構文解析情報を用いたテキストからの数値情報の抽出

(1)

- 1 -

ᵓᩥゎᯒ᝟ሗ䜢⏝䛔䛯䝔䜻䝇䝖䛛䜙䛾ᩘ್᝟ሗ䛾ᢳฟ

Development for Extracting Numerical Information Using Parsing Information

㯮ᅵ㻌 ೺୕

*1

_{᳃ᮏ㻌 ᗣႹ}

*1

_{బ⸨㻌 ⨾Ἃ}

*1

_{ᰗ஭㻌 Ꮥ௓}

*1 Kenzo Kurotsuchi Yasutsugu Morimoto Misa Sato Kohsuke Yanai

*1㻌

_{᪥❧〇సᡤ ◊✲㛤Ⓨ䜾䝹䞊䝥}

Research & Development Group, Hitachi Ltd

For decision support, we developed text mining method to extract information about physical property values related to the decision from a large amount of accumulated technical papers. In numerical expressions, we think that the usefulness will be enhanced by not treating only numerical values, but extracting them from the document by combining the numerical values related to which events (Attribute Value Extraction). We reported a method for extracting pairs of item names and numerical values using StruAP. As a result of the evaluation, we confirmed the effectiveness of this method.

1. 䛿䛨䜑䛻

ㄽᩥ䛺䛹䛾䝔䜻䝇䝖䛜䜴䜵䝤䜔㟁Ꮚⓗ䛺፹య䜢㏻䛨䛶኱㔞䛻 ⵳✚䛥䜜䛶䛔䜛䚹䛣䛾⭾኱䛺䝔䜻䝇䝖᝟ሗ䜢ά⏝䛩䜛ᵝ䚻䛺◊ ✲䛜⾜䜟䜜䛶䛔䜛䚹ே䛷䛿ㄞ䜏ษ䜜䛺䛔኱㔞䛾᝟ሗ䛾୰䛛䜙 ᡤᮃ䛾᝟ሗ䜢ᣦᐃ䛧䛯ᙧᘧ䛷ᢳฟ䛩䜛᝟ሗᢳฟ䠄Information Extraction䠅䛾᳨ウ䜒㐍䜣䛷䛔䜛䚹 ㄽᩥ䛛䜙䛾᝟ሗᢳฟ䛻㛵䛧䛶䠈ᢏ⾡ᡓ␎䜢❧᱌䛩䜛䛸䛝䛻䛭䛾ᢏ⾡ศ㔝䛻䛚䛡䜛㐣ཤ䛾ᢏ⾡䛾䝖䝺䞁䝗䜢୍ぴ໬䛧䠈ᡓ␎❧ ᱌䜢ᨭ᥼䛩䜛ヨ䜏䛜᳨ウ䛥䜜䛶䛔䜛[1]䚹⏕࿨⛉Ꮫศ㔝䛷䛿ㄽ ᩥ䛾᝟ሗ䜢ᇶ䛻ᐇ㦂┠ⓗ䜔⏝䛔䛯ᮦᩱ䠈ᐇ㦂ᡭἲ䜢ᢳฟᩚ⌮ 䛩䜛䛣䛸䛷ᐇ㦂⤖ᯝ䜢᳨⣴䛷䛝䜛䜘䛖䛻䛩䜛ྲྀ䜚⤌䜏䛜᳨ウ䛥䜜 [2]䠈་⸆ศ㔝䛷䛿㑇ఏᏊ䞉໬ྜ≀䞉⑌ᝈ䛾㛫䛾㛵ಀ䜢ศᯒ䛩䜛䛣䛸䛜᳨ウ䛥䜜䛶䛔䜛[3]䚹᝟ሗᢳฟ䛾ᑐ㇟䛸䛺䜛᝟ሗ䛿◊✲ ┠ⓗ䜔ᐇ㦂ᡭἲ䠈ᩘ್᝟ሗ䛺䛹ከᒱ䛻䜟䛯䜚䠈䛭䜜䛮䜜㔜せ䛷䛒䜛䚹ᩘ್᝟ሗ䛻䛴䛔䛶୍ᐃ䛾ᢏ⾡䝖䝺䞁䝗䛜䛒䜛◊✲ศ㔝䛷䛿≉䛻㔜せ䛷䛒䜛䛸➹⪅䛿⪃䛘䜛䚹౛䛘䜀༙ᑟయศ㔝䛷䛿 ༙ᑟయ䛾㞟✚ᗘ䛜18 䞄᭶䛤䛸䛻 2 ಸ䛻䛺䜛 Moore's law[4]䜔䠈 NAND 䝣䝷䝑䝅䝳䝯䝰䝸䛾ᐜ㔞䛜䠍ᖺ䛤䛸䛻 2 ಸ䛻䛺䜛 Hwang's Law[5]䛾䜘䛖䛺ᢏ⾡䝖䝺䞁䝗䛜Ꮡᅾ䛩䜛䚹༙ᑟయ䛾㞟✚ᗘ䛾䜘䛖䛺ᩘ್᝟ሗ䛿≀⌮ἲ๎䜢௓䛧䛶䛭䛾௚䛾ᵝ䚻䛺ᩘ್᝟ሗ䛻ᙳ 㡪䜢୚䛘䛶䛚䜚䠈౛䛘䜀Moore's law 䛿༙ᑟయ⣲Ꮚ䛾䝀䞊䝖㛗䠈䝀䞊䝖ᖜ䠈᥋ྜ῝䛥䠈䝀䞊䝖㠃✚䠈䝀䞊䝖㓟໬⭷ཌ䠈୙⣧≀⃰ᗘ䠈㟁※㟁ᅽ䠈㟁⏺ᙉᗘ䠈ᾘ㈝㟁ὶ䠈䝀䞊䝖㟁Ẽᐜ㔞䠈䝀䞊䝖㐜ᘏ ᫬㛫䠈ᾘ㈝㟁ຊ䠈㟁ຊᐦᗘ䛻㛵䛩䜛ᢏ⾡䝖䝺䞁䝗䜢ண 䛩䜛䛾䛻ᙺ❧䛶䜙䜜䛶䛔䜛䚹䛣䛾䜘䛖䛺✀䚻䛾ᩘ್᝟ሗ䜢ㄽᩥ䛛䜙⥙ ⨶ⓗ䛻ᢳฟ䛧䛶ᩚ⌮䛩䜛䛣䛸䛷䝖䝺䞁䝗䜢ᢕᥱ䛩䜛䛣䛸䛿᭷┈䛰䛸➹⪅䛿⪃䛘䜛䚹䜎䛯䝔䜻䝇䝖䛛䜙䛾ᩘ್᝟ሗ䛾ᢳฟ䛻㛵䛧䛶䠈ᖜᗈ䛟◊✲䛜⾜䜟䜜䛶䛔䜛[6-10]䚹༢䛻ᩘ್䛰䛡䜢ᢳฟ䛩䜛䛾䛷䛿䛺䛟䠈䛹䛾 ஦㇟䠄௨ୗ䠈㡯┠ྡ䛸䛩䜛䠅䛻㛵䛩䜛ᩘ್䛷䛒䜛䛛䜢⤌䜏䛻䛧䛶 ᩥ᭩୰䛛䜙ᢳฟ䛩䜛䛣䛸(Attribute Value Extraction)䛷᭷┈ᛶ䛜㧗䜎䜛䛸⪃䛘䜙䜜䜛 [7]䚹䜎䛯ᩘ್䛿༢఩䜢క䛖䜒䛾䠈౛䛘䜀㡯┠ ྡ䛂㟁Ẽ᢬ᢠ䛃䛻ᑐᛂ䛩䜛䛂10 M䃈䛃䛸䠈༢఩䜢క䜟䛺䛔䜒䛾䠈౛ 䛘䜀㡯┠ྡ䛂ㄏ㟁⋡䛃䛻ᑐᛂ䛩䜛䛂4.4䛃䛻ศ䛡䜙䜜䜛䚹༢఩䜢క 䛖ᩘ್䛿༢఩䜢ᡭ᥃䛛䜚䛸䛧䛶㡯┠ྡ䜢᥎ᐃ䛩䜛䛣䛸䛜ྍ⬟䛷䛒䜛䛾䛻ᑐ䛧䠈༢఩䜢క䜟䛺䛔ᩘ್䛿㡯┠ྡ䛾᥎ᐃ䛜ẚ㍑ⓗ 㞴䛧䛔䛸⪃䛘䜙䜜䜛䚹 ᮏሗ࿌䛷䛿ᢏ⾡ືྥ䛻㛵䛩䜛ุ᩿ᨭ᥼䜢⾜䛖䝅䝇䝔䝮䜢┠ ᶆ䛸䛧䠈ㄽᩥ䛛䜙≀ᛶ್䛻㛵䛩䜛᝟ሗ䠈౛䛘䜀䛂㡯┠ྡ䠖ㄏ㟁⋡䠈 ᩘ್䠖䠐䠊䠐䛃䛺䛹䛾ᢳฟ䛻ྲྀ䜚⤌䜐䚹≀ᛶ್䛾ᢳฟ䛻䛒䛯䛳䛶䛿 ᩘ್䠄౛䛘䜀䠈䠐䠊䠐䠅䛸㡯┠ྡ䠄౛䛘䜀䠈ㄏ㟁⋡䠅䜢⤌䜏䛻䛧䛶ᢳ ฟ䛩䜛䚹ᩘ್䛻䛴䛔䛶䠈༢఩䜢క䜟䛺䛔ᩘ್䛾ᢳฟ䜢᳨ウ䛩䜛䚹䛣䜜䛻䜘䜚౛䛘䜀ᐇ㦂䜢⾜䛖䛸䛝䛻㐣ཤ䛾ᩥ⊩䜢ㄪ䜉䜛䛣䛸䛷䜎䛰ᐇ㦂䛥䜜䛶䛔䛺䛔᮲௳䜢ㄪ䜉䛯䜚䠈ᖺḟ䛤䛸䛻ᛶ⬟್䜢ᩚ⌮ 䛩䜛䛣䛸䛷ḟ䛻䜑䛦䛩䜉䛝ᛶ⬟್䜢᥎ᐃ䛧䛯䜚䛩䜛䛣䛸䛷◊✲㛤 Ⓨάື䜢ᨭ᥼䛩䜛ྲྀ䜚⤌䜏䜢䜑䛦䛩䚹

