acl acl2013 acl2013-81 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Xiaojun Wan
Abstract: The task of review rating prediction can be well addressed by using regression algorithms if there is a reliable training set of reviews with human ratings. In this paper, we aim to investigate a more challenging task of crosslanguage review rating prediction, which makes use of only rated reviews in a source language (e.g. English) to predict the rating scores of unrated reviews in a target language (e.g. German). We propose a new coregression algorithm to address this task by leveraging unlabeled reviews. Evaluation results on several datasets show that our proposed co-regression algorithm can consistently improve the prediction results. 1
Reference: text
sentIndex sentText sentNum sentScore
1 cn Abstract The task of review rating prediction can be well addressed by using regression algorithms if there is a reliable training set of reviews with human ratings. [sent-3, score-1.637]
2 In this paper, we aim to investigate a more challenging task of crosslanguage review rating prediction, which makes use of only rated reviews in a source language (e. [sent-4, score-1.258]
3 English) to predict the rating scores of unrated reviews in a target language (e. [sent-6, score-0.948]
4 We propose a new coregression algorithm to address this task by leveraging unlabeled reviews. [sent-9, score-0.355]
5 Evaluation results on several datasets show that our proposed co-regression algorithm can consistently improve the prediction results. [sent-10, score-0.356]
6 These reviews usually contain valuable information for other people’s reference when they buy the same or similar products. [sent-12, score-0.326]
7 In some applications, it is useful to categorize a review into either positive or negative, but in many real-world scenarios, it is important to provide numerical ratings rather than binary decisions. [sent-13, score-0.341]
8 The task of review rating prediction aims to automatically predict the rating scores of unrated product reviews. [sent-14, score-1.712]
9 It is considered as a finergrained task than the binary sentiment classification task. [sent-15, score-0.197]
10 Review rating prediction has been modeled as a multi-class classification or regression task, and the regression based methods have shown better performance than the multi-class classification based methods in recent studies (Li et al. [sent-16, score-1.451]
11 Traditionally, the review rating prediction task has been investigated in a monolingual setting, which means that the training reviews with human ratings and the test reviews are in the same language. [sent-19, score-1.714]
12 However, a more challenging task is to predict the rating scores of the reviews in a target language (e. [sent-20, score-0.941]
13 German) by making use of the rated reviews in a different source language (e. [sent-22, score-0.361]
14 the rated reviews) for review rating prediction in different languages are imbalanced, it would be very useful to make use of the resources in resource-rich languages to help address the review rating prediction task in resource-poor languages. [sent-27, score-2.088]
15 The task of cross-language review rating prediction can be typically addressed by using machine translation services for review translation, and then applying regression methods based on the monolingual training and test sets. [sent-28, score-1.716]
16 However, due to the poor quality of machine translation, the reviews translated from one language A to another language B are usually very different from the original reviews in language B, because the words or syntax of the translated reviews may be erroneous or non-native. [sent-29, score-0.87]
17 This phenomenon brings great challenges for existing regression algorithms. [sent-30, score-0.298]
18 In this study, we propose a new co-regression algorithm to address the above problem by lever- aging unlabeled reviews in the target language. [sent-31, score-0.528]
19 Our algorithm can leverage both views of the reviews in the source language and the target language to collaboratively determine the confidently predicted ones out of the unlabeled reviews, and then use the selected examples to enlarge the training set. [sent-32, score-0.861]
20 Evaluation results on several datasets show that our proposed coregression algorithm can consistently improve the prediction results. [sent-33, score-0.48]
21 2 Related Work Most previous works on review rating prediction model this problem as a multi-class classification task or a regression task. [sent-34, score-1.391]
22 Various features have been exploited from the review text, including words, patterns, syntactic structure, and semantic topic (Qu et al. [sent-35, score-0.282]
23 c A2s0s1o3ci Aatsiosonc fioartio Cno fmorpu Ctoamtiopnuatalt Lioinngauli Lsitnicgsu,i psatgicess 526–531, ing models, such as SVM, are adopted for rating prediction. [sent-41, score-0.501]
24 (201 1) pro- pose a novel tensor-based learning framework to incorporate reviewer and product information into the text based learner for rating prediction. [sent-43, score-0.602]
