Machine Floriography: Sentiment-inspired Flower Predictions
over Gated Recurrent Neural Networks
Avi Bleiweiss
BShalem Research, Sunnyvale, U.S.A.
Keywords:
Language of Flowers, Gated Recurrent Neural Networks, Machine Translation, Softmax Regression.
Abstract:
The design of a flower bouquet often comprises a manual step of plant selection that follows an artistic style
arrangement. Floral choices for a collection are typically founded on visual aesthetic principles that include
shape, line, and color of petals. In this paper, we propose a novel framework that instead classifies sentences
that describe sentiments and emotions typically conveyed by flowers, and predicts the bouquet content impli-
citly. Our work exploits the figurative Language of Flowers that formalizes an expandable list of translation
records, each mapping a short-text sentiment sequence to a unique flower type we identify with the bouquet
center-of-interest. Records are represented as word embeddings we feed into a gated recurrent neural-network,
and a discriminative decoder follows to maximize the score of the lead flower and rank complementary flower
types based on their posterior probabilities. Already normalized, these scores directly shape the mix weights
in the final arrangement and support our intuition of a naturally formed bouquet. Our quantitative evaluation
reviews both stand-alone and baseline comparative results.
1 INTRODUCTION
Communicating through the use of flower meanings
to express emotions, also known as Floriography, has
been a traditional and well established practice around
the world for centuries. Nonetheless, endorsing this
coded exchange as a Language of Flowers only gai-
ned traction during the Victorian era, and was backed
by publishing a growing compilation body of floral
dictionaries that explained the meanings of flowers.
Shortly thereafter, the language of flowers was fa-
vored as the prime medium to send secret messages
otherwise prohibited in public conversations. Flori-
ography was not only about the simple emotion atta-
ched to an individual flower, but rather what portrayed
in a combination of petals and thrones placed in an
arranged bouquet. The language had since develo-
ped considerably, and today several online resources
(Roof and Roof, 2010; Diffenbaugh, 2011) provide
updated flower interpretations and most authentic sen-
timent transcriptions.
In this research, we investigate the linguistic pro-
perties of the language of flowers using unsupervi-
sed learning of word vector representations (Mikolov
et al., 2013a; Pennington et al., 2014), and modeling
the language after neural machine translation that pre-
dicts a definitive flower type given a sentiment phrase
as input. Furthermore, we extend the single target
perspective of the language and relate the short-text
sentiment sequence to a plurality of flowers that com-
bine both a principal or pivotal flower, with statisti-
cally ranked subordinate flowers to form a bouquet.
Recurrent Neural Networks (RNN) recently be-
came a widespread tool for language modeling tasks
(Sutskever et al., 2014; Hoang et al., 2016; Tran
et al., 2016). In our case study, we feed sequentially
concatenated translation records into a shallow RNN
architecture that consists of an input, hidden, and
output layers (Elman, 1990; Mikolov et al., 2010).
At every time step, the output probability distribu-
tion over the entire language vocabulary renders our
framework for an automatic selection of sentiment-
aware flower species that requires minimal human
counseling. We ran experiments on both a standard
RNN that applies a hyperbolic activation function di-
rectly and through a gated recurrent unit (GRU) (Cho
et al., 2014; Chung et al., 2014), and confirmed GRU
to better sustain vanishing propagating gradient (Ho-
chreiter and Schmidhuber, 1997) and improve our re-
call performance.
The main contribution of our work is an effective
neural translation model we apply to a small cor-
pus comprised of extremely short-text sequences, by
sharing representation power of context both adja-
Bleiweiss, A.
Machine Floriography: Sentiment-inspired Flower Predictions over Gated Recurrent Neural Networks.