2. 㛵㐃◊✲

ᩘ್䛸㡯┠ྡ䜢⤌䜏䛻䛧䛯᝟ሗ䛾ᢳฟ䛻㛵䛧䛶䠈ᢳฟᑐ㇟䛾㡯┠ྡ䛸ᩘ್䠈䛭䛾㛵ಀ䜢ὀ㔘᝟ሗ䛸䛧䛶⏝䛔䜛ᡭἲ䛜ᥦ᱌ 䛥䜜䛶䛔䜛[9]䚹ὀ㔘᝟ሗ䜢ᶵᲔᏛ⩦䜔⤫ィⓗ᧯స䛻䜘䜚ฎ⌮䛧䛶ᢳฟ䝰䝕䝹䜔ᢳฟつ๎䜢సᡂ䛩䜛䛣䛸䛜ከ䛔䠄௨ୗ䠈ᶵᲔᏛ ⩦䝧䞊䝇䛾ᡭἲ䛸䛩䜛䠅䚹⸨⏿䜙䛿᪂⪺グ஦䜢᝟ሗᢳฟ䛾ᑐ㇟䛸䛧䠈ಀ䜚ཷ䛡䛾㛵ಀ䛸᱁ຓモ䜢ᡭ᥃䛛䜚䛸䛧䛶᪂⪺グ஦䛛䜙ᩘ ್᝟ሗ䜢ᢳฟ䛧䛯[7]䚹ᩘ್䜢ྵ䜐ᩥ⠇䜢᳨⣴䛧䠈䛭䛾ᩥ⠇䛜 ಀ䜛⏝ゝ䛾࿘㎶䛾ྡモ䜢≀䠄ᮏሗ࿌䛻䛚䛡䜛㡯┠ྡ䜢ྵ䜐ᴫ ᛕ䠅䛾ೃ⿵䛸䛧䠈ᐇ㝿䛾ᩥ᭩䛻䛚䛡䜛ṇゎ⋡䛸ฟ⌧㢖ᗘ䜢 200 ௳⛬ᗘ䛾᪂⪺グ஦䜢⏝䛔䛶⤫ィฎ⌮䛩䜛䛣䛸䛷సᡂ䛧䛯ᢳฟ つ๎䜢≀ೃ⿵䛻㐺⏝䛧䛶≀䜢ᢳฟ䛧䛯䚹䜎䛯 Ghani 䜙䛿䜰䝟䝺䝹䠄⾰㢮䠅ᑠ኎䜚䛾䜴䜵䝤䝃䜲䝖䛾䝕䞊䝍䜢᝟ሗᢳฟ䛾ᑐ㇟䛸䛧䠈 ༙ᩍᖌ௜䛝Ꮫ⩦ἲ䜢⏝䛔䛶〇ရ᝟ሗ䜢ᢳฟ䛧䛯[6]䚹⣙ 600 ௳ 䛾〇ရ᝟ሗ䜢カ⦎䝕䞊䝍䛸䛧䛶⏝䛔䛶䛔䜛䚹ᶵᲔᏛ⩦䝧䞊䝇䛾 ᡭἲ䛿ẚ㍑ⓗከᩘ䛾カ⦎䝕䞊䝍䜔⤫ィฎ⌮䛾ᑐ㇟䛸䛺䜛䝕䞊䝍䜢⏝䛔䜛䛣䛸䛜ከ䛔䚹 ୍᪉䠈ᩥ⊩[7]䛾䝧䞊䝇䝷䜲䞁ᡭἲ䛿ᩘ್䛛䜙䛾㊥㞳ᩥ⠇ᩘ 䛜᭱䜒㏆䛔ྡモ䜢≀䠄ᮏሗ࿌䛻䛚䛡䜛㡯┠ྡ䜢ྵ䜐ᴫᛕ䠅䛸䛩䜛ᡭἲ䛷䛒䜛䚹䛣䛾ᡭἲ䛿≉䛻カ⦎䝕䞊䝍䜢ᚲせ䛸䛧䛺䛔䛜䠈 ከ䛟䛾ᩘ್䜢ྵ䜐ㄽᩥ䛷䛿⢭ᗘ䛾పୗ䛜ᠱᛕ䛥䜜䜛䚹 ㄽᩥ䛛䜙䛾᝟ሗ䛾ᢳฟ䛻㛵䛧䛶䠈ᕝᮏ䜙䛿ᐇ㦂ᡭἲ䛻㛵䛩䜛⏝ㄒ㎡᭩䛾ᩚഛ䜢⾜䛳䛶䛔䜛[2]䚹䜎䛯䠈Torres 䜙䛿䝋䝣䝖䜴䜵䜰ᕤᏛ䛻㛵䛧䛶䠈┠ⓗ䜔Ⓨぢ䛧䛯䛣䛸䠈௒ᚋ䛾ィ⏬䛾グ㏙⟠ᡤ 䜢ᢳฟ䛩䜛ヨ䜏䜢⾜䛳䛶䛔䜛[11]䚹䛥䜙䛻 Peng 䜙䛿 Conditional Random Fields 䜢⏝䛔䜛䛣䛸䛷ㄽᩥ䛛䜙䝍䜲䝖䝹䠈せ᪨䠈ཧ⪃ᩥ ⊩䛺䛹䜢ᢳฟ䛧䛶䛔䜛[12]䚹 ᡃ䚻䛜ㄪᰝ䛧䛯⠊ᅖ䛷䛿ㄽᩥ䛛䜙䛾᝟ሗᢳฟ䛾◊✲䛿᪂⪺ グ஦䜔䠳䡁䠾䛛䜙䛾᝟ሗᢳฟ䛸ẚ㍑䛧䛶ඛ⾜◊✲䛿ᑡ䛺䛛䛳䛯䚹㐃⤡ඛ䠖㯮ᅵ೺୕䠈᪥❧〇సᡤ◊✲㛤Ⓨ䜾䝹䞊䝥㻌㻌㻌㻌㼗㼑㼚㼦㼛㻚㼗㼡㼞㼛㼠㼟㼡㼏㼔㼕㻚㼝㼟㻬㼔㼕㼠㼍㼏㼔㼕㻚㼏㼛㼙㻌

The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020

(2)

