AI 용어집
인공지능 완전 사전
Aspect-Based Sentiment Analysis (ABSA)
An NLP task that aims to identify specific aspects of an entity and determine the sentiment (positive, negative, neutral) expressed towards each of these aspects.
Aspect Category
A predefined classification of aspects into semantic groups (e.g., 'SERVICE', 'AMBIENCE', 'FOOD' for a restaurant) to structure the analysis.
Aspect Term
The explicit word or phrase in the text that refers to a specific aspect (e.g., 'the pizza', 'the service', 'the delivery time').
Sentiment Polarity
The classification of sentiment associated with an aspect into discrete categories, typically positive, negative, or neutral.
Target Opinion Detection
The task of linking an opinion or sentiment expression to the specific aspect (the target) it refers to in the text.
Dependency Parser
An NLP tool used to build the syntactic tree of a sentence, allowing descriptive words (adjectives) to be linked to their respective nouns (aspects).
Supervised Learning for ABSA
The use of machine learning models trained on manually annotated corpora to predict aspects and sentiment polarities.
Weakly Supervised Learning
Methods that use partial, noisy labels or heuristics to train ABSA models without requiring complete and costly manual annotation.
Sequence-to-Sequence (Seq2Seq) Models
A neural network architecture, often based on LSTM or Transformers, used to generate structured aspect-sentiment pairs from input text.
Convolutional Neural Networks (CNN) for ABSA
The application of CNNs to extract local features (n-grams) in text, effective for capturing relationships between nearby words like an adjective and its noun.
BERT for ABSA
The adaptation of the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers) for ABSA tasks, often by adding a specific classification head.
Implicit Sentiment Analysis
The detection of sentiments that are not expressed by explicit polarity words but inferred from context (e.g., 'I couldn't eat my soup, it was cold' implies a negative sentiment about the 'temperature' aspect).
Annotated Corpus for ABSA
A set of text data where aspects, aspect terms, and sentiment polarities have been manually labeled, serving as a reference for training and evaluating models.
F1-Score Evaluation
A common evaluation metric for ABSA, which calculates the harmonic mean of precision and recall, measuring the model's ability to correctly identify aspect-sentiment pairs.
Implicit Aspect Detection
The identification of aspects that are not directly named but are implied by context (e.g., 'I waited 30 minutes' implies the 'waiting time' aspect).
Sentence-Level vs. Aspect-Level Sentiment Analysis
The distinction between classifying the overall sentiment of a sentence and the granular analysis of ABSA, which can reveal contradictory sentiments within the same sentence.
Quadruple Task in ABSA
An advanced formulation of ABSA that aims to simultaneously identify the opinion holder, aspect category, aspect term, and sentiment polarity in a text.