25 (2012) study the use of automatic text summaries instead of the full reviews for movie review rating prediction. [sent-45, score-1.127]
26 In addition to predicting the overall rating of a full review, multi-aspect rating prediction has also been investigated (Lu et al. [sent-46, score-1.214]
27 All the above previous works are working under a monolingual setting, and to the best of our knowledge, there exists no previous work on cross-language review rating prediction. [sent-51, score-0.809]
28 It is noteworthy that a few studies have been conducted for the task of cross-lingual sentiment classification or text classification, which aims to make use of labeled data in a language for the binary classification task in a different language (Mihalcea et al. [sent-52, score-0.395]
29 However, the binary classification task is very different from the regression task studied in this paper, and the proposed methods in the above previous works cannot be directly applied. [sent-58, score-0.468]
30 3 Problem Definition and Baseline Approaches Let L={(x1, y1), (xi, yi), (xn, yn)} denote the labeled training set of reviews in a source language (e. [sent-59, score-0.417]
31 English), where xi is the i-th review and yi is its real-valued label, and n is the number of labeled examples; Let T denote the test review set in a different target language (e. [sent-61, score-0.698]
32 German); Then the task of cross-language review rating prediction aims at automatically predicting the rating scores of the reviews in T by leveraging the labeled reviews in L. [sent-63, score-2.23]
33 No labeled reviews in the target language are allowed to be used. [sent-64, score-0.424]
34 The task is a regression problem and it is challenging due to the language gap between the labeled training dataset and the test dataset. [sent-65, score-0.48]
35 Fortunately, due to the development of machine translation techniques, a few online machine translation services can be used for review translation. [sent-66, score-0.375]
36 After review translation, the training reviews and the test reviews are now in the same … … …, … … …, 1 http://translate. [sent-68, score-0.889]
37 In this study, without loss of generality, we adopt the widely used regression SVM (Vapnik 1995; Joachims 1999) implemented in the SVMLight toolkit2 as the basic regressor. [sent-73, score-0.324]
38 The features include all unigrams and bigrams in the review texts, and the value of each feature is simply set to its frequency (TF) in a review. [sent-75, score-0.338]
39 Using features in different languages, we have the following baseline approaches for addressing the cross-language regression problem. [sent-76, score-0.298]
40 REG_S: It conducts regression learning and prediction in the source language. [sent-77, score-0.573]
41 REG_T: It conducts regression learning and prediction in the target language. [sent-78, score-0.607]
42 REG_ST: It conducts regression learning and prediction with all the features in both languages. [sent-79, score-0.573]
43 REG STC: It combines REG S and REG T _S _S _T by averaging their prediction values. [sent-80, score-0.212]
44 However, the above regression methods do not perform very well due to the unsatisfactory machine translation quality and the various language expressions. [sent-81, score-0.357]
45 1 Our Proposed Approach Overview Our basic idea is to make use of some amounts of unlabeled reviews in the target language to improve the regression performance. [sent-84, score-0.757]
46 Considering that the reviews have two views in two languages and inspired by the co-training style algorithms (Blum and Mitchell, 1998; Zhou and Li, 2005), we propose a new co-training style algorithm called co-regression to leverage the unlabeled data in a collaborative way. [sent-85, score-0.552]
47 The proposed co-regression algorithm can make full use of both the features in the source language and the features in the target language in a unified framework similar to (Wan 2009). [sent-86, score-0.148]
48 Each review has two versions in the two languages. [sent-87, score-0.282]
49 The source-language features and the target-language features for each review are considered two redundant views of the review. [sent-88, score-0.313]
50 In the training phase, the co-regression algorithm is applied to learn two regressors in the two languages. [sent-89, score-0.261]
51 In the prediction phase, the two regressors are applied to predict two rating scores of the review. [sent-90, score-0.939]
52 org 527 final rating score of the review is the average of the two rating scores. [sent-93, score-1.284]
53 2 Our Proposed Co-Regression Algorithm In co-training for classification, some confidently classified examples by one classifier are provided for the other classifier, and vice versa. [sent-95, score-0.241]
54 Each of the two classifiers can improve by learning from the newly labeled examples provided by the other classifier. [sent-96, score-0.138]
55 However, in the classification scenario, the confidence value of each prediction can be easily obtained through consulting the classifier. [sent-98, score-0.386]
