{
  "document_id": "lb_qasper_c92d96ed55bc5dcd92f963d3c5d26e52661b74e71090f24a",
  "family": "public_qasper",
  "split": "test",
  "text": "Introduction\nIn practice, it is often difficult and costly to annotate sufficient training data for diverse application domains on-the-fly. We may have sufficient labeled data in an existing domain (called the source domain), but very few or no labeled data in a new domain (called the target domain). This issue has motivated research on cross-domain sentiment classification, where knowledge in the source domain is transferred to the target domain in order to alleviate the required labeling effort.\nOne key challenge of domain adaptation is that data in the source and target domains are drawn from different distributions. Thus, adaptation performance will decline with an increase in distribution difference. Specifically, in sentiment analysis, reviews of different products have different vocabulary. For instance, restaurants reviews would contain opinion words such as “tender”, “tasty”, or “undercooked” and movie reviews would contain “thrilling”, “horrific”, or “hilarious”. The intersection between these two sets of opinion words could be small which makes domain adaptation difficult.\nSeveral techniques have been proposed for addressing the problem of domain shifting. The aim is to bridge the source and target domains by learning domain-invariant feature representations so that a classifier trained on a source domain can be adapted to another target domain. In cross-domain sentiment classification, many works BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 utilize a key intuition that domain-specific features could be aligned with the help of domain-invariant features (pivot features). For instance, “hilarious” and “tasty” could be aligned as both of them are relevant to “good”.\nDespite their promising results, these works share two major limitations. First, they highly depend on the heuristic selection of pivot features, which may be sensitive to different applications. Thus the learned new representations may not effectively reduce the domain difference. Furthermore, these works only utilize the unlabeled target data for representation learning while the sentiment classifier was solely trained on the source domain. There have not been many studies on exploiting unlabeled target data for refining the classifier, even though it may contain beneficial information. How to effectively leverage unlabeled target data still remains an important challenge for domain adaptation.\nIn this work, we argue that the information from unlabeled target data is beneficial for domain adaptation and we propose a novel Domain Adaptive Semi-supervised learning framework (DAS) to better exploit it. Our main intuition is to treat the problem as a semi-supervised learning task by considering target instances as unlabeled data, assuming the domain distance can be effectively reduced through domain-invariant representation learning. Specifically, the proposed approach jointly performs feature adaptation and semi-supervised learning in a multi-task learning setting. For feature adaptation, it explicitly minimizes the distance between the encoded representations of the two domains. On this basis, two semi-supervised regularizations – entropy minimization and self-ensemble bootstrapping – are jointly employed to exploit unlabeled target data for classifier refinement.\nWe evaluate our method rigorously under multiple experimental settings by taking label distribution and corpus size into consideration. The results show that our model is able to obtain significant improvements over strong baselines. We also demonstrate through a series of analysis that the proposed method benefits greatly from incorporating unlabeled target data via semi-supervised learning, which is consistent with our motivation. Our datasets and source code can be obtained from https://github.com/ruidan/DAS.\nRelated Work\nDomain Adaptation: The majority of feature adaptation methods for sentiment analysis rely on a key intuition that even though certain opinion words are completely distinct for each domain, they can be aligned if they have high correlation with some domain-invariant opinion words (pivot words) such as “excellent” or “terrible”. Blitzer et al. ( BIBREF0 ) proposed a method based on structural correspondence learning (SCL), which uses pivot feature prediction to induce a projected feature space that works well for both the source and the target domains. The pivot words are selected in a way to cover common domain-invariant opinion words. Subsequent research aims to better align the domain-specific words BIBREF1 , BIBREF5 , BIBREF3 such that the domain discrepancy could be reduced. More recently, Yu and Jiang ( BIBREF4 ) borrow the idea of pivot feature prediction from SCL and extend it to a neural network-based solution with auxiliary tasks. In their experiment, substantial improvement over SCL has been observed due to the use of real-valued word embeddings. Unsupervised representation learning with deep neural networks (DNN) such as denoising autoencoders has also been explored for feature adaptation BIBREF6 , BIBREF7 , BIBREF8 . It has been shown that DNNs could learn transferable representations that disentangle the underlying factors of variation behind data samples.