Abstract: One embodiment of the present invention sets forth a technique for analyzing a transcription of a recording. The technique includes generating features representing transcriptions produced by multiple automatic speech recognition (ASR) engines from voice activity in the recording and a best transcription of the recording produced by an ensemble model from the transcriptions. The technique also includes applying a machine learning model to the features to produce a score representing an accuracy of the best transcription. The technique further includes storing the score in association with the best transcription.

Abstract: One embodiment of the present invention sets forth a technique for performing ensemble modeling of ASR output. The technique includes generating input to a machine learning model from snippets of voice activity in the recording and transcriptions produced by multiple automatic speech recognition (ASR) engines from the recording. The technique also includes applying the machine learning model to the input to select, based on transcriptions of the snippet produced by at least one contributor ASR engine of the multiple ASR engines and at least one selector ASR engine of the multiple ASR engines, a best transcription of the snippet from possible transcriptions of the snippet produced by the multiple ASR engines. The technique further includes storing the best transcription in association with the snippet.

Abstract: One embodiment of the present invention sets forth a technique for analyzing transcriptions of a recording. The technique includes storing per-character differences between a first set of characters from a first transcription of the recording and a second set of characters from a second transcription of the recording in a matrix with a fixed width. The technique also includes encoding the per-character differences in the matrix into a vector of the fixed width. The technique further includes outputting the vector as a representation of a pairwise error rate between the first transcription and the second transcription.

Abstract: One embodiment of the present invention sets forth a technique for analyzing transcriptions of a recording. The technique includes storing per-character differences between a first set of characters from a first transcription of the recording and a second set of characters from a second transcription of the recording in a matrix with a fixed width. The technique also includes encoding the per-character differences in the matrix into a vector of the fixed width. The technique further includes outputting the vector as a representation of a pairwise error rate between the first transcription and the second transcription.

Abstract: One embodiment of the present invention sets forth a technique for analyzing a transcription of a recording. The technique includes generating features representing transcriptions produced by multiple automatic speech recognition (ASR) engines from voice activity in the recording and a best transcription of the recording produced by an ensemble model from the transcriptions. The technique also includes applying a machine learning model to the features to produce a score representing an accuracy of the best transcription. The technique further includes storing the score in association with the best transcription.

Abstract: Computer-readable media, computer systems, and computing methods are provided for classifying search results as either of good quality or of poor quality. Initially, a portion of the search results, such as the highest ranked documents, are selected for evaluation. A level of quality for each of the selected search results is determined using a classification process that includes the following steps: targeting features demonstrated by the selected search results to be evaluated; evaluating the selected features to generate a level-of-quality score for each of the selected search results; comparing the score against a predefined threshold value; and, based on the comparison, assigning each of the selected search results an absolute measurement. The absolute measurement indicates poor quality when the score is less than the threshold value. Upon recognizing that the selected search results are of poor quality, automatically executing a corrective action that reformulates the issued search query.

Abstract: Embodiments of this invention relate to detecting and blurring images. In an embodiment, a system detects objects in a photographic image. The system includes an object detector module configured to detect regions of the photographic image that include objects of a particular type at least based on the content of the photographic image. The system further includes a false positive detector module configured to determine whether each region detected by the object detector module includes an object of the particular type at least based on information about the context in which the photographic image was taken.

Abstract: A method and apparatus for automatically identifying character segments for character recognition is provided. The method involves receiving a plurality of words and a ground truth corresponding to each word of the plurality of words. The plurality of words may be received in a cursive script. Each word of the plurality of words is segmented into one or more character segments based on the ground truth corresponding to each word. Thereafter, the segmentation of each word is refined by iteratively re-segmenting each word based on one or more similar character segments.

Abstract: Systems, methods, and computer storage media having computer-executable instructions embodied thereon for rewriting queries and labeling word pairs. Queries are received and alternate words are identified for word pairs (i.e., query words and alternate words). Word pair links are presented to users and indicators are received based on actions taken by the users. Labels are assigned to the word pairs based on the indicators and communicated to a classifier.

Abstract: Embodiments of this invention relate to detecting and blurring images. In an embodiment, a system detects objects in a photographic image. The system includes an object detector module configured to detect regions of the photographic image that include objects of a particular type at least based on the content of the photographic image. The system further includes a false positive detector module configured to determine whether each region detected by the object detector module includes an object of the particular type at least based on information about the context in which the photographic image was taken.

Abstract: A method and apparatus for automatically identifying character segments for character recognition is provided. The method involves receiving a plurality of words and a ground truth corresponding to each word of the plurality of words. The plurality of words may be received in a cursive script. Each word of the plurality of words is segmented into one or more character segments based on the ground truth corresponding to each word. Thereafter, the segmentation of each word is refined by iteratively re-segmenting each word based on one or more similar character segments.

Abstract: Embodiments of this invention relate to detecting and blurring images. In an embodiment, a system detects objects in a photographic image. The system includes an object detector module configured to detect regions of the photographic image that include objects of a particular type at least based on the content of the photographic image. The system further includes a false positive detector module configured to determine whether each region detected by the object detector module includes an object of the particular type at least based on information about the context in which the photographic image was taken.