- 2 -

3. ᥦ᱌ᡭἲ

ᡃ䚻䛿ㄽᩥ䛛䜙≀ᛶ್䛻㛵䛩䜛ᩘ್䛸㡯┠ྡ䜢ᢳฟ䛩䜛䛣䛸䛻ྲྀ䜚⤌䜣䛰䚹 ㄽᩥ䛸䛧䛶ከ䛟䛾ᩘ್䜢ྵ䜐ᛂ⏝≀⌮䛾ㄽᩥ䜢ᑐ㇟䛸䛧䛯䚹䛣䛾䜘䛖䛺ㄽᩥ䛛䜙㞴᫆ᗘ䛜㧗䛔䛸⪃䛘䜙䜜䜛༢఩䜢క䜟䛺䛔 ᩘ್䜢ᢳฟ䛩䜛䛯䜑䛻䛿⢭ᗘ䛾㧗䛔ᢳฟ᪉ἲ䛜ᚲせ䛷䛒䜛䚹 ㄽᩥ䛿䝤䝻䜾䛺䛹䛸ẚ㍑䛩䜛䛸䛟䛰䛡䛯⾲⌧䛜↓䛟䠈䝔䜻䝇䝖䛾 ရ㉁䛿㧗䛔䛸⪃䛘䜙䜜䜛୍᪉䠈᪂⪺グ஦䛸ẚ㍑䛩䜛䛸ᇳ➹⪅䛜 ⇍⦎䛧䛶䛔䛺䛔䛣䛸䛜䛒䜚䠈⾲⌧䛾ᦂ䜙䛞䛜ぢ䜙䜜䜛䚹౛䛘䜀䠈 ᩘ್䛸༢఩䛸䛾㛫䛻✵ⓑ䜢㛤䛡䜛䛣䛸䛜ከ䛔䛜䠈୍㒊䛾ᩥ䛷䛿䛭䛾✵ⓑ䛜↓䛔䚹䜎䛯 PDF ᙧᘧ䛾ㄽᩥ䜒ከ䛔䚹PDF ᙧᘧ䛾䝣䜯䜲䝹䛛䜙䝔䜻䝇䝖䜢ᢳฟ䛩䜛䛸䛂䛃䛾䜘䛖䛺䝜䜲䝈䛜ΰධ 䛧䠈ṇ䛧䛟䝔䜻䝇䝖䛜ᢳฟ䛷䛝䛺䛔䛣䛸䜒䛒䜛䚹ᩘ್䜢ከ䛟ྵ䜐ㄽ ᩥ䜔ᚲ䛪䛧䜒㧗䛟䛺䛔䝔䜻䝇䝖䛾ရ㉁䛻䛚䛔䛶䠈᝟ሗ䛾ᢳฟ䛜䛹䛾⛬ᗘ䛾⢭ᗘ䛷ྍ⬟䛺䛾䛛䜢ᮏሗ࿌䛷᳨ウ䛩䜛䚹 ᩘ್䛸㡯┠ྡ䛾⤌䜏䜢ᢳฟ䛩䜛ᡭἲ䛻䛴䛔䛶䠈ᮏሗ࿌䛷䛿 ᵓᩥ᝟ሗ䜢ྵ䜐䝹䞊䝹䜢సᡂ䛩䜛ᡭἲ䠄௨ୗ䠈䝹䞊䝹䝧䞊䝇 ᡭἲ䛸䛩䜛䠅䜢⏝䛔䛯䚹䛣䛾⌮⏤䛿䠈䠍䠊ุ᩿ᨭ᥼䝅䝇䝔䝮䛾฼ ⏝⪅䛻᝟ሗᢳฟ䛾ᇶ‽䜢ẚ㍑ⓗㄝ᫂䛧䜔䛩䛔䛯䜑䠈䠎䠊䝹䞊䝹䜢సᡂ䛩䜛䛯䜑䛻ཧ⪃䛻䛩䜛ᩥ䛾ᩘ䛜䠈ᶵᲔᏛ⩦䝧䞊䝇䛾ᡭ ἲ䛷カ⦎ᩥ䛸䛧䛶⏝䛔䜛ᩥ䛾ᩘ䜘䜚ᑡ䛺䛟䛶Ⰻ䛔ྍ⬟ᛶ䛜䛒䜛䛯䜑䠈䠏䠊䝹䞊䝹సᡂ䛻⇍⦎䛧䛯సᴗ⪅䜢☜ಖ䛩䜛䛣䛸䛜䛷䛝䛯䛯䜑䠈䛷䛒䜛䚹䠎䛻䛴䛔䛶䛿సᴗᕤᩘ䛾▷⦰䜢ᮇᚅ䛧䛯䚹 ᶵᲔᏛ⩦䝧䞊䝇䛾ᡭἲ䛸ẚ㍑䛧䛯䝹䞊䝹䝧䞊䝇䛾ᡭἲ䛾฼ Ⅼ䛿ୗグ䛷䛒䜛䚹 x 䝹䞊䝹䜢సᡂ䛩䜛䛯䜑䛻ཧ⪃䛻䛩䜛ᩥ䛾ᩘ䛜䠈ᶵᲔᏛ⩦ 䛷カ⦎ᩥ䛸䛧䛶⏝䛔䜛ᩥ䛾ᩘ䜘䜚䠈ᑡ䛺䛟䛺䜛䛣䛸䛜ከ䛔䚹 x 䝹䞊䝹䛻㐺ྜ䛩䜛ሙྜ䛿☜ᐇ䛻ᢳฟ䛷䛝䜛䛯䜑䠈ᢳฟ䛻 ኻᩋ䛧䛯ሙྜ䛾ཎᅉ䛾ゎᯒ䛜ᶵᲔᏛ⩦䝧䞊䝇䛾ᡭἲ䛸 ẚ㍑䛧䛶ᐜ᫆䛷䛒䜛䚹 ୍᪉䠈ᶵᲔᏛ⩦䝧䞊䝇䛾ᡭἲ䛾฼Ⅼ䛿ୗグ䛷䛒䜛䚹 x ὀ㔘᝟ሗ䛾సᡂసᴗ䛻䛚䛔䛶䠈䝹䞊䝹䝧䞊䝇䛾ᡭἲ䛻 ẚ䜉䛶ᶵᲔᏛ⩦䝧䞊䝇䛾ᡭἲ䛿సᴗ⪅䛻せồ䛥䜜䜛஦ ๓▱㆑䛜ᑡ䛺䛔䛣䛸䚹䝹䞊䝹䝧䞊䝇䛾ᡭἲ䛷䛿సᴗ⪅䛻 ᩥἲⓗ䛺▱㆑䛜せồ䛥䜜䜛䛣䛸䛜ከ䛔䚹 ุ᩿ᨭ᥼䝅䝇䝔䝮䛾┠ⓗ䛻ᛂ䛨䛶䝹䞊䝹䝧䞊䝇䛾ᡭἲ䠈䜒䛧䛟䛿ᶵᲔᏛ⩦䝧䞊䝇䛾ᡭἲ䠈䛭䜜䜙䜢⤌䜏ྜ䜟䛫䛯ᡭἲ䛾୰ 䛛䜙㑅ᢥ䛩䜛䛣䛸䛻䛺䜛䚹 ᅗ䠍䛻௒ᅇ᳨ウ䛧䛯䝅䝇䝔䝮䛾ᵓᡂᅗ䜢♧䛩䚹ㄽᩥ䛾䝔䜻䝇䝖䛻ᑐ䛧䛶ಀ䜚ཷ䛡ゎᯒ䜢⾜䛖䛣䛸䛷ಀ䜚ཷ䛡ᮌ䜢⏕ᡂ䛧䠈䝹䞊䝹䜶䞁䝆䞁䜢⏝䛔䛶ᢳฟ䝹䞊䝹䜢㐺⏝䛩䜛䛣䛸䛻䜘䜚䠈㡯┠ྡ䛸 ᩘ್䜢⤌䜏䛻䛧䛶ᢳฟ䛧䠈ㄽᩥྡ䛸ྜ䜟䛫䛶䝕䞊䝍䝧䞊䝇䛻᱁ ⣡䛩䜛䚹ᢳฟ䝹䞊䝹䛸ᢳฟᑐ㇟䛸䛺䜛㡯┠ྡ䛿ேᡭ䛷సᡂ䛧䛯䚹䝕䞊䝍䝧䞊䝇䛻᱁⣡䛥䜜䜛᝟ሗ䛾౛䜢⾲䠍䛻♧䛩䚹

