» Download the final program.
» Download the proceedings.
List of accepted full papers with abstracts:
(1) Michael Wiegand and Dietrich Klakow. Bootstrapping Supervised Machine-learning Polarity Classifiers with Rule-based Classification [slides]
Abstract. In this paper, we explore the effectiveness of bootstrapping supervised machine-learning polarity classifiers using the output of domain-independent rule-based classifiers. The benefit of this method is that no labeled training data are required. Still, this method allows to capture in-domain knowledge by training the supervised classifier on in-domain features, such as bag of words. We investigate how important the quality of the rule-based classifier is and what features are useful for the supervised classifier. The former addresses the issue in how far relevant constructions for polarity classification, such as word sense disambiguation, negation modeling, or intensification, are important for this self-training approach. We not only compare how this method relates to conventional semi-supervised learning but also examine how it performs under more difficult settings in which classes are not balanced and mixed reviews are included in the dataset.
(6) Dinko Lambov, Gaël Dias and João Graįa. Multi-view Learning for Text Subjectivity Classification
Abstract. In this paper we consider the problem of building models that have high sentiment classification accuracy across domains. For that purpose, we present and evaluate a method based on co-training using both high-level and low-level features. In particular, we show that multi-view learning combining high-level and low-level features with adapted classifiers can lead to improved results over text subjectivity classification. Our experimental results present accuracy levels across domains of 86.4% combining LDA learning models over high level features and SVM over bigrams.
(14) Silke Scheible. The smallest, cheapest, and best: Superlatives in Opinion Mining
Abstract. This paper introduces superlatives as special indicators for product features in customer reviews. The investigation shows that one type of superlative (called 'ISA') is of particular relevance, as instances in this class tend to contain both a feature string and its associated opinion word. An identification of the components of such superlative comparisons can therefore help to solve two Opinion Mining tasks at once: Feature and Opinion Word Identification. The study further introduces and evaluates a novel tool that can reliably identify such superlatives, and extract from them potential product feature strings and opinion words.
(16) Anne Küppers and Lydia-Mai Ho-Dac. Private State in Public Media: Subjectivity in French Traditional and Online News [slides]
Abstract. This paper reports on ongoing work dealing with the linguistic impact of putting the news on-line. In this framework, we investigate differences in one traditional newspaper and two forms of alternative on-line media with respect to the expression of authorial stance. Our research is based on a comparable large-scale corpus of articles published on the websites of the three respective media and aims at answering the question to what extent the presence of the author varies in the different media. 1) Is it a matter of amount and mode of the author's presence? 2) Is it a matter of lexical choice and diversity? 3) If this were the case, what expressions are used in the respective media? Our endeavour will be a methodological one. We firstly present our data, and thus describe the different news media included in our analysis and the diverse computer aided and manual production steps we performed in order to build up the corpus. Secondly, we outline our working hypotheses that are linked to the chosen types of media and describe the theoretical framework within which they are situated. Thirdly, we present our research method as well as some first results and insights gained throughout the pilot study of our data.
(4) Simon Clematide and Manfred Klenner. Evaluation and Extension of a Polarity Lexicon for German. [slides]
Abstract. We have manually curated a polarity lexicon for German, comprising word polarities and polarity strength values of about 8000 words: nouns, verbs and adjectives. The decisions were primarily carried out using the synsets from GermaNet, a WordNet-like lexical database. In an evaluation on German novels, it turned out that the stock of adjectives was too small. We carried out experiments to automatically learn new subjective adjectives together with their polarity orientation and polarity strength. For this purpose, we applied a corpus-based approach which works with pairs of coordinated adjectives extracted from a large German newspaper corpus. In the context of this work, we evaluated two subtasks in detail. First, how good are we at reproducing the polarity classification -- including our three-level strength measure -- contained in our initial lexicon by machine learning methods. Second, because adding of training material did not improve the results at the expected rate, we evaluated the human inter-coder agreement on polarity classifications in an experiment. The results show that judgements about the strength of polarity do vary considerably between different persons. Given these problems related to the design and automatical augmentation of polarity lexicons, we have successfully experimented with a semi-automatically approach where a list of reliable candidate words (here: adjectives) is generated to ease the manual annotation process.
