[๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Aspect-based sentiment analysis via affective knowledge enhanced graph convolutional networks

๐ŸŠ ๋…ผ๋ฌธ ๋งํฌ: https://www.sciencedirect.com/science/article/pii/S0950705121009059

Liang, B., Su, H., Gui, L., Cambria, E., & Xu, R. (2022). Aspect-based sentiment analysis via affective knowledge enhanced graph convolutional networks. Knowledge-Based Systems, 235, 107643.


   Aspect-based-sentiment anlaysis (ABSA)๋Š” ๊ธฐ์กด์˜ ๊ฐ์„ฑ ๋ถ„์„ ๊ธฐ๋ฒ•๊ณผ ๋‹ฌ๋ฆฌ, ํ…์ŠคํŠธ ๋‚ด ๋“ฑ์žฅํ•˜๋Š” ์—ฌ๋Ÿฌ ์†์„ฑ๋“ค์— ๋Œ€ํ•œ ๊ฐ์„ฑ์„ ์ง์ ‘์ ์œผ๋กœ ํ•™์Šตํ•˜๊ณ  ์˜ˆ์ธกํ•˜๋Š” ๊ธฐ๋ฒ•์ด๋‹ค. ์ตœ๊ทผ์—๋Š” ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•์ด ๋ฐœ์ „ ํ•˜๋ฉด์„œ, ์ด๋Ÿฌํ•œ ABSA ์ž‘์—…์— ๊ทธ๋ž˜ํ”„ ๋ชจ๋ธ์ด ๋„๋ฆฌ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด์˜ dependency tree ๊ธฐ๋ฐ˜ ABSA ๋ชจ๋ธ์€, ๋ฌธ๋งฅ ๋‹จ์–ด (ํ…์ŠคํŠธ ๋‚ด ๋‹จ์–ด)์™€ ์ธก๋ฉด ๋‹จ์–ด (์ฃผ๋ชฉํ•˜๊ณ ์ž ํ•˜๋Š” aspect ํ‚ค์›Œ๋“œ ๋‹จ์–ด) ์‚ฌ์ด์˜ ๊ด€๊ณ„(dependency information)๋งŒ์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์œผ๋ฉฐ, aspect์— ๋Œ€ํ•œ ๊ตฌ์ฒด์ ์ธ ๋ฌธ๋งฅ์  ๊ฐ์ • ์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•˜๊ณ  ์žˆ๋‹ค.

 

์ฆ‰ ABSA ํ•™์Šต์„ ์œ„ํ•ด ์‚ฌ์šฉ๋˜๋Š” SemEval ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์˜ˆ์ธก ์ •ํ™•๋„์—๋งŒ ์ง‘์ค‘ํ•˜๊ฒŒ ๋˜์–ด, ์‹ค์ œ ๋ฌธ๋งฅ์— ๋Œ€ํ•œ ๊ฐ์„ฑ ์ •๋ณด๋Š” ์ฃผ๋ชฉ ๋ฐ›์ง€ ๋ชปํ•œ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ, ๋ฌธ์žฅ์˜ common sense๋ฅผ ๊ณ ๋ คํ•˜์ง€ ๋ชปํ•œ๋‹ค๊ณ  ํ‘œํ˜„ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์•ˆํ•˜๋Š” ๋ชจ๋ธ์€, ๊ธฐ์กด ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ aspect ๋‹จ์–ด์™€ ๋ฌธ๋งฅ ๋‹จ์–ด์˜ ์—ฐ๊ฒฐ์„ฑ์„ ๊ณ ๋ คํ•˜๋ฉด์„œ๋„, Senticnet์ด๋ผ๋Š” ์˜คํ”ˆ ์†Œ์Šค ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ฐ˜์ ์ธ ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ๊ฒƒ์ด ํ•ต์‹ฌ์ด๋‹ค.