56 However, in the regression scenario, the confidence value of each prediction is not provided by the regressor. [sent-100, score-0.613]
57 So the key question is how to get the confidence value of each labeled example. [sent-101, score-0.203]
58 In (Zhou and Li, 2005), the assumption is that the most confidently labeled example of a regressor should be with such a property, i. [sent-102, score-0.518]
59 the error of the regressor on the labeled example set (i. [sent-104, score-0.315]
60 the training set) should decrease the most if the most confidently labeled example is utilized. [sent-106, score-0.33]
61 In other words, the confidence value of each labeled example is measured by the decrease of the error (e. [sent-107, score-0.203]
62 mean square error) on the labeled set of the regressor utilizing the information provided by the example. [sent-109, score-0.342]
63 Thus, each example in the unlabeled set is required to be checked by training a new regression model utilizing the example. [sent-110, score-0.46]
64 However, the model training process is usually very time-consuming for many regression algorithms, which significantly limits the use of the work in (Zhou and Li, 2005). [sent-111, score-0.325]
65 Actually, in (Zhou and Li, 2005), only the lazy learning based KNN regressor is adopted. [sent-112, score-0.215]
66 Moreover, the confidence of the labeled examples is assessed based only on the labeled example set (i. [sent-113, score-0.313]
67 the training set), which makes the generalization ability of the regressor not good. [sent-115, score-0.242]
68 Our intuition is that if the two regressors agree on the prediction scores of an example very well, then the example is very confidently labeled. [sent-117, score-0.607]
69 On the contrary, if the prediction scores of an example by the two regressors are very different, we can hardly make a decision whether the example is confidently labeled or not. [sent-118, score-0.707]
70 Therefore, we use the absolute difference value between the prediction scores of the two regressors as the confidence value of a labeled example, and if the example is chosen, its final prediction score is the average of the two prediction scores. [sent-119, score-1.059]
71 Based on this strategy, the confidently labeled examples can be easily and efficiently chosen from the unlabeled set as in the co-training algorithm, and these examples are then added into the labeled set for re-training the two regressors. [sent-120, score-0.614]
72 In the proposed co-regression algorithm, any regression algorithm can be used as the basic regressor to construct Rsource and Rtarand in this study, we adopt the same regression SVM implemented in the SVMLight toolkit with default parameter values. [sent-122, score-0.951]
73 I m can be empirically set acand cording to the total size of the unlabeled set U, and we have I×m≤ |U|. [sent-125, score-0.135]
74 Our proposed co-regression algorithm is much more efficient than the COREG algorithm (Zhou and Li, 2005). [sent-126, score-0.183]
75 If we consider the timeconsuming regression learning process as one get, 528 basic operation and make use of all unlabeled examples in U, the computational complexity of COREG is O(| U|+I). [sent-127, score-0.471]
76 Since |U| is much larger than I, our proposed co-regression algorithm is much more efficient than COREG, and thus our proposed co-regression algorithm is more suitable to be used in applications with a variety of regression algorithms. [sent-129, score-0.526]
77 Moreover, in our proposed co-regression algorithm, the confidence of each prediction is determined collaboratively by two regressors. [sent-130, score-0.366]
78 The selection is not restricted by the training set, and it is very likely that a portion of good examples can be chosen for generalize the regressor towards the test set. [sent-131, score-0.28]
79 It consists of Amazon product reviews for three product categories (i. [sent-133, score-0.422]
80 For each language-category pair there exist three sets of training documents, test documents, and unlabeled documents. [sent-136, score-0.162]
81 The training and test sets comprise 2000 documents each, whereas the number of unlabeled documents varies from 9000 – 170000. [sent-137, score-0.243]
82 The dataset is provided with the rating score between 1 to 5 assigned by users, which can be used for the review rating prediction task. [sent-138, score-1.496]
83 We extracted texts from both the summary field and the text field to represent a review text. [sent-139, score-0.282]
84 We then extracted the rating score as a review’s corresponding real-valued label. [sent-140, score-0.501]
85 html only 8000 unlabeled documents for each product category. [sent-146, score-0.228]
86 For our proposed coregression algorithm, the growth size m is simply set to 50. [sent-151, score-0.198]
87 We implemented the COREG algorithm by replacing the KNN regressor with the regression SVM and the pool size is also set to 50. [sent-152, score-0.582]
88 We can see that on all product categories, the MSE values of our co-regression algorithm and the two component regressors tend to decline over a wide range of I, which means that the selected confidently labeled examples at each iteration are indeed helpful to improve the regressors. [sent-155, score-0.677]