\nAlthough the aforementioned methods aim to reduce the domain discrepancy, they do not explicitly minimize the distance between distributions, and some of them highly rely on the selection of pivot features. In our method, we formally construct an objective for this purpose. Similar ideas have been explored in many computer vision problems, where the representations of the underlying domains are encouraged to be similar through explicit objectives BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 such as maximum mean discrepancy (MMD) BIBREF14 . In NLP tasks, Li et al. ( BIBREF15 ) and Chen et al. ( BIBREF16 ) both proposed using adversarial training framework for reducing domain difference. In their model, a sub-network is added as a domain discriminator while deep features are learned to confuse the discriminator. The feature adaptation component in our model shares similar intuition with MMD and adversary training. We will show a detailed comparison with them in our experiments.\nSemi-supervised Learning: We attempt to treat domain adaptation as a semi-supervised learning task by considering the target instances as unlabeled data. Some efforts have been initiated on transfer learning from unlabeled data BIBREF17 , BIBREF18 , BIBREF19 . In our model, we reduce the domain discrepancy by feature adaptation, and thereafter adopt semi-supervised learning techniques to learn from unlabeled data. Primarily motivated by BIBREF20 and BIBREF21 , we employed entropy minimization and self-ensemble bootstrapping as regularizations to incorporate unlabeled data. Our experimental results show that both methods are effective when jointly trained with the feature adaptation objective, which confirms to our motivation.\nNotations and Model Overview\nWe conduct most of our experiments under an unsupervised domain adaptation setting, where we have no labeled data from the target domain. Consider two sets INLINEFORM0 and INLINEFORM1 . INLINEFORM2 is from the source domain with INLINEFORM3 labeled examples, where INLINEFORM4 is a one-hot vector representation of sentiment label and INLINEFORM5 denotes the number of classes. INLINEFORM6 is from the target domain with INLINEFORM7 unlabeled examples. INLINEFORM8 denotes the total number of training documents including both labeled and unlabeled. We aim to learn a sentiment classifier from INLINEFORM13 and INLINEFORM14 such that the classifier would work well on the target domain. We also present some results under a setting where we assume that a small number of labeled target examples are available (see Figure FIGREF27 ).\nFor the proposed model, we denote INLINEFORM0 parameterized by INLINEFORM1 as a neural-based feature encoder that maps documents from both domains to a shared feature space, and INLINEFORM2 parameterized by INLINEFORM3 as a fully connected layer with softmax activation serving as the sentiment classifier. We aim to learn feature representations that are domain-invariant and at the same time discriminative on both domains, thus we simultaneously consider three factors in our objective: (1) minimize the classification error on the labeled source examples; (2) minimize the domain discrepancy; and (3) leverage unlabeled data via semi-supervised learning.\nSuppose we already have the encoded features of documents INLINEFORM0 (see Section SECREF10 ), the objective function for purpose (1) is thus the cross entropy loss on the labeled source examples DISPLAYFORM0\nwhere INLINEFORM0 denotes the predicted label distribution. In the following subsections, we will explain how to perform feature adaptation and domain adaptive semi-supervised learning in details for purpose (2) and (3) respectively.\nFeature Adaptation\nUnlike prior works BIBREF0 , BIBREF4 , our method does not attempt to align domain-specific words through pivot words. In our preliminary experiments, we found that word embeddings pre-trained on a large corpus are able to adequately capture this information. As we will later show in our experiments, even without adaptation, a naive neural network classifier with pre-trained word embeddings can already achieve reasonably good results.