Complex Sentiment Analysis using Recursive Autoencoders

Shashank Sashihithlu, Siddhi S. Soman

Stanford University
{sashihit, ssoman}@ stanford.edu
Advised by Richard Socher

Abstract
We use a machine learning framework based on
recursive autoencoders to make sentence level
predictions regarding the subject, object and
sentiment of the sentence. Our model learns
vector representations of multi-word phrases.
The syntactic structure of the sentence is
incorporated into our model by using the
Stanford Parser tree thereby enabling it to make
sentiment predictions with respect to subject,
object, etc.
1

Introduction

Deep learning involves learning a hierarchy

of internal representations using a stack of
multiple modules, each of which is trainable and
can implement complex non-linear functions.
Deep learning allows us to go from low level
representations to high level representations and
is more efficient and accurate than shallow
architectures such as kernel machines. A deep
architecture achieves this efficiency by trading
space for time and using sparse feature
encoding. We build a hierarchy of modules,
where stage k measures the compatibility
between features at level k-1 and level k, where
the compatibility may be measured as the
negative of log likelihood or energy. The
objective at each stage is to find a set of features
that would minimize the energy. The input to
stage k+1 is the feature vector of stage k. Deep
belief networks have been shown to perform
significantly better than other techniques for
problems such as sentiment analysis, image
classification and handwriting analysis.

Sentiment analysis (together with opinion

mining) is becoming a promising topic and has
important applications in the fields of social
media, customer relationship management etc.
Our project aims at performing complex
sentiment analysis using deep learning
techniques. Previous work in [2] looked at the
problem of extracting sentiments using
recursive autoencoders. We tried to extend this
work by performing more complex sentiment
analysis which includes extracting the subject,
object and the sentiment. We plan to use data
comprising of 'opinionated text'. The recursive
neural network architecture described in [3]
jointly learns how to parse and how to represent
phrases in a continuous vector space of features.
This allows us to capture syntactic and
compositional semantic information from
natural language sentences. We use this
architecture combined with the auto encoder
scheme to perform complex sentiment analysis.
2

Dataset

We use the MPQA dataset which consists of

opinionated text. The MPQA Opinion Corpus
contains news articles from a wide variety of
news sources manually annotated for opinions
and other private states (i.e., beliefs, emotions,
sentiments, speculations, etc.). The corpus
contains 535 documents, a total of 11,114
sentences. The articles in the corpus are from
187 different foreign and U.S. news sources
which date from June 2001 to May 2002.
The dataset contains annotations of different
types like express subjectivity, direct-subjective,
objective speech event etc. These also have

nested annotations like polarity (attribute for

marking the polarity of the expression, in
context, according to the nested-source),
intensity etc. We try to use the annotations
within this dataset to extract the different parts
of the sentence like subject, object along with
the sentiment.
3

Essentially these same steps are repeated for all

triplets and the reconstruction error at nodes of
the tree is found.

Recursive Autoencoders

We represent words as continuous vectors of

parameters. We represent a sentence (or any ngram) as an ordered list of these vectors
. Now, assume that we are given
a list of word vectors
as
well as a binary tree structure for this input in
the form of branching triplets of parents with
children:
. Each child can be an
input word vector or a non-terminal node in the
tree. For the example in Fig.1, we have the
following
triplets:
In order to be able to
apply the same neural network to each pair of
children, the hidden representations
have to
have the same dimensionality as the s. Given
this tree structure, we can now compute the
parent representations. The first parent vector
is computed from the children (
:

where we multiplied a matrix of parameters

by the concatenation of the two
children. One way of assessing how well this ndimensional vector represents its children is to
try to reconstruct the children in a
reconstruction layer:

During training, the goal is to minimize the

reconstruction errors of this input pair. For each
pair, we compute the Euclidean distance
between the original input and its
reconstruction:

Fig 1: RAE on a binary tree

So far, the RAE was completely unsupervised
and induced general representations that capture
the semantics of multi-word phrases. We extend
RAEs to a semi-supervised setting in order to
predict a sentence or phrase level target
distribution t. This is done by adding on top of
each parent node a simple softmax layer to
predict class distributions:

Assuming there are K labels,

is a Kdimensional multinomial distribution and
. Let
be the kth element of the
multinomial target label distribution t for one
entry. The softmax layers outputs are
interpreted as conditional probabilities
, hence the cross-entropy error is :

Using this cross-entropy error for the label as

well as the reconstruction error, we get the total
error for sentence, label pairs
in the corpus
as:

where E is summed over all nodes for a

particular sentence and is a weighted average of
the reconstruction error and the cross-entropy
error. This error function is minimized with
respect to all the parameters by using L-BFGS.
4

