DL for sentence classification project Write-up

A Study of Deep Learning on Sentence Classification
By: Trinh Hoang-Trieu
Supervisor: Nguyen Le-Minh
Nguyen lab, School of Information Science
Japan Advanced Institute of Science and Technology
thtrieu@apcs.vn
Abstract
The author study different deep learning models on the task of sentence classification. Dataset used is
TREC. Models under investigation are: Convolutional Neural Networks proposed by Yoon Kim, Long
Short Term Memory, Long Short Term Memory on top of a Convolutional layer, and Convolutional
network augmented with max-pooling position. Results show that architectures with time-retaining
mechanism works better than those do not. The author also propose new changes to the conventional
final layer, namely to replace the last fully connected layer with a Linear Support Vector Machine, or to
replace the discriminative nature of the last layer with a generative one.
1. TREC dataset
TREC dataset contains 5500 labeled questions in training set and another 500 for test set. The dataset
has 6 labels, 50 level-2 labels. Average length of each sentence is 10, vocabulary size of 8700. A
sample from TREC is as follow:
DESC:manner How did serfdom develop in and then leave Russia ?
ENTY:cremat What films featured the character Popeye Doyle ?
DESC:manner How can I find a list of celebrities ' real names ?
ENTY:animal What fowl grabs the spotlight after the Chinese Year of the Monkey ?
ABBR:exp What is the full form of .com ?
HUM:ind What contemptible scoundrel stole the cork from my lunch ?
We also experiment on the Vietnamese TREC dataset, which is a direct translation of the TREC dataset
and thus, also have very similar statistics.
2. Word2Vec dataset
In this experiment, we also used another dataset to support training. The dataset consists of pre-trained
300-dimension embeddings of 3 millions words and phrases [1]. These embeddings will be used as the
representation of input for training. This covers 7500 out of 8700 words present in the TREC
vocabulary.
For the Vietnamese TREC dataset, we used another pre-trained dataset, which covers 4600 words out
of approximately 8700 words in the Vietnamese TREC dataset.

3. Convolutional Neural Networks for Sentence Classification
Yoon Kim proposes a simple yet effective convolutional architecture that achieves remarkable result on
several datasets [2]. The network consists of three layers. The first uses 300 convolutional filters with
varying size to detect 300 different features across the sentence's length. The second performs
maximum operation to summarise the previous detection. The last is a fully connected layer with
softmax output.
Figure 1. Convolutional neural network for sentence classification
The details ofS this network is as follow:
• Layer 1: 3 types of window size 3x300, 4x300, 5x300. 100 feature maps each.
• Layer 2: max-pool-over-time.
• Layer 3: Fully connected with softmax output. Weight norm constrained by 3.
▪ Dropout with p = 0.5
4. First experiment: Long Short Term Memory
The authors first experiment with a simple Long Short Term Memory architecture. We used the many-
to-one scheme to extract features before discriminating on these extracted features by a fully-connected
layer.
Figure 2. Many-to-one structure.
Source:
http://karpathy.github.io/2015/05/21/rnn-effectiveness/

The detail of this network is as follow:
- First Layer: Word Embedding
- Second Layer: Long Short Term Memory
- Third Layer: Fully connected with Softmax output
For the second layer, we experimented on three different variations of Long Short Term Memory Cell:
4.a. Gated Recurrent Unit
Figure 3: Gated Recurrent Unit
Source: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
4.b. Basic Long Short Term Memory
Figure 4: Basic Long Short Term Memory

4.c. Long Short Term Memory with peepholes
Figure 5: Long Short Term Memory with peepholes
5. Long Short Term Memory on top of a Convolutional layer
Instead of max-pool-over-time, the author propose a softer version where the maximum operation is
perform over a smaller region of the feature maps. By this, temporal information is better retained,
max-pool layer does not produce a single value but a sequence instead. After this layer, the information
is then feed into a Long Short Term Memory Cell with many-to-one scheme and a fully connected layer
with softmax output. The authors also experimented with three aforementioned types of LSTM Cell.
Figure 6: Long Short Term Memory on top of a Convolutional Layer
6. Convolutional network augmented with max-pooling position
In this network, besides performing maximum operation at the max-pooling layer, we also apply the
argmax function to produce the position in which this maximum value takes place. By supplying this
position information, the authors expect a better performance since this is also a time-retaining
mechanism, only without any recurrent connection.
To be able to propagate gradient information through this argmax operation, the authors approximate
this argmax operation with another differentiable function:

Figure 7: Convolutional network with max-pooling position
7. Hyper-parameter tuning and Experimental details
The authors use 10-fold cross validation on training set for hyper parameter tuning. The word
embedding input is kept static throughout all model except the first one to replicate Yoon Kim's
experiment. Since the average length of the sentence is only 10, while the maximum length of
sentences in the dataset is approximately 40, we also did not use paddings to avoid feeding the
networks too much noise (otherwise, on average there will be 30 padding words for each 10 words in
the dataset).
Words that do not appear in the pre-train embedding dataset is initialised with variance 0.25 to match
with existing words. For training procedure,we hold-out 5% of the data for early stopping since the
dataset is not too big. Training is carried out in batches of size 50 to speed up convergence.
8. Result
The LSTM Cell with peepholes gives best result when a single recurrent layer is used (described in
Section 4), while Gated Recurrent Unit gives best result in the hybrid model of Convolution and
Recurrent connections (described in Section 5).
Performance of the Convolutional Network augmented with position is very unstable due to the
presence of argmax operator. In our implementation, argmax is approximated with a function involves
very big constant, which in turn results in unreliable gradients and poor result. In our model, the final
layer is a fully connected layer, which is also not the best choice since classification is unlikely a linear
function of features and their position. Other results are reported as follow:
TREC TRECvn
CNN (implemented on our system) 93 91.8
CNN-LSTM 94.2 92.8
LSTM 95.4 94.2
Table 1. Accuracy (%) of different models on two datasets

9. Conclusion
As can be seen, models which utilises temporal information gives better results comparing to those do
not.
10. Future work: Support Vector Machine as the final layer
Of all the proposed model, the final layer is always a fully connected layer with softmax output, which
is equivalent to a Linear Classifier. In this case, the previous layers act as a feature extractor. Good
performance as reported indicates that these feature extractor is able to extract useful features that
represent points that can be separated well by a simple Linear Classifier. It is highly likely that these
features are also useful for other kind of Linear Classifier.
The authors propose using Linear Support Vector Machine as the final layer, this can be done by simply
replacing the usual cross-entropy loss function by Linear Support Vector Machine's loss function.
Namely, instead of using:
(Where is the extracted features from the previous layers)
With Support Vector Machine being the final layer, loss function is now:
Where RELU stands for Rectifier Linear Unit function, t is the target that is either -1 or 1 . This loss
function represent the loss when one-vs-rest scheme is used (since TREC is a multiclass dataset). For
Linear Support Vector with a soft-margin, the loss function is:
11. Reference:
[1] Kim, Yoon. "Convolutional neural networks for sentence classification." arXiv preprint
arXiv:1408.5882 (2014).
[2] Mikolov, Tomas, et al. "Distributed representations of words and phrases and their
compositionality." Advances in neural information processing systems. 2013.

DL for sentence classification project Write-up

More Related Content

What's hot (20)

Similar to DL for sentence classification project Write-up (20)

Recently uploaded (20)

DL for sentence classification project Write-up