Tracking Brand-Associated Polarity-Bearing Topics in User Reviews

Runcong Zhao1,2, Lin Gui1, Hanqi Yan2, Yulan He1,2,3
1King’s College London, United Kingdom, 2University of Warwick, United Kingdom,

3The Alan Turing Institute, United Kingdom

runcong.zhao@warwick.ac.uk, yulan.he@kcl.ac.uk

Abstrait

Monitoring online customer reviews is im-
portant for business organizations to measure
customer satisfaction and better manage their
reputations. In this paper, we propose a novel
dynamic Brand-Topic Model (dBTM) lequel
is able to automatically detect and track brand-
associated sentiment scores and polarity-
bearing topics from product reviews organized
in temporally ordered time intervals. dBTM
models the evolution of the latent brand po-
larity scores and the topic-word distributions
over time by Gaussian state space models.
It also incorporates a meta learning strategy
to control the update of the topic-word distri-
bution in each time interval in order to ensure
smooth topic transitions and better brand score
prédictions. It has been evaluated on a data-
set constructed from MakeupAlley reviews
and a hotel review dataset. Experimental re-
sults show that dBTM outperforms a number
of competitive baselines in brand ranking,
achieving a good balance of topic coherence
and uniqueness, and extracting well-separated
polarity-bearing topics across time intervals.1

Introduction

With the increasing popularity of social media
platforms, customers tend to share their personal
experience towards products online. Tracking cus-
tomer reviews online could help business orga-
nizations to measure customer satisfaction and
better manage their reputations. Monitoring brand-
associated topic changes in reviews can be done
through the use of dynamic topic models (Blei
and Lafferty, 2006; Wang et al., 2008; Dieng et al.,
2019). Approaches such as the dynamic Joint
Sentiment-Topic (dJST) model (He et al., 2014)
are able to extract polarity-bearing topics evolved
over time by assuming the dependency of the
sentiment-topic-word distributions across time

1Data and code are available at https://github

.com/BLPXSPG/dBTM.

404

slices. They require the incorporation of word
prior polarity information, cependant, and assume
topics are associated with discrete polarity cat-
egories. En outre, they are not able to infer
brand polarity scores directly.

A recently proposed Brand-Topic Model
(BTM) (Zhao et al., 2021) is able to automat-
ically infer real-valued brand-associated senti-
ment scores from reviews and generate a set of
sentiment-topics by gradually varying its associ-
ated sentiment scores from negative to positive.
This allows users to detect, Par exemple, strongly
positive topics or slightly negative topics. BTM,
cependant, assumes all documents are available
prior to model learning and cannot track topic
evolution and brand polarity changes over time.

In this paper, we propose a novel framework
inspired by Meta-Learning, which is widely used
for distribution adaptation tasks (Suo et al., 2020).
When training the model on temporally ordered
documents divided into time slice, we assume
that extracting polarity-bearing topics and infer-
ring brand polarity scores in each time slice can
be treated as a new sub-task and the goal of model
learning is to learn to adapt the topic-word dis-
tributions associated with different brand polarity
scores in a new time slice. We use BTM as the
base model and store the parameters learned in a
mémoire. At each time slice, we gauge model per-
formance on a validation set based on the model-
generated brand ranking results. The evaluation
results are used for early stopping and dynam-
ically initializing model parameters in the next
time slice with meta learning. The resulting model
is called dynamic Brand Topic Modeling (dBTM).
The final outcome from dBTM is illustrated in
Chiffre 1, in which it can simultaneously track topic
evolution and infer latent brand polarity score
changes over time. De plus, it also enables the
generation of fine-grained polarity-bearing topics
in each time slice by gradually varying brand po-
larity scores. En substance, we can observe topic

Transactions of the Association for Computational Linguistics, vol. 11, pp. 404–418, 2023. https://doi.org/10.1162/tacl a 00555
Action Editor: David Bamman. Submission batch: 9/2022; Revision batch: 11/2022; Published 5/2023.
c(cid:2) 2023 Association for Computational Linguistics. Distributed under a CC-BY 4.0 Licence.

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 1: Brand-associated polarity-bearing topics tracking by our proposed model. We show top words from
an example topic extracted in time slice 1, 4, et 8 along the horizontal axis. In each time slice, we can see a
set of topics generated by gradually varying their associated sentiment scores from −1 (negative) à 1 (positive)
along the vertical axis. For easy inspection, positive words are highlighted in blue while negative ones in red.
We can observe in Time 1, negative topics are mainly centred on the complaint of the chemical smell of a
perfume, while positive topics are about the praise of the look of a product. From Time 1 to Time 8, we can also
see the evolving aspects in negative topics moving from complaining about the strong chemical of perfume to
overpowering sweet scent. In the lower part of the figure, we show the inferred polarity scores of three brands.
Par exemple, Chanel is generally ranked higher than Lancˆome, which in turn scores higher than The Body Shop.

transitions in two dimensions, either along a dis-
crete time dimension, or along a continuous brand
polarity score dimension.

We have evaluated dBTM on a review dataset
constructed from MakeupAlley,2 consisting of
over 611K reviews spanning over 9 années, et
a hotel review dataset sampled from HotelRec
(Antognini and Faltings, 2020), containing re-
views of the most popular 25 hotels over 7 années.
We compare its performance with a number of
competitive baselines and observe that it gener-
ates better brand ranking results, predicts more
accurate brand score time series, and produces
well-separated polarity-bearing topics with more
balanced topic coherence and diversity. More in-
terestingly, we have evaluated dBTM in a more
difficult setup, where the supervised label infor-
mation, c'est, review ratings, is only supplied
in the first time slice, and afterwards, dBTM is
trained in an unsupervised way without the use of
review ratings. dBTM under such a setting can still
produce brand ranking results across time slices
more accurately compared to baselines trained
under the supervised setting. This is a desirable
property as dBTM, initially trained on a small set
of labeled data, can self-adapt its parameters with
streaming data in an unsupervised way.

2https://www.makeupalley.com/.

Our contributions are three-fold:

• We propose a new model, called dBTM, built
on the Gaussian state space model with meta
learning for dynamic brand topic and polarity
score tracking;

• We develop a novel meta learning strategy to
dynamically initialize the model parameters
at each time slice in order to better capture
rating score changes, which in turn generates
topics with a better overall quality;

• Our experimental results show that dBTM
trained with the supervision of review rat-
ings at the initial time slice can self-adapt
its parameters with streaming data in an un-
supervised way and yet still achieve better
brand ranking results compared to supervised
baselines.

2 Related Work

Our work is related to the following research:

2.1 Dynamic Topic Models

Topic models such as the Latent Dirichlet Allo-
cation (LDA) model (Blei et al., 2003) is one of
the most successful approaches for the statistical
analysis of document collections. Dynamic topic
models aim to analyse the temporal evolution of

405

topics in large document collections over time.
Early approaches built on LDA include the dy-
namic topic model (DTM) (Blei and Lafferty,
2006), which uses the Kalman filter to model the
transition of topics across time, and the continu-
ous time dynamic topic model (Wang et al., 2008),
which replaced the discrete state space model of
the DTM with its continuous generalization. More
recently, DTM has been combined with word em-
beddings in order to generate more diverse and
coherent topics in document streams (Dieng et al.,
2019).

