Abstract: A method including embedding, by a trained issue MLM (machine learning model), a new natural language issue statement into an issue vector. An inner product of the issue vector with an actions matrix is calculated. The actions matrix includes centroid-vectors calculated using a clustering method from a second output of a trained action MLM which embedded prior actions expressed in natural language action statements taken as a result of prior natural issue statements. Calculating the inner product results in probabilities associated with the prior actions. Each of the probabilities represents a corresponding estimate that a corresponding prior action is relevant to the issue vector. A list of proposed actions relevant to the issue vector is generated by comparing the probabilities to a threshold value and selecting a subset of the prior actions with corresponding probabilities above the threshold. The list of proposed actions is transmitted to a user device.

Abstract: Provided are a method, system, and computer program product for representing text, in which a text is received and analyzed by utilizing a pre-trained embedding model and a feature vector model, wherein selected words in the text have corresponding weights. Operations whose parameters include weights of a feature vector and an embedding are performed to generate a weighted embedding data structure. A summation is performed of all corresponding columns of a plurality of rows of the weighted embedding data structure to generate a data structure that represents the text. The data structure that represents the text is utilized to generate at least one of a classification metadata for the text and a summarization of the text.

Abstract: The present invention relates to audio coding systems which make use of a harmonic transposition method for high frequency reconstruction (HFR). A system and a method for generating a high frequency component of a signal from a low frequency component of the signal is described. The system comprises an analysis filter bank providing a plurality of analysis subband signals of the low frequency component of the signal. It also comprises a non-linear processing unit to generate a synthesis subband signal with a synthesis frequency by modifying the phase of a first and a second of the plurality of analysis subband signals and by combining the phase-modified analysis subband signals. Finally, it comprises a synthesis filter bank for generating the high frequency component of the signal from the synthesis subband signal.

Abstract: A method, computer system, and a computer program product for masking identifying traits contained in response text is provided. Embodiments may include receiving a request to anonymize response text in response to a predefined respondent interaction, wherein the response text is generated by the respondent and then obtaining the response text, wherein the obtained response text has semantic characteristics. Next, the obtained response text may be input into a natural language processing (NLP) algorithm and thereafter receiving an alternative masking text as output from the NLP algorithm, wherein the received alternative masking text maintains the semantic characteristics of the obtained response text. Finally, the response text may be replaced with the received alternative masking text.

Abstract: An embodiment calculates a point of view (POV) value for a subportion of an item of media content, including evaluating user inputs related to an automatically detected POV of the subportion of the item. The embodiment also calculates a non-factual cost for the subportion of the item indicative of an amount of the subportion of the item that lacks support in a fact-based corpora. The embodiment also performs a cascaded summarization process comprising generating a summary of the item using the subportion of the item and a summarization technique, analyzing the summary using a fact-checking algorithm to determine whether the summary satisfies a factual score threshold, and performing a next iteration of the cascaded summarization process while the summary fails to satisfy the factual score threshold. The embodiment also communicates a summary satisfying the factual score threshold to a user device.

Abstract: An approach is provided that receives a message and applies a deep analytic analysis to the message. The deep analytic analysis results in a set of enriched message embedding (EME) data that is passed to a trained neural network. Based on a set of scores received from the trained neural network, a conversation is identified from a number of available conversations to which the received message belongs. The received first message is then associated with the identified conversation.

Abstract: In some cases, lower quality, large scale training data can be automatically generated by automatic labeling. The generated training data can be used to pre-train a machine learning model. For instance, the model can be a model for detection of verbal harassment. Parameters of the pre-trained model can be refined or updated using another one or more higher-quality sets of training data, with which the model can be subsequently trained.

Abstract: A method of improving output content through iterative generation is provided. The method includes receiving a natural language input, obtaining user intention information based on the natural language input by using a natural language understanding (NLU) model, setting a target area in base content based on a first user input, determining input content based on the user intention information or a second user input, generating output content related to the base content based on the input content, the target area, and the user intention information by using a neural network (NN) model, generating a caption for the output content by using an image captioning model, calculating similarity between text of the natural language input and the generated output content, and iterating generation of the output content based on the similarity.

Abstract: It discloses a speech feature reuse-based storing and calculating compression method for a keyword-spotting CNN, and belongs to the technical filed of calculating, reckoning or counting. If the updated row number of input data is equal to a convolution step size, every time new input data arrive, an input layer of a neural network replaces the earliest part of the input data with the new input data and meanwhile adjusts an addressing sequence of the input data, thereby performing an operation on the input data and corresponding convolution kernels in an arrival sequence of the input data, and an operation result is stored in an intermediate data memory of the neural network to update corresponding data.

Abstract: Systems, computer-implemented methods, and computer program products to transform a lexicon that describes an information asset are provided. According to an embodiment, a system can comprise a memory that stores computer executable components and a processor that executes the computer executable components stored in the memory. The computer executable components can comprise a term validation component that can determine from a subject matter expert, a validated term that can indicate validation of a candidate term that describes an information asset. The computer executable components can further comprise a lexicon transforming component that, based on the validated term, can transform a lexicon that describes the information asset, by incorporating the validated term into the lexicon.