89 Our proposed co-regression algorithm outperforms all the baselines (including COREG) over different iteration members, which verifies the effectiveness of our proposed algorithm. [sent-156, score-0.195]
90 We can also see that the COREG algorithm does not perform well for this cross-language regression task. [sent-157, score-0.367]
91 Overall, our proposed co-regression algorithm can consistently improve the prediction results. [sent-158, score-0.356]
92 6 Conclusion and Future Work In this paper, we study a new task of crosslanguage review rating prediction and propose a new co-regression algorithm to address this task. [sent-159, score-1.15]
93 In future work, we will apply the proposed coregression algorithm to other cross-language or cross-domain regression problems in order to verify its robustness. [sent-160, score-0.536]
94 Comparative experiments on sentiment classification for online product reviews. [sent-178, score-0.236]
95 Beyond the stars: Improving rating predictions using review text content. [sent-182, score-0.783]
96 Integrating collaborative filtering and sentiment analysis: A rating inference approach. [sent-189, score-0.627]
97 Incorporating reviewer and product information for review rating prediction. [sent-193, score-0.884]
98 Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. [sent-220, score-0.6]
99 The bag-of-opinions method for review rating prediction from sparse text patterns. [sent-228, score-0.995]
100 A joint model of text and aspect ratings for sentiment summarization. [sent-247, score-0.196]
wordName wordTfidf (topN-words)
[('rating', 0.501), ('regression', 0.298), ('reviews', 0.29), ('review', 0.282), ('coreg', 0.218), ('regressor', 0.215), ('prediction', 0.212), ('confidently', 0.203), ('regressors', 0.165), ('unlabeled', 0.135), ('coregression', 0.124), ('labeled', 0.1), ('sentiment', 0.099), ('lu', 0.078), ('reg', 0.076), ('zhou', 0.076), ('confidence', 0.075), ('classification', 0.071), ('rated', 0.071), ('algorithm', 0.069), ('banea', 0.068), ('product', 0.066), ('knn', 0.063), ('conducts', 0.063), ('prettenhofer', 0.063), ('unrated', 0.062), ('ratings', 0.059), ('crosslanguage', 0.059), ('svmlight', 0.057), ('ganu', 0.055), ('thumbs', 0.052), ('svm', 0.052), ('mihalcea', 0.051), ('leung', 0.051), ('benjamin', 0.05), ('wan', 0.049), ('blum', 0.045), ('carmen', 0.045), ('li', 0.045), ('proposed', 0.045), ('german', 0.045), ('mse', 0.043), ('saggion', 0.04), ('vapnik', 0.04), ('pang', 0.04), ('stars', 0.039), ('qu', 0.039), ('stein', 0.039), ('aspect', 0.038), ('examples', 0.038), ('titov', 0.037), ('snyder', 0.036), ('buy', 0.036), ('iteration', 0.036), ('chengxiang', 0.035), ('reviewer', 0.035), ('janyce', 0.034), ('collaboratively', 0.034), ('target', 0.034), ('predict', 0.034), ('scenario', 0.033), ('translation', 0.032), ('views', 0.031), ('bin', 0.03), ('shi', 0.03), ('consistently', 0.03), ('growth', 0.029), ('cardie', 0.029), ('claire', 0.029), ('services', 0.029), ('yue', 0.029), ('challenging', 0.028), ('value', 0.028), ('unigrams', 0.028), ('gayatree', 0.027), ('grief', 0.027), ('icdmw', 0.027), ('unsatisfactory', 0.027), ('avrim', 0.027), ('hongning', 0.027), ('lizhen', 0.027), ('nova', 0.027), ('stc', 0.027), ('tsou', 0.027), ('wanxiaoj', 0.027), ('summaries', 0.027), ('collaborative', 0.027), ('lillian', 0.027), ('varies', 0.027), ('square', 0.027), ('documents', 0.027), ('training', 0.027), ('scores', 0.027), ('task', 0.027), ('generality', 0.027), ('movie', 0.027), ('joachims', 0.027), ('monolingual', 0.026), ('rada', 0.026), ('adopt', 0.026)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000001 81 acl-2013-Co-Regression for Cross-Language Review Rating Prediction
Author: Xiaojun Wan
Abstract: The task of review rating prediction can be well addressed by using regression algorithms if there is a reliable training set of reviews with human ratings. In this paper, we aim to investigate a more challenging task of crosslanguage review rating prediction, which makes use of only rated reviews in a source language (e.g. English) to predict the rating scores of unrated reviews in a target language (e.g. German). We propose a new coregression algorithm to address this task by leveraging unlabeled reviews. Evaluation results on several datasets show that our proposed co-regression algorithm can consistently improve the prediction results. 1
2 0.36585605 211 acl-2013-LABR: A Large Scale Arabic Book Reviews Dataset
Author: Mohamed Aly ; Amir Atiya
Abstract: We introduce LABR, the largest sentiment analysis dataset to-date for the Arabic language. It consists of over 63,000 book reviews, each rated on a scale of 1 to 5 stars. We investigate the properties of the the dataset, and present its statistics. We explore using the dataset for two tasks: sentiment polarity classification and rating classification. We provide standard splits of the dataset into training and testing, for both polarity and rating classification, in both balanced and unbalanced settings. We run baseline experiments on the dataset to establish a benchmark.
