Propagation-Based Domain-Transferable Gradual Sentiment Analysis
C
´
elia da Costa Pereira
1 a
, Claude Pasquier
1 b
and Andrea G. B. Tettamanzi
2 c
1
Universit
´
e C
ˆ
ote d’Azur, I3S, CNRS, Sophia Antipolis, France
2
Universit
´
e C
ˆ
ote d’Azur, I3S, Inria, Sophia Antipolis, France
Keywords:
Sentiment Analysis, Fuzzy Polarity Propagation.
Abstract:
We propose a novel refinement of a gradual polarity propagation method to learn the polarities of concepts
and their uncertainties with respect to various domains from a labeled corpus. Our contribution consists of
introducing a positive correction term in the polarity propagation equation to counterbalance negative psycho-
logical bias in reviews. The proposed approach is evaluated using a standard benchmark, showing an improved
performance relative to the state of the art, good cross-domain transfer and excellent coverage.
1 INTRODUCTION
Sentiment analysis aims at determining the global po-
larity (positive, negative or neutral) of a document
based on the polarities of the words in the document.
A supervised method trains a model (classifyer) by
using the datasets of reviews or labeled texts and use
such models to classify the user opinions. However,
the polarity of some words in a review might depend
on the domain knowledge considered (Rexha et al.,
2018; Pirnau, 2018). For example (Yoshida et al.,
2011), the word ‘long’, which has a positive polarity
in the Camera domain, has a negative polarity if we
are characterizing the execution time of a computer
program.
Several solutions have been proposed. Concept-
based approaches include the one proposed by
Schouten et al. (Schouten and Frasincar, 2015), who
show that considering concept-based features instead
of term-based features helps improving the perfor-
mance of multi-domain sentiment analysis methods.
The good quality of the results obtained with this rel-
atively straightforward setup encourages the use of
more advanced ways of handling semantic informa-
tion.
Yoshida et al. (Yoshida et al., 2011) propose a so-
lution to improve transfer learning methods. How-
ever, as it has been rightly pointed out by Abdullah
et al. (Abdullah et al., 2019), the transfer learning ap-
a
https://orcid.org/0000-0001-6278-7740
b
https://orcid.org/0000-0001-7498-395X
c
https://orcid.org/0000-0002-8877-4654
proach imposes the necessity to build a new transfer
model and this limits its generalization capability.
A serie of works (Dragoni et al., 2014; Dragoni,
2015; Dragoni et al., 2015; Dragoni et al., 2016)
use fuzzy logic to model the relationships between
the polarity of concepts and the domain. They use
a two-level graph, where the first level represents
the relations between concepts, whereas the second
level represents the relations between the concepts
and their polarities in the various targeted domains,
the idea being to capture the fact that the same con-
cept can be positive in one domain, but negative in
another. This is accomplished thanks to a polarity
propagation algorithm and without the necessity of
starting the learning process for each different do-
main. The main advantage of that approach, named
MDFSA (Multi-Domain Fuzzy Sentiment Analyzer),
which won the ESWC 2014 Concept-Level Sentiment
Analysis Challenge (Dragoni et al., 2014), is that it
both accounts for the conceptual representation of the
terms in the documents by using WordNet and Sentic-
Net, and proposes a solution avoiding to build a new
model each time a new domain needs to be analysed.
However, MDFSA had several issues, including
the following:
1. It is not possible to discard some of the remain-
ing word ambiguities due to the fact that a synset
corresponds to a group of words (nouns, adjec-
tives, verbs and adverbs) that can be interchange-
able and depending on the type of the term used,
the meaning of the word can change.
520
Pereira, C. C., Pasquier, C. and Tettamanzi, A. G. B.
Propagation-Based Domain-Transferable Gradual Sentiment Analysis.
DOI: 10.5220/0013158200003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 3, pages 520-527
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

2. The same stopping criterion for the propagation
algorithm is used for the different independent do-
mains which could be challenged, i.e. that simul-
taneous stopping of propagation could be prema-
ture for some domains and delayed for others.
3. The propagation of polarities takes place without
taking into account the similarity of related con-
cepts in the graph. Indeed, the more similar the
concepts are, the higher the weights associated
with their relationship in the graph should be.
To solve them, we recently proposed Sental (Pasquier
et al., 2020), an extension of MDFSA, wich we take
as a starting point. In particular, concerning (1), to
decide whether a term occurring in a document is as-
sociated to a synset v or not, Sental looks at its part
of speech (POS) tag and considers it an instance of
v only if its POS tag matches the POS of v. Con-
cerning (2), Sental tests for convergence for each do-
main separately. Finally, concerning (3), Sental uses
a pre-trained word embedding model to complete the
semantic graph with a graded relation of semantic
similarity, in addition to the crisp relations defined in
WordNet.
