Old Wine and Warm Beer: Target-Specific Sentiment Analysis of Adjectives

1. Old Wine and Warm Beer: Target-Specific Sentiment Analysis of Adjectives Author: Angela Fahrni & Manfred Klenner Source: Artificial Intelligence and the Simulation of Behaviour (AISB) 2008

2. Motivation & Goal • Rather than having a prior polarity, adjectives are often bearing a target-specific polarity. • A single adjective even switches polarity depending on the accompanying noun • This study focus on the target-specific polarity determination of adjectives

3. Approach • Propose a two stage model • Identification of domain-specific targets • Construction of a target-specific polarity adjective lexicon by bootstrapping approach • In sentiment analysis, this model outperforms a baseline system that is based on a prior adjective lexicon derived from SentiWordNet

4. Prior polarity • Several resources provide prior polarity • Adjective lists • SentiWordNet • WordNet-Affect • However, the polarity of words is not in any case domain-independent • For example: ‘unpredictable’ • ’unpredictable plot’ in movie domain is a good thing • ’unpredictable boss’ is not • A kind of target-specific sentiment disambiguation seems to be necessary

5. Identify the targets of a domain • Wikipedia’s category system is used to organize the stock of Wikipedia articles • Items in the Wikipedia’s and Wikionary’s category system the used as targets

6. Fast food domain • This study identifies 46,807 targets in fast food domain • Keep the hierarchy in order to propagate polarities • ’cold coca cola’ is positive • then ’cold coca cola cherry’ is positive

7. Vague adjectives • Only a few adjectives do have a prior positive or negative polarity • vague adjectives such as ’big’, ’young’, ’large’, ’deep’ are best understood as bearing neutral prior polarity

8. Three contextual effects • Context acting as intensifiers of the intrinsic positive or negative polarity of a target noun • (e.g. ’deep insight’, ’deep disappointment’) • Combine with a neutral noun to form a non-neutral polarity NP • (e.g. ’old bread’) • A single neutral adjective yields positive or negative polarity depending on the (neutral) noun • e.g. the violation (’cold pizza’) and affirmation (’cold coke’) of intrinsic or common sense properties of target objects (pizza, coke)

9. Inverse effect • Even with the prior polarities • the adjective ’lost’ has a (prior) negative polarity • ’lost virtue’ (virtue=positive) is negative • ’lost glasses’ (glasses=neutral) is negative • ’lost anger’ (anger=negative) is positive

10. Identify the polarity of non-seed adjectives • In the literature, adjectives with a clear prior polarity often been used as a seed list in order to identify the polarity of additional adjectives • The assumption of these approaches was that the augmented list again establishes a set of adjectives having a prior polarity • Contradicting polarities of an adjective encountered in a corpus were interpreted as a kind of noise and are resolved to one (predominant) polarity using statistical measures

11. Building a target specific adjective lexicon by contextual pattern • Using same method Contextual pattern such as coordination to identify the polarity of non-seed • But building a target specific adjective lexicon instead of a domain-independent lexicon

12. Target specific polarity lexicon • Seed adjective lexicon consists of 120 negative and 80 positive adjectives • the polarity is supposed to be domain- and target-independent • Two different corpora are being used • 1600 texts from epinions.com (corpus I) • Tagged • All targets are identified • Because manually choose articles about Fast food, adjectives and targets are relevant to target-specific adjective lexicon • The most frequent targets from corpus I are used to find new texts in corpus II • world wide web (corpus II) • Corpus II is the pool used to identify the polarity of the non-seed adjectives from corpus I with respect to specific targets

13. Target specific polarity lexicon • According to contextual pattern, search both corpora for tag sequences that relate a target and at least two adjectives • the noun or noun sequence is a target • at least one of the adjectives is from the seed list • at least one of the adjectives comes from the stock of target-relevant adjectives

14. Target specific polarity lexicon • Two sequence patterns are considered: • adjective co-ordination • e.g. ’good and tasteful burger’ • Copula constructions, e.g. NP BE Adj Adj+ • e.g. ’the french fries are soggy and rather tasteless’ • Now, assumed that adjectives in such constructions share the polarity • Require a target to be present relative to which the sentiment disambiguation is done • The adjective and target must be of interest according to a reference corpus (corpus I)

15. Polarity tagged pairs • Examples of polarity tagged pairs generated by our systems

16. Polarity values of non-seed adjectives • Polarity values of non-seed adjectives are given as the mean of the polarity values of their peers • Peers: e.g. a seed adjective that occurs together with it in a coordination

17. SentiWordNet • SentiWordNet rely on the seed of paradigm words with a clear polarity • the adjective ’hot’ has 22 senses, • Neutral:7 • Negative:5 • Positive:10 • Only interested in the polarities, so merge the positive, negative and neutral senses into one polarity entry

18. SentiWordNet • the adjective ’hot’ • neutral: 0.6 • positive: 0.28 • negative: 0.12 • Generated an adjective lexicon from SentiWordNet in this way • 21,194 adjective entries has been derived

19. Evaluation • 3,891 manually classified noun phrases • resulting gold standard comprises 1,832 positive, 415 negative and 1,644 neutral instances

20. Three different experimental settings • Compared the polarity decisions of SentiWordNet (our baseline system) and our system for the whole data set (all) • Took only those classifications that received different polarities from the two systems (conflict) • Third, only the instances where both systems agreed in their polarity assignment are taken (agree)

21. Result

22. Result

24. domain-specific polarity • domain-specific polarity: All those adjectives that have a single polarity with all of its targets • If the polarity of an adjective depends on the target, an adjective-target pair is added to the polarity-specific lexicon

25. target-specific polarity of adjectives • The target-specific polarity of adjectives is determined in a corpus-driven manner by searching for combinations of a target-specific adjective with adjectives that have a known prior polarity