Fujitsu Corporation

Tomoya Iwakura

Seishi Okamoto

Takuya Makino

Junko Furukawa

Akiba Miura

Taiki Watanabe

An information processing method in which a computer executes processing includes: acquiring a sentence; specifying a word that appears immediately before or immediately after a first word in the acquired sentence by using a prediction model that predicts a word that appears immediately before or immediately after an input word; determining whether or not an estimated relationship between the first word and the second word in the sentence is appropriate on the basis of the specified word and a rule regarding a unit that corresponds to a relationship between words stored in a storage; and outputting information regarding the estimated relationship between the first word and the second word in a case where it is determined that the relationship is appropriate.

Information processing method, computer-readable recording medium storing information processing program, and information processing device

A learning method to be executed by a computer, the learning method includes when a first input sentence in which a predetermined target is represented by a first named entity is input to a first machine learning model, learning a first parameter of the first machine learning model such that a value output from the first machine learning model approaches correct answer information corresponding to the first input sentence; and when an intermediate representation generated when the first input sentence is input to the first machine learning model and a second input sentence in which the predetermined target is represented by a second named entity are input to a second machine learning model, learning the first parameter and a second parameter of the second machine learning model such that a value output from the second machine learning model approaches correct answer information corresponding to the second input sentence.

Learning method using machine learning to generate correct sentences, extraction method, and information processing apparatus

A translation method, implemented by a computer, includes: converting a text written in a first language into a replacement text in which a named entity in the text is replaced with a predetermined character string; translating the replacement text into a second language by using a text translation model which is a neural network; and translating a named entity corresponding to the predetermined character string in the replacement text into the second language by using a named entity translation model which is a neural network.

Translation method, learning method, and non-transitory computer-readable storage medium for storing translation program to translate a named entity based on an attention score using neural network

A computer-implemented summary generation method includes obtaining input text; generating an initial lattice including serially coupled nodes corresponding to words within the input text; generating a node corresponding to an expression within the initial lattice; adding the generated node to the initial lattice to provide an extended lattice corresponding to the input text; calculating a generation probability of each word within the input text using a dictionary and a machine learning model (model); calculating a generation probability for each node included in the extended lattice based on a hidden state output by a cell corresponding to the node among cells in an encoder of the model and a hidden state updated by a cell in a decoder of the model; and generating an element of a summary of the input text based on the generation probability of each word and the generation probability of each node of the extended lattice.

Summary generation method and summary generation apparatus

An apparatus acquires learning-data, including feature-elements, to which a label is assigned. The apparatus generates a first-set of expanded feature-elements by expanding the feature-elements. With reference to a model where a confidence value is stored in association with each of a second-set of expanded feature-elements, the apparatus updates confidence values associated with expanded feature-elements common between the first- and second-sets of expanded feature-elements, based on the label. Upon occurrence of an error indicating that a score calculated from the updated confidence values is inconsistent with the label, the apparatus sets a feature-size indicating a maximum size of expanded feature-elements to be used to update the model, based on the number of occurrences of the error for the acquired learning-data, and updates the model by adding, out of expanded feature-elements generated according to the set feature-size, expanded feature-elements unmatched with the second-set of expanded feature-elements, to the model.

Efficient updating of a model used for data learning

A present method includes first updating, based on a weight of each training sample (TS), a first score for each of features, which is a cue when extracting a correct structure from each TS, to calculate a model defined by first scores; performing, for each TS, a processing including identifying a maximum score among second scores, each of which is assigned, by the model, to either of candidate structures other than the correct structure among candidate structures derived from the TS; and first calculating a difference between the identified maximum score and a second score assigned by the model to the correct structure; and second calculating a confidence degree based on an upper limit value of errors, which is defined by the differences; second updating the weight of each TS based on the confidence degree and the differences; and repeating the first updating, performing, second calculating and second updating.

Model learning method

A rule learning method in machine learning includes distributing features to a given number of buckets based on a weight of the features which are correlated with a training example; specifying a feature with a maximum gain value as a rule based on a weight of the training example from each of the buckets; calculating a confidence value of the specified rule based on the weight of the training example; storing the specified rule and the confidence value in a rule data storage unit; updating the weights of the training examples based on the specified rule, the confidence value of the specified rule, data of the training example, and the weight of the training example; and repeating the distributing, the specifying, the calculating, the storing, and the updating, when the rule and the confidence value are to be further generated.

Rule learning method, program, and device selecting rule for updating weights based on confidence value

A rule learning method for making a computer perform rule learning processing in machine learning includes firstly calculating an evaluation value of respective features in a training example by using data and weights of the training examples; selecting a given number of features in descending order of the evaluation values; secondly calculating a confidence value for one of the given number of selected features; updating the weights of training example, by using the data and weights of the training examples, and the confidence value corresponding to the one feature; firstly repeating the updating for the remaining features of the given number of features; and secondly repeating, for a given number of times, the firstly calculating, the selecting, the secondly calculating, the updating, and the firstly repeating.

Rule learning by updating training example and a remaining feature weights

This invention provides a technique to correctly inform the human being of content of contents to be published but to prevent machines from collecting part of the contents whose distribution is not desired by the information provider. This invention includes: reading out contents data to be published, which includes text data, and identifying a character string whose output as the text data should be avoided from the contents data; converting the identified character string into substitution data other than the text data so as to maintain content of the identified character string; and generating publication contents data to maintain publication content of the contents data by using data other than the identified character string in the contents data and the substitution data. Thus, by carrying out such a processing, it becomes possible to conceal the character string against machines without changing the publication content for the human being.

Growing community of inventors

Kawasaki, Japan

Tomoya Iwakura