Abstract: A system is provided comprising a processor and a memory storing instructions which configure the processor to process an original sentence structure through an encoder neural network to decompose the original sentence structure into an original semantics component and an original syntax component, process the original syntax component through a syntax variation autoencoder (VAE) to receive a syntax mean vector and a syntax covariance matrix, obtain a sampled syntax value from a syntax Gaussian posterior parameterized by the syntax mean vector and the syntax covariance matrix, process the original semantics component through a semantics VAE to receive a semantics mean vector and a semantics covariance matrix, obtain a sampled semantics vector from the Gaussian semantics posterior parameterized by the semantics mean vector and the semantics covariance matrix, and process the sampled syntax vector and the sampled semantics vector through a decoder neural network to compose a new sentence.

Abstract: Embodiments described herein provide adversarial attacks targeting the cross-lingual generalization ability of massive multilingual representations, demonstrating their effectiveness on multilingual models for natural language inference and question answering. An efficient adversarial training scheme can thus be implemented with the adversarial attacks, which takes the same number of steps as standard supervised training and show that it encourages language-invariance in representations, thereby improving both clean and robust accuracy.

Abstract: One embodiment provides a method, including: accessing, at an information handling device, a dynamic visual media corpus, wherein the dynamic visual media corpus comprises a plurality of dynamic visual media scripts; segmenting each of the plurality of dynamic visual media scripts into scenes; generating, for each of the plurality of dynamic visual media scripts, a character fingerprint identifying topics corresponding to each character within a corresponding dynamic visual media script, wherein the generating comprises (i) extracting both characters and topics from the dynamic visual media script and (ii) associating each of the topics with a corresponding character, wherein the character fingerprint identifies costumes of a given character and a topic corresponding to each costume; and producing, for each scene within each dynamic visual media script, a scene vector identifying (iii) the topics included within a corresponding scene and (iv) a character fingerprint for each character occurring within the sce

Abstract: A method includes obtaining n pairs of translation sentences of a source language and a target language, where each of the n pairs of translation sentences includes a source language sentence and a target language sentence that are translations of each other, extracting a source language segment from each source language sentence in the n pairs of translation sentences using an extraction rule of the source language, extracting a target language segment from each target language sentence in the n pairs of translation sentences, and generating an extraction rule of the target language based on n target language segments extracted from n target language sentences.

Abstract: Introduced here are computer programs and associated computer-implemented techniques for discovering the presence of filler words through tokenization of a transcript derived from audio content. When audio content is obtained by a media production platform, the audio content can be converted into text content as part of a speech-to-text operation. The text content can then be tokenized and labeled using a Natural Language Processing (NLP) library. Tokenizing/labeling may be performed in accordance with a series of rules associated with filler words. At a high level, these rules may examine the text content (and associated tokens/labels) to determine whether patterns, relationships, verbatim, and context indicate that a term is a filler word. Any filler words that are discovered in the text content can be identified as such so that appropriate action(s) can be taken.

Abstract: Techniques for stop word data augmentation for training chatbot systems in natural language processing. In one particular aspect, a computer-implemented method includes receiving a training set of utterances for training an intent classifier to identify one or more intents for one or more utterances; augmenting the training set of utterances with stop words to generate an augmented training set of out-of-domain utterances for an unresolved intent category corresponding to an unresolved intent; and training the intent classifier using the training set of utterances and the augmented training set of out-of-domain utterances. The augmenting includes: selecting one or more utterances from the training set of utterances, and for each selected utterance, preserving existing stop words within the utterance and replacing at least one non-stop word within the utterance with a stop word or stop word phrase selected from a list of stop words to generate an out-of-domain utterance.

Abstract: The present disclosure provides a method, a device, an equipment and a storage medium for mining a topic concept. The method includes: acquiring a plurality of candidate topic concepts based on a query; performing word segmentation on the plurality of candidate topic concepts and performing part-of-speech tagging on words obtained after performing the word segmentation, to obtain a part-of-speech sequence of each of the plurality of candidate topic concepts; and filtering the plurality of candidate topic concepts based on the part-of-speech sequence, to filter out a topic concept corresponding to a target part-of-speech sequence among the plurality of candidate topic concepts, in which a proportion of accurate topic concepts in the target part-of-speech sequence is lower than or equal to a first preset threshold, or a proportion of inaccurate topic concepts in the target part-of-speech sequence is higher than or equal to a second preset threshold.

Abstract: This disclosure describes a tabletop game assistant system configured to ingest and guide tangible games (such as board games, card games, etc.) using natural language interaction and image capture/visual display components. The system can include features enabling a game developer to “teach” the system the rules of a game using natural language, such as written instructions, to reduce or eliminate the need for writing dedicated code. The system may process images of a game board and/or tokens such as game pieces and/or cards to further generate game data in the form of a logical game model. The system can use the game data to guide human players of the game and, in some cases, participate as a player itself. The system may further be configured to observe a game and detect invalid actions, answer questions regarding the rules, and suggest moves. The system may provide additional utilities such as generating a random output (e.g., rolling virtual dice) and learning to recognize new game pieces.

Abstract: An information processing apparatus includes a controller that is configured to identify a first language into which a content of a speech that is input is to be translated, based on first information about a place, estimate an intention of the content of the speech based on the content of the speech that is translated into the first language, select a service to be provided, based on the intention that is estimated, and provide a guide related to the service that is selected, in a language of the speech. The first language is different from the language of the speech.

