=Paper=
{{Paper
|id=Vol-2130/paper2
|storemode=property
|title=Multi-task Emoji Learning
|pdfUrl=https://ceur-ws.org/Vol-2130/paper2.pdf
|volume=Vol-2130
|authors=Francesco Barbieri,Luís Marujo,Pradeep Karuturi,William Brendel
}}
==Multi-task Emoji Learning==
https://ceur-ws.org/Vol-2130/paper2.pdf
Multi-task Emoji Learning
Francesco Barbieri♣ Luı́s Marujo♥ Pradeep Karuturi♥ William Brendel♥
♣
Large Scale Text Understanding Systems Lab, TALN, UPF, Barcelona, Spain
♥
Snap Inc. Research, Venice, California, USA
♣
{name.surname}@upf.edu, ♥ {name.surname}@snap.com
Over the past few years, the interest in emoji re-
search has increased and several studies has been pub-
Abstract lished in the area of distributional semantics [BRS16,
ERA+ 16, WBSD17b, WBSD17a, BCC18], sentiment
Emojis are very common in social media analysis [NSSM15, HGS+ 17, KK17, RPG+ 18] and
and understanding their underlying seman- multimodal systems [CMS15, CSG+ 18, BBRS18]. In
tics is of great interest from a Natural Lan- the past year researchers also focused on the possi-
guage Processing point of view. In this work, bility of predicting emojis in a text message [BBS17,
we investigate emoji prediction in short text FMS+ 17]. The emoji prediction task consists in pre-
messages using a multi-task pipeline that si- dicting the original emoji present in a tweet (or snap
multaneously predicts emojis, their categories caption) given only the non-emoji textual content.
and sub-categories. The categories are either Prior explorations of emoji prediction tended to focus
manually predefined in the unicode standard on less than 2% of the total number (2653) of emojis
or automatically obtained by clustering over in Unicode 6 standard 1 emojis. Another limitation of
word embeddings. We show that using this those papers was that emoji prediction could be am-
categorical information adds meaningful in- biguous. For example, when the model predicts the
formation, thus improving the performance of correct label to be , emojis like , , , or can
emoji prediction task. We systematically an- also be valid predictions.
alyze the performance of the emoji prediction In this work, we extended the emoji prediction task
task by varying the number of training sam- to 300 emojis in order to study a larger number of
ples and also do a qualitative analysis by using emojis along with their unicode standard categories,
attention weights from the prediction task. sub-categories, and the new semantic clusters that we
created. We are not aware of any previous research
1 Introduction work focused on either predicting a large number of
emojis (300), or using a multi-task approach to predict
Emojis are a popular set of ideograms created in the emojis or emoji categories. We also do a systematic
late 1990s to enrich written communication by adding analysis of how the number of training samples affect
nonverbal expressive power to digital communication. the performance of the emoji prediction task. To mit-
These symbols can be used by human readers to con- igate the problem of emoji ambiguity, we concentrate
vey emotions and information in a condensed form. As on broad emoji category prediction in addition to that
Snapchat, Twitter and other social media platforms of individual emoji prediction. We grouped emojis in
have become popular, so has the usage of emojis. two different ways. The first one was defined by the
Despite their popularity, there is very little research Unicode consortium2 , which groups emojis into seven
work in predicting emojis. categories (e.g., “Smileys & People”, “Nature”) and 74
sub-categories (e.g., “face-positive”, “face-negative”).
Copyright c 2018 held by the author(s). Copying permitted for The main categories are commonly found on mobile
private and academic purposes.