3 0.12858947 284 acl-2013-Probabilistic Sense Sentiment Similarity through Hidden Emotions
Author: Mitra Mohtarami ; Man Lan ; Chew Lim Tan
Abstract: Sentiment Similarity of word pairs reflects the distance between the words regarding their underlying sentiments. This paper aims to infer the sentiment similarity between word pairs with respect to their senses. To achieve this aim, we propose a probabilistic emotionbased approach that is built on a hidden emotional model. The model aims to predict a vector of basic human emotions for each sense of the words. The resultant emotional vectors are then employed to infer the sentiment similarity of word pairs. We apply the proposed approach to address two main NLP tasks, namely, Indirect yes/no Question Answer Pairs inference and Sentiment Orientation prediction. Extensive experiments demonstrate the effectiveness of the proposed approach.
4 0.12153488 318 acl-2013-Sentiment Relevance
Author: Christian Scheible ; Hinrich Schutze
Abstract: A number of different notions, including subjectivity, have been proposed for distinguishing parts of documents that convey sentiment from those that do not. We propose a new concept, sentiment relevance, to make this distinction and argue that it better reflects the requirements of sentiment analysis systems. We demonstrate experimentally that sentiment relevance and subjectivity are related, but different. Since no large amount of labeled training data for our new notion of sentiment relevance is available, we investigate two semi-supervised methods for creating sentiment relevance classifiers: a distant supervision approach that leverages structured information about the domain of the reviews; and transfer learning on feature representations based on lexical taxonomies that enables knowledge transfer. We show that both methods learn sentiment relevance classifiers that perform well.
5 0.11800383 2 acl-2013-A Bayesian Model for Joint Unsupervised Induction of Sentiment, Aspect and Discourse Representations
Author: Angeliki Lazaridou ; Ivan Titov ; Caroline Sporleder
Abstract: We propose a joint model for unsupervised induction of sentiment, aspect and discourse information and show that by incorporating a notion of latent discourse relations in the model, we improve the prediction accuracy for aspect and sentiment polarity on the sub-sentential level. We deviate from the traditional view of discourse, as we induce types of discourse relations and associated discourse cues relevant to the considered opinion analysis task; consequently, the induced discourse relations play the role of opinion and aspect shifters. The quantitative analysis that we conducted indicated that the integration of a discourse model increased the prediction accuracy results with respect to the discourse-agnostic approach and the qualitative analysis suggests that the induced representations encode a meaningful discourse structure.
6 0.11346823 350 acl-2013-TopicSpam: a Topic-Model based approach for spam detection
7 0.11318471 188 acl-2013-Identifying Sentiment Words Using an Optimization-based Model without Seed Words
8 0.10223159 309 acl-2013-Scaling Semi-supervised Naive Bayes with Feature Marginals
9 0.094120927 342 acl-2013-Text Classification from Positive and Unlabeled Data using Misclassified Data Correction
10 0.093037046 345 acl-2013-The Haves and the Have-Nots: Leveraging Unlabelled Corpora for Sentiment Analysis
11 0.091005772 147 acl-2013-Exploiting Topic based Twitter Sentiment for Stock Prediction
12 0.088088185 168 acl-2013-Generating Recommendation Dialogs by Extracting Information from User Reviews
13 0.086918429 131 acl-2013-Dual Training and Dual Prediction for Polarity Classification
14 0.085747443 379 acl-2013-Utterance-Level Multimodal Sentiment Analysis
15 0.083474807 253 acl-2013-Multilingual Affect Polarity and Valence Prediction in Metaphor-Rich Texts
16 0.083429016 148 acl-2013-Exploring Sentiment in Social Media: Bootstrapping Subjectivity Clues from Multilingual Twitter Streams
17 0.082573816 173 acl-2013-Graph-based Semi-Supervised Model for Joint Chinese Word Segmentation and Part-of-Speech Tagging
18 0.081834249 248 acl-2013-Modelling Annotator Bias with Multi-task Gaussian Processes: An Application to Machine Translation Quality Estimation
19 0.080657132 294 acl-2013-Re-embedding words
