chunk_id
stringlengths
3
7
chunk
stringlengths
1
823
source_url
stringclasses
416 values
title
stringclasses
416 values
chunk_idx
int64
0
294
chunk_start_char
int64
0
139k
chunk_end_char
int64
303
139k
15_6
Automatic Speech Recognition System Models An end-to-end ASR model maps input feature vectors to an output sequence of vectors of posterior probabilities of tokens without using separate acoustic model, pronunciation model and language model. In this work we implemented two different types of state of art end to end...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
6
3,553
4,123
15_7
Connectionist Temporal Classification (CTC) The main ideas behind CTC based ASR were first introduced in the following papers BIBREF8 , BIBREF9 . In our work we used a single layer gated recurrent unit (GRU) BIBREF10 with 128 hidden units as encoder for the CTC network. The decoder consists of a combination of a den...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
7
4,123
4,578
15_8
The number of time steps of the GRU encoder is equal to product of the sampling frequency of the input features and the length of the input sequence. Since different speakers have different rate of speech, we used dynamic recurrent neural network (RNN) cell. There is no fixed value for time steps of the encoder. Usua...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
8
4,578
5,034
15_9
A RNN based CTC network tries to make length of output tokens equal to T by allowing the repetition of output prediction unit tokens and by introducing a special token called blank token BIBREF8 across all the frames. We used CTC loss function with adam optimizer BIBREF11 and during inference time we used CTC beam sea...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
9
5,034
5,648
15_10
Let the number of time steps of the RNN encoder for ( INLINEFORM6 , INLINEFORM7 ) is INLINEFORM8 . In case of character based CTC model, the RNN predicts a character at every time step. Whereas in word based CTC model, the RNN predicts a word at every time step. For the sake of simplicity, let us assume that length of...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
10
5,648
6,021
15_11
Let the probability vector output by the RNN at each time step INLINEFORM11 be INLINEFORM12 and let INLINEFORM13 value of INLINEFORM14 be denoted by INLINEFORM15 . The probability that model outputs INLINEFORM16 on input INLINEFORM17 is given by INLINEFORM18 . During the training phase, we would like to maximize the c...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
11
6,021
6,433
15_12
In case when the length of INLINEFORM0 is less than INLINEFORM1 , we extend the target vector INLINEFORM2 by repeating a few of its values and by introducing blank token ( INLINEFORM3 ) to create a target vector of length INLINEFORM4 . Let the possible extensions of INLINEFORM5 be denoted by INLINEFORM6 . For example...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
12
6,433
6,923
15_13
We then define INLINEFORM16 as INLINEFORM17 . In our work we used character based CTC ASR model. CTC assumes the conditional independence constraint that output predictions are independent given the entire input sequence. RNN Encoder-Decoder or Attention model RNN encoder - decoder ASR model consists of a RNN encod...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
13
6,923
7,463
15_14
There is no fixed value for time steps in our case. We used dynamic RNN cell. We used a single layer GRU with 128 hidden units for both encoder and decoder. A dense layer followed by softmax activation is used after the decoder GRU to get the prediction probabilities. Dense layer performs an affine transformation. The...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
14
7,463
7,908
15_15
Training objective is to maximize the log probability of the ordered conditionals, ie: INLINEFORM0 , where X is input feature vector, INLINEFORM1 's are the labels for the ordered words present in that training example and INLINEFORM2 is the length of the output label sentence for that example. Cross entropy was used ...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
15
7,908
8,532
15_16
Let the number of times steps of encoder GRU for that example be INLINEFORM2 . The GRU encoder will transform the input features ( INLINEFORM3 ) into hidden output feature vectors ( INLINEFORM4 ). Let INLINEFORM5 word label in INLINEFORM6 (sentence) be INLINEFORM7 , then to predict INLINEFORM8 at decoder time step INL...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
16
8,532
9,064
15_17
INLINEFORM0 can be intuitively seen as a measure of how much attention INLINEFORM1 must pay to INLINEFORM2 , INLINEFORM3 . INLINEFORM4 is mathematically defined as INLINEFORM5 , where INLINEFORM6 is hidden state of the decoder GRU at time step INLINEFORM7 . The way of computing value for INLINEFORM0 depends on the ...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
17
9,064
9,608
15_18
Design of Experiments for building the database We built two types of simultaneous speech EEG recording databases for this work. For database A five female and five male subjects took part in the experiment. For database B five male and three female subjects took part in the experiment. Except two subjects, rest all...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
18
9,608
10,240
15_19