Apart from the commonly used LDA, Poisson
factorization can also be used for topic model-
ing, in which it factorizes a document-word count
matrix into a product of a document-topic matrix
and a topic-word matrix. It can be extended to
analyse sequential count vectors such as a docu-
ment corpus which contains a single word count
matrix with one column per time interval, by cap-
turing dependence among time steps by a Kalman
filter (Charlin et al., 2015), neural networks (Gong
and Huang, 2017), or by extending a Poisson dis-
tribution on the document-word counts as a non-
homogeneous Poisson process over time (Hosseini
et coll., 2018).

While the aforementioned models are typi-
cally used in the unsupervised setting, the Joint
Sentiment-Topic (JST) model (Lin and He, 2009;
Lin et al., 2012) incorporated the polarity word
prior into model
learning, which enables the
extraction of topics grouped under different senti-
ment categories. JST is later extended into a dy-
namic counterpart, called dJST, which tracks both
topic and sentiment shifts over time (He et al.,
2014) by assuming that the sentiment-topic word
distribution at the current time is generated from
the Dirichlet distribution parameterised by the
sentiment-topic word distributions at previous
time intervals.

2.2 Market/Brand Topic Analysis

LDA and its variants have been explored for
marketing research. Examples include user inter-
ests detection by analyzing consumer purchase
behavior (Gao et al., 2017; Sun et al., 2021), le
tracking of the competitors in the luxury mar-
ket among given brands by mining the Twitter
data (Zhang et al., 2015), and identify emerging
app issues from user reviews (Yang et al., 2021).
Matrix factorization, which is able to extract the
global information, is also used to be applied in

product recommendation (Zhou et al., 2020) et
review summarization (Cui and Hu, 2021). Le
interaction between topics and polarities can be
modeled by the incorporation of approximations
by sampling based methods (Lin and He, 2009)
with sentiment prior knowledge such as senti-
ment lexicon (Lin et al., 2012). But such prior
knowledge would be highly domain-specific. Seed
words with known polarities or seed words gen-
erated by morphological information (Brody and
Elhadad, 2010) is another common method to ob-
tain topic polarity. But those methods are focused
on analying the polarity of existing topics. More
recently, the Brand-Topic Model built on Poisson
factorization was proposed (Zhao et al., 2021),
which can infer brand polarity scores and generate
fine-grained polarity-bearing topics. The detailed
description of BTM can be found in Section 3.

2.3 Meta Learning

Meta learning, ou
learning to learn, can be
broadly categorized into metric-based learning
and optimization-based learning. Metric-based
learning aims to learn a distance function between
training instances so that it can classify a test in-
stance by comparing it with the training instances
in the learned embedding space (Sung et al., 2018).
Optimization-based learning usually splits the la-
beled samples into training and validation sets.
The basic idea is to fine-tune the parameters on
the training set to obtain the updated parame-
ters. These are then evaluated on the validation
set to get the error, which is converted as a loss
value for optimizing the original parameters (Finn
et coll., 2017; Jamal and Qi, 2019). Meta learn-
ing has been explored in many tasks, y compris
text classification (Geng et al., 2020), topic mod-
eling (Song et al., 2020), knowledge representa-
tion (Zheng et al., 2021), recommender systems
(Neupane et al., 2021; Dong et al., 2020; Lu et al.,
2020), and event detection (Deng et al., 2020).
Especially, the meta learning based methods have
achieved significant successes in distribution
adaptation (Suo et al., 2020; Yu et al., 2021). Nous
propose a meta learning strategy here to learn
how to automatically initialize model parameters
in each time slice.

3 Preliminary: Brand Topic Model

The Brand-Topic Model (BTM) (Zhao et al.,
2021), as shown in the middle part of Figure 2,

406

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

from a Poisson distribution cdv ∼ Poisson(λdv)
where the rate parameter λdv can be factorized as:
(cid:2)

λdv =

θdkβkv

(1)

Ici, θdk denotes the per-document topic in-
tensity, and βkv represents the topic-word distri-
bution. We have θ ∈ RD×K
, où
+
D is the total number of documents in a corpus, K
is the topic number, and V is the vocabulary size.
Alors, brand-polarity score xbd and topic-word
offset ηkv are added to the model:

, β ∈ RK×V

(cid:2)

λdv =

θdk exp(log βkv + xbdηkv)

(2)

xbd is the brand polarity score for document
d of brand b and we have η ∈ RK×V
, x ∈ R.
The model-normalized brand polarity assignment
à [−1,1] in its output for demonstration purposes.
The intuition behind the above formulation is
that the latent variable xbd, which captures the
brand polarity score, can be either positive or
negative. If a word tends to frequently occur in
reviews with positive polarities, but the polarity
score of the current brand is negative, then the
occurrence count of such a word would be reduced
by making xbd and ηkv to have opposite signs.

A Gamma prior is placed on θ and β, avec
un, b, c, d being hyperparameters, while a normal
prior is placed over the brand polarity score x and
the topic-word count offset η.

θdk ∼ Gamma(un, b), βkv ∼ Gamma(c, d),

xbd

∼ N (0, 1),

ηkv ∼ N (0, je)

BTM makes use of Gumbel-Softmax (Jang
et coll., 2017) to construct document features for
sentiment classification. This is because directly
sampling word counts from the Poisson distribu-
tion is not differentiable. Gumbel-Softmax, lequel
is a gradient estimator with the reparameterization
trick, is used to enable back-propagation of gra-
dients. More details can be found in Zhao et al.
(2021).

4 Dynamic Brand Topic Model (dBTM)

To track brand-associated topic dynamics in cus-
tomer reviews, we split the documents into time
slices where the time period of each slice can be
set arbitrarily at, Par exemple, a week, a month, ou

Chiffre 2: The overall architecture of the dynamic
Brand-Topic Model (dBTM), which extends the Brand-
Topic Model (BTM) shown in the upper box to deal
with streaming documents. En particulier, at time slice
t, the document-topic distribution θt is initialized by
a vanilla Poisson factorization model, the evolution
of the latent brand-associated polarity scores xt and
the polarity-associated topic-word offset ηt is mod-
eled by two separate Gaussian state space models. Le
topic-word distribution βt has its prior set based on
the trend of the model performance on brand ranking
results in the previous two time slices. Lines colored
in gray indicate parameters are linked by Gaussian
state space models, while those colored in green indi-
cate forward calculations.

is trained on review documents paired with their
document-level sentiment class labels (par exemple., ‘Pos-
itive’, ‘Negative’, and ‘Neutral’). It can automat-
ically infer real-valued brand-associated polarity
scores and generate fine-grained sentiment-topics
in which a continuous change of words under
a certain topic can be observed with a gradual
change of its associated sentiment. It was partly
inspired by the Text-Based Ideal Point (TBIP)
model (Vafa et al., 2020), which aims to model
the generation of text via Poisson factorization.
En particulier, for the input bag of words data, le
count for term v in document d is formulated as
term count cdv, which is assumed to be sampled

407

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

a year. In each time slice, we have a stream of M
documents {d1, · · · , dM } ordered by their pub-
lication timestamps. A document d at time slice
t is input as a Bag-of-Words representation. Nous
extend BTM to deal with streaming documents
by assuming that documents at the current time
slice are influenced by documents at past. Le
resulting model is called dynamic Brand-Topic
Model (dBTM), with its architecture illustrated
in Figure 2.