Introduction

  ๋ฌธ์„œ๋‚˜ ๋ฌธ์žฅ ๋‚ด์—์„œ ๋‹จ์ผํ•œ ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ๋ถ„์„ํ•˜๋Š” ์ „ํ†ต์ ์ธ ๊ฐ์„ฑ ๋ถ„์„๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ์†์„ฑ ๊ธฐ๋ฐ˜ ๊ฐ์„ฑ ๋ถ„์„(ABSA)์€ ๋™์ผํ•œ ๋ฌธ์žฅ์—์„œ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์˜ ๊ฐ์„ฑ ๊ทน์„ฑ(eg., positive, neutral or negative)์„ ๊ฐ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, 'The food is good, but the service is terrible'์ด๋Š” ๋ฌธ์žฅ์—์„œ food'์™€ 'service'๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ธก๋ฉด์ด ์–ธ๊ธ‰๋œ๋‹ค. ์—ฌ๊ธฐ์„œ '์Œ์‹'์— ๋Œ€ํ•œ ๊ฐ์„ฑ ๊ทน์„ฑ์€ ๊ธ์ •์ ์ด์ง€๋งŒ '์„œ๋น„์Šค'์— ๋Œ€ํ•œ ๊ฐ์„ฑ ๊ทน์„ฑ์€ ๋ถ€์ •์ ์ด๋‹ค. ABSA์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์˜ˆ์‹œ๋Š” ์•„๋ž˜ ๊ทธ๋ฆผ๊ณผ ๊ฐ™๋‹ค.

Example of Aspect-based Sentiment Analysis

  ABSA์— ๋Œ€ํ•œ ์ดˆ๊ธฐ ์ž‘์—…๋“ค ๋Œ€๋ถ€๋ถ„์€ ์ฃผ์–ด์ง„ ์ธก๋ฉด์— ๋Œ€ํ•œ ๋ฌธ๋งฅ์œผ๋กœ๋ถ€ํ„ฐ ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์˜€๋‹ค. ๋” ๋‚˜์•„๊ฐ€, ์–ดํ…์…˜ ๊ธฐ๋ฐ˜ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์„ ํ†ตํ•ด ๋ฌธ์žฅ์˜ ๋ฌธ๋งฅ๊ณผ ํŠน์ • aspect์™€ ๊ด€๋ จ๋œ ๋ถ€๋ถ„์— ์ดˆ์ ์„ ๋งž์ถ”๋Š” ๋ชจ๋ธ์ด ์ฃผ๋ชฉ์„ ๋ฐ›์•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์–ดํ…์…˜ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ๋ฌธ๋งฅ ๋‹จ์–ด์™€ aspect ๋‹จ์–ด ๊ฐ„์˜ ๊ฐ์„ฑ ์˜์กด์„ฑ (sentiment dependencies)๋ฅผ ์ถ”์ถœํ•˜๊ธฐ ๋ถ€์กฑํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋‹จ์–ด ๊ฐ„์˜ ๊ด€๊ณ„ ์ •๋ณด๋ฅผ ํ‘œํ˜„ํ•˜๋Š” ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋“ค์ด ์ œ์•ˆ๋˜์—ˆ๋‹ค. ๋…ผ๋ฌธ์€ ์ด๋Ÿฌํ•œ ๊ทธ๋ž˜ํ”„ ๋ชจ๋ธ์„ ํ†ตํ•ด ๋ฌธ๋ฒ•์ ์œผ๋กœ ๊ด€๋ จ๋œ ๋ฌธ๋งฅ ๋‹จ์–ด๋“ค์„ aspect ๋‹จ์–ด์™€ ์—ฐ๊ฒฐํ•˜์—ฌ ์žฅ๊ธฐ์ ์ธ ๋ฌธ๋ฒ• ์˜์กด์„ฑ์„ ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋งํ•œ๋‹ค.

 

  ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด์˜ ๊ทธ๋ž˜ํ”„ ๋„คํŠธ์›Œํฌ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋“ค์€ ๋Œ€๋ถ€๋ถ„ ๋ฌธ์žฅ์˜ ๋ฌธ๋ฒ•์  ์˜์กด์„ฑ๋งŒ ๊ณ ๋ คํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋ฌธ์žฅ์„ ๊ตฌ์„ฑํ•˜๋Š” ์ƒ์‹์  ๊ฐ์„ฑ ์ •๋ณด(commeonsense knowledge)๋ฅผ ๋ฌด์‹œํ•œ๋‹ค๋Š” ๊ฒƒ์ด ๋ณธ ๋…ผ๋ฌธ์˜ main research gap์ด ๋œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ์—ฐ๊ตฌ๋Š” ๊ฐ ๋ฌธ์žฅ์— ๋Œ€ํ•ด ์˜์กด์„ฑ ํŠธ๋ฆฌ(dependency tree)์™€ ๊ฐ์„ฑ ์ƒ์‹ ์ง€์‹(affenctive commonsense knowledge)์„ ๋ชจ๋‘ ํ™œ์šฉํ•˜์—ฌ ๊ทธ๋ž˜ํ”„๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•œ๋‹ค. ๋…ผ๋ฌธ์ด ๋ฐํžˆ๋Š” main contribution์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