This data was recorded in presence of background noise of 40 dB (noise generated by room air conditioner fan). We then asked each subject to repeat the same experiment two more times, thus we had 30 speech EEG recording examples for each sentence. For data set B, the 8 subjects were asked to repeat the same previous ...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
19
10,240
10,789
15_20
Our EEG cap had 32 wet EEG electrodes including one electrode as ground as shown in Figure 1. We used EEGLab BIBREF17 to obtain the EEG sensor location mapping. It is based on standard 10-20 EEG sensor placement method for 32 electrodes. For data set A, we used data from first 8 subjects for training the model, remai...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
20
10,789
11,319
15_21
EEG and Speech feature extraction details EEG signals were sampled at 1000Hz and a fourth order IIR band pass filter with cut off frequencies 0.1Hz and 70Hz was applied. A notch filter with cut off frequency 60 Hz was used to remove the power line noise. EEGlab's BIBREF17 Independent component analysis (ICA) toolbox...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
21
11,319
11,799
15_22
We extracted five statistical features for EEG, namely root mean square, zero crossing rate,moving window average,kurtosis and power spectral entropy BIBREF0 . So in total we extracted 31(channels) X 5 or 155 features for EEG signals.The EEG features were extracted at a sampling frequency of 100Hz for each EEG channel...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
22
11,799
12,308
15_23
Similarly zero crossing rate was chosen as it is a commonly used feature both for speech recognition and bio signal analysis. Remaining features were chosen to capture time domain statistical information. We performed lot of experiments to identify this set of features. Initially we used only spectral entropy and zero...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
23
12,308
12,909
15_24
We first extracted MFCC 13 features and then computed first and second order differentials (delta and delta-delta) thus having total MFCC 39 features. The MFCC features were also sampled at 100Hz same as the sampling frequency of EEG features to avoid seq2seq problem. EEG Feature Dimension Reduction Algorithm Details...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
24
12,909
13,465
15_25
We reduced the 155 EEG features to a dimension of 30 by applying Kernel Principle Component Analysis (KPCA) BIBREF19 .We plotted cumulative explained variance versus number of components to identify the right feature dimension as shown in Figure 2. We used KPCA with polynomial kernel of degree 3 BIBREF0 . We further c...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
25
13,465
13,925
15_26
When we used the EEG features for ASR without dimension reduction, the ASR performance went down by 40 %. The non linear dimension reduction of EEG features significantly improved the performance of ASR. Results The attention model was predicting a word and CTC model was predicting a character at every time step, h...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
26
13,925
14,417
15_27
Table i@ and ii@ shows the test time results for attention model for both the data sets when trained using EEG features and concatenation of EEG, acoustic features respectively. As seen from the results the attention model gave lower WER when trained and tested on smaller number of sentences. As the vocabulary size i...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
27
14,417
15,039
15_28
Table iv@ and v@ shows the results obtained using CTC model. The error rates for CTC model also went up with the increase in vocabulary size for both the data sets. However the CTC model was trained for 500 epochs compared to 100 epochs for attention model to observe loss convergence and batch size was set to one for...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
28
15,039
15,555
15_29
Table vi@ shows the CTC model test time results when we trained the model using EEG features from only T7 and T8 sensors on the most noisy data set B. We observed that as vocabulary size increase, error rates were slightly lower than the error rates from Table iv@ where we used EEG features from all 31 sensors with di...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
29
15,555
16,121
15_30
Figures 4 shows the visualization of the attention weights when the attention model was trained and tested using only EEG features for Data set B. The plots shows the EEG feature importance ( attention) distribution across time steps for predicting first sentence and it indicates that attention model was not able to ...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
30
16,121
16,843
15_31
For both attention and CTC model we observed that as the vocabulary size increase, concatenating acoustic features with EEG features will help in reducing the test time error rates. We further plan to publish our speech EEG data base used in this work to help advancement of research in this area. For future work, w...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
31
16,843
17,304
15_32
We will also investigate whether it is possible to improve the attention model results by tuning hyper parameters to improve the model's ability to condition on the input,improve CTC model results by training with more number of examples and by using external language model during inference time. Acknowledgement We...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
