Abstract: Systems, computer-implemented methods, and computer program products to facilitate reducing error propagation when combining generative aligners and transition-based aligners are provided. According to an embodiment, a system can comprise a processor that executed components stored in memory. The computer executable components comprise a generative alignment component, an error propagation component, a discriminative parser, and a stochastic oracle policy component. The error propagation component can compute a posterior distribution over one or more hard alignments of parts given a pair of the generative alignment component. The discriminative parser can be trained via the stochastic oracle policy component to reduce error propagation when combining the generative alignment component with the discriminative parser.

Abstract: Mechanisms are provided for generating instruction tuning training data to train a machine learning computer model. The mechanisms receive a set of seed tasks, each specifying an instruction tuning example having at least an instruction and an output. The mechanisms generate, by a first language model (LM), from instructions of seed tasks in a first portion of the set of seed tasks, a synthetically generated instruction. The mechanisms generate, by each second LM in a plurality of second LMs, based on the synthetically generated instruction and a second portion of the set of seed tasks, a predicted output. The mechanisms score the predicted outputs and select a highest scoring output based on the scoring of the predicted outputs. The mechanisms generate a synthetic instruction tuning example, for inclusion in a training dataset for training the machine learning computer model, based on the synthetic instruction and selected output.

Abstract: Embodiments of the invention provide a computer-implemented method that includes executing, using a generative language model, generative language model operations operable to generate an output sequence responsive to an original input sequence. The generative language model operations include token pruning operations that include performing a base set of token pruning operations on intermediate versions of the original input sequence; and performing token pruning (TP) constraint evaluations. The base set of token pruning operations identify pruning candidate tokens in the intermediate versions of the original input sequence. The TP constraint evaluations determine that at least one of the pruning candidate tokens will be pruned from an associated intermediate version of the original input sequence.

Abstract: Systems and techniques that facilitate semantics-guided domain-specific data augmentation for text-to-graph parsing are provided. In various embodiments, a system can access an annotated training dataset, wherein the annotated training dataset can comprise a set of abstract meaning representation graphs respectively corresponding to a set of natural language sentences. In various aspects, the system can generate an augmented version of the annotated training dataset, based on applying semantics-guided composition operations or semantics-guided substitution operations to the set of abstract meaning representation graphs. In various instances, a lexicon legend can comprise domain-specific graphs respectively representing discrete tokens unique to a domain of the annotated training dataset.

Abstract: Systems, computer-implemented methods, and computer program products to facilitate reducing error propagation when combining generative aligners and transition-based aligners are provided. According to an embodiment, a system can comprise a processor that executed components stored in memory. The computer executable components comprise a generative alignment component, an error propagation component, a discriminative parser, and a stochastic oracle policy component. The error propagation component can compute a posterior distribution over one or more hard alignments of parts given a pair of the generative alignment component. The discriminative parser can be trained via the stochastic oracle policy component to reduce error propagation when combining the generative alignment component with the discriminative parser.

Abstract: A computer system, product, and method are provided. The computer system includes an artificial intelligence (AI) platform operatively coupled to a processor. The AI platform includes tools in the form of a machine learning model (MLM) manager, a metric manager, and a training manager. The MLM manager accesses a plurality of pre-trained source MLMs, and inputs a plurality of data objects of a test dataset into each of the source MLMs. The test dataset includes the plurality of data objects associated with respective labels. For each source MLM, associated labels are generated from the inputted data objects and a similarity metric is calculated. The MLM manager selects a base MLM to be used for transfer learning from the plurality of source MLMs based upon the calculated similarity metric. The training manager trains the selected base MLM with a target dataset for the target domain.

Abstract: An approach for generating synthetic treebanks to be used in training a parser in a production system is provided. A processor receives a request to generate one or more synthetic treebanks from a production system, wherein the request indicates a language for the one or more synthetic treebanks. A processor retrieves at least one corpus of text in which the requested language is present. A processor provides the at least one corpus to a transformer enhanced parser neural network model. A processor generates at least one synthetic treebank associated with a string of text from the at least one corpus of text in which the requested language is present. A processor sends the at least one synthetic treebank to the production system, wherein the production system trains a parser utilized by the production system with the at least one synthetic treebank.

Abstract: A method, computer program product, and/or computer system generate a first adversarial statement via: (1) receiving a question and an original context for the question; (2) converting the question into a statement with a placeholder answer; (3) picking randomly an answer entity from a training text corpus; (4) replacing the placeholder answer with the randomly picked answer entity; and (5) leaving a correct question entity in the statement. The first adversarial statement is inserted into the original context to form a first adversarial context. The question and the first adversarial context as a first pair and the question and the original context as a second pair are input into a question-answer dialog system to train the question-answer dialog system.

Abstract: A method and system are provided for combining models. The method includes forming, by a computer having a processor and a memory, model pairs from a model ensemble that includes a plurality of models. The method further includes comparing the model pairs based on sets of output results produced by the model pairs to provide comparison results. The method also includes constructing, by the computer, a combination model from at least one of the model pairs based on the comparison results. The comparing step is performed using user-generated set-based feedback.

Abstract: A method includes receiving, by a question-answer system, a question in a first language and the question in a second language and predicting, by the question-answer system, a first answer to the question in the first language and a second answer to the question in the second language. The method also includes generating, by the question-answer system, a first vector representing the question in the first language and a second vector representing the question in the second language and adjusting the question-answer system based on the first and second answers and the first and second vectors such that when the question-answer system subsequently generates a third vector representing the question in the first language and a fourth vector representing the question in the second language, a distance between the third and fourth vectors is less than a distance between the first and second vectors.