Nevertheless, Sental still has an important issue:
after a few iterations, its propagation process tends to
drive polarities towards −1. This happens because,
on average, negative sentiments have more impact
and are more resistant to disconfirmation than posi-
tive ones.
The aim of this paper is to solve this problem.
We propose a new formula with a term to correct the
negative bias by exerting a positive pressure which is
strongest for neutral polarities and whose effect di-
minishes for extreme polarities. To the best of our
knowledge it is the first time that this aspect, which
is very well known in psychology (Baumeister et al.,
2001), is taken into account in a sentiment analysis
framework.
The rest of the paper is structured as follows. Sec-
tion 2 presents the original method we are proposing.
Section 3 presents the experiments and discuss the re-
sults. Finally, Section 4 concludes the paper.
2 METHOD
The algorithm used to learn concept polarities for var-
ious domains consists of three phases, namely:
1. Semantic graph construction from background
knowledge;
2. Concept polarity initialization, based on a training
set of documents, associated with a domain and
labeled with a rating;
Figure 1: An illustration of the semantic graph constructed
by the proposed method. Circles represent the vertices of
the graph and solid lines its edges. The documents of the
training set are shown around the semantic graph. Dashed
lines represent the occurrence, in a document, of a term as-
sociated with a vertex of the graph (i.e., a lemma or a Word-
Net synset). Documents are rated (e.g., on a scale from −−
to ++, which may then be mapped to the [−1,1] interval).
3. Propagation of polarity information over the se-
mantic graph.
Its result is an estimation of polarities, represented as
convex fuzzy sets over the [−1,+1] interval, for each
concept and for each domain. Figure 1 illustrates the
idea of a semantic graph, whose construction and use
is detailed below.
2.1 Semantic Graph Construction
The backbone of the semantic graph is based on
WordNet (Miller, 1995) and SenticNet (Cambria
et al., 2010), combined as in (Dragoni et al., 2015).
Both are publicly available lexical databases in which
nouns, verbs, adjectives, and adverbs are organized
into sets of synonyms (synsets), each representing a
lexicalized concept. Synsets are linked by semantic
relationships, including synonymy, antonymy and hy-
pernymy. We distinguish concepts, which are abstract
notions representing meaning, from terms, which are
tangible ways of expressing concepts (in written lan-
guage, they are words or phrases). Now, words, and
typically the most frequent words (Casas et al., 2019),
can be polysemous; a preprocessing phase is thus
needed to link the terms found in the texts as accu-
rately as possible to their corresponding synsets. Un-
like WordNet, SenticNet is specifically built for opin-
ion mining, but the polarities it associates to terms
are not used, because the very assumption on which
our approach is based is that polarity is not an in-
trinsic property of a term, but an extrinsic, domain-
dependent property. Besides covering terms that are
not indexed in WordNet, SenticNet allows to resolve
Propagation-Based Domain-Transferable Gradual Sentiment Analysis
521

a number of cases of ambiguity.
To further reduce ambiguity, we propose to parse
the text and use the POS of a term to associate it to the
correct synset. As an example, the same term ‘light’
can be, according to WordNet 3.1, a noun (‘do you
have a light?’), a verb (‘light a cigarette’), an adjec-
tive (‘a light diet’) or an adverb (‘experienced trav-
elers travel light’), which are represented by distinct
synsets.
In addition to the WordNet and SenticNet re-
lations, already used in the literature, we use a
word embedding model, pre-trained by applying
Word2vec (Mikolov et al., 2013) to roughly 100 bil-
lion words from a Google News dataset,
1
to com-
plete the semantic graph with relationships of seman-
tic similarity between terms.
The semantic graph is constructed as a weighted
graph (V,E,w). Each element of V is either a concept
(i.e., a synset) or the canonical form (lemma) of a term
used in a review; w : E → [−1,1] is a weight function
and the edges in E are created:
• between synsets linked by a hypernym relation-
ship, a synset and its lemmas, and between
lemmas linked by a synonym relationship, with
weight +1;
• between lemmas linked by an antonym relation-
ship, with weight −1;
• between each lemma and the five closest lem-
mas according to the pre-trained word2vec model,
with their cosine similarity as weight.
Each vertex v ∈ V is labeled by a vector ⃗p(v) of polar-
ities, so that p
i
(v) is the polarity of v in the ith domain.
2.2 Concept Polarity Initialization
The initial polarities ⃗p
(0)
i
(v) ∈ [−1,1] of all the ver-
tices v of the semantic graph are computed, for each
domain i, as the average polarity of the documents of
domain i in the training set, in which at least a term
of v occurs. If no term associated to v occurs in a
document of domain i, p
(0)
i
(v) = 0.
