Transfer learning, or sometimes known as domain adaptation,222In this paper, we do not distinguish the conceptual difference between transfer learning and domain adaptation. Domain—in the sense we use throughout this paper—is defined by datasets.
plays an important role in various natural language processing (NLP) applications, especially when we do not have large enough datasets for the task of interest (called thetarget task ). In such scenarios, we would like to transfer or adapt knowledge from other domains (called the source domains/tasks ) so as to mitigate the problem of overfitting and to improve model performance in . For traditional feature-rich or kernel-based models, researchers have developed a variety of elegant methods for domain adaptation; examples include EasyAdapt [Daumé III2007, Daumé III et al.2010], instance weighting [Jiang and Zhai2007, Foster et al.2010], and structural correspondence learning [Blitzer et al.2006, Prettenhofer and Stein2010].
Recently, deep neural networks are emerging as the prevailing technical solution to almost every field in NLP. Although capable of learning highly nonlinear features, deep neural networks are very prone to overfitting, compared with traditional methods. Transfer learning therefore becomes even more important. Fortunately, neural networks can be trained in a transferable way by their incremental learning nature: we can directly use trained (tuned) parameters from a source task to initialize the network in the target task; alternatively, we may also train two tasks simultaneously with some parameters shared. But their performance should be verified by empirical experiments.
Existing studies have already shown some evidence of the transferability of neural features. For example, in image processing, low-level neural layers closely resemble Gabor filters or color blobs [Zeiler and Fergus2014, Krizhevsky et al.2012]; they can be transferred well to different tasks. image suggest that high-level layers are also transferable in general visual recognition; how further investigate the transferability of neural layers in different levels of abstraction.
Although transfer learning is promising in image processing, conclusions appear to be less clear in NLP applications. Image pixels are low-level signals, which are generally continuous and less related to semantics. By contrast, natural language tokens are discrete: each word well reflects the thought of humans, but neighboring words do not share as much information as pixels in images do. Previous neural NLP studies have casually applied transfer techniques, but their results are not consistent. unified apply multi-task learning to SRL, NER, POS, and CHK,333 The acronyms refer to semantic role labeling, named entity recognition, part-of-speech tagging, and chunking, respectively. but obtain only 0.04–0.21% error reduction444Here, we quote the accuracies obtained by using unsupervised pretraining of word embeddings. This is the highest performance in that paper; using pretrained word embeddings is also a common practice in the literature. (out of a base error rate of 16–18%). NLI, on the contrary, improve a natural language inference task from an accuracy of 71.3% to 80.8% by initializing parameters with an additional dataset of 550,000 samples. Therefore, more systematic studies are needed to shed light on transferring neural networks in the field of NLP.
In this paper, we investigate the question “How transferable are neural networks in NLP applications?”
We distinguish two scenarios of transfer: (1) transferring knowledge to a semantically similar/equivalent task but with a different dataset; (2) transferring knowledge to a task that is semantically different but shares the same neural topology/architecture so that neural parameters can indeed be transferred. We further distinguish two transfer methods: (1) using the parameters trained on to initialize (INIT), and (2) multi-task learning (MULT), i.e., training and simultaneously. (Please see Sections 2 and 4). Our study mainly focuses on the following research questions:
How transferable are neural networks between two tasks with similar or different semantics in NLP applications?
How transferable are different layers of NLP neural models?
How transferable are INIT and MULT, respectively? What is the effect of combining these two methods?
We conducted extensive experiments over six datasets on classifying sentences and sentence pairs. We leveraged the widely-used convolutional neural network (CNN) and long short term memory (LSTM)-based recurrent neural network (RNN) as our models.
Based on our experimental results, we have the following main observations, some of which are unexpected.
Whether a neural network is transferable in NLP depends largely on how semantically similar the tasks are, which is different from the consensus in image processing. The output layer is mainly specific to the dataset and not transferable. Word embeddings are likely to be transferable to semantically different tasks. MULT and INIT appear to be generally comparable to each other; combining these two methods does not result in further gain in our study.