\nWe attempt to explicitly minimize the distance between the source and target feature representations ( INLINEFORM0 and INLINEFORM1 ). A few methods from literature can be applied such as Maximum Mean Discrepancy (MMD) BIBREF14 or adversary training BIBREF15 , BIBREF16 . The main idea of MMD is to estimate the distance between two distributions as the distance between sample means of the projected embeddings in Hilbert space. MMD is implicitly computed through a characteristic kernel, which is used to ensure that the sample mean is injective, leading to the MMD being zero if and only if the distributions are identical. In our implementation, we skip the mapping procedure induced by a characteristic kernel for simplifying the computation and learning. We simply estimate the distribution distance as the distance between the sample means in the current embedding space. Although this approximation cannot preserve all statistical features of the underlying distributions, we find it performs comparably to MMD on our problem. The following equations formally describe the feature adaptation loss INLINEFORM2 : DISPLAYFORM0\nINLINEFORM0 normalization is applied on the mean representations INLINEFORM1 and INLINEFORM2 , rescaling the vectors such that all entries sum to 1. We adopt a symmetric version of KL divergence BIBREF12 as the distance function. Given two distribution vectors INLINEFORM3 , INLINEFORM4 .\nDomain Adaptive Semi-supervised Learning (DAS)\nWe attempt to exploit the information in target data through semi-supervised learning objectives, which are jointly trained with INLINEFORM0 and INLINEFORM1 . Normally, to incorporate target data, we can minimize the cross entropy loss between the true label distributions INLINEFORM2 and the predicted label distributions INLINEFORM3 over target samples. The challenge here is that INLINEFORM4 is unknown, and thus we attempt to estimate it via semi-supervised learning. We use entropy minimization and bootstrapping for this purpose. We will later show in our experiments that both methods are effective, and jointly employing them overall yields the best results.\nEntropy Minimization: In this method, INLINEFORM0 is estimated as the predicted label distribution INLINEFORM1 , which is a function of INLINEFORM2 and INLINEFORM3 . The loss can thus be written as DISPLAYFORM0\nAssume the domain discrepancy can be effectively reduced through feature adaptation, by minimizing the entropy penalty, training of the classifier is influenced by the unlabeled target data and will generally maximize the margins between the target examples and the decision boundaries, increasing the prediction confidence on the target domain.\nSelf-ensemble Bootstrapping: Another way to estimate INLINEFORM0 corresponds to bootstrapping. The idea is to estimate the unknown labels as the predictions of the model learned from the previous round of training. Bootstrapping has been explored for domain adaptation in previous works BIBREF18 , BIBREF19 . However, in their methods, domain discrepancy was not explicitly minimized via feature adaptation. Applying bootstrapping or other semi-supervised learning techniques in this case may worsen the results as the classifier can perform quite bad on the target data.\n[t] Pseudocode for training DAS INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 INLINEFORM4 = ensembling momentum, INLINEFORM5 INLINEFORM6 = weight ramp-up function INLINEFORM7 INLINEFORM8 INLINEFORM9 each minibatch INLINEFORM10 , INLINEFORM11 , INLINEFORM12 in\nINLINEFORM0 , INLINEFORM1 , INLINEFORM2 compute loss INLINEFORM3 on INLINEFORM4 compute loss INLINEFORM5 on INLINEFORM6 compute loss INLINEFORM7 on INLINEFORM8 compute loss INLINEFORM9 on INLINEFORM10 INLINEFORM11\nupdate network parameters INLINEFORM0 , for INLINEFORM1 INLINEFORM2 INLINEFORM3\nInspired by the ensembling method proposed in BIBREF21 , we estimate INLINEFORM0 by forming ensemble predictions of labels during training, using the outputs on different training epochs. The loss is formulated as follows: DISPLAYFORM0\nwhere INLINEFORM0 denotes the estimated labels computed on the ensemble predictions from different epochs. The loss is applied on all documents. It serves for bootstrapping on the unlabeled target data, and it also serves as a regularization that encourages the network predictions to be consistent in different training epochs. INLINEFORM1 is jointly trained with INLINEFORM2 , INLINEFORM3 , and INLINEFORM4 . Algorithm SECREF6 illustrates the overall training process of the proposed domain adaptive semi-supervised learning (DAS) framework.