4.1 Initialization

the latent vari-
In the original BTM model,
ables to be inferred include the document-
topic distribution θ, topic-word distribution β,
the brand-associated polarity score x, et le
polarity-associated topic-word offset η. At time
slice 0, we represent all documents in this slice as
a document-word count matrix. We then perform
Poisson factorization with coordinate-ascent vari-
ational inference (Gopalan et al., 2015) to derive
θ and β (see Eq. (1)). The topic-word count off-
set η and the brand polarity score x are sampled
from a standard normal distribution.

4.2 State-Space Model

At time slice t, we can model the evolution of
the latent brand-associated polarity scores xt and
the polarity-associated topic-word offset ηt over
time by a Gaussian state space model:

xt|xt−1 ∼ N (xt−1, σ2
ηt|ηt−1 ∼ N (ηt−1, σ2

xI)
ηI)

(3)

(4)

For the topic-word distribution β, a similar
Gaussian state-space model is adopted except that
log-normal distribution is used:

βt|βt−1 ∼ LN (βt−1, σ2

βI)

(5)

While topic-word distribution could be inher-
ited from a previous time slice, the document-topic
distribution θt needs to be re-initialized at the
start of each time slice since there is a different
set of documents at each time slice. We propose
to run a simple Poisson factorization to derive
the initial values of θt
(p) before we do the model
adaption at each time slice: Ici, the topic-word
distribution in the previous time slice βt−1
(p) être-
comes the prior of the topic-word distribution in
the current time slice βt
(p) as defined in Eq. (5).

We use the subscript (p) to denote that the pa-
rameters are derived in the Poisson factoriation
initialization stage at the start of each time slice.

Essentially, at each time slice t, we initialize
the document-topic distribution θt of the BTM
model as θt
(p), which is obtained by performing
Poisson factorization on the document-word count
matrix in t. For the topic-word distribution, within
BTM, we can set βt to be inherited from βt−1 as
defined in Eq. (5), but additionally, we also have
βt
(p), which is obtained by directly performing
Poisson factorization of the document-word count
matrix in the current time slice. In what follows,
we will present how we initialize the value of
βt through meta learning.

4.3 Meta Learning

We notice that although parameters in each time
interval are linked with parameters in the previ-
ous time interval by Gaussian state-space models,
the results generated at each time interval are not
stable. Inspired by meta learning, we consider la-
tent brand score prediction and sentiment topic
extraction at each time interval as a new sub-task,
and propose a learning strategy to dynamically
initialize model parameters in each interval based
on the brand rating prediction performance on the
validation set of the previous interval. In partic-
ular, we set aside 10% of the training data in
each time interval as the validation set and com-
pare the model-inferred brand ranking result with
the gold standard one using the Spearman’s rank
correlation coefficient. By default, the topic-word
distribution in the current interval, βt, would have
its prior set to βt−1 learned in the previous time
interval. Cependant, if the brand ranking result in
the previous interval is poor, then βt would be
initialized as a weighted interpolation of βt−1
and the topic-word distribution obtained from the
Poisson factorization initialization stage in the cur-
rent interval βt
(p). The rationale is that if the model
performs poorly in the previous interval, then its
learned topic-word distribution should have less
impact on the parameters in the current inter-
val. More concretely, we first evaluate the brand
ranking result returned by the model at time slice
t − 1 on the validation set at t − 1:

ρt−1 = SpearmanRank(ˆrt−1, rt−1)

(6)

where ˆrt−1 denotes the derived brand ranking re-
sult based on the model predicted latent brand

408

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

polarity scores, ˆxt−1, at time slice t − 1, rt−1 is
the gold-standard brand ranking, and ρt−1 is the
Spearman’s rank correlation coefficient. To check
if the brand ranking result gets worse or not,
we compare it with the brand ranking evaluation
result, ρt−2, in the earlier interval. In particu-
lar, we first take Fisher’s z-transformation zρt−1
of ρt−1, which is assumed following a Gaussian
distribution:

(cid:3)

zρt−1 ∼ N

ln(

1 + ρt−1
1 − ρt−1 )0.5,

1
B − 3

(cid:4)

(7)

where B denotes the total number of brands. Alors
we compute the Cumulative Distribution Function
(CDF) of the above normal distribution, denoted
as Φzρt−1 , and calculate Φzρt−1 (ρt−2), which es-
sentially returns Pr(zρt−1 ≤ ρt−2). Lower value
of Φzρt−1 (ρt−2) indicates that the model at t − 1
generates a better brand rank result than that in
the previous time slice t − 2. This is equivalent to
performing a hypothesis test in which we compare
the rank evaluation result ρt−1 with ρt−2 to test
if the model at t − 1 performs better than that at
t − 2. The hypothesis testing result can be used
to set the weight γt to determine how to initialize
the topic-word distribution at t, βt:

(cid:5)
γt = max
βt = (1 − γt)βt−1 + γtβt
(p)

0.05, Φzρt−1 (ρt−2)

(cid:6)

(8)

(9)

The above equations state that if the model trained
at t − 1 generates a better brand ranking result
than that in the previous time slice significantly
(p-value > 0.05), then we are more confident to
initialize βt largely based on βt−1 according to
the estimated probability of Pr(zρt−1 > ρt−2) =
1 − γt. Otherwise, we will have to re-initialize
βt mostly based on the topic-word distribution
obtained from the Poisson factorization initializa-
tion stage in the current interval βt

(p).

