Computational Intelligence - August 2016 - 37

with tailored loss functions to automatically derive sentiment
embeddings from massive texts. These sentiment embeddings
can be used as word features for sentiment analysis without ad
hoc feature engineering. Chen et al. [49] combined a WordNet
[50] glosses composition model and a context clustering model
to learn word sense embeddings which can be used in sentiment analysis tasks.
Socher et al. [17] proposed a recursive neural tensor network (RNTN) for semantic compositionality over a sentiment
treebank which pushed the binary classification accuracy on
the Stanford sentiment tree bank from 80% up to 85.4%.
-Kalchbrenner et al. [18] proposed a dynamic convolutional
neural network (DCNN) to handle input sentences with varying length and induced a feature graph over a sentence that is
capable of explicitly capturing short and long-range relations.
Kim [19] presented two simple CNN models with little hyperparameter tuning which were trained on pre-trained word vectors for sentence-level classification tasks. Tang et al. [51]
introduced a gated recursive neural network to learn continuous document representations for sentiment classification.
2.2 User and Product Modeling for Sentiment Classification

In recent years, there has been growing interests in incorporating the user and product information for sentiment analysis of
product reviews. Seroussi et al. [13] presented a nearest-neighbor collaborative approach for training user-specific classifiers
whose outputs were subsequently combined with user similarity measurement for sentiment inference from text. Li et al. [12]
used the user, product and review features as a three-dimension
tensor, and employed tensor factorization techniques to alleviate the data sparsity problem. Gao et al. [52] referred to user- or
product-specific sentiment polarity biases as user leniency and
product popularity, respectively. They built a model that automatically computed user leniency and product popularity for
sentiment classification. Diao et al. [14] proposed a probabilistic
model based on collaborative filtering and topic modeling to
capture user and product features for sentiment classification.
Li et al. [53] incorporated textual topic and user-word factors
with supervised topic modeling. Zhang et al. [54] formalized
the phrase level sentiment polarity labeling problem in a convex optimization framework, and designed iterative updating
algorithms for leveraging review-level sentiment classification
techniques to boost the performance of phase-level sentiment
polarity labeling. Tang et al. [11], [20] incorporated user, product and review information into a purposely-built neural network model to learn distributed representations of users and
products for document-level sentiment classification.
2.3 Sequence Modeling for Sentiment Analysis

Although sentiment classification has been mostly approached as
a binary or multi-class classification problem, some studies considered sentiment analysis as modeling of sentiment flow
throughout a document, using sequence modeling approaches
for sentiment detection. Mao and Lebanon [55] developed a
variant of CRF for sentiment flow prediction. Liu and Zhou

[56] decomposed a sentence into a series of sub-sequences using
a hidden CRF, and determined the sentence-level polarity by
classifying within sub-sequences and by fusing the obtained subsequence polarities.
Sequence modeling has often been used for fine-grained
sentiment analysis. It aims to detect the subjective expressions
in a text and to characterize their intensity and sentiment as
well as to identify the opinion holder and the target, or topic,
of the opinion [57]. Johansson and Moschitti [58] demonstrated
relational features derived from dependency-syntactic and
semantic role structures are useful for sentiment analysis. Yang
and Cardie [59] proposed a semi-CRF-based approach relaxing
the Markovian assumption inherent to CRFs and operated at
the phrase level rather than the token level, allowing the incorporation of phrase-level features. Irsoy and Cardie [60] applied
deep RNNs to the task of opinion expression extraction formulated as a token-level sequence-labeling task.
In this paper, we proposed to learn user and product
embeddings from the temporal ordered reviews written by a
user or evaluated on a product respectively using a variant of
RNN as a sequence learning model combining user, product
and review content information.
3. Methodology

In this section, we present our approach for learning user and
product distributed representations using a sequence model for
sentiment analysis. An overview of the approach is shown in
Fig. 1. We first describe the document composition model
which produces the distributed representation of each review
document (Section 3.1). Afterwards, we introduce the sequence
model for embedding temporal relations of reviews into user
and product representations learning (Section 3.2). Finally, we
describe the sentiment classification model which encodes user,
product and review information (Section 3.3).
3.1 Modeling Reviews with Multi Filter 1d-CNN

We describe our approach for learning review embeddings
from review content and its rating with a document composition model (multi filter 1d-CNN).
3.1.1 Training Objective
For reviews with respect to 1 - K rating scales (sentiment
strength scores), the training objective of the document composition model is to minimize the ranking loss below:

/ max {0, 1 - g (d) + g (dl )} (1)

d!T

where d is a review document in the training set T with a
certain rating from 1 to K (positive sample); d l is another
review in T with a rating different from d (negative sample);
g ($) is a scoring function which represents the whole neural
network architecture without the last classification layer;
g (d) and g (d l ) are the score of positive and negative sample,
respectively. For each review in the training set, we expect
g (d) to be approximating 1, g (d l ) to be approximating 0,

AUGUST 2016 | IEEE Computational intelligence magazine

Table of Contents for the Digital Edition of Computational Intelligence - August 2016