\nIn Algorithm SECREF6 , INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are weights to balance the effects of INLINEFORM3 , INLINEFORM4 , and INLINEFORM5 respectively. INLINEFORM6 and INLINEFORM7 are constant hyper-parameters. We set INLINEFORM8 as a Gaussian curve to ramp up the weight from 0 to INLINEFORM9 . This is to ensure the ramp-up of the bootstrapping loss component is slow enough in the beginning of the training. After each training epoch, we compute INLINEFORM10 which denotes the predictions made by the network in current epoch, and then the ensemble prediction INLINEFORM11 is updated as a weighted average of the outputs from previous epochs and the current epoch, with recent epochs having larger weight. For generating estimated labels INLINEFORM12 , INLINEFORM13 is converted to a one-hot vector where the entry with the maximum value is set to one and other entries are set to zeros. The self-ensemble bootstrapping is a generalized version of bootstrappings that only use the outputs from the previous round of training BIBREF18 , BIBREF19 . The ensemble prediction is likely to be closer to the correct, unknown labels of the target data.\nCNN Encoder Implementation\nWe have left the feature encoder INLINEFORM0 unspecified, for which, a few options can be considered. In our implementation, we adopt a one-layer CNN structure from previous works BIBREF22 , BIBREF4 , as it has been demonstrated to work well for sentiment classification tasks. Given a review document INLINEFORM1 consisting of INLINEFORM2 words, we begin by associating each word with a continuous word embedding BIBREF23 INLINEFORM3 from an embedding matrix INLINEFORM4 , where INLINEFORM5 is the vocabulary size and INLINEFORM6 is the embedding dimension. INLINEFORM7 is jointly updated with other network parameters during training. Given a window of dense word embeddings INLINEFORM8 , the convolution layer first concatenates these vectors to form a vector INLINEFORM9 of length INLINEFORM10 and then the output vector is computed by Equation ( EQREF11 ): DISPLAYFORM0\nINLINEFORM0 , INLINEFORM1 is the parameter set of the encoder INLINEFORM2 and is shared across all windows of the sequence. INLINEFORM3 is an element-wise non-linear activation function. The convolution operation can capture local contextual dependencies of the input sequence and the extracted feature vectors are similar to INLINEFORM4 -grams. After the convolution operation is applied to the whole sequence, we obtain a list of hidden vectors INLINEFORM5 . A max-over-time pooling layer is applied to obtain the final vector representation INLINEFORM6 of the input document.\nDatasets and Experimental Settings\nExisting benchmark datasets such as the Amazon benchmark BIBREF0 typically remove reviews with neutral labels in both domains. This is problematic as the label information of the target domain is not accessible in an unsupervised domain adaptation setting. Furthermore, removing neutral instances may bias the dataset favorably for max-margin-based algorithms like ours, since the resulting dataset has all uncertain labels removed, leaving only high confidence examples. Therefore, we construct new datasets by ourselves. The results on the original Amazon benchmark is qualitatively similar, and we present them in Appendix SECREF6 for completeness since most of previous works reported results on it.\nSmall-scale datasets: Our new dataset was derived from the large-scale Amazon datasets released by McAuley et al. ( BIBREF24 ). It contains four domains: Book (BK), Electronics (E), Beauty (BT), and Music (M). Each domain contains two datasets. Set 1 contains 6000 instances with exactly balanced class labels, and set 2 contains 6000 instances that are randomly sampled from the large dataset, preserving the original label distribution, which we believe better reflects the label distribution in real life. The examples in these two sets do not overlap. Detailed statistics of the generated datasets are given in Table TABREF9 .\nIn all our experiments on the small-scale datasets, we use set 1 of the source domain as the only source with sentiment label information during training, and we evaluate the trained model on set 1 of the target domain. Since we cannot control the label distribution of unlabeled data during training, we consider two different settings:\nSetting (1): Only set 1 of the target domain is used as the unlabeled set. This tells us how the method performs in a condition when the target domain has a close-to-balanced label distribution. As we also evaluate on set 1 of the target domain, this is also considered as a transductive setting.\nSetting (2): Set 2 from both the source and target domains are used as unlabeled sets. Since set 2 is directly sampled from millions of reviews, it better reflects real-life sentiment distribution.