20 0.078749031 117 acl-2013-Detecting Turnarounds in Sentiment Analysis: Thwarting
topicId topicWeight
[(0, 0.173), (1, 0.144), (2, -0.018), (3, 0.155), (4, -0.021), (5, -0.086), (6, 0.022), (7, -0.029), (8, -0.033), (9, 0.085), (10, 0.098), (11, 0.0), (12, -0.046), (13, -0.002), (14, -0.044), (15, 0.066), (16, -0.079), (17, 0.013), (18, 0.03), (19, 0.114), (20, 0.04), (21, 0.016), (22, -0.019), (23, 0.049), (24, -0.016), (25, -0.025), (26, 0.001), (27, -0.015), (28, -0.057), (29, -0.056), (30, -0.105), (31, 0.023), (32, 0.09), (33, 0.099), (34, 0.032), (35, 0.049), (36, -0.131), (37, 0.041), (38, -0.027), (39, -0.006), (40, -0.01), (41, -0.077), (42, -0.097), (43, 0.126), (44, 0.068), (45, -0.024), (46, -0.003), (47, -0.0), (48, 0.005), (49, 0.037)]
simIndex simValue paperId paperTitle
same-paper 1 0.94531327 81 acl-2013-Co-Regression for Cross-Language Review Rating Prediction
Author: Xiaojun Wan
Abstract: The task of review rating prediction can be well addressed by using regression algorithms if there is a reliable training set of reviews with human ratings. In this paper, we aim to investigate a more challenging task of crosslanguage review rating prediction, which makes use of only rated reviews in a source language (e.g. English) to predict the rating scores of unrated reviews in a target language (e.g. German). We propose a new coregression algorithm to address this task by leveraging unlabeled reviews. Evaluation results on several datasets show that our proposed co-regression algorithm can consistently improve the prediction results. 1
2 0.84463328 211 acl-2013-LABR: A Large Scale Arabic Book Reviews Dataset
Author: Mohamed Aly ; Amir Atiya
Abstract: We introduce LABR, the largest sentiment analysis dataset to-date for the Arabic language. It consists of over 63,000 book reviews, each rated on a scale of 1 to 5 stars. We investigate the properties of the the dataset, and present its statistics. We explore using the dataset for two tasks: sentiment polarity classification and rating classification. We provide standard splits of the dataset into training and testing, for both polarity and rating classification, in both balanced and unbalanced settings. We run baseline experiments on the dataset to establish a benchmark.
3 0.67221487 117 acl-2013-Detecting Turnarounds in Sentiment Analysis: Thwarting
Author: Ankit Ramteke ; Akshat Malu ; Pushpak Bhattacharyya ; J. Saketha Nath
Abstract: Thwarting and sarcasm are two uncharted territories in sentiment analysis, the former because of the lack of training corpora and the latter because of the enormous amount of world knowledge it demands. In this paper, we propose a working definition of thwarting amenable to machine learning and create a system that detects if the document is thwarted or not. We focus on identifying thwarting in product reviews, especially in the camera domain. An ontology of the camera domain is created. Thwarting is looked upon as the phenomenon of polarity reversal at a higher level of ontology compared to the polarity expressed at the lower level. This notion of thwarting defined with respect to an ontology is novel, to the best of our knowledge. A rule based implementation building upon this idea forms our baseline. We show that machine learning with annotated corpora (thwarted/nonthwarted) is more effective than the rule based system. Because of the skewed distribution of thwarting, we adopt the Areaunder-the-Curve measure of performance. To the best of our knowledge, this is the first attempt at the difficult problem of thwarting detection, which we hope will at Akshat Malu Dept. of Computer Science & Engg., Indian Institute of Technology Bombay, Mumbai, India. akshatmalu@ cse .i itb .ac .in J. Saketha Nath Dept. of Computer Science & Engg., Indian Institute of Technology Bombay, Mumbai, India. s aketh@ cse .i itb .ac .in least provide a baseline system to compare against. 1 Credits The authors thank the lexicographers at Center for Indian Language Technology (CFILT) at IIT Bombay for their support for this work. 2
4 0.671004 131 acl-2013-Dual Training and Dual Prediction for Polarity Classification
Author: Rui Xia ; Tao Wang ; Xuelei Hu ; Shoushan Li ; Chengqing Zong
