Transductive Distributional Correspondence Indexing for Cross-Domain Topic Classi cation

Abstract

Obtaining high-quality annotated data for training a classi er for a new domain is often costly. Domain Adaptation (DA) aims at leveraging the annotated data available from a di erent but related source domain in order to deploy a classi cation model for the target domain of interest, thus alleviating the aforementioned costs. To that aim, the learning model is typically given access to a set of unlabelled documents collected from the target domain. These documents might consist of a representative sample of the target distribution, and they could thus be used to infer a general classi cation model for the domain (inductive inference). Alternatively, these documents could be the entire set of documents to be classi ed; this happens when there is only one set of documents we are interested in classifying (transductive inference). Many of the DA methods proposed so far have focused on transductive classi cation by topic, i.e., the task of assigning class labels to a speci c set of documents based on the topics they are about. In this work, we report on new experiments we have conducted in transductive classi cation by topic using Distributional Correspondence Indexing method, a DA method we have recently developed that delivered state-of-the-art results in inductive classi cation by sentiment. The results we have obtained on three popular datasets show DCI to be competitive with the state of the art also in this scenario, and to be superior to all compared methods in many cases.

Keywords

• Transduction
• Cross-Domain Adaptation
• Topic Classi cation
• Distributional Hypothesis

-

As a supervised task, automatic Text Classi cation (TC) is constrained to the availability of high-quality corpora of annotated documents to train a classi er that will then predict the classes of new documents about a given domain of knowledge. In the absence of any such labelled collection for the domain of interest, an additional cost, economical and of time, is to be undertaken in order to collect and annotate the training examples.

Domain Adaptation (DA) is an special case of Transfer Learning (TL)[ 13,14 ] to TC, aimed at reducing, or completely avoiding, such costs, by leveraging on any di erent, but related, source of knowledge for which a training corpus exists already. DA thus challenges one core assumption of machine learning, usually referred to as the iid assumption, according to which the training and test examples are believed to be drawn from the same distribution. Traditionally, two di erent scenarios are considered in DA: (i) cross-domain adaptation [ 2 ], where the source and target domains di er in the topics they are about; and (ii) cross-lingual adaptation [ 15 ], where the source and target domains are expressed in di erent languages, although dealing with the same topics. This article focuses on cross-domain adaptation.

The transference of knowledge is typically attempted by uncovering regularities in examples that are shared across domains. To that aim, a representative (unlabeled) sample from the target distribution is collected and passed to the inference method when learning the decision function. Many of the proposed approaches to cross-domain adaptation so far though, have considered this target sampling to be, at the same time, the test set, i.e., the (only) set of documents one might be interested in classifying (see e.g., [ 4,10,17,18,9,1 ]). This fact leads us distinguish between inductive and transductive cross-domain approaches, depending on the type of inference they carry out3. Accordingly, inductive crossdomain approaches might be viewed as those aiming at deploying a classi cation model that generalizes adequately on the target domain, whereas transductive cross-domain approaches are only requested to deliver an accurate classi cation of the target set at one's disposal [ 16 ].

A general trend one could observe from the related literature of cross-domain adaptation is that the vast majority of the inductive approaches proposed so far have been dedicated to sentiment classi cation (namely, assessing positive or negative labels to opinion-laden texts), while most of the transductive approaches have instead been tested in topic classi cation4. Be that as it may, two welldi erentiated folds of techniques for cross-domain adaptation exist, for which it remains unclear how much e ort it entails for porting one of these methods (say, an inductive one) to the con guration of the other group (say, to the transductive setting), or to more general TL con gurations (e.g., when the source and target tasks are di erent). 3 This distinction is surprisingly overlooked in the related literature though. This is probably so due to the terminology Pan & Yang used in their popular survey [ 13 ], where they categorized as transductive all TL approaches in which the source and target task are the same, but the source and target domains are di erent, while term inductive was instead attributed the opposite meaning, i.e., when the domains are the same, but the source and target tasks di er. 4 This seemingly deliberate partition might rather respond to the characteristics of the most popular benchmark collections available for each problem.

