Literature Review: Predicting the Semantic Orientation of Adjectives

Sentiment Polarity using Adjective

Hatzivassiloglou et al (1997)¹ identified and validated from a large corpus constraints from conjunctions (such as and, but) on the positive or negative semantic orientation of the conjoined adjectives.

Orientation (Polarity) = direction of deviation from the norm

Approach

• Indirect information:
  • Adjectives conjoined by “and” have same polarity, such as “simple and well-received”;
  • Adjectives conjoined by “but” have different polarity, such as “simplistic but well-received”;

Green line represents and, Red dashed line represents but.²

Data collection

21 million word 1987 Wall Street Journal corpus, automatically annotated with part-of-speech tags using the PARTS tagger.

Adjectives data preparation

• Construct a set of adjectives with predetermined orientation labels by taking all adjectives appearing in the corpus 20 times or more and removing adjectives that have no orientation;
• Assign an orientation label (either + or -) to each adjective, using an evaluative approach
  • Criterion: whether the use of this adjective ascribes in general a positive or negative quality to the modified item, making it better or worse than a similar unmodified item.
• Final set contained 1,336 adjectives (657 positive and 679 negative terms).

Adjectives data validation:

They subsequently asked 4 people to independently label a randomly drawn sample of 500 of these 1,336 adjectives, who agreed with us that the positive/negative concept applies to 89.15% of these adjectives on average.

Extract conjunctions between adjectives

By using two-level finite-state grammar, 13,426 conjunctions of adjectives, expanding to a total of 15,431 conjoined adjective pairs are collected.

Test data

15,048 conjunction tokens involve 9,296 distinct pairs of conjoined adjectives (types).

Each conjunction token is classified by the parser according to three variables:

• the conjunction used (and, or, but, either-or, or neither-nor)
• the type of modification (attributive, predicative, appositive, resultative)
• the number of the modified noun (singular or plural)

Validation of the Conjunction Hypothesis

Prediction of link type

• Prediction method 1 - Always predict same orientation:
  • always guessing that a link is of the same- orientation type
• Prediction method 2 - But rule:
  • Method 1 + using but exhibit the opposite pattern
• Prediction method 3 - Log-linear model:
  • $\eta = \mathbf{\omega}^\mathbf{T} \mathbf{x}$, $y = \frac{e^\eta}{1 + e^\eta}$
  • $\mathbf{x}$: the vector of the observed counts in the various conjunction categories
  • $\mathbf{\omega}$: the vector of weights to be learned
  • $y$: the response of the system
  • Using the method of iterative stepwise refinement they selected 9 predictor variables from all 90 possible predictor variables
• Morphological relationships:
  • Adjectives related in form almost always have different semantic orientations
  • Highly accurate (97.06%), but applies only to 1,336 labeled adjectives (891,780 possible pairs)
  • E.g. adequate-inadequate, thoughtful-thoughtless

Cluster

Input

A graph of adjectives connected by dissimilarity links.
Dissimilarity value $d(x, y)$ between 0 and 1:

• Small $d(x, y)$ $\Rightarrow$ same-orientation link between $x$ and $y$
• High $d(x, y)$ $\Rightarrow$ different-orientation link between $x$ and $y$

To partition the graph nodes into subsets of the same orientation, we employ an iterative optimization procedure on each connected component, based on the exchange method, a non- hierarchical clustering algorithm.

Objective function $\Phi$ scoring each possible partition $\mathcal{P}$ of the adjectives into two subgroups $C_1$ and $C_2$ as

$$\Phi (\mathcal{P}) = \sum_{i=1}^{2} (\frac{1}{\left\lvert C_i \right\rvert} \sum_{x,y \in C_i, x \neq y} d(x,y))$$

where $C_i$ stands for the cardinality of cluster $i$

Labeling the Clusters as Positive or Negative

The unmarked member almost always having positive orientation (Lehrer, 1985; Battistella, 1990). Thus:

They computed the average frequency of the words in each group, expecting the group with higher average frequency to contain the positive terms.

Conclusion

They tested how graph connectivity affects the overall performance

---

¹. Hatzivassiloglou, Vasileios, McKeown, Kathleen R (1997). Predicting the semantic orientation of adjectives, 174—181 ↩

². 7 - 4 - Learning Sentiment Lexicons - Stanford NLP - Professor Dan Jurafsky & Chris Manning ↩