4.4 Parameter Inference

We use the mean-field variational distribution
to approximate the posterior distribution of la-
tent variables, je, β, η, X, given the observed
document-word count data c by maximizing the
Evidence Lower-Bound (ELBO):

LELBO = Eqφ[log p(je, β, η, X)]+
log p(c|je, β, η, X) − log qφ(je, β, η, X)]

(10)

où

(cid:7)

qφ(je, β, η, X) =

q(θd)q(βk)q(ηk)q(xb)

d,k,b

(11)

En outre, for each document d, we construct
its representation zd by sampling word counts
using Gumbel softmax from the aforementioned
learned parameters, which is fed to a sentiment
classifier to predict a class distribution ˆyd. Nous
also perform adversarial learning by inverting the
sign of the inferred polarity score of the brand
associated with document d and produce the ad-
versarial representation ˜zd. This is also fed to the
same sentiment classifier which generates another
predicted class distribution ˜yd. We train the model
by minimizing the Wasserstein distance between
the prediction and the actual class distributions.
The final loss function is the combination of the
ELBO and the Wasserstein distance losses:

L = −LELBO +

1
M.

M.(cid:2)

(cid:5)

d=1

LWD(ˆyd, yd)
(cid:6)

+LWD(˜yd, ¯yd)

(12)

where LWD(·) denotes the Wasserstein distance,
yd is the gold-standard class distribution, and ¯yd
is the class distribution derived from the inverted
document rating. By inverting the document rat-
ing, we essentially balance the document rating
distributions that for each positive document, nous
also create a synthetic negative document, et
vice versa.

5 Experimental Setup

Datasets Popular datasets such as Yelp and
Amazon products (Ni et al., 2019) and Multi-
Domain Sentiment dataset (Blitzer et al., 2007) sont
constructed by randomly selecting reviews from
Amazon or Yelp without considering their dis-
tributions over various brands and across different
time periods. Donc, we construct our own
dataset by crawling reviews from top 25 brands
from MakeupAlley, a review website on beauty
des produits. Each review is accompanied with a rat-
ing score, product type, brand, and post time. Nous
consider reviews with the ratings of 1 et 2 as the
negative class, those with the rating of 3 as the
neutral class, and the remaining with the ratings
de 4 et 5 as the positive class, following the la-
bel setting in BTM. The entire dataset contains

409

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Algorithm 1: Training procedure of dBTM
Input

: Number of topics K, number of brands B, time slice t ∈ {0, 1, 2, · · · , T }, a stream of document-word count matrix
C = {c0, · · · , cT }

Output: Document-topic intensity {θt}T

t=0, Topic-word matrix {βt}T

t=0, brand scores {xt

1, · · · , xt
B

t=0

η0 ∼ N (0, Σ2

ηI),

γ0 = 0

1 Initialization:
2 Initialize θ0, β0 by Poisson factorization c0 ∼ Poisson(θ0β0),
3 Update parameters by minimizing the loss defined in Eq. (12)
4 Derive the brand ranking ˆr0 on the validation set based on the inferred brand polarity score {x0
b
5 Calculate the Spearman’s rank correlation coefficient ρ0 = SpearmanRank( ˆr0, r0)
(cid:3)
6 Set the weight γ1 = max
7 Training:
8 for t = 1 to T do
9

sx0 ∼ N (0, je),

0.05, Φzρ0 (0)

(cid:2)

b=1

(p)βt

Pre-training: ct ∼ Poisson(θt
Per epoch initialization:
xt ∼ N (xt−1, σ2
for i = 0 to maximum iterations do

(p))

xI), ηt ∼ N (ηt−1, σ2

ηI), θt = θt

(p), βt = (1 − γt)βt−1 + γtβt

(p)

Update parameters by minimizing the loss defined in Eq. (12)
if checkpoint then

Derive the brand ranking ˆrt on the validation set based on the inferred brand polarity score {xt
b
Calculate the Spearman’s rank correlation coefficient ρt = SpearmanRank( ˆrt, rt)
si (Φzρt (ρt−1) > 0.95) alors

b=1

break # Null hypothesis is rejected by the upper quartile according to Eq. (7)

end
Set the weight γt+1 = max

(cid:2)

0.05, Φzρt (ρt−1)

(cid:3)

end

22
23 end

Dataset

MakeupAlley-Beauty Reviews

Non. of documents per class

Neg / Neu / Pos

Non. of brands
Total no. of documents
Non. of time Slices
Average review length (#words)
Average no. of documents per slice
Vocabulary size

Dataset

Non. of documents per class

Neg / Neu / Pos

Non. of hotels
Total no. of documents
Non. of time Slices
Average review length (#words)
Average no. of documents per slice
Vocabulary size

114,837 / 88,710 / 407,581
25
611,128
9
123
∼ 68k
∼ 4500

HotelRec Reviews

14,600 / 20,629 / 150,265
25
185,496
7
204
∼ 26k
∼ 7000

Tableau 1: Dataset statistics of the reviews.

611,128 reviews spanning over 9 années (2005
à 2013). We treat each year as a time slice and
split reviews into 9 time slices. The average review
length is 123 words. Besides the MakeupAlley-
Beauty, we also run our experiments on HotelRec
(Antognini and Faltings, 2020), by selecting re-
views from the top 25 hotels over 7 années (2012
à 2018). The statistics of our datasets are shown
in Table 1. It can be observed that the dataset
is imbalanced with positive reviews being over

triple the size of negative ones for MakeupAlley-
Beauty and nearly 10 times for HotelRec.

Models for Comparison We conduct experi-
ments using the following models:

• Dynamic Joint Sentiment-Topic (dJST) model
(He et al., 2014), built on LDA, can detect
and track polarity-bearing topics from text
with the word prior sentiment knowledge
incorporated. In our experiments, the MPQA
subjectivity lexicon3 is used to derive the
word prior sentiment information.

• Text-Based Ideal Point (TBIP) (Vafa et al.,
2020), an unsupervised Poisson factorization
model which can infer latent brand sentiment
scores.

• Brand Topic Model (BTM)

(Zhao et al.,
2021), a supervised Poisson factorization
model extended from TBIP with the in-
corporation of document-level sentiment
labels.

• dBTM, our proposed dynamic Brand Topic
model in which the model is trained with the
document-level sentiment labels at each time
slice.

3https://mpqa.cs.pitt.edu/lexicons/.

410

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Time Slice

dJST

TBIP

BTM

O-dBTM

dBTM

Corr

p-value

Corr

p-value

Corr

p-value Corr

p-value

1
2
3
4
5
6
7
8

1
2
3
4
5
6

MakeupAlley-Beauty

−0.249
−0.437
−0.327
−0.127
0.112
−0.118
−0.203
−0.552

0.097
−0.242
−0.112
−0.362
−0.045
0.222

0.230 −0.567
0.527
0.029
0.111 −0.543
0.545 −0.431
0.596 −0.347
0.573 −0.392
0.400
0.330
0.348
0.004

0.121
0.645
0.443
0.244
0.596 −0.392
0.276
0.076
0.292
0.829
0.298
0.285

0.552
0.003
0.488
0.007
0.005 −0.384
0.032 −0.428
0.402
0.089
0.432
0.053
0.417
0.048
0.363
0.089

HotelRec

0.565 −0.508
0.027 −0.337
0.318
0.053
0.301
0.181
0.225
0.156
0.306
0.148

0.004
0.013
0.058
0.033
0.047
0.031
0.038
0.074

0.009
0.100
0.121
0.144
0.279
0.137

0.454
0.459
0.504
0.448
0.438
0.402
0.400
0.359

0.356
0.196
0.419
0.349
0.323
0.294

0.023
0.021
0.010
0.025
0.028
0.047
0.048
0.078

0.081
0.347
0.037
0.087
0.115
0.154

0.402
0.438
0.523
0.453
0.394
0.433
0.402
0.364

0.285
0.382
0.355
0.315
0.364
0.312

0.046
0.029
0.007
0.023
0.051
0.031
0.047
0.074

0.168
0.059
0.082
0.126
0.074
0.130

Tableau 2: Brand ranking results generated by various models trained on time slice t and tested on time
slice t + 1. We report the correlation coefficients corr and its associated two-sided p-values.