᝟ሗᢳฟ䛾䝒䞊䝹䛸䛧䛶䠈 StruAP 䠄Structure-based Abstract Pattern䠅 [13]䜢⏝䛔䛯䚹StruAP 䛿ಀ䜚ཷ䛡ᵓ㐀䛻ᇶ䛵䛟ᢳฟ䝹䞊䝹䜢సᡂ䛩䜛䛣䛸䛷ᢳฟᡭ㡰䜢ᐃ⩏䛧䠈ᢳฟ䝹䞊䝹䛸ಀ䜚ཷ 䛡ᮌ䛾ᮌᵓ㐀䛾䝟䝍䞊䞁䝬䝑䝏䜢㏻䛨䛶㛵ಀᢳฟ䜢⏝䛔䛯᝟ ሗᢳฟ䜢⾜䛖䛣䛸䛜䛷䛝䜛䚹௒ᅇ䛿㡯┠ྡ䛸ᩘ್䛾㛵ಀ䛻ᑐ䛩䜛ᢳฟ䜢⾜䛳䛯䚹ㄽᩥ䛻䛿ᵝ䚻䛺ほⅬ䛛䜙ศᯒྍ⬟䛺᝟ሗ䛜 ྵ䜎䜜䛶䛔䜛䚹᝟ሗᢳฟ䛾✀㢮䛤䛸䛻ಶู䛾ᢳฟ䝅䝇䝔䝮䜢ᵓ ᡂ䛩䜛ᡭἲ䛻ẚ䜉䛶䠈StruAP 䛸䛔䛖୍䛴䛾䝒䞊䝹䜢฼⏝䛧䛶 ✀䚻䛾᝟ሗ䜢ᢳฟ䛩䜛䛣䛸䛷䝅䝇䝔䝮䛾㛤Ⓨᕤᩘ䜢๐ῶ䛩䜛䛣䛸䜢䜑䛦䛩䚹

4. ᐇ㦂

4.1 ᐇ㦂᮲௳ ௒ᅇ౑⏝䛧䛯䝒䞊䝹䜢⾲2 䛻♧䛩䚹ㄽᩥ䛿 PDF ᙧᘧ䛾ㄽᩥ 䜢⏝䛔䛯䚹PDF 䝣䜯䜲䝹䛻ྵ䜎䜜䜛䝔䜻䝇䝖䜢ᢳฟ䛧䠈 Standord CoreNLP 䜢 ⏝ 䛔䛶 ᩥ ศ ๭ 䜢⾜䛳䛯䚹䛭䛾 ᚋ 䠈 ᵓ ᩥ ゎ ᯒ 䜢 Standord CoreNLP 䜢⏝䛔䛶⾜䛳䛯䚹䛥䜙䛻 StruAP 䜢⏝䛔䛶ᩘ ್䛸㡯┠ྡ䛾⤌䜏䛾ᢳฟ䜢⾜䛳䛯䚹