The rest of this paper is organized as follows. Section 2 introduces the datasets that neural models are transferred across; Section 3 details the neural architectures and experimental settings. We describe two approaches (INIT and MULT) to transfer learning in Section 4. We present experimental results in Sections 5–6 and have concluding remarks in Section 7.
|Statistics (# of Samples)|
|Experiment I||Experiment II|
|Examples in Experiment I|
|Sentiment Analysis (IMDB and MR)|
|An idealistic love story that brings out|
|the latent 15-year-old romantic in everyone.|
|Its mysteries are transparently obvious,|
|and it’s too slowly paced to be a thriller.|
|Question Classification (QC)|
|What is the temperature at the center of the earth?||number|
|What state did the Battle of Bighorn take place in?||location|
|Examples in Experiment II|
|Natural Language Inference (SNLI and SICK)|
|Premise||Two men on bicycles competing in a race.|
|People are riding bikes.||E|
|Hypothesis||Men are riding bicycles on the streets.||C|
|A few people are catching fish.||N|
|Paraphrase Detection (MSRP)|
|The DVD-CCA then appealed to the state||Paraphrase|
|The DVD CCA appealed that decision|
|to the U.S. Supreme Court.|
|Earnings per share from recurring operations|
|will be 13 cents to 14 cents.||Non-|
|That beat the company’s April earnings||Paraphrase|
|forecast of 8 to 9 cents a share.|
In our study, we conducted two series of experiments using six open datasets as follows.
Experiment I: Sentence classification
IMDB. A large dataset for binary sentiment classification (positive vs. negative).555https://drive.google.com/file/d/
MR. A small dataset for binary sentiment classification.666https://www.cs.cornell.edu/people/pabo/
QC. A (small) dataset for 6-way question classification (e.g., location, time, and number).777http://cogcomp.cs.illinois.edu/Data/QA/QC/
Experiment II: Sentence-pair classification
SNLI. A large dataset for sentence entailment recognition. The classification objectives are entailment, contradiction, and neutral.888http://nlp.stanford.edu/projects/snli/
SICK. A small dataset with exactly the same classification objective as SNLI.999http://http://alt.qcri.org/semeval2014/task1/
MSRP. A (small) dataset for paraphrase detection. The objective is binary classification: judging whether two sentences have the same meaning.101010http://research.microsoft.com/en-us/downloads/
In each experiment, the large dataset serves as the source domain and small ones are the target domains. Table 1 presents statistics of the above datasets.
We distinguish two scenarios of transfer regarding semantic similarity: (1) semantically equivalent transfer (IMDBMR, SNLISICK), that is, the tasks of and are defined by the same meaning, and (2) semantically different transfer (IMDBQC, SNLIMSRP). Examples are also illustrated in Table 1 to demonstrate semantic relatedness.
It should be noticed that in image or speech processing [Yosinski et al.2014, Wang and Zheng2015], the input of neural networks pretty much consists of raw signals; hence, low-level feature detectors are almost always transferable, even if how manually distinguish artificial objects and natural ones in an image classification task.
Distinguishing semantic relatedness—which emerges from very low layers of either word embeddings or the successive hidden layer—is specific to NLP and also a new insight of our paper. As we shall see in Sections 5 and 6, the transferability of neural networks in NLP is more sensitive to semantics than in image processing.
3 Neural Models and Settings
In each group, we used a single neural model to solve three problems in a unified manner. That is to say, the neural architecture is the same among the three datasets, which makes it possible to investigate transfer learning regardless of whether the tasks are semantically equivalent. Concretely, the neural models are as follows.
b) with a window size of 5 to model local context, and a max pooling layer gathers information to a fixed-size vector. Then the sentence vectors are concatenated and fed to a hidden layer before the softmax output.
In our experiments, embeddings were pretrained by word2vec [Mikolov et al.2013]
; all embeddings and hidden layers were 100 dimensional. We applied stochastic gradient descent with a mini-batch size of 50 for optimization. In each setting, we tuned the hyperparameters as follows: learning rate from, power decay of learning rate from fast, moderate, low
(defined by how much, after one epoch, the learning rate residual is:resp). We regularized our network by dropout with a rate from . Note that we might not run nonsensical settings, e.g., a larger dropout rate if the network has already been underfitting (i.e., accuracy has decreased when the dropout rate increases). We report the test performance associated with the highest validation accuracy.
To setup a baseline, we trained our models without transfer 5 times by different random parameter initializations (Table 2). We have achieved reasonable performance that is comparable to similar models reported in the literature with all six datasets. Therefore, our implementation is fair and suitable for further study of transfer learning.