  • ๋ณธ ์—ฐ๊ตฌ๋Š” ๋ฌธ์žฅ์˜ ๊ตฌ์กฐ์  ์˜์กด์„ฑ๊ณผ ํŠน์ • ์ธก๋ฉด์— ๊ด€๋ จ๋œ ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ๋ชจ๋‘ ํ™œ์šฉํ•œ๋‹ค.
  • ๋ณธ ์—ฐ๊ตฌ๋Š” ๊ธฐ์กด ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ ๊ฐ์„ฑ ์ƒ์‹ ์ง€์‹์„ ํ†ตํ•ด, ๊ทธ๋ž˜ํ”„ ํ•ฉ์„ฑ ์‹ ๊ฒฝ๋ง์„ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•๋ฒ•๋ก ์„ ์ œ์‹œํ•˜์—ฌ ํŠน์ • ์ธก๋ฉด์— ๋Œ€๋‹นํ•˜๋Š” ๊ฐ์„ฑ ์˜์กด์„ฑ์„ ํฌ์ฐฉํ•œ๋‹ค.
  • ๋„ค ๊ฐ€์ง€ ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ์„ธํŠธ์— ์‹คํ—˜ํ•œ ๊ฒฐ๊ณผ, ์ œ์•ˆ ๋ชจ๋ธ์ด ABSA ์ž‘์—…์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค€๋‹ค.

 

๋…ผ๋ฌธ์€ ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๊ตฌ์ฒด์ ์ธ research gap์„ ์ œ์‹œํ•œ๋‹ค. ๋˜ํ•œ ์•„๋ž˜ ๋ฐฉ๋ฒ•๋ก ์—์„œ ๋ณด๊ฒ ์ง€๋งŒ, ์ œ์•ˆ ๋ชจ๋ธ ๊ตฌ์กฐ์—๋„ ๋ฌธ์ œ ์˜์‹์— ๋Œ€ํ•œ ํ•ด๊ฒฐ์ฑ…์ด ๋ช…ํ™•์ด ์ œ์‹œ๋˜์–ด ์žˆ๋‹ค. 

 


Proposed Model

  ์ œ์•ˆ ํ•˜๋Š” ๋ชจ๋ธ์€ GCN (graph convolutional network)๊ณผ SenticNet ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ๊ฒฐํ•ฉํ•œ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ์ด๋‹ค. ์ „์ฒด์ ์ธ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋Š” ์•„๋ž˜ ๊ทธ๋ฆผ๊ณผ ๊ฐ™๋‹ค.

Overall architecture of SenticGCN

  SenticGCN์€ ๋‘ ๊ฐ€์ง€ ํŒŒํŠธ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์ฆ‰ $w_1 ... w_n$์œผ๋กœ ํ‘œํ˜„๋œ ํ…์ŠคํŠธ๊ฐ€ ๋‘ ๊ฐ€์ง€ ํ‘œํ˜„์„ ์–ป๋Š” ๊ฒƒ์ด๋‹ค.

 

๋จผ์ € (1) LSTM ๋ ˆ์ด์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ๋งฅ์  ํ‘œํ˜„์„ ํ•™์Šตํ•œ๋‹ค. ์ด๋Š” ๊ฐ ๋ฌธ์žฅ์˜ ์ž„๋ฒ ๋”ฉ ํ–‰๋ ฌ์„ ์ž…๋ ฅ ๋ฐ›์•„ ์ž…๋ ฅ ๋ฌธ์žฅ์˜ ์ž ์žฌ ๋ฌธ๋งฅ์  ํ‘œํ˜„์„ ๋„์ถœํ•˜๋Š” ๊ฒƒ์ด๋‹ค.