ᑐ㇟䛸䛧䛯ㄽᩥ䛿䠈ᛂ⏝≀⌮䛷䛒䜛༙ᑟయ䝟䝑䜿䞊䝆䞁䜾ᢏ ⾡䛻㛵䛩䜛 ᅜ 㝿 Ꮫ ఍ 䛷䛒䜛 ECTC 2017 (The Electronic Components and Technology Conference)䛷Ⓨ⾲䛥䜜䛯ⱥㄒㄽ ᩥ䛷䛒䜛䚹 ᩘ್䛿༢఩䜢ᣢ䛴䜒䛾䛸ᣢ䛯䛺䛔䜒䛾䛜䛒䜛䚹౛䛘䜀䠈㟁Ẽ ᢬ᢠ䛿䛂M䃈䛃䛺䛹䛾༢఩䜢ᣢ䛴䛾ᑐ䛧䛶䠈ㄏ㟁⋡䛿↓ḟඖ㔞䛷䛒䜚༢఩䜢ᣢ䛯䛺䛔䚹௒ᅇ䛾ホ౯䛷䛿䠈㡯┠ྡ䛸䛧䛶༢఩䜢 ᣢ䛯䛺䛔ᩘ್䛸⤌䜏䛻䛺䜛䜒䛾䜢⏝䛔䛯䚹䛣䜜䛿ணഛᐇ㦂䛻䛚䛔䛶䠈༢఩䜢ᣢ䛯䛺䛔ᩘ್䛸ẚ㍑䛧䛶༢఩䜢ᣢ䛴ᩘ್䛿༢఩䜢 ฼⏝䛩䜛䛣䛸䛷ᐜ᫆䛻ᢳฟ䛩䜛䛣䛸䛜ྍ⬟䛰䛳䛯䛯䜑䛷䛒䜛䚹ᮏ ホ౯䛷䛿䠈㡯┠ྡ䛸䛧䛶ᑐ㇟ㄽᩥ䛻䛚䛔䛶ฟ⌧㢖ᗘ䛾㧗䛛䛳䛯䛂dielectric constant䛃䜢㑅ᢥ䛧䛯䚹 ᑐ㇟ㄽᩥ䛻䛿㡯┠ྡ䛸ᩘ್䜢䛸䜒䛻ྵ䜐ᩥ䛜 46 ᩥᏑᅾ䛧䛯䚹䜎䛪䠈䛭䛾䛖䛱10 ᩥ䜢⏝䛔䛶䝹䞊䝹䜢సᡂ䛧䛯䚹ḟ䛻ṧ䜚 36 ᩥ 䜢⏝䛔䛶ᢳฟ⢭ᗘ䜢ホ౯䛧䛯䚹 ẚ㍑ᡭἲ䛿ᩥ⊩[7]䛾䝧䞊䝇䝷䜲䞁ᡭἲ䜢ཧ⪃䛻䛧䛶䠈㡯┠ ྡ䛛䜙䛾㊥㞳ᩥ⠇ᩘ䛜᭱䜒㏆䛔ᩘ್䜢ᢳฟ䛩䜛䝅䝇䝔䝮䛸䛧䛯䠄䝧䞊䝇䝷䜲䞁䠅䚹 4.2 ⤖ᯝ StruAP 䜢⏝䛔䛯ᥦ᱌ᡭἲ䛸䝧䞊䝇䝷䜲䞁ᡭἲ䛾ホ౯⤖ᯝ䜢⾲3 䛻♧䛩䚹1 ᩥ䛻」ᩘ䛾㡯┠ྡ䛜ྵ䜎䜜䜛䛣䛸䛜䛒䜛䛯䜑䠈ᩘ 䛘䜛ᑐ㇟䛿㡯┠ྡ䛾ᩘ䛸䛧䛯䚹㐺ྜ⋡䛸F ್䛻䛴䛔䛶ᥦ᱌ᡭἲ 䛿䝧䞊䝇䝷䜲䞁ᡭἲ䜢ୖᅇ䜛䛣䛸䜢☜ㄆ䛧䛯䚹 ᇳ➹⪅䛜ᚲ䛪䛧䜒⇍⦎䛧䛶䛔䛺䛔䛣䛸䛻䜘䜛⾲⌧䛾䜖䜙䛞䜔 PDF 䛛䜙䝔䜻䝇䝖䜈䛾ኚ᥮䛻క䛖䝜䜲䝈䛾ΰධ䛻䜘䜛⢭ᗘ䛾పୗ ⾲2㻌 ౑⏝䝒䞊䝹 ಀ䜚ཷ䛡ゎᯒ Stanford CoreNLP 䝹䞊䝹䜶䞁䝆䞁 StruAP [1] ⾲1 㻌 ᢳฟ䛩䜛㡯┠ྡ䛸ᩘ್䛾⤌䜏䛸ㄽᩥྡ䛾౛ 㡯┠ྡ䛸ᩘ್䛾⤌䜏 ㄽᩥྡ 㡯┠ྡ ᩘ್

dielectric constant 4.4 Publication A dielectric constant ~ 2.8 Publication B

䞉䞉䞉䞉䞉䞉䞉䞉䞉 ᅗ1㻌 ᥦ᱌䝅䝇䝔䝮䛾ᵓᡂᅗ ૛ધ ॸय़५ॺ બॉ_ਭऐ ੰෲ ඨ৯੡ध ਯகभੌा؜ ૛ધ੡ ঝشঝ ग़থ४থ ඨ৯੡ બॉਭऐ਽ ྴলঝشঝ

(3)

- 3 - 䛜ᠱᛕ䛥䜜䛯䛜䠈ホ౯䛧䛯⠊ᅖ䛷䛿0.96 䛸㧗䛔 F ್䛜ᚓ䜙䜜䛯䚹䝜䜲䝈䛾ΰධ䛻㛵䛧䛶௒ᅇホ౯䛧䛯36 ᩥ䛾䛖䛱 1 ᩥ䛻䛴䛔䛶䠈 PDF 䛛䜙䝔䜻䝇䝖䜈ኚ᥮䛩䜛䛸䛝䛻グྕ䛂䃈䛃䛜䛂䛃䛻ᩥᏐ໬ 䛡䛧䛶䛔䛯䚹ణ䛧䠈㡯┠ྡ䜔ᩘ್䛛䜙㞳䜜䛯⟠ᡤ䛷⏕䛨䛯ᩥᏐ ໬䛡䛷䛒䛳䛯䛯䜑䠈ᩘ್᝟ሗ䛾ᢳฟ䛻䛿ᙳ㡪䛜↓䛟䠈ᥦ᱌ᡭἲ 䛸䝧䞊䝇䝷䜲䞁ᡭἲ䛸䜒䛻ṇゎ䛷䛒䛳䛯䚹

5. ⪃ᐹ

ᮏᐇ㦂䛻䛚䛔䛶㡯┠ྡ䛸ᩘ್䛾㛵ಀ䛿ୗグ䛾3 ✀㢮䛷䛒䛳䛯䚹䛭䜜䛮䜜䛾౛ᩥ䛸䛸䜒䛻♧䛩䚹 (1) 㡯┠ྡ䛸ᩘ್䛜䛂of䛃䛷᥋⥆䛥䜜䛶䛔䜛䚹

S1: The substrate used for sensor design is Roger’s 3010 high frequency flexible laminates with a thickness of 0.13 mm, dielectric constant of 10.2, and loss tangent (tan δ) of 0.0035. (2) ୺ㄒ䛜㡯┠ྡ䛷䛒䜚䠈ᩘ್䛜㏙ㄒ䛻ಀ䜛䚹

S2: It was found that the dielectric constant was between 2.73 – 2.81 and the loss tangent between 0.012 – 0.05 from 0.1 to 0.8 THz respectively.

(3) 㡯┠ྡ䛸ᩘ್䛜㞄᥋䛧䛶䛔䜛䚹

S3: This result is attributed to the dielectric constant (5.0 – 5.4), and loss tangent (0.007), of the glass core.