DOI: 10.5220/0006583204130421
In Proceedings of the 10th International Conference on Agents and Artificial Intelligence (ICAART 2018) - Volume 2, pages 413-421
ISBN: 978-989-758-275-2
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
413

cent in time and closely related in semantic vector
space. The rest of this paper is organized as fol-
lows. In Section 2, we give a brief review of our
compiled version of the Language of Flowers and the
use of Word2Vec word embeddings to represent sen-
timent sentences and flower names. Section 3 then
overviews the gated recurrent unit (GRU) extension
to a standard RNN, and derives our neural network
architecture for predicting bouquet flower candidacy
from a sentiment phrase. As Section 4 motivates the
order of feeding RNN semantically-close sentiment
vectors to improve accuracy. We proceed to present
our methodology for evaluating system performance
end-to-end, and report extensive quantitative results
over a range of experiments, in Section 5. Summary
and identified prospective avenues for future work are
provided in Section 6.
2 LANGUAGE OF FLOWERS
The online floral dictionaries we obtained sort flo-
wers by name and distribute respective meanings in
instructive alphabetical chapters (Figure 1). Inter-
nally, we represent the language of flowers in a size-l
named-member list of variable-length sentiment phra-
ses, each identified with a single pivotal flower f , as
illustrated in Table 1. To keep the plant names uni-
quely labeled in our implementation, those compo-
sed of multiple words make up a hyphenated com-
pound modifier. In total, we tokenize and lowercase
l = 701 sentiment-flower pairs of distinct flower ty-
pes, as we pare down 3,857 unfiltered words to build
a succinct vocabulary of 1,386 symbols, after remo-
ving stop words and punctuation marks.
Figure 1: Language of Flowers: distributed flower types
arranged in alphabetically sorted buckets of names. Buckets
U and X are evidently empty.
To visualize the distribution of the vocabulary
from several different viewpoints, we used R (R Core
Team, 2013) to render a word cloud (Figure 2) that
depicts the top 150 frequent sentiment labels in the
Table 1: Language of Flowers: a sample of sentiment phra-
ses each identified with a single ground-truth flower-type.
Flower names of multiple words are hyphenated to form a
compound.
Sentiment Phrase Pivotal Flower
endearment, sweet and lovely carnation-white
you pierce my heart gladiolus
pleasant thoughts, think of me pansy
joy, maternal tenderness sorrel-wood
enchantment, sensibility, pray for me verbena
dictionary of the language of flowers. Noticeably of
the highest occurrence count are words of emotio-
nal romantic connotations like ‘love’, ‘beauty’, ‘af-
fection’, ‘friendship’, and ‘heart’. In Figure 3, we
follow to provide term frequencies for each of the top
25 sentiment labels in the vocabulary, as the words
‘love’, ‘beauty’, and ‘affection’ occur 40, 31, and 12
times, respectively. Lastly, the distribution of senti-
ment word lengths is of notable importance to assess
flower prediction performance. To this extent, Figure
4 highlights 330 single-word, 100 four-word, and 80
two-word long sentiment sentences, with a maximum
sequence length of eighteen words.
pleasures
thoughts
beauty
mental
desire
consumed
bravery
fate
secret
declare
sweetness
hope
luck
welcome
domestic
cheerful
night
indifference
rustic
riches
passion
ardour
charity
indiscretion
instability
pretension
good
meeting
modesty
misanthropy
affection
confidence
lasting
knight−errantry
love
attachment
pleasure
delicacy
remembrance
majesty
bonds
unfortunate
industry
heart
death
worthy
presence
reconciliation
change
think
pride
enthusiasm
purity
friendship
appreciation
merit
preference
justice
bashful
amiability
days
touch
truth
snare
music
die
war
memory
true
wit
grief
crown
fire
best
bridal
Figure 2: Language of Flowers: word cloud of the top
150 frequent sentiment labels. Font size is proportional to
the number of word occurrences in the corpus, with ‘love’,
‘beauty’, and ‘affection’ leading.