Model

We use recursive autoencoders to model our

data and for making predictions. The previous
work included a greedy approach for structure
prediction using unsupervised RAE. The
resulting tree structure captures as much of the
single-word information as possible (in order to
allow reconstructing the word vectors) but does
not necessarily follow standard syntactic
constraints. Hence though this approach might
be faster, it does not necessarily capture the
entire syntactic structure.
Owing to this we chose not to use the greedy
approach of tree construction. Instead we use
the Stanford parser tree output as our tree
representation. This representation captures the
syntactic structure of the sentence. This helps us
correlate the syntactic and semantic information
of the sentence and helps us make better
predictions.
e.g.
Sentence:
The Kimberley Provincial Hospital said it would
probably know by Tuesday whether one of its
patients had Congo Fever.
Parsed output:
(ROOT
(S
(NP (DT The) (NNP Kimberley) (NNP
Provincial) (NNP Hospital))
(VP (VBD said)
(SBAR
(S
(NP (PRP it))
(VP (MD would)

(ADVP (RB probably))

(VP (VB know)
(PP (IN by)
(NP (NNP Tuesday)))
(SBAR (IN whether)
(S
(NP
(NP (CD one))
(PP (IN of)
(NP (PRP its) (NNS patients))))
(VP (VBD had)
(NP (NNP Congo) (NN
Fever))))))))))
(. .)))
In our model, words are represented as a
continuous vector of parameters. Our model
learns a vector representation of phrases and
sentences using RAE as shown in the figure.
Our model is also extended to learn a sentence
or phrase level target distribution in a semisupervised setting using an additional softmax
regression layer.
5

Methodology

We used the annotations and information

provided by the MPQA dataset for training our
model. The first step was extracting the
sentences and corresponding annotations from
the MPQA dataset. The data was then preprocessed to the format required by our model.
5.1. Data pre-processing:
The MPQA dataset required substantial preprocessing as we had to extract out the labels we
needed for all the words, phrases and sentences
within the dataset. This helped us in designing
our model to predict sentiments with respect to
the various parts of the sentence like subject and
object.
e.g.
Sentence: The Kimberley Provincial Hospital
said it would probably know by Tuesday
whether one of its patients had Congo Fever.

Pre-processed data:

However since this does not incorporate the

syntactic structure of the sentence in making
predictions we choose not to utilize this
approach. Instead we parse each sentence and
using the Stanford Parser convert each sentence
into its parsed form. We then convert this parsed
tree to a binary form to determine the
hierarchical structure of the sentence. We then
use this in our recursive autoencoder model for
making predictions.

Element Label Word

Basic word

subj

The

the

subj

Kimberley

kimberley

subj

Provincial

provincial

subj

Hospital

hospital

...

said

say

subj

it

it

obj

would

would

obj

probably

probably

obj

know

know

...

by

by

...

Tuesday

tuesday

obj

whether

whether

subj

one

one

...

of

of

...

its

its

...

patients

patient

5.4. Training and Results:

...

had

have

obj

Congo

congo

Let
be the set of our model parameters, then the
gradient becomes:

obj

Fever

fever

punc

5.3. Labelling:

Similarly we processed the dataset to assign

subject, object and sentiment labels to each
word in each sentence for the dataset.
5.2. Tree representation of the sentence:
As mentioned before, the previous approach
utilized a greedy algorithm to find the best
possible tree representation for a sentence.

Using the above tree structure we calculate the

multi label distribution and assign the different
labels. Specifically for this problem we assign
two kinds of labels to the different elements.
The first one is the SOS (subject, object,
sentiment classes) label which holds the
probability of a particular phrase or word
consisting of subject, object and sentiment
clause. The second label gives us more
information about the sentiment if the element
has one. Currently we give it a simple labelling
indicating whether the element has a positive,
negative or neutral sentiment.

To compute this gradient, we first find

reconstruction error at all nodes and then
derivatives for these trees are computed
efficiently via backpropogation through
structure. Because the derivatives of the
supervised cross-entropy error also modify the
matrix
, this objective is not necessarily
continuous and a step in the gradient descent
direction may not necessarily decrease the
objective. However, we intend to use L-BFGS

run over the complete training data to minimize

the objective. It should be noted that the
expected objective of the project is not fully met
yet. However we believe that given more time,
we can tweak the parameters to get good results
as the model and algorithm we use might prove
to significantly outperform previous techniques
used for complex sentiment analysis.

References

1.http://nlp.stanford.edu/pubs/SocherLinNgManning_IC
ML2011.pdf

2.http://www.socher.org/uploads/Main/SocherPennington
HuangNgManning_EMNLP2011.pdf

3.http://www.socher.org/uploads/Main/2010SocherManni
ngNg.pdf

Conclusion and Future Work

We tried to use recursive autoencoder

techniques to make predictions about the
sentence structure as well as the sentiment of the
sentence with respect to the subject and object.
Training the data using deep learning techniques
to make predictions is a hard problem and we
believe that given more time we will really be
able to make a significant contribution towards
solving this problem. Hopefully we should be
able to continue with this work and achieve
good results as we keep on conducting more
experiments. As an example of an additional
factor that could be taken into consideration, we
could take into account the form of the verb
(e.g. active passive voice, tense) while making
predictions. The current model may be easily
extended to include this due to the use of the
Stanford Parser tree output in the model which
incorporates the entire syntactic structure of the
sentence. Further experiments could make use
of these additional factors to make better
predictions.

4.http://www.cs.pitt.edu/mpqa/annotation.html
5.http://cogcomp.cs.illinois.edu/page/resources/data