䝹䞊䝹(1)䛾ᮌᵓ㐀䝟䝍䞊䞁䛸䛭䛾䝹䞊䝹䛷㡯┠ྡ䛸ᩘ್䛾 ⤌䜏䜢ᢳฟ䛷䛝䜛ᩥ䛾ಀ䜚ཷ䛡ᮌ䜢ᅗ2 䛻♧䛩䚹StruAP 䜢⏝䛔䛶䠈ᮌᵓ㐀䝟䝍䞊䞁䛜ಀ䜚ཷ䛡ᮌ䛾ୗ⥺㒊䛸䝬䝑䝏䛩䜛䛣䛸䛷a1 䛷♧䛩㡯┠ྡ䛸a0 䛷♧䛩ᩘ್䛾⤌䜏䜢ᢳฟ䛷䛝䜛䚹 ᥦ᱌ᡭἲ䛷ᢳฟ䛻ᡂຌ䛧䠈䝧䞊䝇䝷䜲䞁ᡭἲ䛷ᢳฟ䛻ኻᩋ䛧䛯౛䜢ୗグ䛻♧䛩䚹

S4: Utilizing a width-height ratio of 1, the dielectric constants of the measured resonances are 3.33.

䝧䞊䝇䝷䜲䞁ᡭἲ䛿ᩘ್䛸䛧䛶㡯┠ྡ䛻㏆䛔䛂1䛃䜢ㄗ䛳䛶㑅 ᢥ䛧䛯䚹

୍᪉䠈ᥦ᱌ᡭἲ䛷ᢳฟ䛻ኻᩋ䛧䠈䝧䞊䝇䝷䜲䞁ᡭἲ䛷ᢳฟ䛻 ኻᩋ䛧䛯౛䜢ୗグ䛾S5 䛻♧䛩䚹

S5: The dielectric constant and loss tangent were found in a previous work to be around 5.5-5.8 and 0.005, respectively [19]. ᩥS5 䛜ᥦ᱌ᡭἲ䛷ᢳฟ䛷䛝䛺䛛䛳䛯⌮⏤䛿ୗグ䛷䛒䜛䚹䠍䠊䝹䞊䝹䜢సᡂ䛩䜛䛻䛒䛯䛳䛶ཧ↷䛧䛯10 ᩥ䛾୰䛻㡯┠ྡ䛜୺ ㄒ䛸䛺䜛ཷືែ䛾ᵓᩥ䛜↓䛟䠈䝹䞊䝹సᡂ⪅䛿䛣䛾ᩥ䛻ྜ⮴䛩䜛䝹䞊䝹䜢 స ᡂ 䛷䛝䛺䛛䛳䛯䛯䜑䠈䠎䠊㡯 ┠ ྡ 䛂dielectric constant䛃䛸ᩘ್䛂5.5-5.8䛃䛾⤌䜏䛸䠈௒ᅇ䛾ᐇ㦂䛷䛿㡯┠ྡ䛸䛧䛺䛛䛳䛯䛜㡯┠ྡ䛻䛺䜚䛖䜛㡯┠ྡೃ⿵䛂loss tangent䛃䛸ᩘ್ೃ ⿵䛂0.005䛃䛾⤌䜏䛸䛔䛖」ᩘ䛾⤌䜏䛜Ꮡᅾ䛩䜛」㞧䛺ᵓᩥ䛷䛒䜛䛯䜑䠈䛷䛒䜛䚹ᩥS5 䛿䠎ಶ䛾㡯┠ྡ䠄ೃ⿵䠅䛸ᩘ್䠄ೃ⿵䠅䛾⤌ 䜏䛜஺ᕪ䛩䜛」㞧䛺ᩥ䛷䛒䜛䛜䠈ཧ↷䛩䜛ᩥ䛾㔞䜢ቑ䜔䛩䠈䜒䛧䛟䛿䠈సᴗ⪅䛜䜘䜚⇍⦎䛩䜛䛣䛸䛷ᢳฟ⢭ᗘ䛜㧗䜎䜛䛣䛸䛜ᮇ ᚅ䛥䜜䜛䚹⾲3 ᩘ್䛸㡯┠ྡ䛾⤌䜏䛾ᢳฟ⤖ᯝ ᡭἲ 㐺ྜ⋡ ෌⌧⋡ F್ ᥦ᱌ᡭἲ 1.00 0.92 0.96 䝧䞊䝇䝷䜲䞁 0.62 1.00 0.77

ᑐ㇟ᩥ䠖 The substrate used for sensor design is Roger’s 3010 high frequency flexible laminates with a thickness of 0.13 mm, dielectric constant of 10.2, and loss tangent (tan δ) of 0.0035.

ಀ僰ཷ傷ᮌ䢢 ᮌᵓ㐀儵儣兠兗䢢 (be ROOT (be S (substrate NP (substrate NP (the DT The) (substrate NN substrate)) (use VP (use VBN used) (for PP (for IN for) (design NP (sensor NN sensor) (design NN design))))) (be VP (be VBZ is) (laminate NP (laminate NP ('s NP (Roger NNP Roger) ('s POS ’s)) (3010 CD 3010) (high JJ high) (frequency NN frequency) (flexible JJ flexible) (laminate NNS laminates)) (with PP (with IN with) (thickness NP (thickness NP (a DT a) (thickness NN thickness)) (of PP (of IN of) (mm NP (0.13 CD 0.13) (mm NN mm))) (, , ,) (constant NP (constant NP (dielectric JJ dielectric) (constant NN constant)) (of PP (of IN of) (10.2 NP (10.2 CD 10.2)))) (, , ,) (and CC and) (tangent NP (tangent NP (loss NN loss) (tangent NN tangent)) (δ PRN (-lrb- -LRB- () (δ NP (tan NN tan) (δ NN δ)) (-rrb- -RRB- ))) (of PP (of IN of) (0.0035 NP (0.0035 CD 0.0035)))))))) (. . .))) ((id . item_of_value)) (#a1. * (.lemma=dielectric *) * (.lemma=constant *) * *) (. * (.lemma=of *) (.POS=NP * (#a0.POS=CD *) * ) ) ᅗ 2 㛵ಀᢳฟ䝒䞊䝹 StruAP 䛾ᮌᵓ㐀䛾䝟䝍䞊䞁䝬䝑䝏౛ The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020