 

(2) ๊ธฐ์กด ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ABSA ๋ชจ๋ธ๊ณผ ๊ฐ™์ด, 'Dependency tree'๋ผ๊ณ  ํ‘œํ˜„๋œ ๋ถ€๋ถ„์—์„œ GCN ๋ ˆ์ด์–ด๋ฅผ ํ†ตํ•ด ๋‹จ์–ด์˜ ๊ด€๊ณ„์„ฑ์„ ํ•™์Šตํ•œ๋‹ค. ๋˜ํ•œ 'Affective'๋ผ๊ณ  ํ‘œํ˜„๋œ ๊ทธ๋ž˜ํ”„๋Š” ๊ฐ์„ฑ ๊ฐ•ํ™” ๊ทธ๋ž˜ํ”„(affective enhanced graph)๋ผ๊ณ  ํ‘œํ˜„ํ•˜๊ณ  SentiNect ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ํ‘œํ˜„ํ•œ ๊ทธ๋ž˜ํ”„ ์ •๋ณด์ด๋‹ค. ์ดํ›„ ์—ฐ๊ฒฐ์„ฑ ๊ทธ๋ž˜ํ”„์™€, ๊ฐ์„ฑ ๊ทธ๋ž˜ํ”„๊ฐ€ ๊ฒฐํ•ฉ๋˜์–ด 'Aspect'๋กœ ํ‘œํ˜„๋œ ์ธก๋ฉด ๊ธฐ๋ฐ˜ ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ํ†ตํ•ด, ๋ฌธ์žฅ์˜ ๊ตฌ๋ฌธ ์ •๋ณด์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๋Š” ๊ธฐ์กด ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ, ํŠน์ • ์ธก๋ฉด๊ณผ ๊ด€๋ จ๋œ ๋ฌธ๋งฅ์  ๊ฐ์„ฑ ์˜์กด์„ฑ์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ์„ฑ ํŠน์ง•์„ ๊ฐ•์กฐํ•˜๊ฒŒ ๋œ๋‹ค.

 

Constructing graph over dependency tree

  ์œ„์—์„œ ์„ค๋ช…ํ•œ (2)๋ถ€๋ถ„์„ ์ž์„ธํžˆ ์‚ดํŽด๋ณด๊ฒ ๋‹ค. ๋จผ์ € GCN์—์„œ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋‹จ์–ด ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ๋‹จ์–ด ๊ด€๊ณ„ ๊ทธ๋ž˜ํ”„๋ฅผ ํ•™์Šตํ•œ๋‹ค. ์ด ๊ทธ๋ž˜ํ”„๋Š” ์•„๋ž˜ ์‹์—์„œ $w_i$์™€ $w_j$ ๋‘ ๋‹จ์–ด๊ฐ€ ์—ฐ๊ฒฐ๋˜์—ˆ๋‹ค๋ฉด 1, ์•„๋‹ˆ๋ฉด 0์œผ๋กœ ํ‘œํ˜„๋œ ๋งคํŠธ๋ฆญ์Šค๋กœ ๋‚˜ํƒ€๋‚œ๋‹ค. ์ดํ›„ GCN model์„ ํ†ตํ•ด ๊ฐ ๋‹จ์–ด ๋…ธ๋“œ์˜ ์ž„๋ฒ ๋”ฉ ์ •๋ณด๊ฐ€ ํ•™์Šต๋œ๋‹ค.

 

  ์—ฌ๊ธฐ์„œ ๊ทธ๋ž˜ํ”„๋Š” ๋‹จ์–ด ๊ฐ„์˜ ์—ฐ๊ฒฐ์„ฑ์„ ๋‚˜ํƒ€๋‚ด๊ธฐ ๋•Œ๋ฌธ์—, ์•„๋ž˜ ๊ทธ๋ฆผ๊ณผ ๊ฐ™์ด ๋ฐฉํ–ฅ์„ฑ์ด ์—†๋Š” undirected edge๋กœ ํ‘œํ˜„๋œ๋‹ค.

 

 

Enhancing graph based on SenticNet

  GCN ๋ ˆ์ด์–ด๋ฅผ ๊ฑฐ์นœ ๋‹จ์–ด๋“ค์€ ๊ฐ๊ฐ dependency ์ •๋ณด๋ฅผ ํ‘œํ˜„ํ•˜๊ณ  ์žˆ๋‹ค. ์ดํ›„ ๋ณธ ๋…ผ๋ฌธ์˜ main idea์™€ ๊ฐ™์ด, ๋ฌธ๋งฅ ๋‹จ์–ด์™€ aspect ๋‹จ์–ด์˜ ๊ฐ์„ฑ์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด SenticNet์—์„œ ๊ฐ์„ฑ ์ ์ˆ˜๋ฅผ ๋”ํ•˜์—ฌ ์ธ์ ‘ ํ–‰๋ ฌ์˜ ํ‘œํ˜„๋ ฅ์„ ํ’๋ถ€ํ•˜๊ฒŒ ํ‘œํ˜„ํ•œ๋‹ค. ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