phone keyboards as shown in Figure 1). Alternatively
In: S. Wijeratne, E. Kiciman, H. Saggion, A. Sheth (eds.): Pro-
ceedings of the 1st International Workshop on Emoji Under-
standing and Applications in Social Media (Emoji2018), Stan- 1 www.unicode.org/emoji/charts/full-emoji-list.html
ford, CA, USA, 25-JUN-2018, published at http://ceur-ws.org 2 www.unicode.org/emoji/charts/emoji-ordering.html
Note that we remove skin color from the emojis3 to
avoid generating very similar labels. Table 2 includes
statistics on the two datasets. We can see that Snap
captions are shorter than tweets, while average word
length is similar. Another important difference be-
tween the two datasets is the most frequent emojis
used. Table 1 shows the 60 most frequent emojis in
each dataset (on the top Twitter and on the bottom
Figure 1: Screenshot of Apple’ Emojis keyboard in- Snap data), along with the number of documents that
cluding a subset of Smileys & People category emojis, include each emoji. In both datasets the frequency is
and buttons to access the remaining categories. very unbalanced: 16% of tweets and 25% of snaps
we also created semantic clusters using embeddings. include one of the three most frequent emojis , ,
We use a Multi-task approach to combine the tasks . Therefore we use a balanced dataset in our ex-
of emoji and category prediction. Multi-task ap- periments, in order to give same importance to each
proaches [Car, Car97, Ben12, CW08] improves gen- emoji, independent of the frequency of usage. We sub-
eralization by transferring information across different sample the most frequent emojis in order to match
tasks and improving each task individually. In partic- the same number of examples of the least represented
ular, multi-task learning with simultaneous training emoji (1,500 examples for Twitter and 3,000 for Snap
on multiple tasks has demonstrated promising results data). We show that using fewer than 1,500 examples
[CW08][FMS+ 17][Ben12]. per emoji leads to a drastic decrease in accuracy of
Our work performs multi-task learning by training the emoji detection (see Figure 3). We focus our ex-
a single model with multiple outputs (the dataset is periments on 300 emojis because we do not have more
annotated with multiple labels) and we evaluate us- than 1,500 tweets per emoji beyond the top 300 emojis
ing our gold standard created from Twitter and Snap in our Twitter dataset. For our experiments we ran-
public posts as described in the Datasets section. domly chose 80% of the documents for training, 10%
The subjectivity of emoji interpretation makes emoji for validation and 10% for testing.
prediction a very challenging task. Nevertheless our
2.1 Twitter Dataset
work shows that simultaneously predicting emojis,
their categories, and sub-categories in a multitask The Twitter dataset contains 50 million tweets re-
framework improves the overall results. It not only trieved using Twitter API. Tweets were posted be-
improves emoji prediction, but it also helps with the tween January 2016 and April 2017 and were geo-
identification of emoji categories, which can be partic- localized in the United States. We removed hyperlinks
ularly more relevant when the emoji prediction model from each tweet, and lowercased all textual content in
is less precise. order to reduce noise and sparsity. Since Twitter data
The remainder of this work is organized in the fol- includes a large percentage of bot data, we filter noise
lowing way: The next section describe the datasets as much as possible, removing repeated tweets (or very
used in our experiments. We then present the Deep similar ones) and selected a maximum of five tweets
Learning Models explored to solve our research prob- per user. From this dataset, we selected tweets includ-
lem, Finally, we discuss the experiments, results and ing anyone of the 300 most frequently occuring emojis
then conclude with future research directions. and at least three tokens (without the emoji), resulting
in a final dataset composed of 2,862,505 tweets.
2 Datasets
2.2 SnapCaption
In this study we explore emoji prediction for two dif-
ferent datasets: Twitter and Snapchat captions. We SnapCaption is an in-house Snapchat internal dataset
select documents (tweets and snaps) that contain a containing only Snapchat captions. A caption is the
single emoji, and at least three tokens apart from the textual overlay component of a snap. These cap-
emoji. We restrict to documents containing a single tions were collected exclusively from snaps submit-
emoji so as to minimize the interference of the pres- ted to public and crowd-sourced stories (as known
ence of other emojis in the emoji prediction task. We as Live Stories or Our Stories). Examples of such
also consider only the documents that include the most public crowd-sourced stories are “New York Story” or
frequent 300 emojis in each dataset. We restrict to the “Thanksgiving Story”. All captions were posted in
top 300 emojis only due to lack of meaningful number one year period and do not contain any image or any
of examples beyond that. A subset of the most fre-
quent emojis for each dataset is reported in Table 1. 3 E.g., are mapped to one single label
Table 1: 60 most frequent emoji for the Twitter (top) and Snap (bottom) datasets.
235 112 111 63 52 50 45 41 38 36 35 33 33 31 31 31 29 29 28 28 27 27 27 27 26 24 24 23 23 23
23 22 20 19 19 18 18 18 17 17 16 16 16 16 16 15 15 15 15 15 15 14 14 14 14 13 13 13 13 12
3343 2425 1734 645 617 578 507 433 418 395 391 380 375 364 356 321 315 313 278 273 267 266 234 226 226 225 222 219 212 197
194 194 191 188 187 186 186 181 174 170 169 167 161 157 154 153 153 150 148 147 147 145 139 136 136 134 134 133 124 123
other associated information. This dataset contains
Table 2: Average, standard deviation and Median
30,004,519 captions.
length of words and characters of the two datasets.