This paper is an extension of our former work in [ 5,11 ], where the Distributional Correspondence Indexing (DCI) method for cross-domain and crosslingual adaptation was proposed. DCI creates words embeddings based in the distributional hypothesis (words with similar meanings tend to co-occur in similar contexts [ 6 ]), and delivered new state-of-the-art results for inductive classication by sentiment recently. We now put to test DCI in a di erent problem setting, i.e., the transductive approach, and report new experiments on a different task, i.e., cross-domain classi cation by topic. Results con rm that our Transductive DCI (hereafter TDCI for short), behaves robustly also in this scenario, delivering comparable classi cation accuracies, and even better in many cases, to state-of-the-art methods, while still being computationally cheap.

The rest of this paper is organized as follows. Section 2 o ers a brief overview of related work. In Section 3 we describe our proposal. Section 4 reports the results of the experiments we have conducted, while Section 5 concludes. 2

Related work

In this section, we brie y review main related methods in the literature of domain adaptation. We will restrict our attention to transductive approaches proposed for topic classi cation. The interested reader can check [ 11 ] for a discussion focusing on inductive methods for sentiment classi cation, and [ 13,14 ] for a more general overview on transfer learning methods.

Transductive Support Vector Machines (TSVMs) for text classi cation where proposed in [ 8 ] as an extension of Suport Vector Machines (SVMs) aiming at minimizing the misclassi cation error in a concrete test set, assuming it accessible when inducing the decision function. Even though it was not particularly designed to deal with DA problems, it has often been reported as a baseline in the related literature. The Co-Clustering approach [ 4 ] uses clusters of words and documents as a bridge to propagate the class structure from the source domain to the target domain. The key idea, is to use the class labels in the source domain as a constraint for the word clusters, that are shared among both domains. The Matrix Tri-factorization [ 18 ] approach follows a somewhat similar assumption, based on the belief that associations between word clusters and classes should remain consistent between the source and target domain. The method thus performs two matrix tri-factorizations, for the source and target domains, in a joint optimization framework subject to sharing the association between word clusters and classes. Topic-bridged Probabilistic Latent Semantic Analysis [ 17 ] is an extension of Probabilistic Latent Semantic Analysis (PLSA) that models the relations between (observed) documents and terms thorough a set of (hidden) latent features, hypothesizing those latent features to be consistent across domains. Along these lines, Topic Correlation Analysis [ 9 ] establishes a distinction between latent features that could be shared between domains, and those that are rather domain speci c. A joint mixture model is rst used to cluster word features into shared and domain-speci c topics. Then, a mapping between the domain-speci c topics from both domains is induced from a correlations analysis, that serves to derive a shared feature space where the transference of supervised knowledge is facilitated. Finally, the Cross Domain Spectral Classi cation [ 10 ] approach formulates the knowledge transference thorough spectral classi cation, via optimizing an objective function aimed at regularizing the supervised information contained in the source domain in order to bring to bear improved consistence with respect to the target domain structure. In [ 1 ] a probabilistic method based on Latent Dirichlet Allocation (LDA) is proposed. The method jointly optimizes the marginal and conditional distributions following a EM algorithm, while also di erentiating between the domain-dependent and domain-independent latent features. 3

Transductive Distributional Correspondence Indexing Loosely speaking, the main challenge one has to face in domain adaptation is to deal with the discrepancy of words relevance that comes about by its particular role in the source domain, and that is not generalizable to the target domain. That is to say, most important words for the source domain, on which the decision surface is likely to hinge upon, are likely not helpful enough in discriminating the positive and negative regions in the target domain.