(5) Muhammad Abdul-mageed and Mohammed Korayem. Automatic Identification of Subjectivity in Morphologically Rich Languages: The Case of Arabic
Abstract. As more user-generated content becomes available online, the need for mining that content becomes increasingly critical. One related area that has been witnessing a flurry of research is that of subjectivity and sentiment analysis. We report our efforts to annotate a corpus of 200 documents from the Penn Arabic Treebank, which is composed of news texts, for subjectivity, along with attempts to automatically classify that data at the sentence level. We investigate the performance of three different machine learning methods on the task with various features and vector settings. We achieve a very high accuracy (i.e., 99.48%) using a support vector machines classifier. We finally briefly discuss issues related to performing text classification on Arabic, a morphologically rich language, and suggest future directions.
(8) Karo Moilanen, Stephen Pulman and Yue Zhang. Packed Feelings and Ordered Sentiments: Sentiment Parsing with Quasi-compositional Polarity Sequencing and Compression
Abstract. Recent solutions proposed for sentence- and phrase-level sentiment analysis have reflected a variety of analytical and computational paradigms that include anything from naīve keyword spotting via machine learning to full-blown logical treatments, either in pure or hybrid forms. As all appear to succeed and fail in different aspects, it is far from evident which paradigm is the optimal one for the task. In this paper, we describe a quasi-compositional sentiment learning and parsing framework that is well-suited for exhaustive, uniform, and principled sentiment classification across words, phrases, and sentences. Using a hybrid approach, we model one fundamental logically defensible compositional sentiment process directly and use supervised learning to account for more complex forms of compositionality learnt from mere flat phrase- and sentence-level sentiment annotations. The proposed framework operates on quasi-compositional sentiment polarity sequences which succinctly capture the sentiment in a given chunk of text without any conventional n-gram features. The results obtained with an initial implementation are highly encouraging and highlight a few surprising observations pertaining to role of sense-level sentiment ambiguity and syntactic information.
(12) Amitava Das and Sivaji Bandyopadhyay. Subjectivity Detection using Genetic Algorithm
Abstract. An opinion classification system on the notion of opinion subjectivity has been reported. The subjectivity classification system uses Genetic-Based Machine Learning (GBML) technique that considers subjectivity as a semantic problem using syntactic simple string co-occurrence rules that involves grammatical construction and linguistic features. Application of machine learning algorithms in NLP generally experiments with combination of various syntactic and semantic linguistic features to identify the most effective feature set. This is viewed as a multi-objective or multi-criteria optimization search problem. The experiments in the present task start with a large set of possible extractable syntactic, semantic and discourse level feature set. The fitness function calculates the accuracy of the subjectivity classifier based on the feature set identified by natural selection through the process of crossover and mutation after each generation. The proposed technique is tested for English and Bengali and for the news, movie review and blog domains. The system evaluation results show precision of 90.22%, and 93.00% respectively for English NEWS and Movie Review corpus and 87.65% and 90.6% for Bengali NEWS and Blog corpus.
(10) Taras Zagibalov and John Carroll. Comparable English-Russian Book Review Corpora for Sentiment Analysis [slides]
Abstract. We present newly-produced comparable corpora of book reviews in English and Russian. The corpora are comparable in terms of domain, style and size. We are using them for cross-lingual experiments in document-level sentiment classification. Quantitative analyses of the corpora and the language differences they exhibit highlight a number of issues that must be considered when developing systems for automatic sentiment classification.
(17) Horacio Saggion, Elena Lloret and Manuel Palomar. Using Text Summaries for Predicting Rating Scales [slides]
Abstract. This paper presents a detailed analysis of a wide range of text summarization approaches within the rating-inference task. This task consists of associating a fine-grained numerical rating to an opinionated document. We collect a small dataset of bank reviews that have been rated from 1 to 5 by real users. Then, we use a SVM classifier to predict the correct rating, employing both the full review and the generated summaries. We mainly suggest three types of summaries (generic, query-focused and sentiment-based) of five compression rates (10% to 50%) to further investigated whether they are useful or not for associating the correct star-rating in comparison to the whole review. Regarding the evaluation, we compute the Mean Squared Error in an attempt to find a specific type of summary and compression rate that performs better over all the rest. The results obtained are very encouraging, and although they are very preliminary to claim a strong tendency for a particular summarization approach, they show that query-focused and sentiment-based summaries may be the most appropriate kinds of summaries for tackling the rating-inference problem.