32
17,304
17,870
15_33
Training loss convergence for CTC model using only EEG features for first 3 sentences from data set B TABLE I WER ON TEST SET FOR ATTENTION MODEL FOR DATA SET A Fig. 3. Training loss convergence for attention model using only EEG features for first 10 sentences from data set A TABLE II WER ON TEST SET FOR ATTENTION...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
33
17,870
18,332
15_34
Visualization of attention weights for the first sentence TABLE IV CER ON TEST SET FOR CTC MODEL FOR DATA SET B TABLE V CER ON TEST SET FOR CTC MODEL FOR DATA SET A TABLE VI CER ON TEST SET FOR CTC MODEL FOR DATA SET B USING EEG FEATURES FROM ONLY T7 AND T8 ELECTRODES TABLE VII CER ON TEST SET FOR CTC MODEL FOR DA...
https://arxiv.org/abs/1906.08871
Advancing Speech Recognition With No Speech Or With Noisy Speech
34
18,332
18,811
16_0
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression BERT is a cutting-edge language representation model pre-trained by a large corpus, which achieves superior performances on various natural language understanding tasks. However, a major blocking issue of applying BERT to online services is that...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
0
0
716
16_1
In this paper, we address this issue by proposing a hybrid solution named LadaBERT (Lightweight adaptation of BERT through hybrid model compression), which combines the advantages of different model compression methods, including weight pruning, matrix factorization and knowledge distillation. LadaBERT achieves state-...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
1
716
1,300
16_2
Ideally, people can start from a pre-trained BERT checkpoint and fine-tune it on a specific downstream task. However, the original BERT models are memory-exhaustive and latency-prohibitive to be served in embedded devices or CPU-based online environments. As the memory and latency constraints vary in different scenari...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
2
1,300
1,967
16_3
For example, DistilBERT BIBREF2 is re-trained on the same corpus as pre-training a vanilla BERT from scratch; and TinyBERT BIBREF3 utilizes expensive data augmentation to fit the distillation target. The costs of these model compression methods are as large as pre-training and unaffordable for low-resource settings. T...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
3
1,967
2,604
16_4
Specifically, LadaBERT is based on an iterative hybrid model compression framework consisting of weighting pruning, matrix factorization and knowledge distillation. Initially, the architecture and weights of student model are inherited from the BERT teacher. In each iteration, the student model is first compressed by ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
4
2,604
3,299
16_5
We conduct extensive experiments on five public datasets of natural language understanding. As an example, the performance comparison of LadaBERT and state-of-the-art models on MNLI-m dataset is illustrated in Figure FIGREF1. We can see that LadaBERT outperforms other BERT-oriented model compression baselines at vari...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
5
3,299
3,883
16_6
The rest of this paper is organized as follows. First, we summarizes the related works of model compression and their applications to BERT in Section SECREF2. Then, the methodology of LadaBERT is introduced in Section SECREF3, and experimental results are presented in Section SECREF4. At last, we conclude this work a...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
6
3,883
4,452
16_7
Therefore, model compression has become an indispensable technique for practice, especially in low-resource settings. In this section, we review the current progresses of model compression techniques briefly, which can be divided into four categories, namely weight pruning, matrix factorization, weight quantization an...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
7
4,452
5,132
16_8
For example, Han et al. BIBREF4 proposed a method to reduce the storage and computation of neural networks by removing unimportant connections, resulting in sparse networks without affecting the model accuracy. Li et al. BIBREF5 presented an acceleration method for convolution neural network by pruning whole filters t...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
8
5,132
5,783
16_9
BIBREF6 alleviated this problem by a data-driven approach that pruned zero-activation neurons iteratively based on intermediate feature maps. Zhu and Gupta BIBREF7 empirically compared large-sparse models with smaller dense models of similar parameter sizes and found that large sparse models performed better consisten...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
9
5,783
6,325
16_10
BIBREF10 learned sparse neural networks through $l_0$ regularization. Related Work ::: Matrix factorization The goal of matrix factorization is to decompose a matrix into the product of two matrices in lower dimensions, and Singular Value Decomposition (SVD) is a popular way of matrix factorization that generalizes ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
10
6,325
6,835
16_11
