![](https://csdnimg.cn/release/download_crawler_static/15634781/bg4.jpg)
246 C. Wan, Y. Peng and K. Xiao et al. / Information Sciences 519 (2020) 243–259
Fig. 1. The Dependency parsing and the POS tagging result of Example 1 . Note : The POS tags “a”, “n”, “d”, “c”, and “wp” refer to adjective, noun, ad-
verb, conjunction and punctuation, respectively. The dependency labels HED, SBV, ADV and COO indicate head, subject-verb, adverbial and coordinate,
respectively.
To sum up, an important issue shared in existing LDA models is that they rarely consider the semantic relationship
of words at the sentence-level. This situation is even more challenging when handling Chinese review texts. Therefore, our
work focuses on addressing the joint extraction task for aspects and opinions by taking the sentence as the unit. We capture
the associations between aspects and aspects, as well as the relationship between aspects and opinions. We enhance the
LDA model by adding the obtained associations as prior and finally receive an improved cohesion and consistency of the
words on the same topic.
3. Association combinations model and definitions
In this section, we introduce three types of word association combinations: non-aspect opinion words and their corre-
sponding default aspects ( NA associations ), low-frequency opinion words and local aspect words ( LO associations ), and the
secondary aspect words and the local aspect words ( SA associations ). This section also describes the rules to detect the
association combinations and the method for computing-related associative strengths.
3.1. NA associations
In product reviews, many sentences contain non-aspect opinion words, such as “very fashionable” and “really good”
etc.
Although the corresponding aspect words are omitted in these sentences, they may appear in other places, such as
“appearance is very fashionable” and “the camera is really good” etc. Therefore, the association combinations of aspect
words and opinion words in other sentences may be used to identify aspects of the current non-aspect opinion words, and
further affect the topic-words allocation. Note that evidence in support of the regular existence of NA associations can be
found in [26] , in which the linguistic concept of lexical collocation is introduced.
We use ( aspect word, opinion word ) to represent an aspect - opinion combination and propose a set of rules to locate it for
handling non-aspect opinion words. Then, we introduce an equation calculating the associative strengths of aspect - opinion
combinations. Given a non-aspect opinion word, the possible associated aspect word can be found from S
ca
1
the candidate
set of aspect - opinion combinations. Finally, the set of NA associations is built as S
NA
= {( default aspect word, non-aspect opinion
word )}. We propose the following three-step process to obtain the NA associations and their associative strengths.
Step 1: We obtain the set of candidate aspect-opinion combinations .
To achieve the task, we formalize the following rule for aspect-opinion combination discovering, and then we provide an
example to demonstrate the process.
Rule 1. If the POS of < parent node, child node > in dependency relation SBV (i.e., subject-verb) satisfies < adjective, noun > ,
then the noun acts as an aspect word, and the adjective acts as an opinion word.
Example 1. “The screen (is) big and the pixel (is) very high”. The result of dependency parsing and POS tagging of the
sentence is shown in Fig. 1 . The sentence includes two aspect - opinion combinations: (screen, big) and (pixel, high).
In Fig. 1 , letters at the bottom indicate the POS of words. Also, we represent a dependency relationship between words
as: “type of dependency < parent node (POS), child node (POS) > ”. For example, SBV < big(a), screen(n) > . As can be seen, the
aspects and opinions in aspect - opinion combinations in the sentence have obvious modification relationship, and they also
have a fixed relationship in terms of POS, e.g., SBV < big(a), screen(n) > and SBV < high(a), pixel(n) > . Therefore, we can use
the dependency parsing to find the typical sentence structure of aspect - opinion combinations. Then, we apply a dependency
relation, denoted by type of dependency < adjective, noun > , to get a candidate aspect - opinion combination, denoted by (noun,
adjective). We represent the set of these candidate aspect - opinion combinations as S
ca 1
.
Step 2: We identify non-aspect opinion words.
We formalize the following rule to obtain non-aspect opinion words and then provide an example to show the usage of
rule.
Rule 2. If the POS of < parent node , child node > in dependency relation ADV (i.e., adverbial) satisfies < adjective, adverb >
and the adjective is not the parent node of SBV, then the adjective acts as a non-aspect opinion word.