As explained above, to decide whether a term oc-
curring in a document is associated to a synset v or
not, we look at its POS tag and we consider it an in-
stance of v only if its POS tag matches the POS of
v.
2.3 Polarity Propagation
In this phase, information about the polarity of ver-
tices is propagated through the edges of the graph, so
1
https://frama.link/google\ word2vec
that concepts for which no polarity information could
be directly extracted from the training set (i.e., those v
such that p
(0)
i
(v) = 0 for some i) can “assimilate”, as
it were, the polarity of their close relatives. In addi-
tion, this propagation process may contribute to cor-
rect or fine-tune the polarity of incorrectly initialized
concepts and, thus, reduce noise.
Polarity propagation through the graph is carried
out iteratively. At each iteration t = 1,2, ..., the po-
larity p
(t)
i
(v) of each vertex v for domain i is updated
taking into account both the values of its neighbors
N(v) = {v
′
| (v,v
′
) ∈ E} and its distinctiveness from
the other terms in the domain.
⃗p
(t+1)
(v) = (1 − λ)⃗p
(t)
(v)
+ λ
1
∥N(v)∥
∑
v
′
∈N(v)
⃗p
(t)
(v
′
)w(v,v
′
) (1)
+ ϕ(1 − |⃗p
(t)
(v)|),
where w(v, v
′
) denotes the weight of the edge between
vertices v and v
′
, 0 < λ < 1, the propagation rate and
0 < ϕ < 1, the positive correction strength are param-
eters of the algorithm.
Notice that the propagation of polarity for one do-
main does not interact with the same process for the
other domains and we can thus consider that polar-
ity propagation is carried out in parallel and indepen-
dently for each domain.
Inspired by simulated annealing (Kirkpatrick
et al., 1983), the propagation rate is decreased at each
iteration, according to a parameter A called annealing
rate. Thus, the value of λ at iteration t is calculated
according to the value of λ at iteration t − 1 as follow:
λ
t
= Aλ
t−1
. (2)
In the method proposed by Dragoni et al. (Dragoni
et al., 2015), the iterative process stops as soon as the
sum of the variations of the polarity for each concept
and domain falls below a fixed threshold. The draw-
back of using a fixed convergence limit is that it de-
pends on the dataset used. Indeed, a dataset composed
of many domains using lots of different terms will
logically generate a greater variation than a smaller
dataset. In our proposed method, we specify a thresh-
old that applies to each domain separately and that is
relative to the number of different nodes composing
the semantic graph of the domain. Thus, for the ith
domain, the polarity propagation stops when the av-
erage polarity variation,
∆
(t)
i
=
1
∥V ∥
∑
v



p
(t)
i
(v) − p
(t−1)
i
(v)



, (3)
falls below a threshold L, which is the convergence
limit. We denote by t
stop
i
the total number of iterations
carried out for the ith domain until ∆
(t)
i
< L.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
522

Figure 2: A fuzzy set with a trapezoidal membership func-
tion, defined by the four parameters (a,b,c,d).
Experience shows that, after a few iterations, the
propagation process described above tends to drive
polarities towards −1. This happens because, on aver-
age, negative sentiments in reviews are more strongly
expressed than positive sentiments. Such difference
of intensity between negative and positive sentiments
has already been observed and is well-known in psy-
chology (Baumeister et al., 2001). That is why we
have added a term to correct this negative bias (third
line of Equation 1), by exerting a positive pressure
which is strongest for neutral polarities and whose
effect tapers off for extreme polarities. The overall
strength of this positive correction term is controlled
by parameter ϕ.
At each iteration t = 0, 1,2,. .., the vectors ⃗p
(t)
(v)
are saved in order to exploit them for the calculation
of the shapes of the fuzzy membership functions de-
scribing the polarity of concept v for each domain. In-
deed, the final polarities are represented as trapezoidal
fuzzy membership functions, whose core is the in-
terval between the initial polarity computed from the
training set, p
(0)
i
(v), and the polarity resulting from
the propagation phase, p
(t
stop
i
)
i
(v) and whose support
extends beyond the core on either side by half the vari-
ance σ
2
v,i
of the distribution of p
(t)
i
(v), t = 0, ... ,t
stop
i
.
To sum up, for each domain i, µ
v,i
is a trapezoid with
parameters (a,b, c,d), like the one depicted in Fig-
ure 2, where
a = min{p
(0)
i
(v), p
(t
stop
i
)
i
(v)},
b = max{p
(0)
i
(v), p
(t
stop
i
)
i
(v)},
c = max{−1,a − σ
2
v,i
/2},
d = min{1,b + σ
2
v,i
/2}.
The idea here is that the most likely values for the po-
larity of v for a domain are those comprised between
the initial and final value of the polarity propagation
phase and the more quickly the polarity values con-
verged during that phase, the least uncertainty there
is about the resulting polarity estimate. Conversely,
a polarity value that converged slowly or with many
fluctuations is going to yield a less reliable, and thus
more uncertain, estimate.