\nLarge-scale datasets: We further conduct experiments on four much larger datasets: IMDB (I), Yelp2014 (Y), Cell Phone (C), and Baby (B). IMDB and Yelp2014 were previously used in BIBREF25 , BIBREF26 . Cell phone and Baby are from the large-scale Amazon dataset BIBREF24 , BIBREF27 . Detailed statistics are summarized in Table TABREF9 . We keep all reviews in the original datasets and consider a transductive setting where all target examples are used for both training (without label information) and evaluation. We perform sampling to balance the classes of labeled source data in each minibatch INLINEFORM3 during training.\nSelection of Development Set\nIdeally, the development set should be drawn from the same distribution as the test set. However, under the unsupervised domain adaptation setting, we do not have any labeled target data at training phase which could be used as development set. In all of our experiments, for each pair of domains, we instead sample 1000 examples from the training set of the source domain as development set. We train the network for a fixed number of epochs, and the model with the minimum classification error on this development set is saved for evaluation. This approach works well on most of the problems since the target domain is supposed to behave like the source domain if the domain difference is effectively reduced.\nAnother problem is how to select the values for hyper-parameters. If we tune INLINEFORM0 and INLINEFORM1 directly on the development set from the source domain, most likely both of them will be set to 0, as unlabeled target data is not helpful for improving in-domain accuracy of the source domain. Other neural network models also have the same problem for hyper-parameter tuning. Therefore, our strategy is to use the development set from the target domain to optimize INLINEFORM2 and INLINEFORM3 for one problem (e.g., we only do this on E INLINEFORM4 BK), and fix their values on the other problems. This setting assumes that we have at least two labeled domains such that we can optimize the hyper-parameters, and then we fix them for other new unlabeled domains to transfer to.\nTraining Details and Hyper-parameters\nWe initialize word embeddings using the 300-dimension GloVe vectors supplied by Pennington et al., ( BIBREF28 ), which were trained on 840 billion tokens from the Common Crawl. For each pair of domains, the vocabulary consists of the top 10000 most frequent words. For words in the vocabulary but not present in the pre-trained embeddings, we randomly initialize them.\nWe set hyper-parameters of the CNN encoder following previous works BIBREF22 , BIBREF4 without specific tuning on our datasets. The window size is set to 3 and the size of the hidden layer is set to 300. The nonlinear activation function is Relu. For regularization, we also follow their settings and employ dropout with probability set to 0.5 on INLINEFORM0 before feeding it to the output layer INLINEFORM1 , and constrain the INLINEFORM2 -norm of the weight vector INLINEFORM3 , setting its max norm to 3.\nOn the small-scale datasets and the Aamzon benchmark, INLINEFORM0 and INLINEFORM1 are set to 200 and 1, respectively, tuned on the development set of task E INLINEFORM2 BK under setting 1. On the large-scale datasets, INLINEFORM3 and INLINEFORM4 are set to 500 and 0.2, respectively, tuned on I INLINEFORM5 Y. We use a Gaussian curve INLINEFORM6 to ramp up the weight of the bootstrapping loss INLINEFORM7 from 0 to INLINEFORM8 , where INLINEFORM9 denotes the maximum number of training epochs. We train 30 epochs for all experiments. We set INLINEFORM10 to 3 and INLINEFORM11 to 0.5 for all experiments.\nThe batch size is set to 50 on the small-scale datasets and the Amazon benchmark. We increase the batch size to 250 on the large-scale datasets to reduce the number of iterations. RMSProp optimizer with learning rate set to 0.0005 is used for all experiments.\nModels for Comparison\nWe compare with the following baselines:\n(1) Naive: A non-domain-adaptive baseline with bag-of-words representations and SVM classifier trained on the source domain.\n(2) mSDA BIBREF7 : This is the state-of-the-art method based on discrete input features. Top 1000 bag-of-words features are kept as pivot features. We set the number of stacked layers to 3 and the corruption probability to 0.5.\n(3) NaiveNN: This is a non-domain-adaptive CNN trained on source domain, which is a variant of our model by setting INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 to zeros.\n(4) AuxNN BIBREF4 : This is a neural model that exploits auxiliary tasks, which has achieved state-of-the-art results on cross-domain sentiment classification. The sentence encoder used in this model is the same as ours.\n(5) ADAN BIBREF16 : This method exploits adversarial training to reduce representation difference between domains. The original paper uses a simple feedforward network as encoder. For fair comparison, we replace it with our CNN-based encoder. We train 5 iterations on the discriminator per iteration on the encoder and sentiment classifier as suggested in their paper.