Matrix factorization was widely studied in the deep learning domain for model compression and acceleration BIBREF13, BIBREF14, BIBREF15. Sainath et al BIBREF13 explored a low-rank matrix factorization method of DNN layers for acoustic modeling. Xu et al. BIBREF14, BIBREF15 applied singular value decomposition to deep ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
11
6,835
7,438
16_12
BIBREF17 compressed the word embedding layer via matrix factorization and achieved promising results in text classification. Winata et al. BIBREF18 carried out experiments for low-rank matrix factorization on different NLP tasks and demonstrated that it was more effective in general than weight pruning. Related Work ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
12
7,438
8,062
16_13
With weight quantization, the weights can be reduced to at most 1-bit binary value from 32-bits floating point numbers. Zhou et al. BIBREF19 showed that quantizing weights to 8-bits does not hurt the performance, and Binarized Neural Networks BIBREF20 contained binary weights and activations of only one bit. Increment...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
13
8,062
8,598
16_14
Variational Network Quantization BIBREF22 formulated the problem of network quantization as a variational inference problem. Moreover, Choi et al. BIBREF23 investigated the drawbacks of conventional quantization methods based on k-means and proposed a Hessian-weighted k-means clustering algorithm as the solution. Rel...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
14
8,598
9,271
16_15
This “synthetic" label is then used to train a smaller network (the student model), which assimilates the function that is learned by the teacher model. Chen et al. BIBREF24 successfully applied knowledge distillation to object detection tasks by introducing several modifications, including a weighted cross-entropy lo...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
15
9,271
9,888
16_16
BIBREF26 proposed online distillation, a variant of knowledge distillation which enabled extra parallelism for training large-scale data. In addition, knowledge distillation is also useful for aggregating model ensembles into a single model by treating the ensemble model as a teacher. Related Work ::: Hybrid approach...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
16
9,888
10,524
16_17
BIBREF28 proposed a unified framework for low-rank and sparse decomposition of weight matrices with feature map reconstructions. Polino et al. BIBREF29 advocated a combination of distillation and quantization techniques and proposed two hybrid models, i.e., quantified distillation and differentiable quantization to ad...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
17
10,524
10,963
16_18
NNCF BIBREF31 provided a neural network compression framework that supported an integration of various model compression methods to generate more lightweight networks and achieved state-of-the-art performances in terms of a trade-off between accuracy and efficiency. In BIBREF32, an AutoML pipeline was adopted for mode...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
18
10,963
11,677
16_19
Most existing works focus on knowledge distillation. For instance, BERT-PKD BIBREF33 is a patient knowledge distillation approach that compresses the original BERT model into a lightweight shallow network. Different from traditional knowledge distillation methods, BERT-PKD enables an exploitation of rich information i...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
19
11,677
12,185
16_20
Distilled BiLSTM BIBREF34 adopts a single-layer BiLSTM as the student model and achieves comparable results with ELMo BIBREF35 through much fewer parameters and less inference time. TinyBERT BIBREF3 reports the best-ever performance on BERT model compression, which exploits a novel attention-based distillation schema ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
20
12,185
12,781
16_21
Both procedures require huge resources and long training times (from several days to weeks), which is cumbersome for industrial applications. Therefore, we are aiming to explore more lightweight solutions in this paper. Lightweight Adaptation of BERT ::: Overview The overall pipeline of LadaBERT (Lightweight Adaptat...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
21
12,781
13,293
16_22
Then, the student model is compressed towards smaller parameter size through a hybrid model compression framework in an iterative manner until the target compression ratio is reached. Concretely, in each iteration, the parameter size of student model is first reduced by $1-\Delta $ based on weight pruning and matrix f...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
22
13,293
13,978
16_23
Moreover, weight pruning and matrix factorization generates better initial and intermediate status of the student model, which improve the efficiency and effectiveness of knowledge distillation. In the following subsections, we will introduce the algorithms in detail. Lightweight Adaptation of BERT ::: Overview ::: M...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
23
13,978
14,462
16_24
Without loss generality, we assume a matrix of parameters ${W} \in \mathbb {R}^{m\times n}$, the singular value decomposition of which can be written as: where ${U} \in \mathbb {R}^{m \times p}$ and ${V} \in \mathbb {R}^{p \times n}$.