2.4 Document Polarity Calculation
Once the model is trained according to the algorithm
described in the previous sections, the (fuzzy) polar-
ity of a novel document D of the ith domain is com-
puted as the average of the fuzzy polarities (repre-
sented by their trapezoidal membership functions) of
all the terms v occurring in the document:
µ
i
=
1
∥V
i
∥
∑
v∈V
i
µ
v,i
, (4)
where V
i
= {v ∈ V | v occurs in D}. This average of
fuzzy sets is computed by applying the extension prin-
ciple, thus yielding, for all x ∈ [−1,1],
µ
i
(x) = sup
x=
1
∥V
i
∥
∑
v∈V
i
x
v
min
v∈V
i
µ
v,i
(x
v
). (5)
However, given that all µ
v,i
are trapezoidal with pa-
rameters
(a
v
,b
v
,c
v
,d
v
),
as pointed out in (Dragoni et al., 2015), µ
i
will always
be trapezoidal as well, with parameters
1
∥V
i
∥
∑
v∈V
i
a
v
,
∑
v∈V
i
b
v
,
∑
v∈V
i
c
v
,
∑
v∈V
i
d
v
!
. (6)
This fuzzy polarity reflects the uncertainty of the esti-
mate obtained by the model. A single polarity figure
can be obtained by applying a defuzzification method.
For the empirical validation of our method, we used
the centroid method for this purpose.
3 EXPERIMENTS AND RESULTS
The proposed system (that we call Sental+) has been
evaluated using the DRANZIERA evaluation proto-
col (Dragoni et al., 2016), a multi-domain sentiment
analysis benchmark, which consists of a dataset con-
taining product reviews from 20 different domains,
crawled from the Amazon web site, as well as guide-
lines allowing the fair evaluation and comparison
of opinion mining systems. In the dataset of the
DRANZIERA benchmark, each domain is composed
of 5,000 positive and 5,000 negative reviews that are
split in five folds containing 1,000 positive and 1,000
negative reviews each.
3.1 Experimental Protocol
As suggested by the guideline of the DRANZIERA
evaluation protocol (Dragoni et al., 2016), the perfor-
mance of the method has been assessed by performing
Propagation-Based Domain-Transferable Gradual Sentiment Analysis
523

a 5-fold cross validation. For each specific domain,
the method was trained on four of the five folds pro-
vided with the benchmark and tested on the remain-
ing fold. The process is repeated five times so that
each fold is in turn used for testing. Experiments were
performed on a computer running Ubuntu 18.04 and
based on a Intel©Core™ i7-7700 @ 3.60GHz with 32
Gb main memory.
The algorithm depends on four different param-
eters: the propagation rate λ, which determines the
diffusion rate of the polarity values between concepts,
the positive correction strength ϕ that maintains the
diversity of polarities, the convergence limit L, which
represents the criterion for stopping the polarity prop-
agation phase for each domain, and the annealing rate
A, used to decrease, at each iteration, the propagation
rate (cf. Equation 2). In order not to bias the method
settings by selecting parameters that could be specif-
ically suited to the DRANZIERA dataset, we carried
out the method tuning using a separate dataset. For
this purpose we used the Blitzer dataset (Blitzer et al.,
2007) which is composed of product reviews belong-
ing to 25 domains. A quick scan of the dataset shows
both that there is a large discrepancy in the number
of reviews available for each domain and that posi-
tive reviews are much more frequent. For example,
the books category contains 975194 reviews (123899
negatives and 851295 positives) while the tools &
hardware domain contains only 14 negative reviews
and 98 positives. In order not to bias the system, we
have taken care to balance the reviews between pos-
itive and negative. We limited the maximum number
of reviews to 1600 (composed by the 800 first pos-
itive and the 800 first negative reviews found in the
XML files). Using a small portion of the dataset, we
have therefore experimented different configurations
of parameters (λ,ϕ, L,A) by varying the propagation
rate between 0.1 and 0.9 in 0.1 steps, by testing all
values for annealing rate between 0.0 and 1.0 in 0.1
steps and using 10
−1
, 10
−2
and 10
−3
as values for the
convergence limit. Our experiments show that using a
propagation rate of 0.3, an annealing rate of 0.5, an
positive correction strength of 0.3 and a convergence
limit of 0.05 lead to the best results. The setting that
leads to the best results is λ = 0.3, ϕ = 0.3, L = 0.05,
and A = 0.5.