• O-dBTM, a variant of our model that is only
trained with the supervised review-level sen-
timent labels in the first time slice (denoted
as the 0-th time slice). In the subsequent time
slices, it is trained under the unsupervised
setting. In such a case, we no longer have
a gold-standard brand ranking in time slices
other than the 0-th one. Instead of directly
calculating the Spearman’s rank correlation
coefficient, we measure the difference of
the brand ranking results in neighboring time
slices and use it to set the weight γt in Eq. (8).

Parameter Setting Frequent bigrams and tri-
grams4 are added as features in addition to
unigrams for document representations. In our
experiments, we train the models using the data
from the current time slice and test the model per-
formance on the full data from the next time slice.
During training, we set aside 10% of data in each
time slice as the validation set. For hyperparam-
eters, we set the batch size to 256, the maximum
training steps to 50,000, the topic number to 50.5
It is worth noting that since topic dynamics are

4Frequent but less informative n-grams such as ‘actually

bought’ were filtered out using NLTK.

5The topic number is set empirically based on the

validation set in the 0-th time slice.

not explicitly modeled in the static models such as
TBIP and BTM, their topics extracted in different
time slices are not directly linked.

6 Experimental Results

Dans cette section, we present the experimental results
in comparison with the baseline models in brand
rating, topic coherence/uniqueness measures, et
qualitative evaluation of generated topics. For fair
comparison, baselines are trained based on all
previous time slices and predict on the current
time slice.

6.1 Brand Rating

TBIP, BTM, and dBTM can infer each brand’s
associated polarity score automatically. For dJST,
we derive the brand rating by aggregating the
label distribution of its associated review docu-
ments and then normalizing over the total number
of brand-related reviews. The average of the
document-level ratings of a brand b at a time
slice t is used as the ground truth of the brand
rating xt
b. We evaluate two aspects of the brand
ratings:

Brand Ranking Results We report in Table 2
the brand ranking results measured by the Spear-
man’s correlation coefficient, showing the cor-
relation of predicted brand rating and the ground

411

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 3: The rating time series for ‘Maybeline New York’. The rating scores are normalized in the range of [−1, 1]
with positive values denoting positive sentiment and negative ones for negative sentiment. In each subfigure, le
dashed curve shows the actual rating scores.

truth, along with the associated two-sided p-values
of the Spearman’s correlations

Topic model variants, such as dJST, TBIP, et
BTM, produced brand ranking results either posi-
tively or negatively correlated with the true rank-
ing results. We can see the correlation of BTM
has switched between positive correlated and neg-
ative rated between time slices. With Gaussian
state space models, our proposed model dBTM
and its variant O-dBTM generate more stable re-
sults. On MakeupAlley-Beauty, dBTM gives the
best results in 4 out of 8 time slices. Intérêt-
franchement, O-dBTM with the supervised information
supplied in only the first time slice outperforms
the static models such as BTM in 3 out of 8 temps
slices, showing the effectiveness of our proposed
architecture in tracking brand score dynamics.
Similar conclusions can be drawn on HotelRec that
O-dBTM gives superior performance compared
to BTM on 5 out of 6 time slices. Both O-dBTM
and dBTM outperform the other baselines except
TBIP in time slice 2.

En résumé, in dBTM, the brand rating score
is treated as a latent variable (c'est à dire., xbd in Eq. (2))
and is directly inferred from the data. On the
contrary, models such as dJST, which require
post-processing to derive brand rating scores by
aggregating the document-level sentiment labels,
are inferior to dBTM. This shows the advantage
of our proposed dBTM over traditional dynamic
topic models in brand ranking.

Brand Rating Time Series The brand rating
time series aims to compare the ability of models
to track the trend of brand rating. For easy com-
parison, we normalize the ratings produced by
each model, so that the plot only reflects the fluc-
tuation of ratings over time. Chiffre 3 shows the
brand rating on the brand ‘Maybeline New York’
generated on the test set of MakeupAlley-Beauty
by various models across time slices. It can be
observed that the brand ratings generated by TBIP
and BTM do not correlate well with the actual
rating scores. dJST shows a better aligned rating

412

Time Slice

dJST

TBIP

BTM

O-dBTM

dBTM

coh

uni

qualité

coh

uni

qualité

coh

uni

qualité

coh

uni

qualité

coh

uni

qualité

MakeupAlley-Beauty

1
2
3
4
5
6
7
8

−3.087 0.564
−3.008 0.513
−3.286 0.552
−3.004 0.515
−3.112 0.560
−3.139 0.542
−3.269 0.521
−3.060 0.560

0.183 −3.653 0.861
0.170 −4.043 0.850
0.168 −3.949 0.843
0.172 −3.629 0.808
0.180 −4.168 0.838
0.173 −4.100 0.841
0.159 −4.049 0.854
0.183 −3.942 0.843

0.236 −3.836 0.862
0.210 −3.867 0.864
0.214 −3.716 0.851
0.223 −3.837 0.846
0.201 −4.023 0.839
0.205 −3.976 0.846
0.211 −3.675 0.845
0.214 −3.715 0.837

0.225 −3.486 0.820
0.223 −3.360 0.807
0.229 −3.369 0.787
0.220 −3.457 0.771
0.208 −3.412 0.793
0.213 −3.433 0.761
0.230 −3.330 0.772
0.225 −3.589 0.789

0.235 −3.685 0.833
0.240 −3.642 0.829
0.234 −3.611 0.823
0.223 −3.549 0.799
0.232 −3.523 0.818
0.222 −3.577 0.814
0.232 −3.667 0.825
0.220 −3.546 0.818

0.226
0.228
0.228
0.225
0.232
0.228
0.225
0.231

Average

−3.120 0.541

0.173 −3.942 0.842

0.214 −3.831 0.849

0.222 −3.430 0.788

0.230 −3.600 0.820

0.228

HotelRec

1
2
3
4
5
6

−3.749 0.615
−4.020 0.633
−3.667 0.593
−4.008 0.644
−3.751 0.691
−3.916 0.697

0.164 −4.024 0.767
0.158 −3.577 0.753
0.162 −3.905 0.817
0.161 −3.747 0.808
0.184 −4.057 0.800
0.178 −3.770 0.810

0.191 −3.935 0.851
0.211 −3.960 0.813
0.209 −4.078 0.844
0.216 −3.946 0.859
0.197 −3.953 0.823
0.215 −4.061 0.855