Despite the relatively concise vocabulary of size
|
V
|
= 1,386 tokens, rather than to use a 1-of-
|
V
|
sparse representation of a one-hot vector ∈ R
|V |×1
,
we map one-hot vectors onto a lower-dimensional
vector space using the Word2Vec (Mikolov et al.,
2013b) embedding technique that encodes semantic
word relationships in simple vector algebra. To train
word vectors effectively, we enabled negative sam-
pling in both the skip-gram and continuous-bag-of-
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
414

Figure 3: Language of Flowers: distribution of the top 25
frequent sentiment labels in the vocabulary.
Figure 4: Language of Flowers: distribution of sentiment
word lengths across the entire train set.
words (CBOW) neural models, and found word vec-
tor dimension, d, a critical hyperparameter to tune and
yield consistent flower-selection predictions in RNN.
3 FLOWER CANDIDACY
Formally, we represent the language of flowers as a
list of translation records, each defines a pair of a
source sentiment phrase and a unique target flower.
We use the notation v
1:k
to describe the sequence
(v
1
, v
2
, ..., v
k
) of k vectors and correspondingly denote
a translation pair as (s
1:k
, f ). A sentiment-flower pair
is decomposed into a k-length sentiment input s
1:k
we
linearly stream into RNN and an output concatenation
(s
2:k
, f ) of k word vectors, and establish a ground-
truth pivotal relation between a flower type and its
immediate preceding context. We note that our lan-
guage corpus incorporates short-text sentiment phra-
ses of length k that ranges from one to eighteen word
vectors (Figure 4). To further align our rendition in-
terface with the RNN architectural notation, we let
x
t
∈ R
d
be the d-dimensional word vector identified
with the t-th word in a text sequence, and denote an
end-to-end translation record as (x
1
, x
2
, ..., x
k
, x
k+1
),
or more compactly x
1:k+1
. We use the Gated Recur-
rent Unit (GRU) (Cho et al., 2014; Chung et al., 2014)
variant of RNN that adaptively captures long term de-
pendencies of different time scales. At each time step
t, the GRU takes an input word vector x
t
and the pre-
vious n-dimensional hidden state h
t−1
to produce the
next hidden state h
t
. Conceptually, the forward GRU
has four basic functional stages that are governed by
the following set of formulas:
z
t
= σ(W
(z)
x
t
+U
(z)
h
t−1
)
r
t
= σ(W
(r)
x
t
+U
(r)
h
t−1
)
˜
h
t
= tanh(r
t
Uh
t−1
+Wx
t
)
h
t
= (1 − z
t
) 
˜
h
t
+ z
t
 h
t−1
where W
(z)
,W
(r)
,W ∈ R
n×d
and U
(z)
,U
(r)
,U ∈ R
n×n
are weight matrices, and the dimensions n and d are
input configurable hyperparameters. The symbols
σ() and tanh() refer to the non-linear sigmoid and
hyperbolic-tangent functions, and  is an element-
wise multiplication. A backward fed GRU defines
←−
h
t
=
←−−
GRU (x
t
), t ∈ [k, 1].
In modeling the language of flowers, our main ob-
jective is to score flower candidacy for a bouquet,
by predicting flower type probabilities based on the
short-text sentiment sequence retained in each trans-
lation record. Serialized sentiment vectors x
1:k
are
streamed to RNN in isolation and independently, but
the inherent persistent-memory nature of GRU sum-
marizes at each time step the newly observed word
vector in a record sequence, x
t
, with the cumula-
tive previous context, h
t−1
. Similarities computed as
inner-products between each of the input converted
word-vectors and the GRU encoded h
t
are then for-
warded to a softmax discriminative decoder. Hence,
the next predicted word is the output probability dis-
tribution ˆy
t
= softmax(W
(S)
h
t
) over the entire lan-
guage vocabulary, where W
(S)
∈ R
|
V
|
×n
and ˆy
t
∈ R
|
V
|
,
and
|
V
|
is the cardinality of the vocabulary. During
the training of RNN, we attach a higher scoring bias
to the ground-truth pivotal flower-word, x
k+1
, that
immediately succeeds the last word of a sentiment
phrase, x
k
. Whereas in evaluation, to generate a se-
lection for a bouquet of flowers we rank all the poste-
rior probabilities of ˆy
t
that were predicted for x
k
, and
from the top-m indices we filter out all sentiment word
instances, and return to the user the remaining flower
tokens.