Abstract: Receive a question via a graphical user interface (GUI), obtain a passage of text potentially relevant to the question, and receive, via the GUI, a selection of a number of question-answering models to be ensembled. Produce a plurality of answers to the question by running a plurality of question-answering models, consistent with the selection of the number of question-answering models to be ensembled, on the passage of text. Produce an ensembled answer by ensembling the plurality of answers according to their respective confidence scores. Display, via the GUI, the ensembled answer in context of the passage of text, with the ensembled answer visually marked in the passage of text. Optionally, repeat these steps for a second passage of text.

Abstract: A method, computer program product, and/or computer system generate a first adversarial statement via: (1) receiving a question and an original context for the question; (2) converting the question into a statement with a placeholder answer; (3) picking randomly an answer entity from a training text corpus; (4) replacing the placeholder answer with the randomly picked answer entity; and (5) leaving a correct question entity in the statement. The first adversarial statement is inserted into the original context to form a first adversarial context. The question and the first adversarial context as a first pair and the question and the original context as a second pair are input into a question-answer dialog system to train the question-answer dialog system.

Abstract: A computer-implemented method according to one embodiment includes converting an input question into a vector form using trained word embeddings; constructing a type similarity matrix using a predetermined ontology; and determining a score for all possible types for the input question, based on the input question in vector form and the type similarity matrix.

Abstract: A method, computer program product, and/or computer system protects a question-answer dialog system from being attacked by adversarial statements that incorrectly answer a question. A computing device accesses a plurality of adversarial statements that are capable of making an adversarial attack on a question-answer dialog system, which is trained to provide a correct answer to a specific type of question. The computing device utilizes the plurality of adversarial statements to train a machine learning model for the question-answer dialog system. The computing device then reinforces the trained machine learning model by bootstrapping adversarial policies that identify multiple types of adversarial statements onto the trained machine learning model. The computing device then utilizes the trained and bootstrapped machine learning model to avoid adversarial attacks when responding to questions submitted to the question-answer dialog system.

Abstract: An approach for generating synthetic treebanks to be used in training a parser in a production system is provided. A processor receives a request to generate one or more synthetic treebanks from a production system, wherein the request indicates a language for the one or more synthetic treebanks. A processor retrieves at least one corpus of text in which the requested language is present. A processor provides the at least one corpus to a transformer enhanced parser neural network model. A processor generates at least one synthetic treebank associated with a string of text from the at least one corpus of text in which the requested language is present. A processor sends the at least one synthetic treebank to the production system, wherein the production system trains a parser utilized by the production system with the at least one synthetic treebank.

Abstract: A system, computer program product, and method are provided for extraction of factual data from unstructured natural language (NL) text. A detection model is applied to convert unstructured NL text in a first language to annotated NL text. The detection model identifies two or more mentions from the unstructured NL text and a logical position of the mentions. The detection model further identifies a sequential position for each of the mentions and attaches a sequential position identifier. A pattern of rules corresponding with the annotated NL text is identified and applied to the annotated NL text, and one or more facts embedded within the annotated NL text are extracted and converted into structured data.

Abstract: A processor may receive a text segment. The processor may analyze the text segment at a plurality of granularity levels wherein each of the plurality of granularity levels has a comparative selection value for identifying one or more objects of interest within the text segment. The processor may select an optimized granularity level with an optimum comparative selection value. The processor may identify the one or more objects of interest within the text segment. The processor may display the one or more objects of interest to a user.

Abstract: A method includes receiving, by a question-answer system, a question in a first language and the question in a second language and predicting, by the question-answer system, a first answer to the question in the first language and a second answer to the question in the second language. The method also includes generating, by the question-answer system, a first vector representing the question in the first language and a second vector representing the question in the second language and adjusting the question-answer system based on the first and second answers and the first and second vectors such that when the question-answer system subsequently generates a third vector representing the question in the first language and a fourth vector representing the question in the second language, a distance between the third and fourth vectors is less than a distance between the first and second vectors.

Abstract: A method, computer program product, and/or computer system protects a question-answer dialog system from being attacked by adversarial statements that incorrectly answer a question. A computing device accesses a plurality of adversarial statements that are capable of making an adversarial attack on a question-answer dialog system, which is trained to provide a correct answer to a specific type of question. The computing device utilizes the plurality of adversarial statements to train a machine learning model for the question-answer dialog system. The computing device then reinforces the trained machine learning model by bootstrapping adversarial policies that identify multiple types of adversarial statements onto the trained machine learning model. The computing device then utilizes the trained and bootstrapped machine learning model to avoid adversarial attacks when responding to questions submitted to the question-answer dialog system.

Abstract: A computer system, product, and method are provided. The computer system includes an artificial intelligence (AI) platform operatively coupled to a processor. The AI platform includes tools in the form of a machine learning model (MLM) manager, a metric manager, and a training manager. The MLM manager accesses a plurality of pre-trained source MLMs, and inputs a plurality of data objects of a test dataset into each of the source MLMs. The test dataset includes the plurality of data objects associated with respective labels. For each source MLM, associated labels are generated from the inputted data objects and a similarity metric is calculated. The MLM manager selects a base MLM to be used for transfer learning from the plurality of source MLMs based upon the calculated similarity metric. The training manager trains the selected base MLM with a target dataset for the target domain.