GRAMMAR RULE FILTER MODEL LEARNING DEVICE, GRAMMAR RULE FILTER DEVICE, SYNTAX ANALYSIS DEVICE, AND PROGRAM

• Main Authors: HAYASHI KATSUHIKO, HIRAO TSUTOMU, KAMIGAITO HIDETAKA, OKUMURA MANABU
• Format: Patent
• Language: English; Japanese
• Published: 24.08.2017
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

PROBLEM TO BE SOLVED: To reduce grammar rules in order to fast perform syntax analysis of a sentence.SOLUTION: A rule filter model learning section 26 generates a simplified syntax tree yfor each syntax tree (y) of learning data, extracts a set Rof grammar rules r, and generates a conversion function of grammar rules. Based on a vector f(x, y) of a feature quantity extracted from a sentence (x) of the learning data and the simplified syntax tree y, a weight vector wis updated. Based on the set Rof the grammar rules r, the learnt weight vector wand the input sentence (x), regarding each of the grammar rules r, a rule filter section 30 converts the grammar rule rinto a non-converted grammar rule (r) in accordance with a calculated score and adds the converted grammar rule to a set R' of grammar rules (r).SELECTED DRAWING: Figure 3 【課題】文の構文解析を高速に行うために文法規則を減らすことができる。【解決手段】規則フィルターモデル学習部２６は、学習データの構文木ｙの各々について、簡単化された構文木ｙ￣を生成し、文法規則ｒ￣の集合Ｒ￣を抽出すると共に、文法規則の変換関数を生成する。また、学習データの文ｘと簡単化された構文木ｙ￣とから抽出される特徴量のベクトルｆ￣（ｘ，ｙ￣）に基づいて、重みベクトルｗ￣を更新する。規則フィルター部３０は、文法規則ｒ￣の集合Ｒ￣と、学習された重みベクトルｗ￣と、入力文ｘとに基づいて、文法規則ｒ￣の各々について、計算されるスコアに応じて、文法規則ｒ￣を、変換前の文法規則ｒに変換して、文法規則ｒの集合Ｒ´に追加する。【選択図】図３