4 SHARED REPRESENTATION
In their recent work, Lee and Dernoncourt (2016)
show that the chronology of sequences of short-text
Machine Floriography: Sentiment-inspired Flower Predictions over Gated Recurrent Neural Networks
415

Figure 5: Language of Flowers: combining Multidimensi-
onal Scaling and k-means clustering to visualize sentiment
phrase relatedness in semantic vector space. Shown five
collections with only a few outliers.
representations fed to RNN, improves sentence clas-
sification accuracy. Motivated by their results, we
sought after a sequencing model that schedules sen-
timent sentences to RNN based on plausible seman-
tic relatedness between phrases. However, our online
acquisition of translation records that are enumerated
in alphabetical bins based on flower names (Figure
1), implied a partition type that constrains the more
interesting information about sentiment semantic si-
milarity.
To address this shortcoming, we first had to avoid
non-conforming representations in dot-product simi-
larity computations by reshaping the varying dimen-
sionality of sentiment clauses into uniform sized fe-
ature vectors. We chose to leverage a basic convolu-
tional architecture practice and applied mean pooling
to each of the encoded sentiment sequences x
1:k
, and
averaged the word vectors to yield a single vector re-
presentation s =
1
k
∑
k
t=1
x
t
, where s ∈ R
d
. From the
set of single-vector formatted sentiments s
1:l
, where
l = 701, we follow by constructing a distance matrix
D ∈ R
l×l
, and use Multidimensional Scaling (MDS)
(Torgerson, 1958; Hofmann and Buhmann, 1995)
to project the large dissimilarity matrix onto a two-
dimensional embedding space. By combining MDS
with k-means (Kaufman and Rousseeuw, 1990), we
produced visualization of clusters that group seman-
tically close sentiment phrases (Figure 5). Surpri-
singly, only a few outliers persist and most sentiment
sequences notably gather consistently. Results we re-
port next were obtained by scheduling sentiment se-
quences to RNN in the order prescribed by their pro-
jected coordinates, from left-to-right and bottom-to-
top. This is motivated by sharing representation po-
wer of similar context, where the final hidden state
of the current encoded sentiment is fed as the initial
hidden state of the next spatially closest sentiment se-
quence.
5 EVALUATION
Our workflow for evaluation is straightforward. First,
the user enters our system a desired count of plant ty-
pes ranging from 1 to m, along with an arbitrary sen-
timent composition of words that are part of the voca-
bulary of the language of flowers (Figure 2). The user
text sequence is transformed to a single word-vector
representation and follows cosine similarity calcula-
tions (Salton et al., 1975; Baeza-Yates and Ribeiro-
Neto, 1999) with each of the trained mean-pooled
sentiment-sentences, s
i
. For the semantically closest
translation pair, we query the pivotal and supporting
flowers and return to the user distinct flower images
and proportional weights that are used for final bou-
quet arrangement. To compare the predictive perfor-
mance of our GRU-based RNN system against, we
chose a baseline that we feed with our mean-pooled
representation of a sentiment sentence, and use soft-
max regression for the unsupervised learning algo-
rithm. We cross validated our baseline on both a
held out development set and an exclusively genera-
ted test-set. Figure 6 illustrates an architectural over-
view of the pipelines for both the main and baseline
computational paths.