It should be mentioned that the goal of this paper is not to outperform state-of-the-art results; instead, we would like to conduct a fair comparison of different methods and settings for transfer learning in NLP.
|Dataset||Avg acc.std.||Related model|
|IMDB||87.0||89.3 (Non-NN, Dong,2015)|
|MR||77.7 (RAE, Socher, 2013)|
|QC||90.2 (RNN, Zhao,2015)|
|SNLI||77.6 (RNN, Bowman,2015)|
|SICK||71.3 (RNN, Bowman,2015)|
|MSRP||69.6 (Arc-I CNN, Hu,2014)|
4 Transfer Methods
Transfer learning aims to use knowledge in a source domain to aid the target domain. As neural networks are usually trained incrementally with gradient descent (or variants), it is straightforward to use gradient information in both source and target domains for optimization so as to accomplish knowledge transfer. Depending on how samples in source and target domains are scheduled, there are two main approaches to neural network-based transfer learning:
Parameter initialization (INIT). The INIT approach first trains the network on , and then directly uses the tuned parameters to initialize the network for . After transfer, we may fix () the parameters in the target domain [Glorot et al.2011], i.e., no training is performed on . But when labeled data are available in , it would be better to fine-tune () the parameters.
INIT is also related to unsupervised pretraining such as word embedding learning [Mikolov et al.2013]
and autoencoders[Bengio et al.2006]. In these approaches, parameters that are (pre)trained in an unsupervised way are transferred to initialize the model for a supervised task [Plank and Moschitti2013]. However, our paper focuses on “supervised pretraining,” which means we transfer knowledge from a labeled source domain.
where and are the individual cost function of each domain. (Both and are normalized by the number of training samples.) is a hyperparameter balancing the two domains.
It is nontrivial to optimize Equation 1 in practice by gradient-based methods. One may take the partial derivative of and thus goes to the learning rate [Liu et al.2016], but the model is then vulnerable because it is likely to blow up with large learning rates (multiplied by or ) and be stuck in local optima with small ones.
unified alternatively choose a data sample from either domain with a certain probability (controlled by) and take the derivative for the particular data sample. In this way, domain transfer is independent of learning rates, but we may not be able to fully use the entire dataset of if is large. We adopted the latter approach in our experiment for simplicity. (More in-depth analysis may be needed in future work.) Formally, our multi-task learning strategy is as follows.
Switch to with prob. , or to with prob. .
Compute the gradient of the next data sample in the particular domain.
Further, INIT and MULT can be combined straightforwardly, and we obtain the third setting:
Combination (MULT+INIT). We first pretrain on the source domain for parameter initialization, and then train and simultaneously.
From a theoretical perspective, INIT and MULT work in different ways. In the MULT approach, the source domain regularizes the model by “aliasing” the error surface of the target domain; hence the neural network is less prone to overfitting. In INIT, ’s error surface remains intact. Before training on the target dataset, the parameters are initialized in such a meaningful way that they contain additional knowledge in the source domain. However, in an extreme case where ’s error surface is convex, INIT is ineffective because the parameters can reach the global optimum regardless of their initialization. In practice, deep neural networks usually have highly complicated, non-convex error surfaces. By properly initializing parameters with the knowledge of , we can reasonably expect that the parameters are in a better “catchment basin,” and that the INIT approach can transfer knowledge from to .
5 Results of Transferring by INIT
We first analyze how INIT behaves in NLP-based transfer learning. In addition to two different transfer scenarios regarding semantic relatedness as described in Section 2, we further evaluated two settings: (1) fine-tuning parameters , and (2) freezing parameters after transfer . Existing evidence shows that frozen parameters would generally hurt the performance [Peng et al.2015], but this setting provides a more direct understanding on how transferable the features are (because the factor of target domain optimization is ruled out). Therefore, we included it in our experiments. Moreover, we transferred parameters layer by layer to answer our second research question.
Through Subsections 5.1–5.3, we initialized the parameters of with the ones corresponding to the highest validation accuracy of . In Subsection 5.4, we further investigated when the parameters are ready to be transferred during the training on .
5.1 Overall Performance
Table 3 shows the main results of INIT. A quick observation is that, in both groups, transfer learning of semantically equivalent tasks (IMDBMR, SNLISICK) appears to be successful with an improvement of 6%. The results are not surprising and also reported in NLI.
