Information Technology Reference
In-Depth Information
to classify English news quotations, it is an important requirement to develop
approaches that are applicable with as little effort to other languages, since the authors
incorporate their results into the Europe Media Monitor (EMM) news engine.
18
EMM
collects and processes news articles from multilingual news sources. In [
2
] Balahur
and Steinberger present reflections on how sentiment analysis applied to news arti-
cles differs from sentiment analysis on highly subjective texts like online reviews.
The authors find that the task of sentiment analysis on news articles can be decom-
posed into three subtasks: determining the sentiment target, distinguishing between
“good and bad news content” and “good and bad sentiment expressed on the target”,
and classifying explicitly expressed sentiments on the sentiment target that does not
require any world knowledge or the interpretation by the reader. The inter-annotator
agreement increases if the task is clearly defined in advance. Finally, the authors work
out three possible perspectives on news articles: the author's (news bias research), the
reader's (interpretation by readers influenced by their backgrounds), and the text's
view. Each view requires a different approach to sentiment analysis and the authors
limit their work to identifying sentiments concretely expressed in the text. In [
3
]
Balahur et al. provide a comparison of different sentiment resources and classifi-
cation strategies to categorize news quotations as positive, negative, and neutral. In
their studies, the authors also examine how a preceding subjectivity detection step
affects the classification results. They conclude that using large sentiment lexicon and
a previous subjectivity filtering improves the results considering vocabulary-based
methods. Straightforward bag-of-words approaches are limited and not effective
enough for sentiment analysis on news quotations. The authors also conclude that
exploiting sentiment annotations based on single topics are not suitable for the open-
domain sentiment analysis on news. Thus, they propose a topic-dependent sentiment
analysis with specialized models.
The work of Balahur et. al in [
4
] analyzes two aspects of sentiment analysis
for English-language quotations in news articles. First, the authors examine how
different word windows around an opinion target influence sentiment classification
accuracy. Second, they exclude sentiment-bearing words that are category-specific
words at the same time, in order to separate good or bad news content from positive
and negative sentiments toward the opinion targets. The sentiment score is calculated
by summing up the sentiment scores of all quotation words. As a result, the authors
argue that taking into account only a word window around the target entity instead
of including the entire quotation text yields better results. Considering the lexicons,
the authors find that there are large differences between their performances and that
combining them helps. However, the accuracy of the approach does not only depend
on a large lexicon.
Sentiment analysis on German-language texts has been applied in [
27
,
31
,
46
].
Momtazi presents a rule-based approach to classify sentiments toward celebrities
mentioned in short German social media texts. In order to label the short texts as
positive or negative and assign the strength of the sentiment, the author creates
and applies a sentiment dictionary and a list of booster and negation words. Mom-
18
Search WWH ::
Custom Search