5.1 Experimental Setup
To evaluate our system in practice, we have imple-
mented our own versions of a GRU-based RNN mo-
dule and the Word2Vec embedding technique, both
natively in R (R Core Team, 2013) for better integra-
tion with our software framework. Considering our
self-sustained corpus of irregular context, we chose to
collectively initialize our word vectors randomly and
learn them purely from the dictionary data. Instead of
obtaining pre-trained word vectors on large vocabula-
ries of external corpora that miss most our flower na-
mes. Skip-gram and CBOW performed almost iden-
tically in terms of flower prediction accuracy, both
are challenged by a language that combines short-
text sentences with a small number of samples, re-
spectively. Our model is trained to maximize the log-
likelihood of predicting the ground-truth target flower
for the language set of translation records, using mini-
batch stochastic gradient descent (SGD) and perfor-
ming error back-propagation. At every SGD itera-
tion, GRU weight matrices and word vectors are up-
dated, using the AdaDelta parameter update rule (Zei-
ler, 2012).
In Table 2, we list our experimental choices
of hyperparameters that control both the RNN and
Word2Vec subsystems. Given the succinct nature of
text fragments that constitute our sentiment phrase,
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
416

RNN
RNN
RNN
𝑥
𝑘
𝑥
1
𝑥
2
⋯
Mean
Pooling
⋯
⋯
ℎ
1
ℎ
2
ℎ
𝑘
ℎ
𝑘−1
𝑦
𝑘
Softmax
Activation
𝑦
𝑘
flower
predictions
sentiment phrase word-vectors
𝑠
Softmax
Regression
flower
classification
Figure 6: Architecture overview of computational pipelines for the mainline feedforward GRU-based RNN (right) and softmax
regression baseline (left).
Table 2: Experimental choices of tuned hyperparameters
applied to both RNN and Word2Vec modules.
Hyperparameter Notation Value
Hidden state dimension n 18
Word vector dimension d 100
Context window c
w
5
Negative samples n
s
10
Mini-batch size b 10
Top ranked probabilities m 10
we held the context window size c
w
at five words sym-
metrically, and assigned the number of negative sam-
ples n
s
to its recommended default value for small da-
tasets (Mikolov et al., 2013b). Correspondingly, we
allocated a reasonable example fraction of our train
set size for the SGD mini-batch size b. To optimize
the score for predicting the flower types, we modi-
fied one dimensional parameter at a time and kept
the other fixed. This culminated in setting the pre-
ferred hidden-state dimension n to the maximal word
length of a sentiment phrase (Figure 4), as we noti-
ced accuracy performance improvement by 6.7 per-
centage points when modifying the word vector size d
from 10 to 100. However, we observed a diminishing
return in embeddings larger than the hundred dimen-
sions. Unless stated otherwise, the results we report
were obtained using the skip-gram neural model and
a unidirectional GRU-based RNN, with m = 10.
5.2 Experimental Results
We chose a simple baseline model that replaces RNN
with softmax regression for learning, and as fea-
ture vectors uses a sentiment representation of mean-
pooled word embeddings, s ∈ R
d
. Each short-text
sentiment sequence of the train set is assigned one-
of-five categorical labels (Figure 5), as we turn our
problem to perform multi-class flower classification.
0.15
0.20
0.25
0.30
0.35
1e−06 1e−03 1e+00
Regularization
Precision
dev
test
Figure 7: Baseline softmax regression results on an inclu-
sive development set and exclusive test set. Showing pre-
cision as a function of a logarithmic-scaled regularization
parameter.
For validation evaluation, we formed a two-way
data split of 80/20 by proportionally sampling random
sentiment sentences from each of the five training
classes into a held out development set. To construct
our test set and allow for the study of our problem
domain itself, we leveraged the generative capacity
of an RNN-based language model (Sutskever et al.,
2011), and created entirely new text sequences that
closely resemble the emotional context at the core of
the language of flowers. The test collection we built
comprises 140 synthetic sentiment sentences of va-
riable length, about 20 percent of the train set size,
and is evaluated against the unsplit train-set size of
l = 701. Each of these extended sentiment phrases
is paired with one of the existing flowers in the cor-
pus and hence implicitly inherits a ground-truth multi-
class label. The pairing of a flower to a generated test
Machine Floriography: Sentiment-inspired Flower Predictions over Gated Recurrent Neural Networks
417

(a) Forward RNN (b) Backward RNN
Figure 8: Distribution of flower predictions that are sampled across the entire train set at discrete recall steps. Contrasting (a)
forward x
1:k
with (b) backward x
k:1
propagation of a sentiment phrase in RNN.