For IMDBQC and SNLIMSRP, however, there is no improvement of transferring hidden layers (embeddings excluded), namely LSTM-RNN units and CNN feature maps. The EHO setting yields a slight degradation of 0.2–0.4%, .5x std. The incapability of transferring is also proved by locking embeddings and hidden layers (EHO). We see in this setting, the test performance is very low in QC or even worse than majority-class guess in MSRP. By further examining its training accuracy, which is 48.2% and 65.5%, respectively, we conclude that extracted features by LSTM-RNN and CNN models in are almost irrelevant to the ultimate tasks (QC and MSRP).
Although in previous studies, researchers have mainly drawn positive conclusions about transfer learning, we find a negative result similar to ours upon careful examination of unified, and unfortunately, their results may be somewhat misinterpreted. In that paper, the authors report transferring NER, POS, CHK, and pretrained word embeddings improves the SRL task by 1.91–3.90% accuracy (out of 16.54–18.40% error rate), but their gain is mainly due to word embeddings. In the settings that use pretrained word embeddings (which is common in NLP), NER, POS, and CHK together improve the SRL accuracy by only 0.04–0.21%.
The above results are rather frustrating, indicating for RQ1 that neural networks may not be transferable to NLP tasks of different semantics. Transfer learning for NLP is more prone to semantics than the image processing domain, where even high-level feature detectors are almost always transferable [Donahue et al.2014, Yosinski et al.2014].
5.2 Layer-by-Layer Analysis
To answer RQ2, we next analyze the transferability of each layer. First, we freeze both embeddings and hidden layers (EH). Even in semantically equivalent settings, if we further freeze the output layer (O), the performance in both IMDBMR and SNLISICK drops, but by randomly initializing the output layer’s parameters (O), we can obtain a similar or higher result compared with the baseline (EHO). The finding suggests that the output layer is mainly specific to a dataset. Transferring the output layer’s parameters yields little (if any) gain.
Regarding embeddings and hidden layers (in the settings EHOEHO vs. E⌧HO), the IMDBMR experiment suggests both of embeddings and the hidden layer play an important role, each improving the accuracy by 3%. In SNLISICK, however, the main improvement lies in the hidden layer. A plausible explanation is that in sentiment classification tasks (IMDB and MR
), information emerges from raw input, i.e., sentiment lexicons and thus their embeddings, but natural language inference tasks (SNLI and SICK) address more on semantic compositionality and thus hidden layers are more important.
Moreover, for semantically different tasks (IMDBQC and SNLIMSRP), the embeddings are the only parameters that have been observed to be transferable, slightly benefiting the target task by 2.7x and 1.8x std, respectively.
5.3 How does learning rate affect transfer?
NLI suggest that after transferring, a large learning rate may damage the knowledge stored in the parameters; in their paper, they transfer the learning rate information (AdaDelta) from to in addition to the parameters.
Although the rule of the thumb is to choose all hyperparameters—including the learning rate—by validation, we are curious whether the above conjecture holds. Estimating a rough range of sensible hyperparameters can ease the burden of model selection; it also provides evidence to better understand how transfer learning actually works.
We plot the learning curves of different learning rates in Figure 2 (IMDBMR and SNLISICK, EHO). (In the figure, no learning rate decay is applied.) As we see, with a large learning rate like , the accuracy increases fast and peaks at earlier epochs. Training with a small learning rate (e.g., ) is slow, but its peak performance is comparable to large learning rates when iterated by, say, 100 epochs. The learning curves in Figure 2
are similar to classic speed/variance trade-off, and we have the following additional discovery:
In INIT, transferring learning rate information is not necessarily useful. A large learning rate does not damage the knowledge stored in the pretrained hyperparameters, but accelerates the training process to a large extent. In all, we may need to perform validation to choose the learning rate if computational resources are available.
5.4 When is it ready to transfer?
In the above experiments, we transfer the parameters when they achieve the highest validation performance on . This is a straightforward and intuitive practice.
However, we may imagine that the parameters well-tuned to the source dataset may be too specific to it, i.e., the model overfits and thus may underfit . Another advantage of early transfer lies in computational concerns. If we manage to transfer model parameters after one or a few epochs on , we can save much time especially when is large.
