ABSA 소개

( 참고 : https://velog.io/@jonas-jun/ASC-1 )

Contents

1. ABSA
2. ASC with BERT

[ 1. ABSA ]

(1) ABSA 내의 다양한 task들

SemEval 2014

( SemEval : Semantic Evaluation )

1. aspect TERM extraction

   • aspect term이라 할 만한 단어 찾기

   • ex) I liked the the service and the staff, but not the food.

     output : staff, food

2. aspect TERM polarity

   • “aspect term이 주어졌을 때 “, 해당 감정 도출

   • ex) I hated their fajitas, but their salads were great

     output : {fajitas: neg., salads: pos.}

3. aspect CATEGORY detection

   • “pre-defined” aspect categories {e.g., price, food}가 들어있는지

   • ex) The restaurant was expensive, but the menu was great.

     output : {price, food}

4. aspect CATEGORY polarity

SemEval 2016

1. Sentence level ABSA

   1. aspect category detection :

      • pre-defined된 Entity/Attribute 쌍이 있는지

        ( E#A가 하나의 카테고리 )

   2. opinion target expression (OTE) :

      • category(E#A)와 관련된 단어의 위치를 파악

   3. sentiment polarity:

      • E#A와 OTE가 있을 때, 감성 분류

   example 1 :

   • input : “The So called laptop runs to slow and I hate it!”
   • output :
     • {LAPTOP#OPERATION_PERFORMANCE, negative}
     • {LAPTOP#GENERAL, negative}

   example 2 :

   • input : “Chow fun was dry; pork shu mai was more than usually greasy and had to share a table with loud and rude family.”
   • output :
     • {FOOD#QUALITY, “Chow fun”, negative, from=0, to=8}
     • {FOOD#QUALITY, “pork shu mai”, negative, from=18, to=30}

2. Text level ABSA

   • category(E#A)와 감정 찾아내기

3. Out-of-domain ABSA

   • 학습한 모델을 “다른 도메인 data”에서 테스트하기

(2) 용어

• Aspect Term :
  • ex) “The beef was tender and melted in my mouth”
  • aspect term : “beef”

• Aspect Category : “LAPTOP#GENERAL”

• Polarity : (긍/부/중)

• Opinion Target Expression :

  • Aspect Term과 유사하지만,

  • 어떤 카테고리에 대해 직접적으로 표현하는 단어

    ( category(E#A)와 관련된 단어 )

• AE (Aspect Extraction)

  • Aspect Term을 발견/추출하는 task

• ASC (Aspect Sentiment Classification)

  • Aspect Term에 대한 감정을 구분

.

[ 2. ASC with BERT ]

Introduction

• BERT : Transformer의 encoder 부분만 사용하여 pre-train

  ( GPT : Transformer의 decoder ~ )

• ASC : Aspect Sentiment Classification

• (기본적으로) Sequence Classification
• 문장 2개가 들어왔을 떄, 유사한지/다른지 (1/0) 판별
• ABSA를 위해선…
  • 문장 1 : text
  • 문장 2 : aspect term

Utilizing BERT Intermediate Layers for Aspect Based Sentiment Analysis and Natural Language Inference ( Youwei Song, 2020 )