0
25
50
75
100
1 2 3 4 5 6 7 8 9 10
Flower Types
Recall
rnn
gru
Figure 9: Comparing standard to GRU-based RNN on per-
centage recall of flower predictions as a function of increa-
sed flower type count ∈ (1, 2, . . . , m = 10). Curves present
a non-linear decline in performance.
phrase g is obtained by computing the similarity with
each of the train set sequences s
i
, and finding the clo-
sest record in argmax
1≤i≤l
sim(g, s
i
). Figure 7 shows
our baseline performance results as a function of a
logarithmic-scaled regularization parameter, on both
the development and test data sets with an average
precision of 0.22 and 0.23, respectively.
In evaluating our GRU-based RNN system, our
main goal is to quantitatively assess the quality of flo-
wer predictions looked at from differing perspectives.
The top-m ranked probabilities presented in the net-
work output layer may index any combination of flo-
wer and sentiment token types. We let t
f
denote the
number of flower tokens and t
s
the sentiment token
count in this m-sized window of descending probabi-
lities. Thus, the performance interpretation of t
f
and
t
s
in our model amounts to the true-positive and false-
negative observations made in each flower selection
query, respectively. For our flower prediction metric,
we follow the recall measure of relevance that is defi-
ned as the ratio
t
f
t
f
+t
s
, or more compactly
t
f
m
.
Figure 8 provides distribution of flower predicti-
ons in recall bins of 0.1 increments for both forward
Figure 10: Accumulated true-positives of flower predicti-
ons as a function of the word length of the sentiment
phrase. The data produced matches the vocabulary distri-
bution shown in Figure 4.
and backward sentiment-phrase propagation in RNN
(Schuster and Paliwal, 1997), and in Figure 9, we
translate the discrete data to a continuous recall curve
as a function of a non-descending number of flower
types in a bouquet. The average recall for the sweet-
spot content of one to five flower types is 87.8% for
GRU-based, and 81.5% for plain RNN, both outper-
forming the softmax regression baseline by a factor of
3.8X and 3.5X, respectively. In Figure 10, we present
the collective number of true-positives as a function
of a non-descending word length of the sentiment text
sequence. Flower predictions closely correlate with
the word length distribution in the vocabulary (Figure
4), mainly owing to streaming sentiment context to
RNN in an order prescribed by spatial proximity in
semantic vector space.
High-end bouquets created of tens of flower ty-
pes are considered a rarity, but despite their limited
presence and small floral market-share, they make an
important case for our system evaluation. In Figure
11, we review the distribution of flower predictions
for high-end bouquets with the system hyperparame-
ter m = 50. Distinctly, incidents of unpredictable bou-
quet selections do transpire and display a fairly sym-
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
418

Table 3: Visualization of statistically created flower selections that were prompted by user-supplied sentiment expressions,
shown at the top. We outline side-by-side three bouquets of five, seven, and eight flower candidates, respectively. Depicted
are thumbnail images for both the pivotal flower, on the far right, and top-ranked supporting flowers to its left, along with
individual flower names and their proportional mix weights.
Sentiments
light heartedness wit ill-timed beauty is your only attraction
Flowers
Proportions
Figure 11: Distribution of flower predictions for high-end
bouquets (m = 50). Shown sampled at discrete recall steps
of five flower type bundles, as unpredictable selections
transpire symmetrically.
metrical behavior with sizes ranging from 1 to 14, ex-
cluding a bouquet size of ten plants with two predicti-
ons, and from 35 to 50 assuming no predictions. Con-
versely, for bouquets of mid-range sizes from 20 to
29 species that possess each at least twenty predicti-
ons, the total predictions amounts to 608 and repre-
sent about 86.7 percent of the training set dimension.