Words Chars
2.3 Categories and Clusters of Emojis Dataset Avg. Std. Median Avg. Std. Median
We also consider broader classes of emojis, such as Twitter 12.73 4.23 12 91.14 24.67 92
Snap 5.01 2.29 4 25.62 11.75 23
unicode categories and semantic clusters. The unicode
consortium defines a set of 7 categories categories and
74 sub-categories. 3 Models
The problem with Unicode categories and sub- Our main architecture, illustrated in Fig. (2), starts
categories is that they fail to accurately capture se- with our character and word embedding modules
mantically related emojis. Emojis like and are whose outputs are fused by our feature attention unit
both in the sub-category neutral faces even though and the word attention unit. Finally the fully con-
they clearly indicate different emotions. Another ex- nected layers and the softmax play the role of the final
ample is and that are semantically similar, but multi-task classifier.
they appear in different categories (“Smiling Faces”
Previous approaches [BBS17, FMS+ 17] have suc-
and “Emotions”) even though they address nearly
cessfully learned LSTM models for emoji predic-
identical meanings. To overcome this limitation, we
tion tasks. We experimented different plain LSTMs,
propose a second approach to automatically organize
stacked LSTMs [FMS+ 17], and different word repre-
emojis by clustering them using pre-trained word em-
sentations before solidifying on our final model archi-
beddings similar to emoji2vec [?]. These clusters have
tecture Fig. (2). In addition, we explored single task
the advantage of better capturing the semantic infor-
models and multi-task models. In the case of the
mation of emojis. For example and are in the
multi-task models, the entire network is shared and
same cluster. These clusters are an important aspect
the specialization only occurs at the final stage to pre-
to consider because they are based on how emojis co-
dict specific labels of each task. This specialization is
occur in short text messages from tweets and captions
accomplished through specific linear transformations.
of public snaps. We pretrained two different sets of
Finally we used a cross entropy loss function for all
skip-gram embeddings [MLS13] for Twitter and Snap.
classification tasks. In the case of multitask learning,
The first skip-gram model was trained on a dataset
the final loss is the sum of each single loss4 . In the
of about 70 million tweets and the second skip-gram
following subsections, we detail each stage of our main
model was trained on about 100 million Snap captions.
architecture.
Using the embeddings of the 300 most frequent emo-
jis of each dataset, we created two sets of 30 clusters
using a k-means algorithm. The number of clusters 3.1 Word Representation
was defined based on qualitative analysis (clusters that
seemed to better organize emojis by semantics). In ad- The word embeddings are learned together with the
dition, the number of clusters was selected such that updates to the model. For out-of-vocabulary words
each cluster has a similar number of emojis that are (OOVWs), we used a fixed representation that is han-
usually displayed on a mobile keyword. As a result, we dled as a separate word. In order to train the fixed
would be able to just provide an icon to access directly representation for OOVWs, we stochastically replace
each cluster in a similar way as the Figure 1 shows for (with p = 0.5) each word that occurs only once in the
the top categories. The resulting clusters will group training data. When we use pre-trained word embed-
semantically similar emojis (like in [BKRS16] where dings, that are concatenated with the learned vector.
11 cluster are created for 100 emojis), grouping love,
sad faces, hand/gestures, animals, food, drinks, par- 4 We also experimented weighted sum, with various weights,
ties, Christmas, and so on. but the best results are obtained with a simple sum of the losses.
3.4 Feature Attention
The feature attention module aims to linearly fuse
multiple input signals instead of simply concatenating
them. In our architecture, this module learns a uni-
fied word representation space, i.e. it produces a sin-
gle vector representation with aggregated knowledge
among our multiple input word representations, based
on their weighted importance. We can motivate this
module from the following observations.