3.2 Results of DRANZIERA Evaluation
When applied to the DRANZIERA dataset with the
settings previously identified,
the average precision obtained over all 20 domains
is 0.8191, which constitutes a significant improve-
ment over MDFSA (Dragoni et al., 2015), MDFSA-
Table 1: Average precision, recall, F1 score and standard
deviation of the F1 score obtained on the 20 domains of the
DRANZIERA dataset.
Method Precision Recall F1 score SD
MDFSA 0.6832 0.9245 0.7857 0.0000
MDFSA-NODK 0.7145 0.9245 0.8060 0.0001
IRSA 0.6598 0.8742 0.7520 0.0002
IRSA-NODK 0.6784 0.8742 0.7640 0.0003
Sental w/ embedding 0.7527 0.9942 0.8551 0.0446
Sental w/ POS 0.7617 0.9947 0.8612 0.0435
Sental+ 0.8191 0.9941 0.8975 0.0275
NODK (Dragoni et al., 2015), IRSA (Dragoni, 2015)
and IRSA-NODK (Dragoni et al., 2016) which obtain
a precision of 0.6832, 0.7145, 0.6598 and 0.6784 re-
spectively.
It should be noted that this improvement in pre-
cision is not detrimental to the recall value, which
is higher than the other methods. As a result, the
proposed method obtains an even greater improve-
ment with respect to the other methods if perfor-
mance is measured in terms of the F1 score, in par-
ticular a 9.15% improvement with respect to the best
of them, MDFSA-NODK. Table 1 provides a sum-
mary of the comparison of the results obtained by
our method with four other methods evaluated on
the DRANZIERA dataset, whose results are provided
in (Dragoni et al., 2016). Using word embedding
(Sental w/ embedding) provides an improvement over
existing methods; incorporating POS tags (Sental w/
POS) further improves precision and recall, but the
best results are achieved by adding a positive cor-
rection term to the polarity propagation equation
(Sental+).
The breakdown of the results obtained by our
method implementing the four proposed solutions on
the 20 domains of the DRANZIERA dataset is pre-
sented in Table 2.
3.3 Cross-Domain Transfer Experiment
To test the generalization capabilities of our ap-
proach, we have strived to use our model, trained
with DRANZIERA data, on other datasets. Sen-
timent analysis has become extremely popular but
datasets available for use by multi-domain sentiment
analysis are still scarce. Two datasets, proposed by
Hutto and Gilbert (Hutto and Gilbert, 2014), cor-
respond to domains that could benefit from being
processed with our method. The first one, ‘Prod-
uct’, contains 3708 customer reviews of five elec-
tronics products. The second one, ‘Movie’, con-
tains 10,605 sentence-level snippets from the site http:
//rotten.tomatoes.com reanalyzed by 20 independent
human raters. ‘Product’ reviews should be efficiently
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
524

Table 2: Detail of the results obtained on the 20 domains of
the DRANZIERA dataset.
Domain Precision Recall F1 score SD
Amazon Video 0.7253 0.9946 0.8389 0.0310
Automotive 0.8402 0.9943 0.9108 0.0219
Baby 0.7357 0.9945 0.8457 0.0702
Beauty 0.8523 0.9947 0.9180 0.0085
Books 0.7910 0.9911 0.8798 0.0410
Clothing 0.8807 0.9973 0.9354 0.0401
Electronics 0.7998 0.9950 0.8868 0.0293
Health 0.8388 0.9947 0.9101 0.0208
Home Kitchen 0.8142 0.9951 0.8956 0.0399
Movies TV 0.8064 0.9934 0.8902 0.0273
Music 0.7764 0.9933 0.8716 0.0333
Office Products 0.8256 0.9949 0.9024 0.0168
Patio 0.8378 0.9945 0.9094 0.0201
Pet Supplies 0.7970 0.9912 0.8836 0.0265
Shoes 0.9183 0.9965 0.9558 0.0123
Software 0.8090 0.9943 0.8921 0.0216
Sports Outdoors 0.8320 0.9915 0.9048 0.0179
Tools Home Impr. 0.8430 0.9947 0.9126 0.0248
Toys Games 0.8581 0.9950 0.9215 0.0265
Video Games 0.7998 0.9916 0.8854 0.0201
Average 0.8191 0.9941 0.8975 0.0275
analyzed with our method trained on the ‘Electron-
ics’ domain of DRANZIERA dataset. Although the
DRANZIERA dataset does not contain the ‘Movie’
domain as such, it includes some related domains,
for example ‘Movies TV’ or ‘Amazon Instant Video’,
whose reviews may be used to infer the rating of
movies.
Our method, trained on each of the domains of
the DRANZIERA dataset was used to predict the ori-
entation of each ‘Movie’ and ‘Product’ review be-
tween positive and negative. Table 3 lists the pre-
cision values obtained on ‘Movie’ and ‘Product’ re-
views according to the category of the DRANZIERA
dataset used for training. Precisions displayed in
bold highlight the best score obtained for each do-
main. Recall values are not displayed because the
variation is small; they range from 0.9916 to 0.9944
for ‘Movie’ and from 0.9889 to 0.9964 for ‘Product’.