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
24
14,462
14,698
16_25
${\Sigma } =diag(\sigma _1,\sigma _2,\ldots ,\sigma _p)$ is a diagonal matrix composed of singular values and $p$ is the full rank of $W$ satisfying $p \le min(m, n)$. To compress this weight matrix, we select a lower rank $r$. The diagonal matrix ${\Sigma }$ is truncated by selecting the top $r$ singular values.
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
25
14,698
15,014
16_26
i.e., ${\Sigma }_r =diag(\sigma _1, \sigma _2,\ldots ,\sigma _r)$, while ${U}$ and ${V}$ are also truncated by selecting the top $r$ columns and rows respectively, resulting in ${U}_r \in \mathbb {R}^{m\times r}$ and ${V}_r \in \mathbb {R}^{r\times n}$.
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
26
15,014
15,268
16_27
Thus, low-rank matrix approximation of ${W}$ can be formulated as: In this way, the original weight matrix $W$ is decomposed by the multiplication of two smaller matrices, where ${A}={U}_r\sqrt{{\Sigma }_r} \in \mathbb {R}^{n\times r}$ and ${B}={V}_r\sqrt{{\Sigma }_r} \in \mathbb {R}^{m\times r}$. These two matrices...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
27
15,268
15,654
16_28
Given a rank $r \le min(m, n)$, the compression ratio of matrix factorization is defined as: Therefore, for a target model compression ratio $P_{svd}$, the desired rank $r$ can be calculated by: Lightweight Adaptation of BERT ::: Overview ::: Weight pruning Weight pruning BIBREF4 is an unstructured compression met...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
28
15,654
16,037
16_29
For a neural network $f({x; \theta })$ with parameters $\theta $, weight pruning finds a binary mask ${M} \in \lbrace 0, 1\rbrace ^{|\theta |}$ subject to a given sparsity ratio, $P_{weight}$. The neural network after pruning will be $f({x; M \cdot \theta })$, where the non-zero parameter size is $||{M}||_1 = P_{weigh...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
29
16,037
16,436
16_30
For example, when $P_m = 0.3$, there are 70% zeros and 30% ones in the mask ${m}$. We adopt a simple pruning strategy in our implementation: the binary mask is generated by setting the smallest weights to zeros BIBREF36. To combine the benefits of weight pruning with matrix factorization, we leverage a hybrid approac...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
30
16,436
16,839
16_31
Following Equation (DISPLAY_FORM12), SVD-based matrix factorization for any weight matrix ${W}$ can be written as: ${W}_{svd}={A}_{m\times r}{B}_{n\times r}^T$. Then, weight pruning is applied on the decomposed matrices ${A} \in \mathbb {R}^{m \times r}$ and ${B} \in \mathbb {R}^{n \times r}$ separately.