$$S_{i,j} = SenticNet_{(wi)} + SenticNet_{(wj)}$$

  ์—ฌ๊ธฐ์„œ $SenticNet_{(wi)}$์€ SenticNet์— ์กด์žฌํ•˜๋Š” ๊ฐ์„ฑ ์Šค์ฝ”์–ด๋กœ [-1, 1] ์‚ฌ์ด์˜ ๊ฐ’์„ ๊ฐ–๋Š”๋‹ค. ๋งŒ์•ฝ $SenticNet_{(wi)} = 0$์ด๋ฉด ํ•ด๋‹น ๋‹จ์–ด๋Š” ์ค‘๋ฆฝ์ ์ธ ๊ฐ์„ฑ ์ •๋ณด๋ฅผ ๊ฐ–๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•œ๋‹ค.

  ๋งˆ์ง€๋ง‰์œผ๋กœ ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ABSA๋ฅผ ์œ„ํ•œ GCN ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋“ค์ด ์ผ๋ฐ˜์ ์œผ๋กœ ๊ทธ๋ž˜ํ”„๋ฅผ ๊ตฌ์„ฑํ•  ๋•Œ, aspect์— ๋Œ€ํ•œ ์ถฉ๋ถ„ํ•œ ๊ณ ๋ ค๊ฐ€ ์ด๋ฃจ์ง€์ง€ ์•Š๋Š”๋‹ค๊ณ  ๋งํ•œ๋‹ค. ์ œ์•ˆ ๋ชจ๋ธ์€ ์•„๋ž˜ ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ๋งฅ ๋‹จ์–ด์™€ ์ธก๋ฉด ๋‹จ์–ด ๊ฐ„์˜ ๊ฐ์„ฑ ์˜์กด์„ฑ์„ ๋”์šฑ ๊ฐ•ํ™”ํ•œ๋‹ค.

 

๋”ฐ๋ผ์„œ ์ตœ์ข…์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๊ฐ•ํ™”๋œ ์ธ์ ‘ ํ–‰๋ ฌ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ‘œํ˜„๋œ๋‹ค.

$$A_{i,j} = D_{i,j} × (S_{i,j} + T_{i,j} + 1)$$

 

์ด๋ ‡๊ฒŒ ๋ณธ ๋…ผ๋ฌธ์€ ์ œ์•ˆํ•˜๋Š” research gap์„ ๋ช…ํ™•ํ•˜๊ฒŒ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํ™•์‹คํ•œ ๋ชจ๋ธ๋ง์„ ์ œ์‹œํ•œ๋‹ค. ์ดํ›„ SenticGCN์€ GCN ๋ฐฉ์‹์„ ํ†ตํ•ด ํ•™์Šต๋˜์–ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ ๊ทธ๋Ÿผ๊ณผ ๊ฐ™์ด positive, negative, neutral์— ๋Œ€ํ•œ ์ตœ์ข… ๊ฐ’์„ ๋„์ถœํ•œ๋‹ค. ํ•™์Šต์„ ์œ„ํ•œ ๋ชฉ์  ํ•จ์ˆ˜๋กœ๋Š” cross-entropy loss with $L_2$ regularization ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค.

 


Experiments

  ์‚ฌ์šฉํ•˜๋Š” ๋ฐ์ดํ„ฐ์„ธํŠธ๋Š” SemEval 2014 task4 (Rest14, Laptop14), SemEval 2015 task 12 (Rest15), SemEval 2016 task 5 (Rest16)์ด๋‹ค. ์„ฑ๋Šฅ ๋น„๊ต๋ฅผ ์œ„ํ•œ ๋ฒค์น˜๋งˆํฌ๋Š” ๋„ˆ๋ฌด ๋งŽ๊ธฐ ๋•Œ๋ฌธ์—, ๋…ผ๋ฌธ์˜ ์บก์ฒ˜ ์‚ฌ์ง„์œผ๋กœ ์ œ์‹œํ•œ๋‹ค.

 

 

์‹คํ—˜ ๊ฒฐ๊ณผ ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์—์„œ, ์ œ์•ˆํ•˜๋Š” ๋ชจ๋ธ์ด ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋ฉฐ, SenticGCN๊ณผ BERT๋ฅผ ๊ฒฐํ•ฉํ•œ ๋ชจ๋ธ๋„ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค.

์ž์„ธํ•œ ๊ฒฐ๊ณผ ํ•ด์„์€ ๋…ผ๋ฌธ ์ฐธ๊ณ .

๋ฐ˜์‘ํ˜•