We therefore made efforts in studying when the neural model is ready to be transferred. Figures 3a and 3c plot the learning curves of the source tasks. The accuracy increases sharply from epochs 1–5; later, it reaches a plateau but is still growing slowly.
In IMDBMR, the source performance and transferring performance align well. The SNLISICK experiment, however, produces interesting yet unexpected results. Using the second epoch of SNLI’s training yields the highest transfer performance on SICK, i.e., 78.98%, when the SNLI performance itself is comparatively low (72.65% vs. 76.26% at epoch 23). Later, the transfer performance decreases gradually by 2.7%. The results in these two experiments are inconsistent and lack explanation.
6 MULT, and its Combination with INIT
To answer RQ3, we investigate how multi-task learning performs in transferring knowledge, as well as the effect of the combination of MULT and INIT. In this section, we applied the setting: sharing embeddings and hidden layers (denoted as EHO), analogous to EHO in INIT. When combining MULT and INIT, we used the pretrained parameters of embeddings and hidden layers on to initialize the multi-task training of and , visually represented by EHO.
In both MULT and MULT+INIT, we had a hyperparameter balancing the source and target tasks (defined in Section 4). was tuned with a granularity of 0.1. As a friendly reminder, refers to using only; refers to using only. After finding that a small yields high performance of MULT in the IMDB+MR and SNLI+SICK experiments (thick blue lines in Figures 4a and 4c), we further tuned the from 0.01 to 0.09 with a fine-grained granularity of 0.02.
The results are shown in Figure 4. From the green curves in the 2nd and 4th subplots, we see MULT (with or without INIT) does not improve the accuracy of target tasks (QC and MSRP); the inability to transfer is cross-checked by the INIT method in Section 5. For MR and SICK, on the other hand, transferability of the neural model is also consistently positive (blue curves in Figures 4a and 4c), supporting our conclusion to RQ1 that neural transfer learning in NLP depends largely on how similar in semantics the source and target datasets are.
In MULT+INIT (EHO), the transfer performance of MULT+INIT remains high for different values of . Because the parameters given by INIT have already conveyed sufficient information about the source task, MULT+INIT consistently outperforms non-transferring by a large margin. Its peak performance, however, is not higher than MULT or INIT. In summary, we answer our RQ3 as follows: in our experiments, MULT and INIT are generally comparable; we do not obtain further gain by combining MULT and INIT.
7 Concluding Remarks
In this paper, we addressed the problem of transfer learning in neural network-based NLP applications. We conducted two series of experiments on six datasets, showing that the transferability of neural NLP models depends largely on the semantic relatedness of the source and target tasks, which is different from other domains like image processing. We analyzed the behavior of different neural layers. We also experimented with two transfer methods: parameter initialization (INIT) and multi-task learning (MULT). Besides, we reported two additional studies in Sections 5.3 and 5.4 (not repeated here). Our paper provides insight on the transferability of neural NLP models; the results also help to better understand neural features in general.
How transferable are the conclusions in this paper? We have to concede that empirical studies are subject to a variety of factors (e.g., models, tasks, datasets), and that conclusions may vary in different scenarios. In our paper, we have tested all results on two groups of experiments involving 6 datasets and 2 neural models (CNN and LSTM-RNN). Both models and tasks are widely studied in the literature, and not chosen deliberately. Results are mostly consistent (except Section 5.4). Along with analyzing our own experimental data, we have also collected related results in previous studies, serving as additional evidence in answering our research questions. Therefore, we think the generality of this work is fair and that the conclusions can be generalized to similar scenarios.
Future work. Our work also points out some future directions of research. For example, we would like to analyze the effect of different MULT strategies. More efforts are also needed in developing an effective yet robust method for multi-task learning.
We thank all reviewers for their constructive comments, Sam Bowman for helpful suggestion, and Vicky Li for discussion on the manuscript. This research is supported by the National Basic Research Program of China (the 973 Program) under Grant No. 2015CB352201 and the National Natural Science Foundation of China under Grant Nos. 61232015, 91318301, 61421091, 61225007, and 61502014.
- [Bengio et al.2006] Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. 2006. Greedy layer-wise training of deep networks. In Advances in Neural Information Processing Systems, pages 153–160.
