Abstract: Techniques for determining a similarity between text segments within a document comprising textual references are described. According to an example, a system comprises 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: an identification component that identifies a reference associated with a set of text and an extraction component that extracts the reference from the set of text. The computer executable components can also comprise an embedding component that replaces the reference with a corresponding vector.

Abstract: The methods, systems, and computer program products described herein provide ways to generate an informative training corpus of samples for use in machine training a high-quality sentiment analysis computer model. In some aspects, a method is disclosed including receiving a plurality of training samples, extracting semantic and sentiment elements of one or more of the training samples, generalizing the semantic and sentiment elements of the one or more of the training samples, generating an informative ranking score for one or more of the training samples based on the generalized semantic and sentiment elements, selecting informative training samples from the plurality of training samples based at least in part on the generated informative ranking scores, and adding the selected informative training samples to an informative training corpus.

Abstract: Techniques are provided for training machine learning model. According to one aspect, a training data is received by one or more processing units. The machine learning model is trained based on the training data, wherein the training comprises: optimizing the machine learning model based on stochastic gradient descent (SGD) by adding a dynamic noise to a gradient of a model parameter of the machine learning model calculated by the SGD.

Abstract: A method is presented for error correction of tabular data in document conversion. The method includes identifying errors from tabular data transformation by employing an error/invalidation checking module and correcting the identified errors from the tabular data transformation by employing an error correction module. The error correction module includes identifying a main structure pattern from common row structures, concatenating separate keywords according to natural language processing models employing training data obtained from a plurality of candidate tabular data, adjusting cells in the tabular data based on a domain-specific knowledge database including the training data in combination with linguistic and semantic knowledge, merging partial tabular data pieces, and generating an adjusted table as output on a display of a computing device.

Abstract: Techniques for detection of document similarity are provided. The computer-implemented method can comprise identifying, by an electronic device operatively coupled to a processing unit, a first pragmatic association of a first segment in a first document portion, the first pragmatic association indicating meaning of the first segment specific to a context of the first segment in the first document portion. The computer-implemented method can also comprise generating a first intermediate document portion from the first document portion by using the first pragmatic association to replace the first segment. The computer-implemented method can further comprise determining a similarity degree between the first document portion and a second document portion by comparing the first intermediate document portion with the second document portion.

Abstract: Techniques are provided for automated privacy scoring of user information. In one example, a system comprises a memory that stores computer executable components, and a processor that executes computer executable components stored in the memory. The computer executable components can comprise a privacy scoring component that employs a privacy identification model to generate a privacy score for a user and a product in the particular context based on information associated with the user and the product in the particular context. The computer executable components can also comprise a privacy enforcement component that implements one or more privacy features on the information based on the privacy score.

Abstract: A method, system, and computer program product, include receiving a request for registration from a service provider, upon the service receiver having authorized the request for registration, registering characteristic information of the service call in a user device of a service receiver, and upon a lapse of time, deregistering the characteristic information from the user device.

Abstract: A topic guidance method, system, and computer program product for suggesting, via a processor on a computer, a conversation topic for the agent to engage the customer based on a learned conversation topic model, the conversation model being a static model.

Abstract: Systems and methods for estimating battery-powered driving distance for a vehicle, including training a relative model for a battery using input historical battery temperature data and historical battery-external factors, and predicting a future battery temperature based on the relative model and one or more of current or future battery-external factors. A battery power capacity is determined using the predicted future battery temperature and input manufacturer specifications for the battery, and a remaining battery powered driving distance is calculated based on input vehicle power consumption data and the determined battery power capacity.

Abstract: Methods and systems for natural language processing include generating respective feature vectors, for each word in an input, based on sequences of input words of different respective lengths. The respective feature vectors for each word in the input are combined to form a combined vector for each word. A hidden state is determined for each word in the input based on the combined vector. The hidden states for all words in the input are combined to form a multi-resolution context vector. A natural language processing action is performed using the multi-resolution context vector.

Abstract: A method is presented for error correction of tabular data in document conversion. The method includes identifying errors from tabular data transformation by employing an error/invalidation checking module and correcting the identified errors from the tabular data transformation by employing an error correction module. The error correction module includes identifying a main structure pattern from common row structures, concatenating separate keywords according to natural language processing models employing training data obtained from a plurality of candidate tabular data, adjusting cells in the tabular data based on a domain-specific knowledge database including the training data in combination with linguistic and semantic knowledge, merging partial tabular data pieces, and generating an adjusted table as output on a display of a computing device.

