11_Sentiment_Analysis

Relationships between items can be a rich source of information about for performing classification or regression on the items.

• Nearby sentences can share the same subjectivity status, subjective or objective
• Mentions separated by “and” usually have similar sentiment labels
• Those separated by “but” usually have contrasting labels

• Domain-dependent polarity
  • Sentiment features for one domain often don’t generalize to another
    •  Cold beer vs. cold pizza
    • “unpredictable” (movie plots vs. car’s steering) 
• Indirect opinion
  • No opinion word
  • After taking the drug, my pain has gone.
• Comparative opinion
  • Phone is better than Blackberry.
• No explicit opinion
  • We brought the mattress yesterday, and a body impression has formed.
  • Fact implicitly expresses negative opinion
  •  “body impression” is negative for mattresses

How do we identify subjective items?

1. Manually compile a list of subjective items

2. Assume that contexts are coherent (corpus-based methods)

• adjectives linked by conjunctions (e.g. and) have the same orientation
  • Staff was very nice and helpful.

3. Assume that alternatives are similarly subjective (“plug into” subjective contexts) (lexical resource-based methods)

• Sentiment classification at both the document and sentence (or clause) levels are useful, but
  • they do not find what specifically people liked and disliked.
  • they do not identify the targets of opinions, i.e., entities and their aspects.
• ->Without knowing targets, opinions are of limited use.

Target Entity Identification is similar to named entity recognition

• E.g., one wants to study opinions on phones given Motorola and Nokia, find all phone brands and models in a corpus, e.g., Samsung, Moto.

The pictures coming out of this camera are amazing. The 'auto' feature takes great pictures most of the time. And with digital, you're not wasting film. …

• Target entity ->  camera
• Aspect -> picture, ‘auto’

• Goal: Given an opinion corpus, extract all aspects
• Frequency-based approaches: nouns (NN) that are frequently talked about are likely to be true aspects (called frequent aspects)
  • Because different reviewers tell different stories
  • Because those are the main aspects

Further Work on Opinion Targets

• Identifying implicit aspects: most adjectives modify or describe some specific attributes of entities.
  • “expensive” ⇒ aspect “price”
  • “beautiful” ⇒ aspect “appearance”
  • “heavy” ⇒ aspect “weight”
• Identifying aspect synonyms
  • “power usage” and “battery life”
• Linking opinion expressions to the corresponding aspects and identifying negation scope in case of negation

• Subjectivity, sentiment, opinion etc. is context and domain dependent.
  • Supervised approaches proven to be effective, but manually annotated corpora in each new domain is not easy to find.
  • -> Approaches addressing domain adaptability
• Problems with different approaches for creating opinion lexicons:
  • Lexical resource-based: high quality graph structure, sense annotated, but
    • of little use without successful word sense disambiguation component
    • coverage problem
    • not many languages have such complete resources
    • ->Approaches producing high coverage and quality, multi-lingual applicability