Online media, such as blogs and social networking sites, generate massive
volumes of unstructured data of great interest to analyze the opinions and
sentiments of individuals and organizations. Novel approaches beyond Natural
Language Processing are necessary to quantify these opinions with polarity
metrics. So far, the sentiment expressed by emojis has received little
attention. The use of symbols, however, has boomed in the past four years.
About twenty billion are typed in Twitter nowadays, and new emojis keep
appearing in each new Unicode version, making them increasingly relevant to
sentiment analysis tasks. This has motivated us to propose a novel approach to
predict the sentiments expressed by emojis in online textual messages, such as
tweets, that does not require human effort to manually annotate data and saves
valuable time for other analysis tasks. For this purpose, we automatically
constructed a novel emoji sentiment lexicon using an unsupervised sentiment
analysis system based on the definitions given by emoji creators in Emojipedia.
Additionally, we automatically created lexicon variants by also considering the
sentiment distribution of the informal texts accompanying emojis. All these
lexica are evaluated and compared regarding the improvement obtained by
including them in sentiment analysis of the annotated datasets provided by
Kralj Novak et al. (2015). The results confirm the competitiveness of our
approach.

在线媒体产生了大量的非结构化数据，我们提出了一种新的方法来预测在线文本消息中表达的表情符号的情感，并构建了一个情感词典进行评估和比较。

从无监督的情感分析中创建表情符号词库

Creating emoji lexica from unsupervised sentiment analysis of their  descriptions

Unsupervised sentiment analysis is traditionally performed by counting those
words in a text that are stored in a sentiment lexicon and then assigning a
label depending on the proportion of positive and negative words registered.
While these "counting" methods are considered to be beneficial as they rate a
text deterministically, their classification rates decrease when the analyzed
texts are short or the vocabulary differs from what the lexicon considers
default. The model proposed in this paper, called Lex2Sent, is an unsupervised
sentiment analysis method to improve the classification of sentiment lexicon
methods. For this purpose, a Doc2Vec-model is trained to determine the
distances between document embeddings and the embeddings of the positive and
negative part of a sentiment lexicon. These distances are then evaluated for
multiple executions of Doc2Vec on resampled documents and are averaged to
perform the classification task. For three benchmark datasets considered in
this paper, the proposed Lex2Sent outperforms every evaluated lexicon,
including state-of-the-art lexica like VADER or the Opinion Lexicon in terms of
classification rate.

本文提出了一种名为 Lex2Sent 的模型，使用 Doc2Vec 模型训练文档嵌入，并根据嵌入的距离来改进情感词典方法的情感分类任务，该模型在三个基准数据集上的表现均优于一流的词典。