be discarded. A Canny edge detection algorithm may be used.

Abstract: Enhancing image quality of an initial full image of a document includes subdividing the initial full image of the document into a number of segments, determining at least one of the segments for enhancement, directing a user to take photographs with a smartphone to provide a separate segment image for the at least one of the segments by presenting on a screen of the smartphone a view of the document having lines superimposed thereon corresponding to the segments, and merging the separate segment image to provide an enhanced full image. The number of segments may depend upon a size of text in the initial full image and a desired height of text in the enhanced full image. The view of the document having lines superimposed thereon corresponding to the segments may provide a target to the user for photographing the at least one of the segments.

Abstract: Detecting an aspect ratio of an image captured with a smartphone includes detecting at least one convex quadrangle of arbitrary shape on the image and generating a plurality of additional convex quadrangles having vertices in a pre-determined vicinity of vertices of the quadrangle on the image. A linear projective mapping matrix is generated for mapping each of the quadrangle and the plurality of additional quadrangles onto a unit square. A plurality of estimated focal lengths of the camera of the smartphone is determined according to matrixes corresponding to the linear projective mappings onto a unit square of the quadrangle and each of the plurality of additional quadrangles. The quadrangle is used to determine the aspect ratio of the image in response to a range of the plurality of estimated focal lengths including a true value of the focal length of the camera of the smartphone.