Prior work, [BBS17] combines both word rep-
resentation x(w) and character-level representation
x(c) by simply concatenating the word and charac-
ter embeddings at each LSTM decoding step ht =
(w) (c)
LSTM([xt ; xt ]). However, this naive concatena-
tion results in inaccurate decoding, specifically for un-
known word token embeddings, e.g., an all-zero vector
(w) (w)
xt = 0 or a random vector xt = ∼ U (−σ, +σ), or
even for out-of-vocabulary words. While this concate-
nation approach does not cause significant errors for
Figure 2: Final architecture of our model well-formatted text, we observe that it induces perfor-
3.2 Char Representation mance degradation for our social media post datasets
which contain a significant number of slang words, i.e.,
In addition, we use a character based embedding misspelled or out-of-vocabulary words. As a result,
[LLM+ 15, ?] stacked with a B-LSTM [GS05], produc- we use a feature attention module, that adaptively
ing a character-based word embedding that focuses on emphasizes each feature representation in a global
word spelling variants. Indeed, the character-based manner at each decoding step t. This process pro-
word embedding learned similar representations for duces a soft-attended context vector xt as an input
words that are orthographically similar, and thus are token for the next stacked B-LSTM that takes care
expected to handle different alternatives of the same of the sentences embedding. [RCP16] introduced a
word types that normally occur in social media. similar approach, where the character embeddings are
weighted with an attention module. We use the fol-
3.3 Bi-directional LSTMs lowing method:
Our bi-directional LSTM modules, [GS05] named B- (w) (c) (w) (c)
LSTM in Fig. (2), consists of a forward LSTM that [at , at ] = σ Wm · [xt ; xt ] + bm
processes an input message from left to right, while the (m)
(m) exp(at ) (1)
αt = ∀m ∈ {w, c}
backward LSTM processes it in the reverse direction. P (m0 )
exp(at )
As a result, the message representation s is based on m0 ∈{w,c}
both the forward and backward LSTM encoding:
(w) (c)
where αt = [αt ; αt ] ∈ R2 is an attention vector at
s = max {0, W[hfw ; hbw ] + d} each decoding step t, and xt is a final context vector
at t that maximizes information gain for xt . Note that
where W is a learned parameter matrix, fw is the for-
this feature attention requires each feature representa-
ward LSTM encoding of the message, bw is the back- (w) (c)
ward LSTM encoding of the message, and d is a bias tion to have the same dimension (e.g. xt , xt ∈ Rp ),
term, and we use a component-wise ReLU as the non- and that the transformation via Wm essentially en-
linear unit. We use B-LSTM modules for both word forces each feature representation to be mapped into
and sentence representations, namely Char B-LSTM the same unified subspace, with the output of the
and Words B-LSTMs in our architecture Fig. (2). Char transformation encoding weighted discriminative fea-
B-LSTM takes a sequence of characters and outputs tures for classification of emojis.
a word embedding vector. This output is mixed with
another word representation via our feature attention 3.5 Word Attention
module. Then, the stacked Words LSTMs receive se-
quences of word representations from the attention Not all the words have the same importance in the
module, and output sentence embedding vectors. representation of a document. We use the attention
mechanism introduced in [YYD+ 16]: 36
uit = tanh(Ww hit + bw ) 34
exp(u>it uw ) (2)
Top 5 Accuracy
X
αit = P ; di = αit hit 32
>
t exp(uit uw ) t
30
where the final document representation di is a lstm
weighted average of the hidden representation hi t of 28 lstm + att
the LSTM. The weights αit are learned by the use of a 2 lstm + att
26 char + 2 lstm + att
Multi-Layer Perceptron (linear transformation W and
biases b) with tanh as non-linear operation, and a soft- 24 char + lstm
char + lstm + att
max to compute the probability of each word.
22
100 500 1,000 1,500 2,000 2,500 3,000
4 Experiments And Results Training examples per class
We use two main variations for experiments: Single-
Task Prediction of emojis, unicode categories, and Figure 3: Acc@top 5 of the same algo. but variable
emoji clusters, and Multi-Task Prediction, where nr. of training instances per class (from 100 to 3000
we combine the single tasks in one single model. We examples for each emoji) on SnapCaptions. Test and
also evaluate the impact of our different modules in- validation set are fixed for each experiment.
cluding the combination of word/char LSTMs and the
Table 3). There are several reasons that could explain
word attention unit. Finally we investigate the influ-
this difference in results. One reason is the length of
ence of the number of layers for the LSTMs.
the text messages, since in Twitter there are on aver-
age twelve words per message, while on Snap has only
4.1 Single-Task Prediction five (see Table 2). Another reason could be the miss-
We explore three different tasks: (i) the emoji pre- ing visual context of Snap posts5 , while only a small
diction task proposed by [BBS17], (ii) prediction of percentage of tweets is complemented with a visual
unicode emoji categories (the emoji in the text be- content. For this reason, tweets contain typically less
long to the faces, animal, objects) and sub-categories semantic ambiguity.
