Sense Exploring of Emojis with Word2Vec Model and Using Emojis to Modify Short-text Sentiments Classification

step1-Data collection

- period: 2022-0415-2022-0528
- filter: retweet or media
- querry: tweeets that contain at least one concerned emoji

Our dataset consists of Tweets which contain at least one of the following emoji: The baseline selection of common-used emoji refers to Ian D. Wood, Sebastian Ruder, 2016

The tag of emoji is besed on human intuition accroding to the original paper, instead of meaning in context.

step2-Data cleaning

2.1-First EDA

2.2-Data cleaning

For English tweets, we follow the following steps:

1. remove url
2. remove user names
3. remove punctuations
4. remove stopwords
5. lower the words
6. lemmatization
7. keep only english characters and emojis

For Chinese tweets, we:

1. remove url
2. remove user names
3. cut the words
4. remove stopwords (en & zh)
5. transform to simplified Chinese
6. remove punctuations (en & zh)
7. remove sensitive Chinese words
8. remove non-chinese words

2.3-Extract pure emojis and text

And, we generate 3 new coloumns which:

• remove all emojis

---

• include all emojis

---

• only include concerned emojis

For English tweets, we follow the following steps:

1. remove url
2. remove user names
3. remove punctuations
4. remove stopwords
5. lower the words
6. lemmatization

And, we generate 3 new coloumns which:

    - 1) remove all emojis
    - 1) remove non-english words
    - 2) all emojis
    - 3) only concerned emojis

For Chinese tweets, we follow the following steps:

1. remove url
2. remove user names
3. cut the words
4. remove stopwords (en & zh)
5. transform to simplified Chinese
6. remove punctuations (en & zh)
7. remove sensitive Chinese words
8. remove english words

And, we generate 3 new coloumns which:

    - 1) remove all emojis
    - 2) all emojis
    - 3) only concerned emojis

step3-EDA & Word embedding

3.1 data exploring

1. plot the most frequent emojis

3.2 word embedding

1. Word2Vec word embedding with only emoji
2. word embedding with both emoji and words

step4-Sentiment Classification

4.1 Method 1: only text

4.2 Method 2: text + emojis' description names

4.3 Method 3: text + emojis' most similar text tokens

4.4 method 4: text + pre-trained word embedding model