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
31
16,839
17,145
16_32
The weight matrix after hybrid compression is denoted by: where ${M_A}$ and ${M_B}$ are binary masks derived by the weight pruning algorithm with compression ratio $P_{weight}$. The compression ratio of this hybrid approach can be calculated by: In LadaBERT, the hybrid compression produce is applied to each layer of...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
32
17,145
17,493
16_33
Given an overall model compression target $P$, the following constraint should be satisfied: where $|\theta |$ is the total number of model parameters and $P$ is the target compression ratio; $|\theta _{embd}|$ denotes the parameter number of embedding layer, which has a relative compression ratio of $P_embd$, and $|...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
33
17,493
17,941
16_34
The classification layer (often MLP layer with Softmax activation) has a small parameter size ($|\theta _{cls}|$), so it is not modified in the model compression procedure. In the experiments, these fine-grained compression ratios can be optimized by random search on the validation data. Lightweight Adaptation of BER...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
34
17,941
18,546
16_35
Various types of knowledge distillation can be employed at different sub-layers. Generally, all types of knowledge distillation can be modeled as minimizing the following loss function: Where $x$ indicates a sample input and $\mathcal {X}$ is the training dataset. $f^{(s)}({x})$ and $f^{(t)}({x})$ represent intermedi...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
35
18,546
19,074
16_36
We follow the recent technique proposed by TinyBERT BIBREF3, which applies knowledge distillation constraints upon embedding, self-attention, hidden representation and prediction levels. Concretely, there are four types of knowledge distillation constraints as follows: Embedding-layer distillation is performed upon t...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
36
19,074
19,586
16_37
Mean Squared Error (MSE) is adopted as the loss function $L(\cdot )$. Attention-layer distillation is performed upon the self-attention sub-layer. $f({x}) = \lbrace a_{ij}\rbrace \in \mathbb {R}^{n \times n}$ represents the attention output for each self-attention sub-layer, and $L(\cdot )$ denotes MSE loss function....
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
37
19,586
20,014
16_38
$f({x})$ denotes the output representation of the corresponding sub-layer, and $L(\cdot )$ also adopts MSE loss function. Prediction-layer distillation makes the student model to learns the predictions from a teacher model directly. It is identical to the vanilla form of knowledge distillation BIBREF1. It takes the s...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
38
20,014
20,493
16_39
Experiments ::: Datasets & Baselines We compare LadaBERT with state-of-the-art model compression approaches on five public datasets of different tasks of natural language understanding, including sentiment classification (SST-2), natural language inference (MNLI-m, MNLI-mm, QNLI) and pairwise semantic equivalence (Q...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
39
20,493
21,129
16_40
Hybrid pruning is a combination of matrix factorization and weight pruning, which conducts iterative weight pruning on the basis of SVD-based matrix factorization. It is performed iteratively until the desired compression ratio is achieved. BERT-FT, BERT-KD and BERT-PKD are reported in BIBREF33, where BERT-FT direct...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
40
21,129
21,638
16_41
The student model is composed of 3 Transformer layers, resulting in a $2.5\times $ compression ratio. Each layer has the same hidden size as the pre-trained teacher, so the initial parameters of student model can be inherited from the corresponding teacher. TinyBERT BIBREF3 instantiates a tiny student model, which ha...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
41
21,638
22,094
16_42
For a fair comparison, we reproduce the TinyBERT pipeline without general distillation and data augmentation, which is time-exhaustive and resource-consuming. BERT-SMALL has the same model architecture as TinyBERT, but is directly pre-trained by the official BERT pipeline. The performance values are inherited from BI...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
42
22,094
22,661
16_43
This model requires a expensive pre-training process using the knowledge distillation constraints. Experiments ::: Setup We leverage the pre-trained checkpoint of base-bert-uncased as the initial model for compression, which contains 12 layers, 12 heads, 110M parameters, and 768 hidden units per layer. Hyper-paramet...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
43
22,661
23,190
16_44
For a comprehensive evaluation, we experiment with four settings of LadaBERT, namely LadaBERT-1, -2, -3 and -4, which reduce the model parameters of BERT-Base by 2.5, 5, 7.5 and 10 times respectively. In our experiment, we take the batch size as 32, learning rate as 2e-5. The optimizer is BertAdam with default settin...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
