Sentiment analysis – short Q&A
20 questions and answers on sentiment analysis, including polarity and subjectivity, lexicon-based and machine learning methods, aspect-based sentiment and evaluation.
What is sentiment analysis in NLP?
Answer: Sentiment analysis, or opinion mining, aims to detect the emotional tone of text—typically classifying it as positive, negative, neutral or more fine-grained attitudes about an entity or topic.
What is the difference between polarity and subjectivity?
Answer: Polarity refers to whether sentiment is positive, negative or neutral, while subjectivity measures how opinionated or factual a text is, distinguishing personal views from objective statements.
How do lexicon-based sentiment methods work?
Answer: Lexicon-based approaches use dictionaries of words associated with sentiment scores; they aggregate scores of words in a text, sometimes with heuristics for negation or intensifiers, to estimate overall sentiment.
What are common sentiment lexicons?
Answer: Popular lexicons include SentiWordNet, AFINN, NRC Emotion Lexicon, VADER and various domain-specific lexicons that list words with polarity or emotion scores for sentiment tasks.
What limitations do lexicon-based methods have?
Answer: They often ignore context, sarcasm, compositional effects and domain-specific sentiment shifts, making them less accurate than supervised or neural methods on complex, noisy real-world data.
How does supervised sentiment classification work?
Answer: Supervised methods train classifiers (logistic regression, SVMs, neural networks) on labeled examples of texts with known sentiment, using features like n-grams or embeddings to learn patterns associated with each class.
What is aspect-based sentiment analysis (ABSA)?
Answer: ABSA identifies specific aspects or features of an entity (e.g. “battery life” or “screen”) and determines sentiment toward each aspect, rather than only predicting overall document sentiment.
How do transformers like BERT improve sentiment analysis?
Answer: Transformers capture long-range context, subtle linguistic cues and domain-specific usage when fine-tuned, enabling more accurate sentiment predictions, especially on complex sentences or nuanced opinions.
What datasets are commonly used to benchmark sentiment analysis?
Answer: Popular datasets include IMDb movie reviews, SST (Stanford Sentiment Treebank), Amazon or Yelp reviews, Twitter sentiment datasets and domain-specific corpora for ABSA like SemEval restaurant or laptop reviews.
What metrics are used to evaluate sentiment models?
Answer: Accuracy, precision, recall and F1-score are standard; for rating regression tasks, mean squared error or Pearson correlation may be used, and class-imbalance calls for macro-averaged metrics.
How does sarcasm affect sentiment analysis?
Answer: Sarcasm often uses positive words to express negative opinions or vice versa, confusing literal word-based models; detecting sarcasm requires understanding pragmatics, context and sometimes user history or emojis.
What is domain adaptation in sentiment analysis?
Answer: Domain adaptation adjusts a model trained on one domain (e.g. movie reviews) to work well on another (e.g. product reviews), addressing vocabulary shifts and different expression patterns of sentiment.
How can we interpret neural sentiment models?
Answer: Interpretation techniques include attention visualization, gradient-based saliency, integrated gradients, LIME or SHAP, which highlight which tokens or phrases contribute most strongly to predicted polarity.
What are emotion classification tasks related to sentiment analysis?
Answer: Emotion classification extends beyond polarity to detect specific emotions like joy, anger, sadness or fear, often using multi-label setups and lexicons or models tuned for fine-grained affective categories.
How does sentence-level sentiment differ from document-level sentiment?
Answer: Sentence-level tasks classify individual sentences, which may contain mixed or context-dependent polarity, while document-level tasks aggregate information across sentences to predict an overall sentiment label or score.
How is sentiment analysis used in industry?
Answer: Companies use sentiment analysis for brand monitoring, social media listening, customer feedback mining, product reviews analysis and routing or prioritizing support tickets based on user emotion.
What ethical issues arise with sentiment analysis?
Answer: Risks include privacy concerns, biased models misinterpreting language from certain groups, misuse for manipulation or surveillance and lack of transparency about how predictions are used in decision-making.
How do we handle emojis and hashtags in sentiment models?
Answer: Emojis and hashtags often carry strong sentiment cues; preprocessing can map them to tokens or features, and modern subword-based models can learn their effect directly from labeled social media data.
What is distant supervision for building sentiment datasets?
Answer: Distant supervision automatically labels data using heuristic signals such as star ratings, positive/negative emojis or hashtags, enabling large but noisy training sets for sentiment models without manual annotation.
Are large language models sufficient to solve sentiment analysis?
Answer: Large models achieve strong performance but still require careful fine-tuning, domain adaptation and evaluation to avoid biases and ensure that predictions are reliable for specific real-world applications.
🔍 Sentiment analysis concepts covered
This page covers sentiment analysis: polarity and subjectivity, lexicon-based and supervised methods, transformer-based sentiment models, aspect-based sentiment and practical considerations for real deployments.