0.216 −4.051 0.812
0.205 −3.851 0.803
0.207 −3.861 0.819
0.218 −3.637 0.814
0.208 −3.705 0.804
0.210 −3.510 0.800

0.201 −3.716 0.818
0.209 −3.696 0.809
0.212 −3.854 0.820
0.224 −3.681 0.794
0.217 −3.547 0.817
0.228 −3.705 0.821

0.220
0.219
0.213
0.216
0.230
0.222

Average

−3.852 0.645

0.168 −3.847 0.793

0.206 −3.989 0.841

0.211 −3.769 0.809

0.215 −3.700 0.813

0.220

Tableau 3: Topic coherence (coh) and uniqueness (uni) measures of the results generated by various
models. We also combine the two scores to derive the overall quality of the extracted topics.

trend, but its prediction missed some short-term
changes such as the peak of brand rating at time
slice 7. Par contre, dBTM correctly predicts the
general trend of the brand rating. The weakly-
supervised O-dBTM is able to follow the general
trend but misses some short-term changes such as
the upward trend from the time slice 1 à 2, et
from the slice 6 à 7.

6.2 Topic Evaluation Results

We use the top 10 words of each topic to calculate
the context-vector-based topic coherence scores
(R¨oder et al., 2015) as well as topic uniqueness
(Nan et al., 2019) which measures the ratio of
word overlap across topics. We want to achieve
balanced topic coherence and diversity. En tant que tel,
topic coherence and topic diversity are combined
to give an overall quality measure of topics (Dieng
et coll., 2020). Since the results for topic coherence
is negative in our experiment, c'est, smaller
absolute values are better, we define the overall
quality of a topic as q = topic uniqueness
|topic coherence| . Tableau 3
shows the topic evaluation results. En général,
there is a trade-off between topic coherence and
topic diversity. En moyenne, dJST has the high-
est coherence but the lowest uniqueness scores,
while TBIP has quite high uniqueness but the low-
est coherence values. Both O-dBTM and dBTM
achieve a good balance between coherence and

uniqueness and outperform other models in over-
all quality.

6.3 Example Topics across Time Periods

We illustrate some representative topics gener-
ated by dBTM in various time slices. For easy
inspection, we retrieve a representative sentence
from the corpus for each topic. For a sentence, nous
derive its representation by averaging the GloVe
embeddings of its constituent words. For a topic,
we also average the GloVe embeddings of its asso-
ciated top words, but weighted by the topic-word
probabilities. The sentence with the highest cosine
similarity is selected.

Example of generated topics relating to ‘Eye
Products’ and ‘Skin Care’ from MakeupAlley-
Beauty is shown in Figure 4. We can observe
that for the topic ‘Eye Products’, the top words
of negative comments for ‘eye cleanser’ evolve
from the reaction of skin (par exemple., ‘sting’, ‘burned’) à
the cleaning ability (par exemple., ‘remove’, ‘residue’). Nous
could also see that the positive topics gradually
change from praising the ability of the product
for ‘dark circle’ in time slice 1 to the quality of
eye shadow in time slice 4 and eye primer in time
slice 8. De plus, we observe the brand name
M.A.C. in the positive topic in time slice 4, lequel
aligns with its ground truth rating. For the topic
‘Skin Care’, it can be observed that negative top-
ics gradually move from the complaint of a skin

413

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 4: Example of generated topics shown as a list of top associated words (underlined) in different time slices
from the MakeupAlley dataset. For easy inspection, we also show the most representative sentence under each
topic. The negative, neutral and positive topics in each time slice are generated by varying the brand polarity score
from −1 to 0, and to 1. Positive words/phrases are highlighted in blue, negative words/phrases are in red, alors que
brand names are in bold.

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Chiffre 5: Example of generated topics shown as a list of top associated words (underlined) in different time slices
from the HotelRec dataset. The representative sentence for each topic is also shown for easy inspection.

cleanser to the thickness of a sunscreen, alors que
positive topics are about the praise of the cover-
age of the M.A.C. foundation more consistently
au fil du temps. The results show that dBTM can gen-
erate well-separated polarity-bearing topics and
it also allows the tracking of topic changes over
temps.

Example of generated topics relating to ‘Room
Condition’ and ‘Food’ from HotelRec is shown
in Figure 5. We can see that for the topic ‘Room
Condition’, top words gradually shift from the
expression of cleanliness (par exemple., ‘clean’ in positive
and ‘dirty’ in negative comments) to the descrip-
tion of the type and size of the rooms (par exemple., ‘ex-

ecutive’ and ‘villa’ in positive reviews, et le
concern of ‘small’ room size in negative com-
ments). For the topic ‘Food’, the concerned food
changes across time from drinks (par exemple., ‘coffee’,
‘tea’) to meals (par exemple., ‘eggs’, ‘toast’). Negative
reviews mainly focus on the concern of food qual-
ville, (par exemple., ‘cold’), while positive reviews contain
a general praise of food and services (par exemple., ‘like’,
‘nice’).

6.4 Ablation Study

We investigate the contribution of the meta learn-
ing component (c'est à dire., Eq. (8) et (9)) by conduct-
ing an ablation study and the results are shown in

414

Time Slice

dBTM

dBTM (no meta learniing)

cor

coh

uni

qualité

cor

coh

uni

qualité

MakeupAlley-Beauty

0.402 −3.685
0.438 −3.642
0.523 −3.611
0.453 −3.549
0.394 −3.523
0.433 −3.577
0.402 −3.667
0.364 −3.546

0.285 −3.716
0.382 −3.696
0.355 −3.854
0.315 −3.681
0.364 −3.547
0.312 −3.705

0.833
0.829
0.823
0.799
0.818
0.814
0.825
0.818

0.818
0.809
0.820
0.794
0.817
0.821

0.435 −3.972
0.226
0.189 −3.704
0.228
0.228 −0.162 −3.873
0.225 −0.042 −3.745
0.086 −3.990
0.232
0.228 −0.029 −3.958
0.225 −0.042 −3.587
0.231 −0.125 −3.920

HotelRec

0.220
0.219
0.213
0.216
0.230
0.222

0.222 −3.559
0.210 −3.796
0.285 −3.763
0.408 −3.597
0.362 −3.657
0.262 −3.694

0.861
0.840
0.828
0.849
0.832
0.856
0.842
0.847

0.791
0.801
0.790
0.817
0.793
0.809

0.217
0.227
0.214
0.227
0.209
0.216
0.235
0.216

0.222
0.211
0.210
0.227
0.217
0.219

1
2
3
4
5
6
7
8

1
2
3
4
5
6

Tableau 4: Results of dBTM with and without the
meta learning component.

Tableau 4. We can observe that in general, remov-
ing meta learning leads to a significant reduc-
tion in brand ranking correlations across all time
slices for the MakeupAlley-Beauty dataset. Dans
terms of topic quality, we observe reduced co-
herence scores, but slightly increased uniqueness
scores without meta learning, leading to an over-
all reduction of topic quality scores in most time
slices.