44
23,190
23,604
16_45
Experiments ::: Performance Comparison The evaluation results of LadaBERT and state-of-the-art approaches are listed in Table TABREF40, where the models are ranked by parameter sizes for feasible comparison. As shown in the table, LadaBERT consistently outperforms the strongest baselines under similar model sizes. I...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
45
23,604
24,092
16_46
With model size of $2.5\times $ reduction, LadaBERT-1 performs significantly better than BERT-PKD, boosting the performance by relative 8.9, 8.1, 6.1, 3.8 and 5.8 percentages on MNLI-m, MNLI-mm, SST-2, QQP and QNLI datasets respectively. Recall that BERT-PKD initializes the student model by selecting 3 of 12 layers i...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
46
24,092
24,446
16_47
It turns out that the discarded layers have huge impact on the model performance, which is hard to be recovered by knowledge distillation. On the other hand, LadaBERT generates the student model by iterative pruning on the pre-trained teacher. In this way, the original knowledge in the teacher model can be preserved t...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
47
24,446
24,942
16_48
As shown in the results, TinyBERT does not work well without expensive data augmentation and general distillation, hindering its application to low-resource settings. The reason is that the student model of TinyBERT is distilled from scratch, so it requires much more data to mimic the teacher's behaviors. Instead, Lad...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
48
24,942
25,512
16_49
Moreover, LadaBERT-3 also outperforms BERT-SMALL on most of the datasets, which is pre-trained from scratch by the official BERT pipeline on a $7.5 \times $ smaller architecture. This indicates that LadaBERT can quickly adapt to a smaller model size and achieve competitive performance without expansive re-training on ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
49
25,512
26,021
16_50
Nevertheless, the performance of LadaBERT-4 is competitive on larger datasets such as MNLI and QQP. This is impressive as LadaBERT is much more efficient without exhaustive re-training on a large corpus. In addition, the inference speed of BiLSTM is usually slower than transformer-based models with similar parameter s...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
50
26,021
26,346
16_51
Experiments ::: Learning curve comparison To further demonstrate the efficiency of LadaBERT, we visualize the learning curves on MNLI-m and QQP datasets in Figure FIGREF42 and FIGREF42, where LadaBERT-3 is compared to the strongest baseline, TinyBERT, under $7.5 \times $ compression ratio. As shown in the figures, L...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
51
26,346
26,757
16_52
After training $2 \times 10^4$ steps (batches) on MNLI-m dataset, the performance of LadaBERT-3 is already comparable to TinyBERT after convergence (approximately $2 \times 10^5$ steps), achieving nearly $10 \times $ acceleration. And on QQP dataset, both performance improvement and training speed acceleration is very...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
52
26,757
27,322
16_53
Experiments ::: Effect of low-rank + sparsity In this paper, we demonstrate that a combination of matrix factorization and weight pruning is better than single solutions for BERT-oriented model compression. Similar phenomena has been reported in the computer vision scenarios BIBREF28, which shows that low-rank and s...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
53
27,322
27,903
16_54
The errors can be calculated by $\mathop {Error}=||\hat{{M}}-{M}||_1$, where $\hat{{M}}$ denotes the weight matrix after pruning. The yellow line in Figure FIGREF44 shows the distribution of errors generated by pure weight pruning, which has a sudden drop at the pruning threshold. The orange line represents for pure ...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
54
27,903
28,414
16_55
First, we apply SVD-based matrix factorization to reduce 60% of total parameters. Then, weight pruning is applied on the decomposed matrices by 50%, resulting in only 20% parameters while the error distribution changes slightly. As a result, it has smaller mean and deviation than pure matrix factorization. In addition...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
55
28,414
29,002
16_56
Existing model compression methods for BERT need to be re-trained on a large corpus to reserve its original performance, which is inapplicable in low-resource settings. In this paper, we propose LadaBERT to address this problem. LadaBERT is a lightweight model compression pipeline that generates adaptive BERT model ef...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
56
29,002
29,687
16_57
Therefore, LadaBERT can be easily plugged into various applications with competitive performances and little training overheads. In the future, we would like to apply LadaBERT to large-scale industrial applications, such as search relevance and query recommendation. Figure 1: Accuracy comparison on MNLI-m dataset Fi...