Abstract: Bag-of-words (BOW) is now the most popular way to model text in machine learning based sentiment classification. However, the performance of such approach sometimes remains rather limited due to some fundamental deficiencies of the BOW model. In this paper, we focus on the polarity shift problem, and propose a novel approach, called dual training and dual prediction (DTDP), to address it. The basic idea of DTDP is to first generate artificial samples that are polarity-opposite to the original samples by polarity reversion, and then leverage both the original and opposite samples for (dual) training and (dual) prediction. Experimental results on four datasets demonstrate the effectiveness of the proposed approach for polarity classification. 1
5 0.66871935 309 acl-2013-Scaling Semi-supervised Naive Bayes with Feature Marginals
Author: Michael Lucas ; Doug Downey
Abstract: Semi-supervised learning (SSL) methods augment standard machine learning (ML) techniques to leverage unlabeled data. SSL techniques are often effective in text classification, where labeled data is scarce but large unlabeled corpora are readily available. However, existing SSL techniques typically require multiple passes over the entirety of the unlabeled data, meaning the techniques are not applicable to large corpora being produced today. In this paper, we show that improving marginal word frequency estimates using unlabeled data can enable semi-supervised text classification that scales to massive unlabeled data sets. We present a novel learning algorithm, which optimizes a Naive Bayes model to accord with statistics calculated from the unlabeled corpus. In experiments with text topic classification and sentiment analysis, we show that our method is both more scalable and more accurate than SSL techniques from previous work.
6 0.64388144 342 acl-2013-Text Classification from Positive and Unlabeled Data using Misclassified Data Correction
7 0.64321822 318 acl-2013-Sentiment Relevance
8 0.63571286 350 acl-2013-TopicSpam: a Topic-Model based approach for spam detection
9 0.61990118 188 acl-2013-Identifying Sentiment Words Using an Optimization-based Model without Seed Words
10 0.55849516 79 acl-2013-Character-to-Character Sentiment Analysis in Shakespeare's Plays
11 0.54607522 63 acl-2013-Automatic detection of deception in child-produced speech using syntactic complexity features
12 0.52971196 284 acl-2013-Probabilistic Sense Sentiment Similarity through Hidden Emotions
13 0.52741307 379 acl-2013-Utterance-Level Multimodal Sentiment Analysis
14 0.50126249 91 acl-2013-Connotation Lexicon: A Dash of Sentiment Beneath the Surface Meaning
15 0.50005287 148 acl-2013-Exploring Sentiment in Social Media: Bootstrapping Subjectivity Clues from Multilingual Twitter Streams
16 0.49781421 345 acl-2013-The Haves and the Have-Nots: Leveraging Unlabelled Corpora for Sentiment Analysis
17 0.49206433 294 acl-2013-Re-embedding words
18 0.47776324 168 acl-2013-Generating Recommendation Dialogs by Extracting Information from User Reviews
19 0.46261147 14 acl-2013-A Novel Classifier Based on Quantum Computation
20 0.45857626 298 acl-2013-Recognizing Rare Social Phenomena in Conversation: Empowerment Detection in Support Group Chatrooms
topicId topicWeight
[(0, 0.067), (4, 0.02), (6, 0.027), (11, 0.065), (15, 0.012), (20, 0.225), (24, 0.041), (26, 0.085), (35, 0.068), (42, 0.045), (48, 0.035), (70, 0.045), (88, 0.041), (90, 0.034), (95, 0.1)]
simIndex simValue paperId paperTitle
1 0.80971777 330 acl-2013-Stem Translation with Affix-Based Rule Selection for Agglutinative Languages
Author: Zhiyang Wang ; Yajuan Lu ; Meng Sun ; Qun Liu
Abstract: Current translation models are mainly designed for languages with limited morphology, which are not readily applicable to agglutinative languages as the difference in the way lexical forms are generated. In this paper, we propose a novel approach for translating agglutinative languages by treating stems and affixes differently. We employ stem as the atomic translation unit to alleviate data spareness. In addition, we associate each stemgranularity translation rule with a distribution of related affixes, and select desirable rules according to the similarity of their affix distributions with given spans to be translated. Experimental results show that our approach significantly improves the translation performance on tasks of translating from three Turkic languages to Chinese.