Abstract: An information processing system includes: a first device configured to acquire a user's uttered voice, transfer the user's uttered voice to at least one of a second and a third devices each actualizing a voice interaction agent, when a control command is acquired, convert a control signal based on the acquired control command to a control signal that matches the second device, and transmit the converted control signal to the second device; a second device configured to recognize the uttered voice transferred from the first device, and output, to the first device, a control command regarding a recognition result obtained by recognizing the uttered voice and response data based on the control signal; and a third device configured to recognize the uttered voice transferred from the first device, and output, to the first device, a control command regarding a recognition result obtained by recognizing the uttered voice.

Abstract: A method of detecting occupancy in an area includes obtaining, with a processor, an audio sample from an audio sensor and determining, with the processor, feature functional values of a set of selected feature functionals from the audio sample. The determining of the feature functional values includes extracting features in the set of selected feature functionals from the audio sample, and determining the feature functional values of the set of selected features from the extracted features. The method further includes determining, with the processor, occupancy in the area using a classifier based on the determined feature functional values.

Abstract: Techniques for automatically extracting data from documents using multiple deep learning models are provided. According to one set of embodiments, a computer system can receive a document in an electronic format and can segment, using an image segmentation deep learning model, the document into a plurality of segments, where each segment corresponds to a visually discrete portion of the document and is classified as being one of a plurality of types. The computer system can then, for each segment in the plurality of segments, retrieve text in the segment using optical character recognition (OCR) and extract data in the segment from the retrieved text using a named entity recognition (NER) deep learning model, where the retrieving and the extracting are performed in a manner that takes into account the segment's type.

Abstract: A system is provided for determining which component of a speech processing system is the cause of an undesired response to a user input. The system processes ASR data and NLU data to determine the component likely to cause the undesired response. Based on which component is the cause of the undesired response, the system performs an appropriate conversation recovery technique to confirm the speech processing results with the user.

Abstract: Provided is an encoder, system and method for metaphor detection in natural language processing. The system comprises an encoding module configured to convert words included in a sentence into BiLSTM representation vectors; a first encoder configured to generate a first entire representation vector of a WSD resolving task; a second encoder configured to generate a second entire representation vector of an MD task; and a multi-task learning module configured to perform knowledge transfer between the first and second encoders. Wherein, each of the first and second encoders includes a graph convolutional neural network (GCN) module configured to encode a link between a target word and a core word to generate GCN representation vectors; a control module configured to regulate the GCN representation vectors to generate an entire representation vector.

Abstract: Disclosed of the present application is relation to deep learning based text classification. The training corpus is screened by key clauses according to the weights of clauses in the training corpus, so as to keep the complete sentence and the original word order as much as possible according to the language habits. Thus, the deep learning model can learn normal semantic features. In addition, the subsample sets corresponding to different preset word length intervals is obtained from the training sample set, and each subsample set is putted into the deep learning model for training, so that several text classification models corresponding to different preset word length intervals can be obtained for text classification. Therefore, the deep learning models can be self-adaptively selected to classify texts based on the above mentioned multiple word length intervals and multi-model training method, to improve text classification accuracy.

Abstract: A transcription of audio speech included in electronic content associated with a meeting is created by an ASR model trained on speech-to-text data. The transcription is post-processed by modifying text included in the transcription, for example, by modifying punctuation, grammar, or formatting introduced by the ASR model and by changing or omitting one or more words that were included in both the audio speech and the transcription. After the transcription is post-processed, output based on the post-processed transcription is generated in the form of a meeting summary and/or template.

Abstract: A method for employing a supervised graph sparsification (SGS) network to use feedback from subsequent graph learning tasks to guide graph sparsification is presented. The method includes, in a training phase, generating sparsified subgraphs by edge sampling from input training graphs following a learned distribution, feeding the sparsified subgraphs to a prediction/classification component, collecting a predication/classification error, and updating parameters of the learned distribution based on a gradient derived from the predication/classification error. The method further includes, in a testing phase, generating sparsified subgraphs by edge sampling from input testing graphs following the learned distribution, feeding the sparsified subgraphs to the prediction/classification component, and outputting prediction/classification results to a visualization device.

Abstract: An information processing apparatus includes an acquisition unit that acquires voice data and image data, respectively, a display control unit that performs control to display the image data acquired by the acquisition unit in synchronization with the voice data, a reception unit that receives a display element to be added for display to a specific character in the image data displayed by the display control unit, and a setting unit that sets a playback period in which the specific character in the voice data is played back, as a display period of the display element received by the reception unit in the image data.

Abstract: A method includes receiving a training example for a listen-attend-spell (LAS) decoder of a two-pass streaming neural network model and determining whether the training example corresponds to a supervised audio-text pair or an unpaired text sequence. When the training example corresponds to an unpaired text sequence, the method also includes determining a cross entropy loss based on a log probability associated with a context vector of the training example. The method also includes updating the LAS decoder and the context vector based on the determined cross entropy loss.