(positive faces, animal-mammal), and (iii) prediction Table 3 highlights the best performing systems on
of automatic clusters that we previously generated us- the emoji prediction task. For the two datasets state of
ing pre-trained word embeddings. the art systems are outperformed by the combination
of additional components. For example, adding a word
4.1.1 Predicting Emojis attention module improves the baseline of [BBS17].
Finally, there is an important difference when predict-
Given a set of documents, each document contain- ing 20 and 300 emojis. We plot on the left of Figure 3
ing only one emoji class, the task consists of pre- the accuracy of same model architecture (Char + 2
dicting the emoji from the text. For this task, we LSTM + word attention) on the emoji prediction task
tested the influence of the number of emoji classes for different numbers of emojis (20 to 300). Best ac-
and the number of examples per class. More pre- curacy at top 5 (a@5) drops from 20 to 100, and then
cisely, for each experiment, we extract a balanced remains constant. We observe the same drop using F1
dataset of Nclass emoji classes, and Ndata examples per (that only considers whether an emoji is predicted as
class, with Nclass = {20, 50, 100, 200, 300 and Ndata first option), however, having more than 100 classes
= {100, 500, 1000, 1500, 2000, 2500, 3000}. Nclass and results in improvement. This is probably due the type
Ndata are tested independently: when we vary Nclass , of the more rare emoji classes added after the most
we fix Ndata to 3000, and when we vary Ndata we fix 100 frequent ones, that are more specific (like , ,
Nclass to 300. Figure 3 shows our experiments with the or ) hence easier to predict.
Snapchat dataset. It is clear that using more examples
per class improves our model by around 1% absolute 4.1.2 Predicting Unicode Emoji Categories
point from 1500 to 3000 examples. For >2000 exam- and Sub-categories
ples the system converges to its optimum.
From Figure 4, we observe that Twitter data is eas- We predict Unicode emojis categories and sub-
ier to model than Snap data. In the 300 emoji predic- categories using the text message that includes only
tion task the best accuracy at top 5 (a@5) on Twitter 5 Snap text messages are captions of images and videos posted
data is 40.05% while on Snap data it is 34.25% (see by Snapchat users, see Datasets section.
Table 3: Emoji prediction results using multiple number of emojis (20 to 300) and different models. We use the
original implementation of [FMS+ 17], while we implement [BBS17].
20 50 100 200 300
Dataset Models F1 a@5 F1 a@5 F1 a@5 F1 a@5 F1 a@5
Twitter LSTM 25.32 62.17 19.51 46.69 18.03 40.49 19.44 39.11 19.86 38.44
LSTM + Att. 25.51 62.82 19.49 47.21 18.3 40.56 19.61 39.10 19.38 38.41
2 LSTM + Att. [FMS+ 17] 24.47 62.38 19.64 47.01 18.36 40.60 19.60 39.09 20.09 38.70
Char + LSTM [BBS17] 26.81 63.37 20.21 47.83 18.87 41.41 20.49 40.14 21.27 40.06
Char + LSTM + Att. 27.37 64.33 20.91 48.23 18.88 41.8 21.19 40.65 21.59 40.06
Char + 2 LSTM + Att. 26.85 64.12 20.36 48.51 18.82 41.92 20.66 40.51 20.53 39.22
Snap LSTM 25.46 53.51 18.62 43.96 15.06 34.72 16.08 32.96 17.57 32.44
LSTM + Att. 25.58 53.67 18.86 44.44 15.40 35.20 16.47 32.95 17.46 32.63
2 LSTM + Att. [FMS+ 17] 24.30 53.01 18.64 43.59 15.40 35.03 16.73 33.26 18.07 32.94
Char + LSTM [BBS17] 24.84 53.37 19.26 44.75 15.50 35.38 17.39 33.89 18.80 33.98
Char + LSTM + Att. 26.01 54.34 19.39 45.1 15.58 35.61 17.44 34.26 18.64 33.86
Char + 2 LSTM + Att. 25.72 53.81 18.95 45.05 16.18 36.03 17.51 33.97 18.87 34.25
65 30 Table 4: Results for single and multi-task predic-
Top 55 acc:
Top acc: Twitter
Twitter tion of emojis including main unicode categories, sub-
60 Top 55 acc:
Top acc: Snap
Snap categories, and clusters.
Macro F1: Twitter Pred. Twitter Snap
Task Loss F1 A@5 F1 A@5
Top 5 Accuracy
55 Macro F1: Snap 25 Macro F1 Main Main 46.56 84.70 45.23 87.90
Category Main + Sub 48.34 85.87 45.07 87.79
50 Main + Emoji 48.17 85.52 44.54 87.83
Main + sub + Emoji 48.52 85.90 44.64 95.52
45 Sub Sub 31.62 51.02 32.15 53.81
20 Category Sub + Main 31.84 51.86 31.72 53.43
40 Sub + Emoji 32.00 52.23 31.99 53.72
Sub + Main + Emoji 32.24 52.40 31.88 65.19
35 Semantic Clusters 34.10 53.56 34.77 53.22
Clusters Clusters + Emoji 35.42 55.90 34.90 53.64
15 Emoji Emoji 21.59 40.06 18.64 33.86
20 50 100 200 300
Emoji + Main 21.62 37.80 19.05 34.24
Number of labels Emoji + Sub 21.58 37.91 18.75 34.27
Emoji + Main + Sub 21.44 37.81 18.78 34.05
Figure 4: F1 and Acc. @ top 5 for the model “Char Emoji + Clusters 21.30 37.90 19.05 29.78
+ 2 LSTM + Word Att.” on Twitter and Snap data.
sub-categories of Unicode Standard.
one emoji as we did in the emoji prediction task.
