GENERATION METHOD, LEARNING METHOD, GENERATION PROGRAM, AND GENERATION DEVICE

• Main Author: MAKINO TAKUYA
• Format: Patent
• Language: English; Japanese
• Published: 08.10.2020
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

To include an unknown word in a summary sentence by an appropriate expression.SOLUTION: A generation method includes: calculating, for each encoder time that corresponds to the word string of an input sentence, the hidden state of encoder time from an immediately preceding hidden state on the basis of the words of the input sentence and the label of named entity; inputting the hidden state outputted from the encoder to a decoder; calculating, for each decoder time that corresponds to the word string of a summary sentence outputted by the decoder, the hidden state of a decoder time from an immediately preceding hidden state on the basis of the words of a summary sentence generated immediately before and the label of unique expression; calculating a first probability distribution, on the basis of the hidden state of the decoder time and the hidden state of each encoder time, with which each word in the word string of the input sentence is copied as the words of the summary sentence at the decoder time; calculating a second probability distribution, on the basis of the hidden state of the decoder time, with which each word of a model dictionary is generated as the words of the summary sentence at the decoder time; and generating the words of an output summary sentence at the decode time on the basis of the first and second probability distributions.SELECTED DRAWING: Figure 8 【課題】未知語を適切な表現で要約文へ含める。【解決手段】生成方法は、入力文の単語列に対応するエンコーダ時刻毎に、入力文の単語及び固有表現のラベルに基づいて１つ前の隠れ状態からエンコーダ時刻の隠れ状態を算出し、エンコーダにより出力される隠れ状態をデコーダへ入力し、デコーダが出力する要約文の単語列に対応するデコーダ時刻毎に、１つ前に生成された要約文の単語及び固有表現のラベルに基づいて１つ前の隠れ状態からデコーダ時刻の隠れ状態を算出し、デコーダ時刻の隠れ状態と各エンコーダ時刻の隠れ状態とに基づいて入力文の単語列の各単語がデコーダ時刻における要約文の単語としてコピーされる第１確率分布を算出し、デコーダ時刻の隠れ状態に基づいて、モデル辞書の各単語がデコーダ時刻における要約文の単語として生成される第２確率分布を算出し、第１確率分布と第２確率分布に基づいてデコーダ時刻における出力要約文の単語を生成する。【選択図】図８