Abstract: A computer-implemented method, computer program product, and computer processing system are provided for computing a trajectory-based Point of Interest recommendation. The method includes generating, by a processor device, a set of embeddings. Each of the embeddings in the set relates to a respective different trajectory contextual element of a user trajectory. The method further includes computing, by the processor device based on the set of embeddings, an activity representation that includes a set of POI candidate embeddings. The method also includes composing, by the processor device, a stop embedding based on the activity representation and the embeddings in the set and corresponding to a given stop in the user trajectory. The method additionally includes computing, by the processor device, the trajectory-based POI recommendation using an attention-based, user-specific, multi-stop trajectory, Recurrent Neural Network (RNN) model applied to the stop embedding.

Abstract: A computer-implemented method for implementing separated attention on like and dislike items for personalized ranking includes performing an element-wise product on a user embedding and a final like item embedding to generate a first vector. The method further includes performing an element-wise product on the user embedding and a final dislike item embedding to generate a second vector. The method further includes computing a probability that the user prefers the like item to the dislike item based on the first and second vectors, and generating one or more item recommendations including one or more electronic images for the user using the probability.

Abstract: A computer-implemented method, a computer program product, and a computer processing system are provided for residual value prediction of an item. The method includes predicting, by a processor device, features of the item from unstructured data and structured data. The method further includes predicting, by the processor device, a residual value of the item using the predicted features. The method also includes generating, by the processor device on an interactive user display device, an interactive display interface that includes a prediction of the residual value of the item and provides a set of user selectable actions for performing relative to the prediction.

Abstract: A method is provided for role-oriented risk analysis in a contract. The method generates, using deep semantic association analysis, a report specifying a set of potential risks relating to explicit and hidden roles of contract parties. The generating step categorizes input statements of the contract into respective obligation/right pairs according to a deep semantic association distribution thereof. Each pair includes a respective obligation and a respective right. The generating step detects deep semantic differences between the respective pairs and a set of reference obligation/right pairs. The generating step identifies the explicit and hidden roles of the involved parties in the respective obligations/rights pairs according to domain-specific use scenarios and multidimensional local and global context clues in the contract. The generating step identifies the set of potential risks by applying a deep semantic role-oriented risk entailment model to the deep semantic differences.

Abstract: A method is provided for clause analysis in a legal domain. The method builds a coherence graph from a set of labeled training documents by (a) creating entity nodes from and of a same type as entities extracted from the set of labeled training documents, (b) creating clause nodes from labeled clauses in the set of labeled training documents, (c) forming bi-directional edges (i) between each of the clause nodes and the entity nodes belonging thereto, (ii) among parent-child clause nodes from among the clause nodes, and (iii) among same-level sibling clause nodes from among the clause nodes. The method merges nodes, from among the entity and clause nodes, that have a same semantic meaning. The method weights the bi-directional edges using a coherence metric. The method identifies a clause structure of a new document by matching the new document against the coherence graph using a node-covering algorithm.

Abstract: A method, a device and a computer program product for managing a shared vehicle are proposed. In the method, first information about a user is obtained. The first information indicates a speed of at least one shared vehicle used by the user. A target vehicle is determined from the at least one shared vehicle based on the first information. Second information about the target vehicle is determined. The second information indicates respective speeds of the target vehicle moving at a plurality of time intervals. An abnormal state of the target vehicle is detected based on the second information. The abnormal state indicates that a malfunction occurs in the target vehicle.

Abstract: A computer-implemented method is presented for implementing multi-aspect sentiment analysis by collaborative attention allocation. The method includes extracting a sequence of word vectors from a sentence received from a data stream, feeding the sequence of word vectors to long short-term memory (LSTM) neural networks to generate a sequence of hidden states corresponding to the sequence of word vectors, generating a plurality of aspect embedding vectors for each aspect, employing an attention mechanism to determine attention weight vectors concurrently for all aspects, and outputting predicted sentiments for each aspect of the sentence to a user interface of a computing device.

Abstract: This disclosure provides a computer-implemented method. The method may include extracting one or more features based on a first utterance from a first interlocutor in a dialog and a second utterance from a second interlocutor in the dialog. The method may further include inferring one or more personality traits of the first interlocutor based on the one or more extracted features from the dialog.

Abstract: A method, computer system, and a computer program product for automatic equation transformation from text is provided. The present invention may include receiving a text document. The present invention may then include identifying a mathematical formula expressed in the received text document. The present invention may then include removing a plurality of superfluous language from the received text document based on the identified mathematical formula. The present invention may also include transforming the identified mathematical formula into a symbolic representation based on a trained model. The present invention may finally include outputting the symbolic representation.