We note that unpredictable selection sizes are of very
low probability for mainline bouquets (m ≤ 10), and
for high-end bouquets, to ensure the requested num-
ber of flower types issued by the user is satisfied, our
system supplies the user a precompiled list of prefer-
red bouquet sizes to chose from, for each m.
In Table 3, we present end-to-end visualization
of our approach to statistically-generated flower se-
lections from decoded sentiment clauses. Outlined
Figure 12: A random sample distribution of sentiment sen-
sible bouquets, each shown with proportional weights of an
assembled size that ranges from one to eight flower types.
side-by-side are three bouquets of five, seven, and
eight flower candidates, respectively. The incoming
short-text sentiment sequences are first evaluated for
the most similar sentiment phrase retained in one
of the language translation records to enter our pi-
peline, shown at the top. We then depict compositi-
onal thumbnail images of the pivotal flower, placed
on the far right, and the supplemental flowers to its
left, along with individual names of plants and their
respective proportional blend-weights. Stacked rela-
tively, slices are shown in a non-ascending manner,
starting from the principal flower at the bottom and up
to the lowest ranked flower. In Figure 12, we highlight
a sample distribution of twenty sentiment-perceived
bouquets, each with a unique center-of-attention flo-
wer. Selections are shown for varying sizes ranging
from one to eight plant types.
Machine Floriography: Sentiment-inspired Flower Predictions over Gated Recurrent Neural Networks
419

6 CONCLUSIONS
In this work, we have demonstrated the plausible po-
tential in composing a bouquet made of a lead and
supporting flowers by attending to a set of unquali-
fied sentiment expressions we stream over RNN. We
confirmed that GRU-based RNN improved our flo-
wer prediction quality by about ten percentage points
compared to a standard RNN, and feeding RNN with
spatiotemporal sentiment context prove particularly
beneficial to the performance of short-text sequences.
Yet using bidirectional propagation of sentiment word
vectors that enter RNN was less instinctive, and con-
tributed to an inconsequential prediction gain. Our
proposed simple workflow offers on average high pre-
diction recall to hundreds of mix choices for a main-
line bouquet, and sends a more cohesive emotional
message that is made of semantically related senti-
ments.
To the extent of our knowledge, the work we pre-
sented is first to apply computational linguistic mo-
deling to the language of flowers. We contend that
floriography is an important NLP discipline to pur-
sue from both its rooted historical impact on society
culture, and the prospect to influence areas of critical
theory and sentiment analysis. In its current state, the
corpus we used in this paper is small and challenging,
but we anticipate the language to expand sentiment
translation to thousands of flower plants and further
merit our statistically reasoned system. A direct pro-
gression of our work is to evolve to a task that matches
flower-sentiment pairs from unstructured full text and
not just from a set of prescribed sentiment phrases,
and have a profound practical importance to impact
a much broader scope of application domains that in-
clude cryptography and secured communication.
ACKNOWLEDGEMENTS
We would like to thank the anonymous reviewers for
their insightful suggestions and feedback.
REFERENCES
Baeza-Yates, R. and Ribeiro-Neto, B., editors (1999).
Modern Information Retrieval. ACM Press Se-
ries/Addison Wesley, Essex, UK.
Cho, K., van Merrienboer, B., Bahdanau, D., and Ben-
gio, Y. (2014). On the properties of neural machine
translation: Encoder-decoder approaches. CoRR,
abs/1409.1259. http://arxiv.org/abs/1409.1259.
Chung, J., G
¨
ulc¸ehre, C¸ ., Cho, K., and Bengio, Y. (2014).
Empirical evaluation of gated recurrent neural net-
works on sequence modeling. CoRR, abs/1412.3555.
http://arxiv.org/abs/1412.3555.
