Audio Features - Intelligent Audio Analysis - page 78

a sequence, T-expressions cannot be built, and a fall back strategy is applied: This
second form is the binary expression (B-expression) [ 88 ] and is a co-occurrence of
adjective plus target in the sequence. As an example, the B-expression for “an/DT
excellent/JJ setting/NN” would be . The candidates for targets

are identified from NPs. This stems from the observation of feature terms being
nouns, as in [ 88 ] for sentiment analysis. NPs in the following form are considered
for target identification: NN ; NN, NN ; JJ, NN ; NN, NN, NN ; JJ, NN, NN ; JJ, JJ, NN .
Words of other POS classes (DT, RB) not contributing to the target identification and
punctuation are removed. The personal pronoun “it” is considered as reference to
the subject of the sequence and accordingly used as target.
Next, target sources need to be identified for each target, i.e., words conveying
the actual information of interest such as affect, gender, or personality, etc.To ensure
that a given target source is being directed to the target in question, the search space
needs to be restricted. This can be accomplished by finding border indications that
appear between clauses or phrases within a sequence. These border indications are
subordinating conjunctions, prepositional phrases, coordinating conjunctions, and
punctuation such as commas or colons. The sequence is thus broken down into units
of statements, and a target source is only associated to a given target if and only if
both occur in the same section without a border indication separating them.
In the ongoing, an arbitrary target is exemplified by the concept of valence, such
as in the review of movie critics shown later on. However, this target can be easily
exchanged by other semantic concepts such as personality of a speaker or mood in
music analysis, etc. All verbs and adjectives are selected from the target section,
and General Inquirer is used to determine its value v .Aword w i is assigned a value
v

< setting, excellent >

1ifitis
tagged Negativ . Should a word not exist in General Inquirer, WordNet synsets are help
to lookup its synonyms, until a match is found. Words for which a valence was found
then are the target words. If a target word is an adjective, a B-expression is built from
it and stored in the result set. If it is a verb, its siblings —the direct neighbours—are
first scanned for target adverbs. Given a match, a T-expression of the form

( w i ) = 1 if it has the General Inquirer tag Positiv , and a value v ( w i ) =−

< target,

verb, adverb is generated. If no match was found, the adverb part is left out. Non-

target verbs are processed in the same fashion, and a T-expression is built if a target
adverb is located within its siblings. Thus, e.g., “a/DT carefully/RB designed/VB
plot/NN” , with “designed” not being a target verb, also results in an expression.
According to the POS class of the target word, T-expressions and B-expressions are
generated. It seems intuitive to use the distance between the target word and the target
of an expression as measure for the strength of their relation. A maximum distance
can be enforced to decide upon overall relation [ 89 , 90 ], but [ 91 ] showed that this
can degrade the performance. As an alternative, following [ 92 ], feature-opinion pairs
similar to B-expressions are extracted. Then, the output value per pair is computed
by the multiplicative inverse of the distance between the two words. Then, only the
expression with the shortest distance between the target word and its target is kept
for further processing. By this, one assumes a target word mostly being directed at
the 'closer' target. This choice also reduces the probability of associating a target
word with an unrelated target. Note that multiple expressions can exist per target,

>