Table 4 shows the prediction results using macro- 4.2 Multi-Task Predictions
F1 and a@5 evaluation metrics. In the first two blocks
(main and sub lines), we predict the main category and In Table 4 we show the multi-task prediction re-
sub-category respectively. The third block details the sults. We considered multiple multi-task combina-
clusters’ evaluation results, and the last block presents tions. Learning more than one objective task simul-
the emoji prediction results. In the first line of each taneously helps in the main category prediction, as
block are the single-task results and the remaining macro F1 improves from 46.56% to 48.52% (4.2% rela-
lines include the ones using a multi-task framework. tive improvement) when adding also sub-category and
emoji losses. Sub-categories prediction also improves
when it is learned together with main categories and
emojis.
4.1.3 Predicting Clusters
On Snap data, category and sub-category predic-
Given a text message containing an emoji e we pre- tion tasks do not improve using a multitask approach
dict the cluster that emoji e belongs to. Cluster cre- in terms of macro F1, but we obtain a relative im-
ation is described in the dataset section. Cluster re- provement of 8.67% and 21.14% using a@5.
sults are reported in Table 4, in the lines correspond- The clusters prediction tasks also benefit from
ing to “Semantic Clusters”. The results are better on multi-task learning when combined with the emoji pre-
Snap than Twitter for broader classes and our clus- diction. However, emoji prediction seems not to im-
ters capture better semantics than the categories and prove much in a multi-task setting for Twitter. Emoji
G: P: 0.26 , 0.15 , 0.13 , 0.10 , 0.04
we are having a cat party in room 135 #austinblueroos #bethedifference
we are having a cat party in room 135 #austinblueroos #bethedifference
G: P: 0.75 , 0.02 , 0.02 , 0.01 , 0.01
Feeling slightly pregnant but it was worth it
Feeling slightly pregnant but it was worth it
G: P: 0.97 , 0.01 , 0.002 , 0.001 , 0.001
It’s official , I have found my #unicorn ! Look at this geometric tribal print !
It’s official , I have found my #unicorn ! Look at this geometric tribal print !
Figure 5: Word and Feature attention visualization. The first line highlights in blue word attention, while the
second line shows the feature attention. Uncolored words mean almost zero attention over them.
have a specific meaning and become difficult to model.
Table 5: Top and bottom 10 emojis with best accuracy
on the Twitter (top) and on Snap (bottom).
4.3.2 Feature and Word Attention
87.5 83.87 83.53 83.33 81.33 80.88 79.86 79.86 79.56 78.95 We previously described the two types of attention ex-
plored. The Feature Attention approach gives more
4.71 4.57 4.43 3.97 1.94 1.94 1.52 1.24 0.63 0.58
importance to either the character or word represen-
89.82 78.27 77.19 76.15 75.72 74.1 73.48 73.03 71.5 71.46
tation of a word. The Word Attention approach in-
creases the importance of more discriminative words,
3.46 3.37 3.13 2.92 2.82 2.41 1.86 1.01 0.29 0 for example the word “key” to predict the emoji .
Figure 5 visualizes the weights of each of these two
prediction on Snap improves from 33.86% to 34.27% attention modules using three example messages. For
or 1.21% relative improvement in terms of a@5 when each of them, we list the gold label (“G”) and the pre-
it is learned together with Unicode sub-categories. dicted labels (“P”), along with their prediction proba-
bility. i.e. the output of the softmax layer. The inter-
4.3 Qualitative Analysis nal weights of the two attention modules are visualized
We analyzed in detail our emoji prediction approach using text highlights. Darker color indicates more at-
(char + 2 LSTM + attention) based on the best per- tention over word (αit from Formula 2 of each word in
forming system described in the previous section. This the message). In second line of each message the red
analysis enumerates the emojis that are easier and highlight shows the weights of the feature attention (α
harder to predict. We also include some visualization of Formula 1). Bold text formatting indicate the out
examples of where the attention module obtains more of vocabulary words.