For HotelRec, we can see that removing meta
learning also leads to a reduction in brand rank-
ing results, but the impact is smaller compared to
MakeupAlley-Beauty. For topic quality, we ob-
serve increased coherence but worse uniqueness,
resulting in slightly worse topic quality results
without meta learning in most time slices. Un
main reason is that unlike makeup brands where
new products are introduced over time, leading
to the change of discussed topics in reviews, le
topic-word distribution does not change much
across different
time slices for hotel reviews.
Donc, the results are less impacted with or
without meta learning.

6.5 Training Time Complexity

All experiments were run on a single GeForce
1080 GPU with 11GB memory. The training time
for each model across time slices is shown in
Chiffre 6. It can be observed that with the increas-
ing number of time slices, the training time of
dJST and BTM grows quickly. Both TBIP and
dBTM take significantly less time to train. TBIP
simply performs Poisson factorization indepen-
dently in each time slice and fails to track topic/
sentiment changes over time. On the contrary,
our proposed dBTM and O-dBTM are able to

Chiffre 6: Training time of models across time slices.

monitor topic/sentiment evolvement and yet take
even less time to train compared to TBIP. Un
main reason is that dBTM and O-dBTM can au-
tomatically adjust the number of iterations with
our proposed meta learning and hence can be
trained more efficiently.

7 Conclusion

We have presented dBTM, which is able to
automatically detect and track brand-associated
topics and sentiment scores. Experimental evalua-
tion based on the reviews from MakeupAlley and
HotelRec demonstrates the superiority of dBTM
over previous models in brand ranking and dy-
namic topic extraction. The variant of dBTM,
O-dBTM, trained with document-level sentiment
labels in the first time slice only, outperforms
baselines in brand ranking and achieves the best
overall result in topic quality evaluation. Ce
shows the effectiveness of the proposed architec-
ture in modeling the evolution of brand scores and
topics across time intervals.

Our model currently only considers review
ratings, but real-world applications potentially in-
volve additional factors (par exemple., user preference). UN
possible solution is to explore simultaneous mod-
eling of user preferences to extract personalised
brand polarity topics.

Remerciements

This work was supported in part by the UK
Engineering and Physical Sciences Research

415

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Council (grant no. EP/T017112/1, EP/V048597/1,
EP/X019063/1). YH is supported by a Turing
AI Fellowship funded by the UK Research and
Innovation (grant no. EP/V020579/1).

Les références

Diego Antognini and Boi Faltings. 2020. Hotel-
rec: A novel very large-scale hotel recom-
mendation dataset. In Proceedings of the 12th
Language Resources and Evaluation Confer-
ence, pages 4917–4923.

David M. Blei and John D. Lafferty. 2006.
Dynamic topic models. In Proceedings of the
23rd International Conference on Machine
Apprentissage, pages 113–120.

Jordan. 2003. Latent dirichlet

David M. Blei, Andrew Y. Ng, and Michael
je.
alloca-
tion. Journal of Machine Learning Research,
3:993–1022.

John Blitzer, Mark Dredze, and Fernando Pereira.
2007. Biographies, bollywood, boom-boxes
and blenders: Domain adaptation for sentiment
classification. In Proceedings of the 45th An-
nual Meeting of the Association of Computa-
tional Linguistics, pages 440–447.

Samuel Brody and No´emie Elhadad. 2010. Un
unsupervised aspect-sentiment model for online
reviews. In Proceedings of the 2010 Annual
Conference of the North American Chapter of
the Association for Computational Linguistics,
pages 804–812.

Laurent Charlin, Rajesh Ranganath,

James
McInerney, and David M. Blei. 2015. Dynamic
poisson factorization. In Proceedings of the
9th ACM Conference on Recommender Sys-
thèmes, pages 155–162.

Peng Cui and Le Hu. 2021. Topic-guided
abstractive multi-document summarization. Dans
Findings of the Association for Computational
Linguistics: EMNLP 2021, pages 1463–1472.
https://doi.org/10.18653/v1/2021
.findings-emnlp.126

ACM International Conference on Web Search
and Data Mining, pages 151–159.

Adji B. Dieng, Francisco J. R.. Ruiz, and David
M.. Blei. 2020. Topic modeling in embedding
In Proceedings of Transactions of
les espaces.
the Association for Computational Linguistics,
pages 439–453. https://est ce que je.org/10.1162
/tacl a 00325

Adji B. Dieng, Francisco J. R.. Ruiz, and David
M.. Blei. 2019. The dynamic embedded topic
model. arXiv preprint arXiv:1907.05545.

Manqing Dong, Feng Yuan, Lina Yao, Xiwei
Xu, and Liming Zhu. 2020. MAMO: Mémoire-
augmented meta-optimization for cold-start
recommendation. In Proceedings of the 26th
ACM SIGKDD Conference on Knowledge
Discovery and Data Mining, pages 688–697.
https://doi.org/10.1145/3394486
.3403113

Chelsea Finn, Pieter Abbeel, and Sergey Levine.
2017. Model-agnostic meta-learning for fast
adaptation of deep networks. In Proceedings
of the 34th International Conference on Ma-
chine Learning, pages 1126–1135.

Li Gao, Jia Wu, Chuan Zhou, and Yue Hu.
2017. Collaborative dynamic sparse topic re-
gression with user profile evolution for item
recommendation. In Proceedings of the 31st
AAAI Conference on Artificial Intelligence,
pages 1316–1322. https://est ce que je.org/10
.1609/aaai.v31i1.10726

Ruiying Geng, Binhua Li, Yongbin Li, Jian Sun,
and Xiaodan Zhu. 2020. Dynamic memory
induction networks for few-shot text classifica-
tion. In Proceedings of the 58th Annual Meeting
of the Association for Computational Linguis-
tics, pages 1087–1094. https://doi.org
/10.18653/v1/2020.acl-main.102

Chengyue Gong and Win-bin Huang. 2017. Deep
dynamic poisson factorization model. En Pro-
ceedings of the 31st International Conference
on Neural Information Processing Systems,
pages 1665–1673.

Shumin Deng, Ningyu Zhang, Jiaojian Kang,
Yichi Zhang, Wei Zhang, and Huajun Chen.
2020. Meta-learning with dynamic-memory-
based prototypical network for few-shot event
the Thirteenth
detection. In Proceedings of

Prem Gopalan, Jake Hofman, and David Blei.
2015. Scalable recommendation with hierar-
chical poisson factorization. In Proceedings of
the 31th Uncertainty in Artificial Intelligence,
pages 326–335.

