Information Technology Reference
In-Depth Information
speech indicator in a wrong context. By automatically determining reporting verbs
and phrases we aim to improve the recall of indirect quotations. We also intend to
complement our co-reference resolution approach with a state-of-the-art component.
We want to determine co-reference chains and then choose the correct one as quo-
tation speaker [
39
]. In order to consolidate quotations uttered by one speaker across
different news documents, we plan to link speakers to Wikipedia entries by apply-
ing a named entity disambiguation approach. Our future work will also include the
incorporation of supervised approaches. On the one hand we plan to treat quotation
extraction as a sequence labeling task and on the other hand we plan to train a binary
classifier that predicts quotations at sentence level. Our goal is to identify appropriate
features for German-language texts and to let an ensemble combine the output of the
rule-based approach with the output of the new classifiers.
1.4 Sentiment Analysis
Publicly available texts such as product reviews, social media contributions, or news
articles discuss almost every thinkable entity and topic. Besides transporting facts,
the texts often cover opinions as well, and even more than facts, the expressed opin-
ions may influence readers. Regardless of whether someone wants to buy a new
camera or wants to find out which party to vote in the next election, people in general
read first what other people think and what experiences they have had before making
their own decisions. That is the reason why companies are interested in a positive
perception of their products and services in the media. Here, well-analyzed opinion-
ated texts may serve as a basis for a multitude of sentiment-related applications like
reputation monitoring or opinion summarization systems. Sentiment analysis may
also be performed to improve other natural language processing tasks that rely on
factual data. Separating opinionated text from objective text turned out to be bene-
ficial for information extraction [
43
]. In this section we show how opinions can be
extracted from news articles. We focus our work on news articles because they are
a reliable information source and mirror the opinions of newsworthy people, which
often serve as role models. We first introduce the term “sentiment analysis” and
“opinion mining” and then present our comprehension of “opinions”. The main part
describes our supervised approach to sentiment analysis. We limit our approach to
quotations because we assume that quotations are the most subjective parts of news
articles.
1.4.1 Introduction
Sentiment analysis aims at identifying subjective language in texts and determining
the orientation and strength of expressed opinions or sentiments toward the corre-
sponding targets. Often the term “opinion mining” denotes the same task and is
Search WWH ::
Custom Search