same-paper 2 0.76651907 81 acl-2013-Co-Regression for Cross-Language Review Rating Prediction
Author: Xiaojun Wan
Abstract: The task of review rating prediction can be well addressed by using regression algorithms if there is a reliable training set of reviews with human ratings. In this paper, we aim to investigate a more challenging task of crosslanguage review rating prediction, which makes use of only rated reviews in a source language (e.g. English) to predict the rating scores of unrated reviews in a target language (e.g. German). We propose a new coregression algorithm to address this task by leveraging unlabeled reviews. Evaluation results on several datasets show that our proposed co-regression algorithm can consistently improve the prediction results. 1
3 0.71627724 7 acl-2013-A Lattice-based Framework for Joint Chinese Word Segmentation, POS Tagging and Parsing
Author: Zhiguo Wang ; Chengqing Zong ; Nianwen Xue
Abstract: For the cascaded task of Chinese word segmentation, POS tagging and parsing, the pipeline approach suffers from error propagation while the joint learning approach suffers from inefficient decoding due to the large combined search space. In this paper, we present a novel lattice-based framework in which a Chinese sentence is first segmented into a word lattice, and then a lattice-based POS tagger and a lattice-based parser are used to process the lattice from two different viewpoints: sequential POS tagging and hierarchical tree building. A strategy is designed to exploit the complementary strengths of the tagger and parser, and encourage them to predict agreed structures. Experimental results on Chinese Treebank show that our lattice-based framework significantly improves the accuracy of the three sub-tasks. 1
4 0.64845729 131 acl-2013-Dual Training and Dual Prediction for Polarity Classification
Author: Rui Xia ; Tao Wang ; Xuelei Hu ; Shoushan Li ; Chengqing Zong
Abstract: Bag-of-words (BOW) is now the most popular way to model text in machine learning based sentiment classification. However, the performance of such approach sometimes remains rather limited due to some fundamental deficiencies of the BOW model. In this paper, we focus on the polarity shift problem, and propose a novel approach, called dual training and dual prediction (DTDP), to address it. The basic idea of DTDP is to first generate artificial samples that are polarity-opposite to the original samples by polarity reversion, and then leverage both the original and opposite samples for (dual) training and (dual) prediction. Experimental results on four datasets demonstrate the effectiveness of the proposed approach for polarity classification. 1
5 0.64506531 174 acl-2013-Graph Propagation for Paraphrasing Out-of-Vocabulary Words in Statistical Machine Translation
Author: Majid Razmara ; Maryam Siahbani ; Reza Haffari ; Anoop Sarkar
Abstract: Out-of-vocabulary (oov) words or phrases still remain a challenge in statistical machine translation especially when a limited amount of parallel text is available for training or when there is a domain shift from training data to test data. In this paper, we propose a novel approach to finding translations for oov words. We induce a lexicon by constructing a graph on source language monolingual text and employ a graph propagation technique in order to find translations for all the source language phrases. Our method differs from previous approaches by adopting a graph propagation approach that takes into account not only one-step (from oov directly to a source language phrase that has a translation) but multi-step paraphrases from oov source language words to other source language phrases and eventually to target language translations. Experimental results show that our graph propagation method significantly improves performance over two strong baselines under intrinsic and extrinsic evaluation metrics.
7 0.64223564 70 acl-2013-Bilingually-Guided Monolingual Dependency Grammar Induction
8 0.63925648 369 acl-2013-Unsupervised Consonant-Vowel Prediction over Hundreds of Languages
9 0.63923317 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
10 0.63911456 173 acl-2013-Graph-based Semi-Supervised Model for Joint Chinese Word Segmentation and Part-of-Speech Tagging
11 0.63857198 196 acl-2013-Improving pairwise coreference models through feature space hierarchy learning
12 0.6379208 117 acl-2013-Detecting Turnarounds in Sentiment Analysis: Thwarting
13 0.63734424 333 acl-2013-Summarization Through Submodularity and Dispersion
14 0.63657987 318 acl-2013-Sentiment Relevance
15 0.63655591 267 acl-2013-PARMA: A Predicate Argument Aligner
16 0.63650578 97 acl-2013-Cross-lingual Projections between Languages from Different Families
17 0.63630307 134 acl-2013-Embedding Semantic Similarity in Tree Kernels for Domain Adaptation of Relation Extraction
18 0.63578922 164 acl-2013-FudanNLP: A Toolkit for Chinese Natural Language Processing
19 0.63575554 316 acl-2013-SenseSpotting: Never let your parallel data tie you to an old domain
20 0.63443911 343 acl-2013-The Effect of Higher-Order Dependency Features in Discriminative Phrase-Structure Parsing