DCI builds upon (i) the concept of pivots terms [ 3 ], namely, frequent and discriminant words which behave expectedly in a similar way in the source and target domains; and (ii) the distributional hypothesis [ 6 ], which states that terms with similar meanings tend to co-occur in similar contexts. Our idea is to model each term as a word embedding where each dimension quanti es its relative semantic similarity to a xed set of pivots. The expectation is that words with equivalent role across domains might end up lying close to each other in the new embedding space, as they are expected to present similar distributions to the pivots in their respective knowledge domains. Take as an example a classi er by genre (sci- , drama, horror, romantic, ...), that is trained with documents from a source domain of lms, but intended to classify documents from a target domain of books. Note that role equivalences between, e.g., `director'-`writer', `duration'`length', or ` lm'-`book' might be uncovered by inspecting their co-occurrence distribution to some pivots like `plot', `character', or `story', which are expected to be approximately invariant across domains. As a result, the surface decision boundary found for the source domain will likely generalize well in the target domain. DCI is an instantiation of this model that implements a pivot selection strategy (section 3.2), and quanti es the similarity of meaning of two words thorough a Distributional Correspondence Function (DCF, section 3.3). 3.1

Preliminaries

Given a source (S) and a target (T ) domain of documents, with di erent marginal distributions, for which a training set of annotated documents T rS exists exclusively for S, cross-domain classi cation by topic might be formalized as the task of assigning class labels C = fc1; : : : ; cjCjg to target documents in a test set T eT by means of a classi er trained on T rS which is also given access to a sample of (non annotated) documents UT from T (and, optionally, to a sample US from S), where classes in C represent prede ned topics of discussion, such as e.g., \politics", \economics", or \computers".

We will here restrict our attention to the binary case C = fc; cg, that is, deciding whether a document discusses a given topic c, or not. We will also adhere to the aforementioned \transductive setting", in which the sample of target documents given to corresponds also to the unique set of documents we might be interested in classifying, i.e., UT = T eT , and there is not any sample US from the source collection other than the training set T rS . 3.2

Pivot Selection

According to [ 3 ], pivots are frequent and discriminant terms that behave similarly in both the source and target domain. Regarding frequency, and as was done in [ 15 ], we restrict the set of pivot candidates to those which occur in at least = 30 document in the source and target corpora. Following [ 2,15 ], we use the mutual information between the term and the classes fc; cg to asses the degree of discrimination of a given feature in the training set (i.e., exclusively in the source domain). Finally, we apply the cross-consistency heuristic de ned in [ 11 ] which allows the model to be aware of the prevalence5 drift across the source and target domains. 3.3

Distributional Correspondence Functions

DCFs are a family of real-valued functions that quantify the deviation of correspondence between two terms with respect to the expected correspondence due to chance. Di erent interpretations of correspondence could be plugged into the de nition, leading to di erent implementations of DCF. In this work, we will restrict our attention to the cases in which correspondence is measured as the cosine similarity (Eq. 1), the Asymmetric Mutual Information (AMI { Eq. 2), the Pointwise Mutual Information (PMI { Eq. 3), and linear (Eq. 4), as discussed in [ 11 ].

Correspondence between two terms f i and f j in a given domain is measured by comparing their context distribution vectors f i and f j . Context distribution vectors are extracted from the co-occurrence matrix of the domain, and model how a term relates to a set of contexts (e.g., documents).

Cosine(f i; f j ) = hf i; f j i kf ikkf j k ppipj AM I(f i; f j ) = (f i; f j )

X

X x2ffi;fig y2ffj;fjg

P (x; y) P (x; y) log2 P (x)P (y) 5 The prevalence of a term is typically de ned as the proportion of documents in which a term appears in a corpus. (1) (2)

Where pi denotes the prevalence (proportion of occurrences in the total number of contexts) of feature f i, P (x) denotes the probability that feature x occurs in a random context, P (x) is the probability that x does not occur in a random context, and (x; y) is a function that changes the sign when x and y are negatively correlated 6. 3.4