\n(6) MMD: MMD has been widely used for minimizing domain discrepancy on images. In those works BIBREF9 , BIBREF13 , variants of deep CNNs are used for encoding images and the MMDs of multiple layers are jointly minimized. In NLP, adding more layers of CNNs may not be very helpful and thus those models from image-related tasks can not be directly applied to our problem. To compare with MMD-based method, we train a model that jointly minimize the classification loss INLINEFORM0 on the source domain and MMD between INLINEFORM1 and INLINEFORM2 . For computing MMD, we use a Gaussian RBF which is a common choice for characteristic kernel.\nIn addition to the above baselines, we also show results of different variants of our model. DAS as shown in Algorithm SECREF6 denotes our full model. DAS-EM denotes the model with only entropy minimization for semi-supervised learning (set INLINEFORM0 ). DAS-SE denotes the model with only self-ensemble bootstrapping for semi-supervised learning (set INLINEFORM1 ). FANN (feature-adaptation neural network) denotes the model without semi-supervised learning performed (set both INLINEFORM2 and INLINEFORM3 to zeros).\nMain Results\nFigure FIGREF17 shows the comparison of adaptation results (see Appendix SECREF7 for the exact numerical numbers). We report classification accuracy on the small-scale dataset. For the large-scale dataset, macro-F1 is instead used since the label distribution in the test set is extremely unbalanced. Key observations are summarized as follows. (1) Both DAS-EM and DAS-SE perform better in most cases compared with ADAN, MDD, and FANN, in which only feature adaptation is performed. This demonstrates the effectiveness of the proposed domain adaptive semi-supervised learning framework. DAS-EM is more effective than DAS-SE in most cases, and the full model DAS with both techniques jointly employed overall has the best performance. (2) When comparing the two settings on the small-scale dataset, all domain-adaptive methods generally perform better under setting 1. In setting 1, the target examples are balanced in classes, which can provide more diverse opinion-related features. However, when considering unsupervised domain adaptation, we should not presume the label distribution of the unlabeled data. Thus, it is necessary to conduct experiments using datasets that reflect real-life sentiment distribution as what we did on setting2 and the large-scale dataset. Unfortunately, this is ignored by most of previous works. (3) Word-embeddings are very helpful, as we can see even NaiveNN can substantially outperform mSDA on most tasks.\nTo see the effect of semi-supervised learning alone, we also conduct experiments by setting INLINEFORM0 to eliminate the effect of feature adaptation. Both entropy minimization and bootstrapping perform very badly in this setting. Entropy minimization gives almost random predictions with accuracy below 0.4, and the results of bootstrapping are also much lower compared to NaiveNN. This suggests that the feature adaptation component is essential. Without it, the learned target representations are less meaningful and discriminative. Applying semi-supervised learning in this case is likely to worsen the results.\nFurther Analysis\nIn Figure FIGREF23 , we show the change of accuracy with respect to the percentage of unlabeled data used for training on three particular problems under setting 1. The value at INLINEFORM0 denotes the accuracies of NaiveNN which does not utilize any target data. For DAS, we observe a nonlinear increasing trend where the accuracy quickly improves at the beginning, and then gradually stabilizes. For other methods, this trend is less obvious, and adding more unlabeled data sometimes even worsen the results. This finding again suggests that the proposed approach can better exploit the information from unlabeled data.\nWe also conduct experiments under a setting with a small number of labeled target examples available. Figure FIGREF27 shows the change of accuracy with respect to the number of labeled target examples added for training. We can observe that DAS is still more effective under this setting, while the performance differences to other methods gradually decrease with the increasing number of labeled target examples.\nCNN Filter Analysis\nIn this subsection, we aim to better understand DAS by analyzing sentiment-related CNN filters. To do that, 1) we first select a list of the most related CNN filters for predicting each sentiment label (positive, negative neutral). Those filters can be identified according to the learned weights INLINEFORM0 of the output layer INLINEFORM1 . Higher weight indicates stronger relatedness. 2) Recall that in our implementation, each CNN filter has a window size of 3 with Relu activation. We can thus represent each selected filter as a ranked list of trigrams with highest activation values.