International Business Machines Corporation

David Contreras

Robert Christian Sizemore

Corville O Allen

Roberto Delima

Brendan Bull

Sterling Richardson Smith

Keith P Biegert

Krishna Mahajan

Aysu Ezen Can

Chris M Mwarabu

Kandhan Sekar

Andrew Ronald Freed

Brien H Muschett

Paul Lewis Felt

Bob Delima

Thomas Hay Rogers

An approach is provided that receives a document and a document type of the document. The document type identifies a document category to which the received document belongs. A set of linguistic metrics are retrieved that correspond to the document type. A quality of the received document is automatically determined based on a set of linguistic features found in the document as compared to the retrieved set of linguistic metrics. The document is then ingested into a corpus that is utilized by a question-answering (QA) system. The ingestion of the document is based on the determined quality.

Document type-specific quality model

A phrase that includes a trigger word that modifies a meaning within the phrase is received. The trigger word is identified. The words of the phrase that are modified by the trigger word are identified by analyzing features of the phrase that link the trigger word to other words. The phrase is interpreted by modifying the second subset of words according to the modification of the trigger word.

Contextual span framework

Provided are a computer program product, system, and method for classifying text to determine a goal type used to select machine learning algorithm outcomes. Natural language processing of text is performed to determine features in the text and their relationships. A classifier classifies the text based on the relationships and features to determine a goal type. The determined features and relationships from the text are inputted into a plurality of different machine learning algorithms to generate outcomes. For each of the machine learning algorithms, a determination is made of performance measurements resulting from the machine learning algorithms generating the outcomes. A determination is made of at least one machine learning algorithm having performance measurements that are highly correlated to the determined goal type. An outcome is determined from at least one of the outcomes.

Classifying text to determine a goal type used to select machine learning algorithm outcomes

A method, system, and computer program product include providing a list of triggers, training the natural language processor with the list of triggers, providing to the natural language processor a text including one trigger, selecting nodes in the text to create an original potential span, predicting whether the original potential span includes another trigger, and adjusting, in response to predicting that the original potential span includes another trigger, the original potential span to exclude the another trigger to create a new potential span.

Automatic detection of context switch triggers

Methods, systems and computer readable media are provided for semantic evaluation of tentative triggers based on contextual triggers. Contextual triggers are identified within text. A parse tree comprising a plurality of nodes is generated corresponding to the text. Tentative triggers are identified within the text. A determination is made as to whether one or more nodes of the parse tree corresponding to the tentative trigger is within a context of one or more nodes of the parse tree corresponding to the contextual triggers. Based on the determination, the tentative trigger type is assigned to a contextual trigger type.

Semantic evaluation of tentative triggers based on contextual triggers

Aspects of the present disclosure relate to identifying semantic relationships. Natural language content is received. A part of speech is determined for respective terms within the natural language content. A semantic type is determined for each of two or more terms within the natural language content. A parse tree representation containing a plurality of nodes is then generated based on the natural language content, each of the plurality of nodes corresponding to at least one term within the natural language content, wherein visual characteristics of respective nodes of the plurality of nodes within the parse tree representation depend on the part of speech and semantic type of the respective terms. A bounding box identifying a semantic relationship is then generated around a set of nodes on the parse tree representation, the set of nodes including the two or more terms.

Identifying semantic relationships using visual recognition

Aspects of the present disclosure relate to identifying spans within unstructured electronic text. Natural language content is received. A part of speech and slot name of each word within the natural language content is identified. A parse tree representation is then generated based on the natural language content, wherein visual characteristics of each node of a plurality of nodes within the parse tree representation depend on the part of speech and slot name of each word. A bounding box identifying a span category is then generated around a set of nodes on the parse tree representation by a machine learning model.

Identifying spans using visual recognition

A phrase may be received that includes a plurality of tokens in a natural language format. A plurality of levels relating to dependencies between tokens of the plurality of tokens within the phrase is determined. A matrix structure is generated for the phrase. The matrix structure utilizes a plurality of rows and a plurality of columns to store data of the phrase. The plurality of rows and the plurality of columns each indicate one of an order of tokens of the plurality of tokens or levels of the plurality of levels.

Natural language processing matrices

A method, computer system, and a computer program product for relevancy-based document quality assessment is provided. The present invention may include computing a document quality score based on at least one container relevancy score determined based on at least one domain link to a domain knowledge base.

Relevancy as an indicator for determining document quality

First content containing a plurality of list items in one or more lists can be parsed for conjunctions and implied list indicators. One or more modifications can occur at one or more conjunctions or implied list indicators. The one or more modifications can comprise one or more of expanding text, contracting text, and replacing text. The one or more modifications can generate second content conducive to natural language processing operations.

List manipulation in natural language processing

Techniques for cognitive document quality determination and automated heuristic generation are provided. A plurality of documents is received, where each of the plurality of documents contains natural language text. A plurality of values is determined for a first plurality of predefined attributes of the plurality of documents. A plurality of quality scores is generated for the plurality of documents by processing the plurality of values using a machine learning model, where the plurality of quality scores indicate a suitability of each of the plurality of documents to be processed using a target processing operation. A subset of documents is identified from the plurality of documents having respective quality scores below a predefined threshold. The subset of documents is flagged for further processing. At least one document of the plurality of documents that is not flagged is selectively processed using the target processing operation.

Cognitive document quality determination with automated heuristic generation

Performing an operation comprising determining that a parse of an input string comprising a plurality of tokens is incomplete, generating, based on a machine learning (ML) model: (i) a plurality of candidate addition tokens for adding to the input string, and (ii) a plurality of candidate removal tokens for removing from the input string, selecting, from the plurality of candidate addition tokens and the plurality of candidate removal tokens, a first candidate token, and modifying the input string based on the first candidate token to facilitate a complete parse of the modified input string by a parser.

Deep learning approach to grammatical correction for incomplete parses

A mechanism is provided for improving predicate parses (or logical representations of a passage) using semantic knowledge. In response to encountering an ambiguous decision point during a syntactic analysis of a portion of natural language content, a candidate meaning of the ambiguous decision point is generated. Characteristics of the ambiguous decision point are evaluated based on a semantic knowledge base to determine a semantic meaning associated with the ambiguous decision point. A determination is made as to whether the semantic meaning supports or refutes the candidate meaning. In response to determining that the semantic meaning refutes the candidate meaning, the candidate meaning of the ambiguous decision point is overridden based on the semantic meaning to include the semantic meaning as a final meaning for the ambiguous decision point. The portion of natural language content is then processed based on the final meaning for the ambiguous decision point.

Predicate parses using semantic knowledge

A technique of improving input text processing in a data processing system that answers questions includes receiving a question from a user. The question has associated supporting original input text that is utilized in answering the question. Corrective measures are applied to the supporting original input text to create supporting revised input text. The supporting revised input text documents corrective measures applied to the supporting original input text. An answer to the question is generated based on, at least in part, the supporting revised input text. The supporting original input text is recreated from the supporting revised input text based on the documented corrective measures in the supporting revised input text subsequent to generating the answer. The answer and the supporting original input text as recreated are transmitted to the user.

Growing community of inventors

Willow Spring, NC, United States of America

David Contreras