Diffenbaugh, V. (2011). Victoria’s dictionary of flo-
wers. http://aboutflowers.com/images/stories/Florist/
languageofflowers-flowerdictionary.pdf.
Elman, J. L. (1990). Finding structure in time. Cognitive
Science, 14(2):179–211.
Hoang, C. D. V., Cohn, T., and Haffari, G. (2016). In-
corporating side information into recurrent neural net-
work language models. In Human Language Techno-
logies: North American Chapter of the Association
for Computational Linguistics (NAACL), pages 1250–
1255, San Diego, California.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term
memory. Neural Computation, 9(8):1735–1780.
Hofmann, T. and Buhmann, J. (1995). Multidimensional
scaling and data clustering. In Advances in Neural
Information Processing Systems (NIPS), pages 459–
466. MIT Press, Cambridge, MA.
Kaufman, L. and Rousseeuw, P. J., editors (1990). Finding
Groups in Data: An Introduction to Cluster Analysis.
Wiley, New York, NY.
Lee, J. Y. and Dernoncourt, F. (2016). Sequential short-text
classification with recurrent and convolutional neural
networks. In Human Language Technologies: North
American Chapter of the Association for Computatio-
nal Linguistics (NAACL), pages 515–520, San Diego,
California.
Mikolov, T., Chen, K., Corrado, G. S., and Dean, J.
(2013a). Efficient estimation of word represen-
tations in vector space. CoRR, abs/1301.3781.
http://arxiv.org/abs/1301.3781.
Mikolov, T., Karafi
´
at, M., Burget, L., Cernock
´
y, J., and
Khudanpur, S. (2010). Recurrent neural network ba-
sed language model. In International Speech Commu-
nication Association (INTERSPEECH), pages 1045–
1048, Chiba, Japan.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., and
Dean, J. (2013b). Distributed representations of words
and phrases and their compositionality. In Advances in
Neural Information Processing Systems (NIPS), pages
3111–3119. Curran Associates, Inc., Red Hook, NY.
Pennington, J., Socher, R., and Manning, C. D. (2014).
GloVe: Global vectors for word representation. In
Empirical Methods in Natural Language Processing
(EMNLP), pages 1532–1543, Doha, Qatar.
R Core Team (2013). R: A Language and Environment
for Statistical Computing. R Foundation for Sta-
tistical Computing, Vienna, Austria. http://www.R-
project.org/.
Roof, L. and Roof, A. H. (2010). Language of flowers pro-
ject. http://languageofflowers.com/index.htm.
Salton, G. M., Wong, A., and Yang, C. S. (1975). A vector
space model for automatic indexing. Communications
of the ACM, 18(11):613–620.
Schuster, M. and Paliwal, K. K. (1997). Bidirectio-
nal recurrent neural networks. Signal Processing,
45(11):2673–2681.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
420

Sutskever, I., Martens, J., and Hinton, G. (2011). Genera-
ting text with recurrent neural networks. In Internati-
onal Conference on Machine Learning (ICML), pages
1017–1024.
Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence
to sequence learning with neural networks. In Advan-
ces in Neural Information Processing Systems (NIPS),
pages 3104–3112. Curran Associates, Inc., Red Hook,
NY.
Torgerson, W. S. (1958). Theory and Methods of Scaling.
John Wiley and Sons, New York, NY.
Tran, K. M., Bisazza, A., and Monz, C. (2016). Recur-
rent memory networks for language modeling. In Hu-
man Language Technologies: North American Chap-
ter of the Association for Computational Linguistics
(NAACL), pages 321–331, San Diego, California.
Zeiler, M. D. (2012). ADADELTA: an adaptive
learning rate method. CoRR, abs/1212.5701.
http://arxiv.org/abs/1212.5701.
Machine Floriography: Sentiment-inspired Flower Predictions over Gated Recurrent Neural Networks
421