information. These examples provide us with a better Based on the three examples, and some additional
understanding of the importance of the character and that we manually evaluated, we verified how these two
word features in our results. attention approaches work. The Word Attention mod-
ule (blue highlight) give us insights on the recognition
of emojis. In the first example the most important
4.3.1 What emoji is difficult to predict?
word is “cat” and the predictions are indeed about
Table 5 shows a list of the top and bottom 10 emo- cats, apart from the fifth predicted emoji . This
jis based on the prediction accuracy. We investigated emoji is triggered (probably) because of the presence
what emojis are difficult to predict, and we found in- of the token “135” as the word attention module also
teresting patterns. As expected, the emojis that are focuses on this token. In the second example, the at-
easier to predict describe specific objects without mul- tention goes to the word “pregnant”, but in this case
tiple meanings (such as, and ) or topics (e.g., this word misleads the network that incorrectly pre-
and ). These emojis, as suggested in [BRS16], could dicts baby emojis. However, the correct emoji is then
easily be replaced by a word, such as by key), or are predicted as fourth option. In the last example, the
used when specific words occur in a text message in- network correctly classifies the emoji , based on the
cluding Christmas for and ). In both datasets, hashtag “#unicorn”.
subjective emojis including and obtained low- Regarding the Feature Attention over the word or
est accuracy values. These subjective emojis describe character representation of each token in a message,
emotional information, and they can be interpreted we observed that the character representation seems
differently among different users and based on the sur- to gain importance on long and less frequent tokens,
rounding context. Hence, these emojis do not seem to namely numbers, hashtags, and as expected, out of
vocabulary words (“135” and “#austinblueroos”). Modifiers Affect Emoji Semantics in
Twitter. In Proceedings of the 7th
5 Conclusion Joint Conference on Lexical and Com-
putational Semantics (*SEM 2018), New
In this paper, we explored emoji prediction in two so-
Orleans, LA, United States, 2018.
cial media platforms, Twitter and Snapchat. We ex-
tended the emoji prediction task to a large number of [Ben12] Yoshua Bengio. Deep learning of repre-
emojis and showed that the prediction performance sentations for unsupervised and transfer
drastically drops between 50 and 100 emojis, while learning. In ICML Workshop, 2012.
the addition of more emojis keeps the accuracy of the
model somehow constant (even if it has to predict more [BKRS16] Francesco Barbieri, German Kruszewski,
emojis). We attribute these results to the specificity Francesco Ronzano, and Horacio Sag-
of the less-used emojis. We also proposed a novel task gion. How Cosmopolitan Are Emojis?
that predicts broader classes of emojis, grouping emo- Exploring Emojis Usage and Meaning
jis in automatic clusters or predefined categories, as de- over Different Languages with Distribu-
fined by the Unicode consortium. These new tasks al- tional Semantics. In Proceedings of the
low us to better evaluate the predictions of the model, 2016 ACM on Multimedia Conference,
since plain emoji prediction may be ambiguous. We pages 531–535, Amsterdam, Netherlands,
also carried out an extensive qualitative analysis in or- October 2016. ACM.
der to understand the importance of the character en-
[BRS16] Francesco Barbieri, Francesco Ronzano,
coding of words in noisy social media text, the number
and Horacio Saggion. What does this
of training examples, and the difficulties in modeling
emoji mean? a vector space skip-gram
specific emojis.
model for t.emojis. In LREC, 2016.
Finally, we proposed a multi-task approach to pre-
dict emojis and emoji group affiliation at the same [Car] R. Caruana. Multitask learning: A
time. We showed that the model obtains significant knowledge-based source of i.b. In
improvements in the Twitter dataset, while more in- ICML’93.
vestigation is needed for the Snapchat dataset.
[Car97] Rich Caruana. Multitask learning. Mach.
Acknowledgments Learn., 28(1):41–75, July 1997.
This work was done when Francesco B. interned at [CMS15] Spencer Cappallo, Thomas Mensink, and
Snap Inc. Francesco B. acknowledge support also Cees GM Snoek. Image2emoji: Zero-shot
from the TUNER project (TIN2015-65308-C5-5-R, emoji prediction for visual media. In Pro-
MINECO/FEDER, UE) and the Maria de Maeztu ceedings of the 23rd ACM international
Units of Excellence Programme (MDM-2015-0502). conference on Multimedia, pages 1311–
1314. ACM, 2015.
References [CSG+ 18] Spencer Cappallo, Stacey Svetlichnaya,
[BBRS18] Francesco Barbieri, Miguel Ballesteros, Pierre Garrigues, Thomas Mensink, and
Francesco Ronzano, and Horacio Sag- Cees GM Snoek. The new modality:
gion. Multimodal emoji prediction. In Emoji challenges in prediction, antici-
Proceedings of NAACL: Short Papers, pation, and retrieval. arXiv preprint
New Orleans, US, 2018. Association for arXiv:1801.10253, 2018.