Overall, there is a decrease in performance when the
method, trained on DRANZIERA domains, is applied
to reviews originating from different datasets; which
seems perfectly logical. We can notice, in Table 3,
that the domain used for training has a great influence
on the results.
For a transfer to ‘Movie’, the best precision was
obtained when the training was performed with the
DRANZIERA reviews belonging to the ‘Movie TV’
domain. Although ‘Music’ is not exactly the category
for which one would have expected to obtain the best
cross-domain transfer quality, the result is still con-
sistent. The other domains for which the transfer of
the learned model is effective are ‘Books’, ‘Amazon
Table 3: Precision and recall of cross-domain transfer from
the 20 categories of the DRANZIERA dataset to indepen-
dant ‘Movie’ and ‘Product’ reviews.
DRANZIERA precision precision
domain on ‘Movie’ on ‘Product’
Amazon Instant Video 0.6962 0.5704
Automotive 0.6008 0.6490
Baby 0.5830 0.5735
Beauty 0.5992 0.6812
Books 0.7021 0.6196
Clothing Accessories 0.5885 0.6889
Electronics 0.5728 0.7309
Health 0.5899 0.6821
Home Kitchen 0.5875 0.6656
Movies TV 0.7134 0.6000
Music 0.6770 0.6507
Office Products 0.6135 0.6813
Patio 0.6099 0.6885
Pet Supplies 0.5960 0.6720
Shoes 0.5795 0.6843
Software 0.6154 0.7098
Sports Outdoors 0.5797 0.6891
Tools Home Improvement 0.5910 0.7086
Toys Games 0.6200 0.6103
Video Games 0.6595 0.6942
Instant Video’ and ‘Music’. Overall, the reviews that
most closely mirror those of the ‘Movie’ reviews are
more oriented towards cultural goods, while the most
distant ones concern more tangible goods (‘Shoes’,
‘Sport Outdoors’ and ‘Electronics’). The result is
exactly the opposite for the transfer to ‘Product’ re-
views. The best precision is obtained when the train-
ing was performed on ‘Electronics’ and the trans-
fer works better overall when the method has been
trained on reviews about concrete objects (‘Software’
and ‘Tools Home Imrpovement’ are both domains that
allow to exceed a precision of 70%.).
The precision therefore varies by nearly 0.15 be-
tween a transfer done from the ‘Music’ domain and
a transfer from the ‘Electronics’ domain. However,
the difference can also be partly explained by the
intrinsic score obtained on each DRANZIERA do-
main. We can indeed notice in Table 3, that the dif-
ference in precision between the ‘Music’ domain and
the ‘Baby’ domain is slightly more than 0.14. How-
ever, we can observe significant differences accord-
ing to the domains. Some cross-domain transfers
go very well, such as the transfer from domains like
‘Books’, ‘Movies TV’ or ‘Amazon Instant Video’ to
movie reviews since the accuracy decline is less than
0.03. Other cross-domain transfers are more prob-
lematic, such as those from ‘Electronics’, ‘Beauty’,
‘Sports Outdoors’, ‘Clothing Accessories’ or ‘Shoes’
to movies reviews since the accuracy falls by more
than 0.2. These observations also seem to make per-
fect sense.
Propagation-Based Domain-Transferable Gradual Sentiment Analysis
525

Table 4: A comparison of our method (trained on the
‘Movies TV’ domain of DRANZIERA) with the best meth-
ods benchmarked by Ribeiro et al., when applied to the
‘Movie’ dataset. The highest score of each column is high-
lighted in boldface.
Method Precision Recall F1 score
AFINN 0.6593 0.7259 0.6910
LIWC15 0.6335 0.6608 0.6469
Opinion Lexicon 0.6977 0.7728 0.7333
Pattern.en 0.6784 0.6559 0.6670
Semantria 0.6964 0.6880 0.6922
SenticNet 0.9630 0.6941 0.8067
SO-CAL 0.7165 0.8910 0.7943
Stanford DM 0.8270 0.9192 0.8707
VADER 0.6519 0.8270 0.7291
Sental+ 0.7134 0.9944 0.8308
To assess the performance of our method we com-
pared it with the benchmark of 24 sentiment analysis
methods performed by Ribeiro et al. (Ribeiro et al.,
2016).
Since the setting we have chosen leads to a high
recall value at the expense of precision, we list in ta-
bles 4 and 5 the prediction performance of all meth-
ods that achieve, in either domain, an F1 score greater
than 0.5 and a higher precision than our method.