Word Embeddings and Document Representation

The feature representations of DCI might be though as a generalization of CoOccurrence vectors (see, e.g., [ 12 ]), where the co-occurrence metric is any of the DCF, and the context window is set to the document length. Once a set of m pivots P = fp1; p2; : : : ; pmg and a DCF have been selected, each term f in the source and target domains is modeled as an m-dimensional vector

P (f ijf j ) (3) (4) (5) (6) !f = ( (f ; p1); (f ; p2); : : : ; (f ; pm)) where f and pi are the context distribution vectors of the term f and the ith pivot, respectively. Note that, because we are operating in the transductive regime, the context distribution vectors f and pi are taken from the cooccurrence matrix of the training set when modeling the source terms, and from the co-occurrence matrix of the test set when modeling the target terms 7.

Finally, train and test documents are indexed in the embedding space via a weighted sum of all word embeddings of the terms composing the documents. That is, document di is represented as the m-dimensional vector ! d i =

X wij !fj fj2di where wij is the weight of term fj in document di (we used the standard cosinenormalized tf idf ), and !fj is the word embedding for term fj .

Once the training and test matrices have been represented in the embedding space, the classi er is learned. As the classi er we adopt the Transductive SVM [ 8 ], that also takes into account the structure of the test data while modeling the decision function. We used the linear-kernel which have consistently delivered good accuracy in text classi cation so far [ 7 ]. 6 That is, when the true positive rate plus the true negative rate as obtained from the 4-cell contingency table of x and y is lower than 0. 7 In this case, and di erently from [ 5,11 ] we do not apply uni cation to the common features, because during preliminary tests we observed most of the features to appear simultaneously in the source and target domains, causing thus most of the words in the vocabulary to get uni ed. This contradicts the rationale behind the uni cation process, originally proposed to consolidate the representations of shared words across languages, such as proper nouns in cross-lingual adaptation. (7) 4

Experiments

In this section, we report on the experiments we run to test the e ectiveness of our TDCI method in cross-domain topic classi cation.

As the evaluation measure we adopt standard accuracy, i.e., the ratio between the number of correctly labeled documents over the total number of documents sued to the classi er, i.e.,

Acc =

T P + T N

T P + F P + F N + T N where T P , T N , F P , and F N stand for the numbers of true positives, true negatives, false positives, and false negatives, respectively. Note this choice is perfectly valid given that all datasets are approximately balanced with respect to the positive and negative classes,

In order to gain in reproducibility and to facilitate a comparison of performance with other methods, we consider most commonly used benchmarks in the related literature, including the Reuters-21578, SRAA, and 20 Newsgroups collections. Aside from being well-known benchmarks collections in the reign of topic classi cation, their class codes are organized hierarchically, and some representative subsets could thus be taken in order to generate new benchmarks that are well-suited for domain adaptation as well8.

Reuters-21578: is one of the most used collections in TC research. Reuters21578 is a set of 21,578 news stories appeared in the Reuters newswire in 1987. Documents in the collection are assigned to 5 top classes, among which, orgs, people, and places classes have commonly been selected in other works for experimenting in domain adaptation, leading to three datasets, orgs vs people, orgs vs places, and people vs places ; a preprocessed version could be found in9.

SRAA: consists of 73,218 Usenet posts about simulated autos, simulated aviation, real autos, and real aviation, accessible in10. In this dataset, the pairs of classes real vs simulated, and auto vs aviation have been used to instantiate two di erent domain adaptation problems. For example, in real vs simulated, documents about aviation were used as the source domain, while documents about autos constitute the target domain; the binary decision problem consists thus in discerning between real and simulated topics. In a similar vein, auto vs aviation is created, where documents about simulated vehicles act as source domain examples, and documents about real vehicles as the target ones. 8 This procedure consists in taking two top classes, say, A and B, with subclasses A:1 : : : A:x and B:1 : : : B:y, respectively. Then, two disjoint folds are taken for the source (S) and target (T ) sides in each class; e.g., AS = [1 i<lA:i and AT = [l i xA:i, represent the source and target splits for the class A. Finally, the training and test sets are de ned as T rS = AS [ BS and T eT = AT [ BT , where documents in A are labeled as positives, and documents in B are labeled as negatives. 9 http://www.cse.ust.hk/TL/dataset/Reuters.zip 10 http://people.cs.umass.edu/~mccallum/data/sraa.tar.gz

