| CPC G06F 16/90335 (2019.01) [G06F 40/242 (2020.01); G06F 40/289 (2020.01); G06F 40/40 (2020.01); G06N 20/00 (2019.01)] | 17 Claims |
|
1. An information processing apparatus comprising:
a memory storing a learned model having learned correlation between character strings included in a search history of words and/or phrases; and
at least one processor inputting information of a search history of words and/or phrases by a target user to the learned model and creating a word and/or a phrase for learning by the target user,
wherein:
the learned model learns a time series of a plurality of character strings included in the search history;
the at least one processor inputs search histories by the target user to the learned model in a time-series manner to output a word and/or a phrase that is not included in the input search histories in a time-series manner;
the learned model is a recurrent neural network (RNN) learning correlation between character strings of the time series included in the search history;
the learned model includes an input layer, an intermediate layer, and an output layer;
the input layer converts a character string input as the search history into a vector quantity, and outputs a character string vector acquired by conversion to the intermediate layer;
the intermediate layer is an RNN block which: receives the character string vector from the input layer and receives a previous output of the RNN block of a previous time, inputs the character string vector from the input layer and the previous output of the RNN block to a preset function f, and outputs output of the function f as a current RNN output of a current time, whereby the intermediate layer outputs an RNN output reflecting time-series input change;
the output layer converts the RNN output from the intermediate layer into a value of distribution of appearance probability of a character based on the input character string;
the processor creates the word and/or phrase by combining characters based on the value of distribution of appearance probability of each character, determines whether the created word and/or phrase is included in dictionary content stored in the memory, and controls an output device to output the created word and/or phrase determined to be included in the dictionary content stored in the memory, wherein the processor executes creation of a word and/or phrase again if it is determined that the created word and/or phrase is not included in the dictionary content stored in the memory; and
in executing learning, the at least one processor (a) inputs to the learned model a partial character string of the character string based on the search history, the partial character string including a predetermined amount of characters, and (b) inputs a next character in the character string as teacher data, and (c) the model predicts the appearance probability for each supposed character string, and (d) regulates a parameter for the function f in the RNN block to decrease a difference between the prediction result and the teacher data, wherein learning is performed by repeatedly executing (a)-(d), wherein, for each subsequent execution of (a)-(d), the at least one processor inputs, as the partial character string, a partial character string obtained by shifting a range of the character string by a predetermined number of characters to obtain a new partial character string overlapping in part with the previous partial character string.
|