Methods meeting these criteria are AFINN (Nielsen,
2011), LIWC15 (Tausczik and Pennebaker, 2010),
Opinion Lexicon (Hu and Liu, 2004),
Pattern.en (Smedt and Daelemans, 2012), Seman-
tria (Lexalytics, 2015), SenticNet (Cambria et al.,
2018),
SO-CAL (Taboada et al., 2011), Stanford
DM (Socher et al., 2013), and VADER (Hutto and
Gilbert, 2014).
Regarding the ‘Movie’ domain, the best method,
consisting in using only the polarities reported in Sen-
ticNet, obtains a high precision but only on a rela-
tively small part of the reviews since its coverage is
69.41%. The second and third best methods, Stanford
DM and SO-CAL, have a coverage of the same or-
der, 91.92% and 89.10% respectively, which are both
below Sental+’s coverage value, 99.44%.
All other methods are outperformed by Sental+,
by both criteria.
When recall is considered together with preci-
sion (cf. Table 4), our method obtains an F1 score of
83.08%, which is higher than the F1 score of the most
precise method, 80.67%, but short of Stanford DM,
which has the highest F1 score with 87.07%, while
SO-CAL is lower, at 79.43%.
Regarding the ‘Product’ domain,
a group of methods (AFINN, LIWC15, Opinion
Lexicon, Pattern.en and Semantria) achieves a preci-
sion close to 80% with a recall ranging from 57% to
Table 5: A comparison of our method (trained on the ‘Elec-
tronic’ domain of DRANZIERA) with the best methods
benchmarked by Ribeiro et al., when applied to the ‘Prod-
uct’ dataset. The highest score of each column is high-
lighted in boldface.
Method Precision Recall F1 score
AFINN 0.7869 0.6280 0.6985
LIWC15 0.7674 0.5645 0.6505
Opinion Lexicon 0.8082 0.6715 0.7335
Pattern.en 0.7571 0.5953 0.6665
Semantria 0.8159 0.5945 0.6878
SenticNet 0.6991 0.9748 0.8142
SO-CAL 0.7823 0.7152 0.7472
Stanford DM 0.6853 0.8028 0.7394
VADER 0.7705 0.7133 0.7408
Sental+ 0.7309 0.9956 0.8430
67%. SO-CAL and VADER score well in both preci-
sion and recall. In contrast to the ‘Movie’ domain,
SenticNet and Stanford DM achieve modest preci-
sions of less than 70%, compensated by high recall
values.
The precision of 73.09% scored by Sental+, (i.e.
better than 70%, as for the ‘Movie’ domain) and its
recall of 99.56% allows the method to obtain the best
F1 score. Overall, Sental+ appears to have a more
consistent performance. It also features an excellent
coverage, which is always above 99%, meaning that
less than 1% of the reviews are not classified.
4 CONCLUSION
We have proposed Sental+, an extension of the Sen-
tal method inspired by the MDFSA approach, which
incorporates an original solution to overcome a ma-
jor issue of Sental: a positive correction term in the
polarity propagation phase effectively counteracts the
negative bias of the reviews.
The resulting method has been validated using a
standard evaluation protocol, showing a significant
improvement with respect to the state of the art. We
have also tested the cross-domain generalization ca-
pabilities of our approach with very promising results.
Injecting more linguistic background knowledge,
as Sental does (here, POS tagging and word embed-
ding) into the semantic graph appears to improve both
the precision and the coverage of the method. This
suggests focusing future work in this direction, for in-
stance by exploiting large language models.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
526

REFERENCES
Abdullah, N. A., Feizollah, A., Sulaiman, A., and Anuar,
N. B. (2019). Challenges and recommended solutions
in multi-source and multi-domain sentiment analysis.
IEEE Access, 7:144957–144971.
Baumeister, R. F., Bratslavsky, E., Finkenauer, C., and
Vohs, K. D. (2001). Bad is stronger than good. Review
of general psychology, 5(4):323–370.
Blitzer, J., Dredze, M., and Pereira, F. (2007). Biographies,
bollywood, boom-boxes and blenders: Domain adap-
tation for sentiment classification. In ACL, pages 187–
205.
Cambria, E., Poria, S., Hazarika, D., and Kwok, K. (2018).
Senticnet 5: Discovering conceptual primitives for
sentiment analysis by means of context embeddings.
In AAAI, pages 1795–1802. AAAI Press.
Cambria, E., Speer, R., Havasi, C., and Hussain, A. (2010).
Senticnet: A publicly available semantic resource for
opinion mining. In AAAI Fall Symposium: Common-
sense Knowledge, volume FS-10-02 of AAAI Techni-
cal Report, pages 14–18. AAAI Press.
Casas, B., Hern
´
andez-Fern
´
andez, A., Catal
`
a, N., Ferrer-
i Cancho, R., and Baixeries, J. (2019). Polysemy
and brevity versus frequency in language. Computer
Speech & Language, 58:19–50.