Dataset ISVM TSVM CoCC TPLSA CDSC MTrick TCA PSCC Linear AMI PMI Cosine orgs vs places 0.721 0.740 0.680 0.653 0.682 0.768 0.730 0.742 0.792 0.787 0.797 0.793 orgs vs people 0.737 0.793 0.764 0.763 0.768 0.808 0.792 0.807 0.782 0.810 0.799 0.805 people vs places 0.595 0.614 0.826 0.805 0.798 0.690 0.626 0.690 0.700 0.703 0.676 0.700 real vs simulated 0.737 0.920 0.880 0.889 0.812 - - - 0.962 0.966 0.967 0.958 auto vs aviation 0.799 0.949 0.932 0.947 0.880 - - - 0.974 0.972 0.978 0.976 comp vs sci 0.699 0.842 0.870 0.989 0.902 - 0.891 0.900 0.906 0.879 0.910 0.858 rec vs talk 0.722 0.971 0.965 0.977 0.908 0.950 0.962 0.962 0.978 0.981 0.978 0.959 rec vs sci 0.803 0.945 0.945 0.951 0.876 0.955 0.879 0.955 0.974 0.969 0.974 0.959 sci vs talk 0.783 0.913 0.946 0.962 0.956 0.937 0.940 0.947 0.946 0.943 0.943 0.931 comp vs rec 0.782 0.903 0.958 0.951 0.958 - 0.940 0.958 0.913 0.909 0.914 0.909 comp vs talk 0.955 0.909 0.980 0.977 0.976 - 0.967 0.967 0.932 0.935 0.944 0.913 Reuters (ave.) 0.684 0.716 0.757 0.740 0.749 0.755 0.716 0.746 0.758 0.767 0.757 0.766 SRAA (ave.) 0.768 0.935 0.906 0.918 0.846 - - - 0.968 0.969 0.973 0.967 20News (ave.) 0.799 0.914 0.944 0.968 0.929 - 0.930 0.948 0.942 0.936 0.944 0.922

20 Newsgroups: is a publicly available11 text collection of approximately 20,000 Usenet discussion groups, which are nearly evenly partitioned across 20 di erent newsgroups. Following the common practice in the related literature, we restrict our attention to the 4 most frequent top classes in the dataset (comp, sci, rec, and talk ). The data is then split by their sub-classes for certain pairs of top-classes. We generated 6 di erent domain adaptation problems by following the same classes split as de ned in [ 17 ].

We compare the performance of TDCI12 with the following baselines (discussed in section 2): Co-Clustering (CoCC{[ 4 ]), Topic-bridged PLSA (TPLSA[ 17 ]), Cross Domain Spectral Classi cation (CDSC{[ 10 ]), Matrix Trifactorization (MTrick{[ 18 ]), Topic Correlation Analysis (TCA{[ 9 ]), and Partially Supervised Cross-Collection LDA (PSCC{[ 1 ]). Additionally, we report experiments on a lower bound baseline that simply classi es the target documents using an Inductive SVM13 trained on the source domain without carrying out any sort of adaptation (ISVM{[ 7 ]), and its transductive version (TSVM{[ 8 ]).