416

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Yulan He, Chenghua Lin, Wei Gao, and Kam-Fai
Wong. 2014. Dynamic joint sentiment-topic
model. ACM Transactions on Intelligent Sys-
tems and Technology, 5:1–21. https://est ce que je
.org/10.1145/2594452

Seyed Abbas Hosseini, Ali Khodadadi, Keivan
Alizadeh, Ali Arabzadeh, Mehrdad Farajtabar,
Hongyuan Zha, and Hamid R. Rabiee. 2018.
Recurrent poisson factorization for temporal
recommendation. IEEE Transactions on Knowl-
edge and Data Engineering, 32(1):121–134.
https://doi.org/10.1109/TKDE.2018
.2879796

Muhammad Abdullah Jamal and Guo-Jun Qi.
2019. Task agnostic meta-learning for few-shot
learning. In Processing of the IEEE Confer-
ence on Computer Vision and Pattern Recog-
nition, pages 11719–11727.

Eric Jang, Shixiang Gu, and Ben Poole. 2017.
Categorical reparameterization with gumbel-
softmax. In International Conference on Learn-
ing Representations.

Chenghua Lin and Yulan He. 2009.

Joint
sentiment/topic model for sentiment analy-
sis. In Proceedings of the ACM Conference
on Information and Knowledge Management,
pages 375–384.

Chenghua Lin, Yulan He, Richard Everson,
and Stefan Ruger. 2012. Weakly supervised
joint sentiment-topic detection from text. IEEE
Transactions on Knowledge and Data En-
gineering, 24(6):1134–1145. https://est ce que je
.org/10.1109/TKDE.2011.48

Yuanfu Lu, Yuan Fang, and Chuan Shi. 2020.
Meta-learning on heterogeneous information
networks for cold-start recommendation. Dans
Proceedings of the 26th ACM SIGKDD Con-
ference on Knowledge Discovery and Data
Mining, pages 1563–1573.

Feng Nan, Ran Ding, Ramesh Nallapati, et
Bing Xiang. 2019. Topic modeling with wasser-
le
stein autoencoders.
57th Annual Meeting of the Association for
Computational Linguistics, pages 6345–6381.

In Proceedings of

Krishna Prasad Neupane, Ervine Zheng, and Qi
learn-
Yu. 2021. Metaedl: Meta evidential
ing for uncertainty-aware cold-start recom-
mendations. In Proceedings of the 21th IEEE

International Conference on Data Mining,
pages 1258–1263. https://est ce que je.org/10
.1109/ICDM51629.2021.00154

Justifying

Jianmo Ni, Jiacheng Li, and Julian McAuley.
en utilisant
2019.
distantly-labeled reviews and fined-grained as-
pects. In Proceedings of Empirical Methods in
Natural Language Processing, pages 188–197.

recommendations

Michael R¨oder, Andreas Both, and Alexander
Hinneburg. 2015. Exploring the space of topic
coherence measures. In Proceedings of the 8th
ACM International Conference on Web Search
and Data Mining, pages 399–408.

Yuanfeng Song, Yongxin Tong, Siqi Bao, Di
Jiang, Hua Wu, and Raymond Chi-Wing
Wong. 2020. Topicocean: An ever-increasing
topic model with meta-learning. In Proceedings
of the 20th IEEE International Conference on
Data Mining, pages 1262–1267. https://est ce que je
.org/10.1109/ICDM50108.2020.00161

Wu-Jiu Sun, Xiao Fan Liu, and Fei Shen. 2021.
Learning dynamic user interactions for online
forum commenting prediction. In Proceedings
of the 21th IEEE International Conference on
Data Mining, pages 1342–1347.

Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang,
Philip H. S. Torr, and Timothy M. Hospedales.
2018. Learning to compare: Relation network
for few-shot learning. In Proceedings of the
IEEE Conference on Computer Vision and Pat-
tern Recognition, pages 1199–1208. https://
doi.org/10.1109/CVPR.2018.00131

Qiuling Suo, Jingyuan Chou, Weida Zhong, et
Aidong Zhang. 2020. Tadanet: Task-adaptive
network for graph-enriched meta-learning. Dans
Proceedings of the 26th ACM SIGKDD Con-
ference on Knowledge Discovery and Data
Mining, pages 1789–1799.

the Conference of

Keyon Vafa, Suresh Naidu, and David M. Blei.
2020. Text-based ideal points. In Proceedings
de
the Association for
Computational Linguistics, pages 5345–5357.
https://doi.org/10.18653/v1/2020
.acl-main.475

Chong Wang, David Blei, and David Heckerman.
2008. Continuous time dynamic topic mod-
le. In Proceedings of the 24th Conference

417

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi

/
t

un
c
je
/

un
r
t
je
c
e
–
p
d

F
/

d
o

je
/

1
0
1
1
6
2

/
t

un
c
_
un
_
0
0
5
5
5
2
0
8
6
3
1
7

/
t

un
c
_
un
_
0
0
5
5
5
p
d

b
oui
g
toi
e
s
t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

on Uncertainty
pages 579–586.

in Artificial

Intelligence,

Tianyi Yang, Cuiyun Gao, Jingya Zang, David
Lo, and Michael R. Lyu. 2021. Tour: Dynamic
topic and sentiment analysis of user reviews
for assisting app release. In Proceedings of
the Web Conference 2021, pages 708–712.
https://doi.org/10.1145/3442442
.3458612

Runsheng Yu, Yu Gong, Xu He, Yu Zhu, Qingwen
Liu, Wenwu Ou, and Bo An. 2021. Person-
alized adaptive meta learning for cold-start
user preference prediction. In Proceedings of
the 35th AAAI Conference on Artificial Intel-
ligence, pages 10772–10780. https://est ce que je
.org/10.1609/aaai.v35i12.17287

Hao Zhang, Gunhee Kim, and Eric P. Xing.
2015. Dynamic topic modeling for monitoring
market competition from online text and image
data. In Proceedings of the 21th International
Conference on Knowledge Discovery and Data
Mining, pages 1425–1434. https://est ce que je
.org/10.1145/2783258.2783293

Runcong Zhao, Lin Gui, Gabriele Pergola, et
Yulan He. 2021. Adversarial learning of pois-
son factorisation model for gauging brand sen-
timent in user reviews. In Proceedings of the
16th Conference of the European Chapter of
the Association for Computational Linguistics,
pages 2341–2351.

Wenbo Zheng, Lan Yan, Chao Gou, et
Fei-Yue Wang. 2021. Knowledge is power:
Hierarchical-knowledge
embedded meta-
learning for visual reasoning in artistic do-
the 27th ACM
mains.
SIGKDD Conference on Knowledge Dis-
covery and Data Mining, pages 2360–2368.
https://doi.org/10.1145/3447548
.3467285

In Proceedings of

Kun Zhou, Yuanhang Zhou, Wayne Xin
Zhao, Xiaoke Wang, and Ji-Rong Wen. 2020.
Towards topic-guided conversational recom-
mender system. In Proceedings of the 28th
International Conference on Computational
Linguistics, pages 4128–4139. https://est ce que je
.org/10.18653/v1/2020.coling-main
.365