Dragoni, M. (2015). Shellfbk: an information retrieval-
based system for multi-domain sentiment analysis. In
Proceedings of the 9th International Workshop on Se-
mantic Evaluation (SemEval 2015), pages 502–509.
Dragoni, M., Tettamanzi, A. G. B., and da Costa Pereira,
C. (2014). A fuzzy system for concept-level senti-
ment analysis. In SemWebEval@ESWC, volume 475
of Communications in Computer and Information Sci-
ence, pages 21–27. Springer.
Dragoni, M., Tettamanzi, A. G. B., and da Costa Pereira,
C. (2015). Propagating and aggregating fuzzy polar-
ities for concept-level sentiment analysis. Cognitive
Computation, 7(2):186–197.
Dragoni, M., Tettamanzi, A. G. B., and da Costa Pereira,
C. (2016). DRANZIERA: an evaluation protocol
for multi-domain opinion mining. In Proceedings of
the 10th International Conference on Language Re-
sources and Evaluation (LREC 2016), Paris, France,
pages 267–272, Paris, France. European Language
Resources Association (ELRA).
Hu, M. and Liu, B. (2004). Mining and summarizing
customer reviews. In Proceedings of the tenth ACM
SIGKDD international conference on Knowledge dis-
covery and data mining, pages 168–177.
Hutto, C. J. and Gilbert, E. (2014). Vader: A parsimonious
rule-based model for sentiment analysis of social me-
dia text. In Eighth international AAAI conference on
weblogs and social media.
Kirkpatrick, S., Jr., D. G., and Vecchi, M. P. (1983).
Optimization by simulated annealing. SCIENCE,
220(4598):671–680.
Lexalytics (2015). Sentiment extraction - measuring the
emotional tone of content. Technical Report.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).
Efficient estimation of word representations in vector
space. arXiv preprint arXiv:1301.3781.
Miller, G. A. (1995). Wordnet: a lexical database for en-
glish. Communications of the ACM, 38(11):39–41.
Nielsen, F.
˚
A. (2011). A new anew: Evaluation of a
word list for sentiment analysis in microblogs. arXiv
preprint arXiv:1103.2903.
Pasquier, C., da Costa Pereira, C., and Tettamanzi, A.
G. B. (2020). Extending a fuzzy polarity propaga-
tion method for multi-domain sentiment analysis with
word embedding and POS tagging. In ECAI, volume
325 of Frontiers in Artificial Intelligence and Applica-
tions, pages 2140–2147. IOS Press.
Pirnau, M. (2018). Sentiment analysis for the tweets that
contain the word “earthquake”. In 2018 10th Inter-
national Conference on Electronics, Computers and
Artificial Intelligence (ECAI), pages 1–6.
Rexha, A., Kr
¨
oll, M., Dragoni, M., and Kern, R. (2018).
The CLAUSY system at ESWC-2018 challenge on se-
mantic sentiment analysis. In SemWebEval@ESWC,
volume 927 of Communications in Computer and In-
formation Science, pages 186–196. Springer.
Ribeiro, F. N., Ara
´
ujo, M., Gonc¸alves, P., Gonc¸alves, M. A.,
and Benevenuto, F. (2016). Sentibench-a benchmark
comparison of state-of-the-practice sentiment analysis
methods. EPJ Data Science, 5(1):23.
Schouten, K. and Frasincar, F. (2015). The benefit of
concept-based features for sentiment analysis. In
SemWebEval@ESWC, volume 548 of Communica-
tions in Computer and Information Science, pages
223–233. Springer.
Smedt, T. D. and Daelemans, W. (2012). Pattern for
python. Journal of Machine Learning Research,
13(Jun):2063–2067.
Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning,
C. D., Ng, A., and Potts, C. (2013). Recursive deep
models for semantic compositionality over a senti-
ment treebank. In Proceedings of the 2013 conference
on empirical methods in natural language processing,
pages 1631–1642.
Taboada, M., Brooke, J., Tofiloski, M., Voll, K. D., and
Stede, M. (2011). Lexicon-based methods for senti-
ment analysis. Computational Linguistics, 37(2):267–
307.
Tausczik, Y. R. and Pennebaker, J. W. (2010). The psy-
chological meaning of words: Liwc and computerized
text analysis methods. Journal of language and social
psychology, 29(1):24–54.
Yoshida, Y., Hirao, T., Iwata, T., Nagata, M., and Mat-
sumoto, Y. (2011). Transfer learning for multiple-
domain sentiment analysis—identifying domain de-
pendent/independent word polarity. In AAAI, pages
1286–1291.
Propagation-Based Domain-Transferable Gradual Sentiment Analysis
527