Table 1 reports the results of our experiments in terms of accuracy for the Reuters-21578, SRAA, and 20 Newsgroup datasets. Reported score values for CoCC, TPLSA, CDSC, MTrick, TCA, and PSCC where taken from the original papers. Columns Linear, AMI, PMI, and Cosine correspond to our TDCI with di erent DCFs (see Section 3.3); in this experiment, we set the number of pivots to 100, following [ 5,11 ]. 11 http://qwone.com/~jason/20Newsgroups/ 12 The code implementing or method is integrated in JaTeCS, and available in https: //github.com/jatecs/jatecs. A stand-alone version could also be accessed in http: //hlt.isti.cnr.it/dciext/ 13 We used the popular Joachims' implementation in http://svmlight.joachims.org/

Overall, the results of these experiments indicate that TDCI is competitive in transductive cross-domain classi cation by topic. In average, all con gurations of TDCI outperform all baselines in terms of accuracy in Reuters-21578 and SRAA datasets. In 20 Newsgroup TDCI performs comparably in average, without surpassing though the best averaged score obtained by TPLSA. When AMI is used as the DCF, TDCI beats all baselines in 6 out of 11 cases, obtaining two best global results (orgs vs people, and rec vs talk ), and the best average in Reuters-21578. The PMI function also obtained promising results, surpassing all baselines in 5 out of 11 cases, with four best global results, and the best average in SRAA. It is also noticeably that TDCI outperforms all comparison methods in SRAA, in all cases irrespective of the DCF, and by a signi cant margin.

When TDCI did not achieve the best score, its performance could still be considered aligned with respect to the baselines, with the sole exceptions of comp vs rec and comp vs talk cases, where TDCI performed comparatively worse. This could be explained by observing the relative performance of ISVM and TSVM. In both cases the improvements in accuracy brought about by transduction represent the lowest ones in the entire 20 Newsgroup benchmark; note it even degrades signi cantly in comp vs talk. Such observation prompted us to confront TDCI with its inductive version (noted IDCI for consistency). IDCI outperforms TDCI only in these two cases (and performed signi cantly worse in the rest of cases, that we omit for the sake of brevity). More precisely, accuracy scores delivered by IDCI ranged from 0.943 (Cosine) to 0.961 (Linear) in comp vs rec (which is now comparable to the baselines performances), and from 0.983 (PMI) to 0.992 (Linear) in comp vs talk (which surpasses the best accuracy score of 0.980 attributed to CoCC). This sensible variation in performance suggests further investigations are needed in order to shade light on deciding whether it is advisable to maintain an inductive strategy even when the structure of the test set is observable.

We also investigated the impact in performance due to variations in the number of pivots, i.e., the dimensionality of the embedding space. Table 1 shows two representative plots that summarizes well the casuistic we found in our experiments.

The plot for rec vs talk exempli es the most frequent case in our experiments, in which accuracy improves smoothly as more pivots are selected. The case of comp vs talk exempli es a less frequent case in which the performance is somehow unstable. These uctuations seem to depend on the number of pivots, and on the DCF at hand. For example, the performance trend is increasing for AMI and decreasing for the Cosine DCFs; while PMI and Linear delivered competitive results even with only 30 pivots. This result seem to indicate the order in which pivots should be selected could, in some cases, depend also on the DCF under consideration, something we plan to clarify in future research.

REC VS TALK

COMP VS TALK In this article, we have explored the performance e ciency of Distributional Correspondence Indexing, a method originally proposed for cross-domain and cross-lingual inductive classi cation by sentiment, in a di erent problem setting, i.e., by considering the topic classi cation task and the transductive inference. Results show our transductive version, dubbed TDCI, to be comparable and even better in many cases to the state of the art on three extensively used datasets. Our experiments also revealed more investigations are still required in order to automatically determine the optimal number of pivots to select, so as to nd more stable distributional correspondence functions.