- [Blitzer et al.2006] John Blitzer, Ryan McDonald, and Fernando Pereira. 2006. Domain adaptation with structural correspondence learning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 120–128.
- [Bowman et al.2015] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 632–642.
- [Bromley et al.1993] Jane Bromley, James W Bentz, Léon Bottou, Isabelle Guyon, Yann LeCun, Cliff Moore, Eduard Säckinger, and Roopak Shah. 1993. Signature verification using a “Siamese” time delay neural network. , 7(04):669–688.
[Collobert and Weston2008]
Ronan Collobert and Jason Weston.
A unified architecture for natural language processing: Deep neural
networks with multitask learning.
Proceedings of the 25th International Conference on Machine Learning, pages 160–167.
- [Daumé III et al.2010] Hal Daumé III, Abhishek Kumar, and Avishek Saha. 2010. Frustratingly easy semi-supervised domain adaptation. In Proceedings of the Workshop on Domain Adaptation for Natural Language Processing, pages 53–59.
- [Daumé III2007] Hal Daumé III. 2007. Frustratingly easy domain adaptation. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 256–263.
- [Donahue et al.2014] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. 2014. DeCAF: A deep convolutional activation feature for generic visual recognition. In Proceedings of the 31st International Conference on Machine Learning, pages 647–655.
- [Dong et al.2015] Li Dong, Furu Wei, Shujie Liu, Ming Zhou, and Ke Xu. 2015. A statistical parsing framework for sentiment classification. Computational Linguistics, 41(2):293–336.
- [Foster et al.2010] George Foster, Cyril Goutte, and Roland Kuhn. 2010. Discriminative instance weighting for domain adaptation in statistical machine translation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 451–459.
[Glorot et al.2011]
Xavier Glorot, Antoine Bordes, and Yoshua Bengio.
Domain adaptation for large-scale sentiment classification: A deep learning approach.In Proceedings of the 28th International Conference on Machine Learning, pages 513–520.
- [Hochreiter and Schmidhuber1997] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735–1780.
- [Hu et al.2014] Baotian Hu, Zhengdong Lu, Hang Li, and Qingcai Chen. 2014. Convolutional neural network architectures for matching natural language sentences. In Advances in Neural Information Processing Systems, pages 2042–2050.
- [Jiang and Zhai2007] Jing Jiang and ChengXiang Zhai. 2007. Instance weighting for domain adaptation in NLP. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 264–271.
- [Krizhevsky et al.2012] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. ImageNet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 1097–1105.
- [Liu et al.2016] Yang Liu, Sujian Li, Xiaodong Zhang, and Zhifang Sui. 2016. Implicit discourse relation classification via multi-task neural networks. In Proceedings of the 30th AAAI Conference on Artificial Intelligence, pages 2750–2756.
- [Mikolov et al.2013] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, pages 3111–3119.
- [Peng et al.2015] Hao Peng, Lili Mou, Ge Li, Yunchuan Chen, Yangyang Lu, and Zhi Jin. 2015. A comparative study on regularization strategies for embedding-based neural networks. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 2106–2111.
- [Plank and Moschitti2013] Barbara Plank and Alessandro Moschitti. 2013. Embedding semantic similarity in tree kernels for domain adaptation of relation extraction. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1498–1507.
- [Prettenhofer and Stein2010] Peter Prettenhofer and Benno Stein. 2010. Cross-language text classification using structural correspondence learning. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1118–1127.
- [Socher et al.2011] Richard Socher, Jeffrey Pennington, Eric H Huang, Andrew Y Ng, and Christopher D Manning. 2011. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 151–161.
- [Wang and Zheng2015] Dong Wang and Thomas Fang Zheng. 2015. Transfer learning for speech and language processing. In Proceedings of the Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, pages 1225–1237.
- [Yosinski et al.2014] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. 2014. How transferable are features in deep neural networks? In Advances in Neural Information Processing Systems, pages 3320–3328.
[Zeiler and Fergus2014]
Matthew D Zeiler and Rob Fergus.
Visualizing and understanding convolutional networks.
Proceedings of 13th European Conference on Computer Vision, pages 818–833.
- [Zhao et al.2015] Han Zhao, Zhengdong Lu, and Pascal Poupart. 2015. Self-adaptive hierarchical sentence model. In International Joint Conference on Artificial Intelligence, pages 4069–4076.