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
57
29,687
30,258
16_58
Figure 5: Distribution of pruning errors
https://arxiv.org/abs/2004.04124
LadaBERT: Lightweight Adaptation of BERT through Hybrid Model Compression
58
30,258
30,300
17_0
Neural Summarization by Extracting Sentences and Words Traditional approaches to extractive summarization rely heavily on human-engineered features. In this work we propose a data-driven approach based on neural networks and continuous sentence features. We develop a general framework for single-document summarization...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
0
0
620
17_1
Experimental results on two summarization datasets demonstrate that our models obtain results comparable to the state of the art without any access to linguistic annotation. Introduction The need to access and digest large amounts of textual data has provided strong impetus to develop automatic summarization systems...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
1
620
1,329
17_2
These include surface features such as sentence position and length BIBREF0 , the words in the title, the presence of proper nouns, content features such as word frequency BIBREF1 , and event features such as action nouns BIBREF2 . Sentences are typically assigned a score indicating the strength of presence of these f...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
2
1,329
1,889
17_3
In this work we propose a data-driven approach to summarization based on neural networks and continuous sentence features. There has been a surge of interest recently in repurposing sequence transduction neural network architectures for NLP tasks such as machine translation BIBREF8 , question answering BIBREF9 , and ...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
3
1,889
2,574
17_4
We develop a general framework for single-document summarization which can be used to extract sentences or words. Our model includes a neural network-based hierarchical document reader or encoder and an attention-based content extractor. The role of the reader is to derive the meaning representation of a document bas...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
4
2,574
3,301
17_5
Similar neural attention architectures have been previously used for geometry reasoning BIBREF12 , under the name Pointer Networks. One stumbling block to applying neural network models to extractive summarization is the lack of training data, i.e., documents with sentences (and words) labeled as summary-worthy. Insp...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
5
3,301
3,962
17_6
Using a number of transformation and scoring algorithms, we are able to match highlights to document content and construct two large scale training datasets, one for sentence extraction and the other for word extraction. Previous approaches have used small scale training data in the range of a few hundred examples. O...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
6
3,962
4,659
17_7
Our word-based model is similar in spirit, however, it operates over continuous representations, produces multi-sentence output, and jointly selects summary words and organizes them into sentences. A few recent studies BIBREF14 , BIBREF15 perform sentence extraction based on pre-trained sentence embeddings following a...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
7
4,659
5,369
17_8
In contrast, our model summarizes documents rather than individual sentences, producing multi-sentential discourse. A major architectural difference is that our decoder selects output symbols from the document of interest rather than the entire vocabulary. This effectively helps us sidestep the difficulty of searching...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
8
5,369
5,837
17_9
Gu:ea:16 and gulcehre2016pointing propose a similar “copy” mechanism in sentence compression and other tasks; their model can accommodate both generation and extraction by selecting which sub-sequences in the input sequence to copy in the output. We evaluate our models both automatically (in terms of Rouge) and by hu...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
9
5,837
6,460
17_10
Problem Formulation In this section we formally define the summarization tasks considered in this paper. Given a document $D$ consisting of a sequence of sentences $\lbrace s_1, \cdots , s_m\rbrace $ and a word set $\lbrace w_1, \cdots , w_n\rbrace $ , we are interested in obtaining summaries at two levels of granul...
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
10
6,460
6,924
17_11
We do this by scoring each sentence within $D$ and predicting a label $y_L \in {\lbrace 0,1\rbrace }$ indicating whether the sentence should be included in the summary.
https://arxiv.org/abs/1603.07252
Neural Summarization by Extracting Sentences and Words
11
6,924
7,093