(4)

- 4 -

6. 䛚䜟䜚䛻

ᮏሗ࿌䛷䛿≉䛻ᛂ⏝≀⌮䛾ㄽᩥ䛻ぢ䜙䜜䜛༢఩䜢క䜟䛺䛔 ᩘ್᝟ሗ䛾ᢳฟ䜢ㄢ㢟䛸䛧䠈ಀ䜚ཷ䛡ᵓ㐀䛻ᇶ䛵䛟䝟䝍䞊䞁䝬䝑䝏䛻䜘䜛ᢳฟ䜢฼⏝䛧䛶㡯┠ྡ䛸ᩘ್䛾⤌䜏䜢ᢳฟ䛩䜛ᡭἲ 䜢ᥦ᱌䛧䛯䚹ⱥㄒㄽᩥ䜢ᑐ㇟䛸䛧䛯ホ౯䜢⾜䛔䠈ᮏᡭἲ䛾᭷ຠ ᛶ䜢☜ㄆ䛧䛯䚹௒ᚋ䛿䠈ከᵝ䛺䝔䜻䝇䝖䛾⾲⌧䜈䛾ᑐᛂ䛻ྲྀ䜚 ⤌䜐䛣䛸䛷䠈ศᯒᑐ㇟䛸䛺䜛ᩥ❶䛾✀㢮䜢ㄽᩥ௨እ䛻ᣑ඘䛧䠈 ุ᩿ᨭ᥼䛾᰿ᣐ䛸䛺䜛᝟ሗ䜢ቑ䜔䛩䛣䛸䜢䜑䛦䛧䛶䛔䛟䚹 ཧ⪃ᩥ⊩

[1] Said A. SalloumEmail authorMostafa Al-EmranAzza Abdel MonemKhaled Shaalan, “Using text mining techniques for extracting information from research articles”, Intelligent Natural Language Processing: Trends and Applications, pp. 373-397, doi: 10.1007/978-3-319-67056-0_18 [2] ᕝᮏ⚈Ꮚ䠈Ⲩᮌḟ㑻䠈⸨ᒣ⛅బኵ䠈Ⳣཎ⚽᫂䠈኱ஂಖබ ⟇䠈䛂⏕࿨⛉Ꮫศ㔝䛾ᐇ㦂ᡭἲ䛻䜘䜛ㄽᩥ䛾ศ㢮䛸䜸䞁䝖䝻䝆䞊䛾฼⏝䛻㛵䛩䜛◊✲䛃䠈JSAI 2017 [3] ᑠᯘ⩏⾜䠈୰Ụ⿱ᶞ䠈䛂᪥❧〇సᡤ䛜ᥦ᱌䛩䜛་⸆ྥ䛡䝔䜻䝇䝖䝬䜲䝙䞁䜾䛃䠈⸆Ꮫᅗ᭩㤋 48(4), pp. 253-257, 2003 [4] G. E. Moore, “Cramming more components onto integrated

circuits”, Proceedings of the IEEE, vol. 86, no. 1, pp. 82–85, 1998.

[5] C. Hwang, “Nanotechnology enables a new memory growth model,” Proceedings of the IEEE, vol. 91, no. 11, pp. 1765– 1771, Nov. 2003., doi: 10.1109/JPROC.2003.818323 [6] Rayid Ghani, Katharina A Probst, Yanxi Liu, Marko Krema,

Andrew Ernest Fano, “Text Mining for Product Attribute Extraction”, ACM SIGKDD Explorations Newsletter June 2006, doi: 10.1145/1147234.1147241 [7] ⸨⏿຾அ䠈ᚿ㈡ṇ⿱䠈᳃㎮๎䠈䛂ಀ䜚ཷ䛡䛾ไ⣙䛸ඃඛつ๎ 䛻ᇶ䛵䛟ᩘ㔞⾲⌧ᢳฟ䛃䠈FI, ᝟ሗᏛᇶ♏◊✲఍ሗ࿌ 64, pp. 119-125 [8] ᩧ⸨බ୍䠈㏕⏣᫛ே䠈୰Ụᐩே䠈ᒾ஭⚞ᗈ䠈⏣ᮧ┤Ⰻ䠈୰ ᕝ⿱ᚿ䠈䛂ᩘ್᝟ሗ䜢䜻䞊䛸䛧䛯᪂⪺グ஦䛛䜙䛾᝟ሗᢳฟ䛃䠈 NL, ⮬↛ゝㄒฎ⌮◊✲఍ሗ࿌ 125, pp. 63-70

[9] Katharina Probst, Rayid Ghani, Marko Krema, Andrew Ernest Fano, Yan Liu, “Semi-Supervised Learning of Attribute-Value Pairs from Product Descriptions”, IJCAI 2007, Proceedings of the 20th International Joint Conference on Artificial Intelligence

[10] Seth Kulick, Ann Bies, Mark Liberman, Mark Mandel, Ryan McDonald, Martha Palmer, Andrew Schein, Lyle Ungar, “Integrated Annotation for Biomedical Information Extraction”, HLT/NAACL 2004 Workshop: Biolink 2004 [11] José Alberto S. Torres, Daniela S. Cruzes, Laís do

Nascimento Salvador, “Automatic Results Identification in Software Engineering Papers. Is it possible?”, 2012 12th International Conference on Computational Science and Its Applications, doi: 10.1109/ICCSA.2012.27

[12]Fuchun Peng, Andrew McCallum, “Accurate Information Extraction from Research Papers using Conditional Random Fields”, Information Processing & Management 42, Issue 4, pp.963-979, doi: 10.1016/j.ipm.2005.09.002

[13]K. Yanai, M. Sato, T. Yanase, K. Kurotsuchi, Y. Koreeda, and Y. Niwa, “StruAP: A Tool for Bundling Linguistic Trees through Structure-based Abstract Pattern,” EMNLP 2017 -

Conf. Empir. Methods Nat. Lang. Process. Syst. Demonstr. Proc., pp. 31–36, 2017, doi: 10.18653/v1/d17-2006.