Abstract: An automatic interpretation method performed by a correspondent terminal communicating with an utterer terminal includes receiving, by a communication unit, voice feature information about an utterer and an automatic translation result, obtained by automatically translating a voice uttered in a source language by the utterer in a target language, from the utterer terminal and performing, by a sound synthesizer, voice synthesis on the basis of the automatic translation result and the voice feature information to output a personalized synthesis voice as an automatic interpretation result. The voice feature information about the utterer includes a hidden variable including a first additional voice result and a voice feature parameter and a second additional voice feature, which are extracted from a voice of the utterer.

Abstract: A method includes obtaining a plurality of training data sets each associated with a respective native language and includes a plurality of respective training data samples. For each respective training data sample of each training data set in the respective native language, the method includes transliterating the corresponding transcription in the respective native script into corresponding transliterated text representing the respective native language of the corresponding audio in a target script and associating the corresponding transliterated text in the target script with the corresponding audio in the respective native language to generate a respective normalized training data sample.

Abstract: A framework ranks multiple hypotheses generated by one or more ASR engines for each input speech utterance. The framework jointly implements ASR improvement and NLU. It makes use of NLU related knowledge to facilitate the ranking of competing hypotheses, and outputs the top-ranked hypothesis as the improved ASR result together with the NLU results of the speech utterance. The NLU results include intent detection results and the slot filling results.

Abstract: An illustrative embodiment includes a method for training an end-to-end (E2E) spoken language understanding (SLU) system. The method includes receiving a training corpus comprising a set of text classified using one or more sets of semantic labels but unpaired with speech and using the set of unpaired text to train the E2E SLU system to classify speech using at least one of the one or more sets of semantic labels. The method may include training a text-to-intent model using the set of unpaired text; and training a speech-to-intent model using the text-to-intent model. Alternatively or additionally, the method may include using a text-to-speech (TTS) system to generate synthetic speech from the unpaired text; and training the E2E SLU system using the synthetic speech.

Abstract: The present disclosure provides a corpus cleaning method and a corpus entry system. The method includes: obtaining an input utterance; generating a predicted value of an information amount of each word in the input utterance according to the context of the input utterance using a pre-trained general model; and determining redundant words according to the predicted value of the information amount of each word, and determining whether to remove the redundant words from the input utterance. In such a manner, the objectivity and accuracy of corpus cleaning can be improved.

Abstract: In a frictionless handoff of audio content playing, a client device listens for ultrasonic audio. The client hears a playing of a modified audio content by another client device, which includes audio content and an ultrasonic audio quick response (QR) code overlaid on the audio content. The ultrasonic audio QR code includes location information corresponding to a location in the audio content. The client device extracts the ultrasonic audio QR code from the modified audio content. After determining that the playing of the modified audio content has stopped, the client device receives a command to resume playing of the audio content on the client device. In, response to the command, the client device retrieves location information in a last extracted ultrasonic audio QR code and plays the audio content starting at a location in the audio content corresponding to the location information in the last extracted ultrasonic audio QR code.

Abstract: In general, techniques are described by which to correlate scene-based audio data for psychoacoustic audio coding. A device comprising a memory and one or more processors may be configured to perform the techniques. The memory may store a bitstream including a plurality of encoded correlated components of a soundfield represented by scene-based audio data. The one or more processors may perform psychoacoustic audio decoding with respect to one or more of the plurality of encoded correlated components to obtain a plurality of correlated components, and obtain, from the bitstream, an indication representative of how the one or more of the plurality of correlated components were reordered in the bitstream. The one or more processors may reorder, based on the indication, the plurality of correlated components to obtain a plurality of reordered components, and reconstruct, based on the plurality of reordered components, the scene-based audio data.

Abstract: A method for automatically translating raw data into real human voiced audio content is provided according to an embodiment of the present disclosure. The method may comprise ingesting data, separating the data into or associating the data with a data type, and creating a list of descriptive data associated with the data type. In some embodiments, the method further comprises compiling audio phrases types associated with the descriptive data, associating a pre-recorded audio file with each audio phrase, and merging a plurality of pre-recorded audio files to create a final audio file.

Abstract: A keyboard instrument includes at least one processor that determines a first pattern of intonation to be applied to a first time segment of a voice data on the basis of a first user operation on a first operation element, causes a first singing voice for the first time segment to be digitally synthesized from the first segment data in accordance with the determined first pattern of intonation, determines a second pattern of intonation to be applied to the second time segment of the voice data on the basis of a second user operation on a second operation element, and causes a second singing voice for the second time segment to be digitally synthesized from the second segment data in accordance with the determined second pattern of intonation.

Abstract: The present application relates to a system language switching method, a computer readable storage medium, a terminal device, and a device. The method includes first obtaining a preset image for setting a system language of a target terminal, then extracting text information in the image and determining a target language corresponding to the text information, and finally switching the system language of the target terminal to the target language. Through the present application, the user only needs to prepare an image for setting the system language of the target terminal in advance, for example, a piece of paper with Chinese written, and a system can obtain the text information on the image through the processes of image acquisition, text information extraction, and the like, determine that the text message is Chinese, and finally switch the system language of the target terminal to Chinese.

Abstract: System, method, and various embodiments for providing contextualized character recognition system are described herein. An embodiment operates by determining a plurality of predicted words of an image. An accuracy measure or each of the plurality of predicted words is identified and a replaceable word with an accuracy measure below a threshold is identified. A plurality of candidate words associated with the replaceable word are identified and a probability for each of the candidate words is calculated based on a contextual analysis. One of the candidate words with a highest probability is selected. The plurality of predicted words including the selected candidate word with the highest probability replacing the replaceable word is output.