\nWe analyze the CNN filters learned by NaiveNN, FANN and DAS respectively on task E INLINEFORM0 BT under setting 1. We focus on E INLINEFORM1 BT for study because electronics and beauty are very different domains and each of them has a diverse set of domain-specific sentiment expressions. For each method, we identify the top 10 most related filters for each sentiment label, and extract the top trigrams of each selected filter on both source and target domains. Since labeled source examples are used for training, we find the filters learned by the three methods capture similar expressions on the source domain, containing both domain-invariant and domain-specific trigrams. On the target domain, DAS captures more target-specific expressions compared to the other two methods. Due to space limitation, we only present a small subset of positive-sentiment-related filters in Table TABREF34 . The complete results are provided in Appendix SECREF8 . From Table TABREF34 , we can observe that the filters learned by NaiveNN are almost unable to capture target-specific sentiment expressions, while FANN is able to capture limited target-specific words such as “clean” and “scent”. The filters learned by DAS are more domain-adaptive, capturing diverse sentiment expressions in the target domain.\nConclusion\nIn this work, we propose DAS, a novel framework that jointly performs feature adaptation and semi-supervised learning. We have demonstrated through multiple experiments that DAS can better leverage unlabeled data, and achieve substantial improvements over baseline methods. We have also shown that feature adaptation is an essential component, without which, semi-supervised learning is not able to function properly. The proposed framework could be potentially adapted to other domain adaptation tasks, which is the focus of our future studies.\nResults on Amazon Benchmark\nMost previous works BIBREF0 , BIBREF1 , BIBREF6 , BIBREF7 , BIBREF29 carried out experiments on the Amazon benchmark released by Blitzer et al. ( BIBREF0 ). The dataset contains 4 different domains: Book (B), DVDs (D), Electronics (E), and Kitchen (K). Following their experimental settings, we consider the binary classification task to predict whether a review is positive or negative on the target domain. Each domain consists of 1000 positive and 1000 negative reviews respectively. We also allow 4000 unlabeled reviews to be used for both the source and the target domains, of which the positive and negative reviews are balanced as well, following the settings in previous works. We construct 12 cross-domain sentiment classification tasks and split the labeled data in each domain into a training set of 1600 reviews and a test set of 400 reviews. The classifier is trained on the training set of the source domain and is evaluated on the test set of the target domain. The comparison results are shown in Table TABREF37 .\nNumerical Results of Figure \nDue to space limitation, we only show results in figures in the paper. All numerical numbers used for plotting Figure FIGREF17 are presented in Table TABREF38 . We can observe that DAS-EM, DAS-SE, and DAS all achieve substantial improvements over baseline methods under different settings.\nCNN Filter Analysis Full Results\nAs mentioned in Section SECREF36 , we conduct CNN filter analysis on NaiveNN, FANN, and DAS. For each method, we identify the top 10 most related filters for positive, negative, neutral sentiment labels respectively, and then represent each selected filter as a ranked list of trigrams with the highest activation values on it. Table TABREF39 , TABREF40 , TABREF41 in the following pages illustrate the trigrams from the target domain (beauty) captured by the selected filters learned on E INLINEFORM0 BT for each method.\nWe can observe that compared to NaiveNN and FANN, DAS is able to capture a more diverse set of relevant sentiment expressions on the target domain for each sentiment label. This observation is consistent with our motivation. Since NaiveNN, FANN and other baseline methods solely train the sentiment classifier on the source domain, the learned encoder is not able to produce discriminative features on the target domain. DAS addresses this problem by refining the classifier on the target domain with semi-supervised learning, and the overall objective forces the encoder to learn feature representations that are not only domain-invariant but also discriminative on both domains.",
  "char_len": 33297,
  "document_hash": "1a1ac0aba08ac80f31ff063766cd1b7b9f43563ae53f956f8c4c9d3f730d7664",
  "questions": [
    {
      "question_id": "q00",
      "qtype": "qasper",
      "question": "What are the source and target domains?"
    }
  ]
}