References

1. Bao , Y. , Collier , N. , Datta , A. : A partially supervised cross-collection topic model for cross-domain text classi cation . In: Proceedings of the 22nd ACM international conference on Conference on information & knowledge management . pp. 239 { 248 . ACM ( 2013 )

2. Blitzer , J. , Dredze , M. , Pereira , F. : Biographies , Bollywood, boom -boxes and blenders: Domain adaptation for sentiment classi cation . In: Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics (ACL 2007 ). pp. 440 { 447 . Prague, CZ ( 2007 )

3. Blitzer , J. , McDonald , R. , Pereira , F. : Domain adaptation with structural correspondence learning . In: Proceedings of the 4th Conference on Empirical Methods in Natural Language Processing (EMNLP 2006 ). pp. 120 { 128 . Sydney , AU ( 2006 )

4. Dai , W. , Xue , G.R. , Yang , Q. , Yu , Y. : Co-clustering based classi cation for outof-domain documents . In: Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining . pp. 210 { 219 . ACM ( 2007 )

5. Esuli , A. , Moreo Fernandez , A. : Distributional correspondence indexing for crosslanguage text categorization . In: Proceedings of the 37th European Conference on Information Retrieval (ECIR 2015 ). pp. 104 { 109 . Wien , AT ( 2015 )

6. Harris , Z.S.: Distributional structure . Word 10 ( 23 ), 146 { 162 ( 1954 )

7. Joachims , T. : Text categorization with support vector machines: Learning with many relevant features . In: Proceedings of the 10th European Conference on Machine Learning (ECML 1998 ). pp. 137 { 142 . Chemnitz, DE ( 1998 )

8. Joachims , T. : Transductive inference for text classi cation using support vector machines . In: Proceedings of the 16th International Conference on Machine Learning (ICML 1999 ). pp. 200 { 209 . Bled , SL ( 1999 )

9. Li , L. , Jin , X. , Long , M. : Topic correlation analysis for cross-domain text classication . In: Proceedings of the 26th AAAI Conference on Arti cial Intelligence (AAAI 2012 ). pp. 998 { 1004 . Toronto, CA ( 2012 )

10. Ling , X. , Dai , W. , Xue , G.R. , Yang , Q. , Yu , Y. : Spectral-domain transfer learning . In: Proceedings of the 14th ACM International Conference on Knowledge Discovery and Data Mining (KDD 2008 ). pp. 488 { 496 . Las Vegas , US ( 2008 )

11. Moreo Fernandez , A. , Esuli , A. , Sebastiani , F. : Distributional correspondence indexing for cross-lingual and cross-domain sentiment classi cation . Journal of Arti cial Intelligence Research 55 , 131 { 163 ( 2016 )

12. Niwa , Y. , Nitta , Y. : Co-occurrence vectors from corpora vs. distance vectors from dictionaries . In: Proceedings of the 15th conference on Computational linguisticsVolume 1 . pp. 304 { 309 . Association for Computational Linguistics ( 1994 )

13. Pan , S.J. , Yang , Q. : A survey on transfer learning . IEEE Transactions on Knowledge and Data Engineering 22 ( 10 ), 1345 { 1359 ( 2010 )

14. Pan , W. , Zhong , E. , Yang , Q. : Transfer learning for text mining . In: Aggarwal, C.C. , Zhai , C. (eds.) Mining Text Data , pp. 223 { 258 . Springer, Heidelberg, DE ( 2012 )

15. Prettenhofer , P. , Stein , B. : Cross-lingual adaptation using structural correspondence learning . ACM Transactions on Intelligent Systems and Technology 3 ( 1 ), Article 13 ( 2011 )

16. Vapnik , V.: Statistical Learning Theory . Wiley, New York, US ( 1998 )

17. Xue , G.R. , Dai , W. , Yang , Q. , Yu , Y. : Topic-bridged PLSA for cross-domain text classi cation . In: Proceedings of the 31st ACM International Conference on Research and Development in Information Retrieval (SIGIR 2008 ). pp. 627 { 634 . Singapore , SN ( 2008 )

18. Zhuang , F. , Luo , P. , Xiong , H. , He , Q. , Xiong , Y. , Shi , Z. : Exploiting associations between word clusters and document classes for cross-domain text categorization . Statistical Analysis and Data Mining 4 ( 1 ), 100 { 114 ( 2011 )