Abstract: OCR errors are identified and corrected through learning. An error probability estimator is trained using ground truths to learn error probability estimation. Multiple OCR engines process a text image, and convert it into texts. The error probability estimator compares the outcomes of the multiple OCR engines for mismatches, and determines an error probability for each of the mismatches. If the error probability of a mismatch exceeds an error probability threshold, a suspect is generated and grouped together with similar suspects in a cluster. A question for the cluster is generated and rendered to a human operator for answering. The answer from the human operator is then applied to all suspects in the cluster to correct OCR errors in the resulting text. The answer is also used to further train the error probability estimator.

Abstract: Various technologies and techniques are disclosed that improve handwriting recognition operations. Handwritten input is received in training mode and run through several base recognizers to generate several alternate lists. The alternate lists are unioned together into a combined alternate list. If the correct result is in the combined list, each correct/incorrect alternate pair is used to generate training patterns. The weights associated with the alternate pairs are stored. At runtime, the combined alternate list is generated just as training time. The trained comparator-net can be used to compare any two alternates in the combined list. A template matching base recognizer is used with one or more neural network base recognizers to improve recognition operations. The system provides comparator-net and reorder-net processes trained on print and cursive data, and ones that have been trained on cursive-only data. The respective comparator-net and reorder-net processes are used accordingly.

Abstract: A method and system for segmenting a text into a plurality of sections is provided. The text may be received in the form of an image. The method involves receiving one or more input labels from a user corresponding to one or more segmentation points of a plurality of segmentation points of the text. The plurality of segmentation points of the text are obtained by applying one or more segmentation heuristics over the text. The one or more input labels provided by the user are utilized to label the plurality of segmentation points of the text. In response to labeling, validation is performed to identify whether a segmentation point of the plurality of segmentation points is a valid segmentation point. Thereafter, based on the validation, a set of valid segmentation points is updated with one or more segmentation points of the plurality of segmentation points. The set of valid segmentation points facilitates segmentation of the text for recognizing the plurality of sections.

Abstract: Systems, methods, and computer storage media having computer-executable instructions embodied thereon for rewriting queries and labeling word pairs. Queries are received and alternate words are identified for word pairs (i.e., query words and alternate words). Word pair links are presented to users and indicators are received based on actions taken by the users. Labels are assigned to the word pairs based on the indicators and communicated to a classifier.

Abstract: Computer-readable media, computer systems, and computing methods are provided for classifying search results as either of good quality or of poor quality. Initially, a portion of the search results, such as the highest ranked documents, are selected for evaluation. A level of quality for each of the selected search results is determined using a classification process that includes the following steps: targeting features demonstrated by the selected search results to be evaluated; evaluating the selected features to generate a level-of-quality score for each of the selected search results; comparing the score against a predefined threshold value; and, based on the comparison, assigning each of the selected search results an absolute measurement. The absolute measurement indicates poor quality when the score is less than the threshold value. Upon recognizing that the selected search results are of poor quality, automatically executing a corrective action that reformulates the issued search query.

Abstract: Computer-readable media, computer systems, and computing methods are provided for classifying search results as either of good quality or of poor quality. Initially, a portion of the search results, such as the highest ranked documents, are selected for evaluation. A level of quality for each of the selected search results is determined using a classification process that includes the following steps: targeting features demonstrated by the selected search results to be evaluated; evaluating the selected features to generate a level-of-quality score for each of the selected search results; comparing the score against a predefined threshold value; and, based on the comparison, assigning each of the selected search results an absolute measurement. The absolute measurement indicates poor quality when the score is less than the threshold value. Upon recognizing that the selected search results are of poor quality, a corrective action that reformulates the issued search query is automatically invoked.

Abstract: In embodiments consistent with the subject matter of this disclosure, a user may input strokes as digital ink to a processing device. The processing device may partition the input strokes into multiple regions of strokes. A first recognizer and a second recognizer may score grammar objects included in regions and represented by chart entries. The scores may be converted to a converted score, which may have at least a near standard normal distribution. The processing device may present a recognition result based on highest converted scores according to a recurrence formula. The processing device may receive a correction hint with respect to misrecognized strokes and may add a penalty score with respect to chart entries representing grammar objects breaking the correction hint. Incremental recognition may be performed when a pause is detected during inputting of strokes.

Abstract: Various technologies and techniques are disclosed that improve handwriting recognition operations. Handwritten input is received in training mode and run through several base recognizers to generate several alternate lists. The alternate lists are unioned together into a combined alternate list. If the correct result is in the combined list, each correct/incorrect alternate pair is used to generate training patterns. The weights associated with the alternate pairs are stored. At runtime, the combined alternate list is generated just as training time. The trained comparator-net can be used to compare any two alternates in the combined list. A template matching base recognizer is used with one or more neural network base recognizers to improve recognition operations. The system provides comparator-net and reorder-net processes trained on print and cursive data, and ones that have been trained on cursive-only data. The respective comparator-net and reorder-net processes are used accordingly.