Computational Linguistics.
[CW08] R. Collobert and J. Weston. A unified
[BBS17] Francesco Barbieri, Miguel Ballesteros, architecture for natural language process-
and Horacio Saggion. Are emojis pre- ing: Deep neural networks with multitask
dictable? In Proceedings of the 15th Con- learning. In ICML. ACM, 2008.
ference of the European Chapter of the
Association for Computational Linguis- [ERA+ 16] Ben Eisner, Tim Rocktäschel, Isabelle
tics: Volume 2, Short Papers, pages 105– Augenstein, Matko Bosnjak, and Sebas-
111, Valencia, Spain, April 2017. Associ- tian Riedel. emoji2vec: Learning emoji
ation for Computational Linguistics. representations from their description.
In Proceedings of The Fourth Interna-
[BCC18] Francesco Barbieri and Jose Camacho- tional Workshop on Natural Language
Collados. How Gender and Skin Tone Processing for Social Media, pages 48–54,
Austin, TX, USA, November 2016. Asso- [WBSD17a] Sanjaya Wijeratne, Lakshika Balasuriya,
ciation for Computational Linguistics. Amit Sheth, and Derek Doran. Emojinet:
An open service and api for emoji sense
[FMS+ 17] Bjarke Felbo, Alan Mislove, Anders
discovery. International AAAI Confer-
Søgaard, Iyad Rahwan, and Sune
ence on Web and Social Media (ICWSM
Lehmann. Using millions of emoji oc-
2017). Montreal, Canada, 2017.
currences to learn any-domain represent.
for detecting sentiment, emotion and [WBSD17b] Sanjaya Wijeratne, Lakshika Balasuriya,
sarcasm. In EMNLP, 2017. Amit Sheth, and Derek Doran. A
semantics-based measure of emoji simi-
[GS05] Alex Graves and Juergen Schmidhuber.
larity. Web Intelligence, 2017.
Framewise phoneme classification with
bidirectional lstm and other neural net- [YYD+ 16] Zichao Yang, Diyi Yang, Chris Dyer,
work architectures. Neural Networks, 18, Xiaodong He, Alex Smola, and Eduard
2005. Hovy. Hierarchical attention networks for
[HGS+ 17] Tianran Hu, Han Guo, Hao Sun, Thuy- document classification. In Proceedings of
vy Thi Nguyen, and Jiebo Luo. Spice the 2016 Conference of the North Ameri-
up Your Chat: The Intentions and Sen- can Chapter of the Association for Com-
timent Effects of Using Emoji. Proc. of putational Linguistics: Human Language
ICWSM 2017, 2017. Technologies, pages 1480–1489, 2016.
[KK17] Mayu Kimura and Marie Katsurai. Au-
tomatic construction of an emoji sen-
timent lexicon. In Proceedings of the
2017 IEEE/ACM International Confer-
ence on Advances in Social Networks
Analysis and Mining 2017, pages 1033–
1036. ACM, 2017.
[LLM+ 15] W. Ling, T. Luı́s, L. Marujo, R.F. As-
tudillo, S. Amir, C. Dyer, A.W. Black,
and I. Trancoso. Finding function in
form: Compositional character models
for open vocabulary word representation.
EMNLP, 2015.
[MLS13] Tomas Mikolov, Quoc V Le, and Ilya
Sutskever. Exploiting similarities among
languages for machine translation. arXiv
preprint arXiv:1309.4168, 2013.
[NSSM15] Petra Kralj Novak, Jasmina Smailović,
Borut Sluban, and Igor Mozetič.
Sentiment of emojis. PloS one,
10(12):e0144296, 2015.
[RCP16] Marek Rei, Gamal KO Crichton, and
Sampo Pyysalo. Attending to characters
in neural sequence labeling models. In
Coling, 2016.
[RPG+ 18] David Rodrigues, Marı́lia Prada, Rui
Gaspar, Margarida V Garrido, and Di-
niz Lopes. Lisbon emoji and emoti-
con database (leed): Norms for emoji
and emoticons in seven evaluative dimen-
sions. Behavior research methods, pages
392–405, 2018.