LEXICON ADAPTATION AS CONSTRAINT OPTIMIZATION
Polarity lexicons have been a valuable resource for sentiment analysis and opin- ion mining. In particular, they have been an essential ingredient for fine-grained sentiment analysis (e.g., Kim and Hovy (2004), Kennedy and Inkpen (2005), Wil- son et al. (2005)). Even though the polarity lexicon plays an important role, it has received relatively less attention in previous research. In most cases, po- larity lexicon construction is discussed only briefly as a preprocessing step for a sentiment analysis task (e.g., Hu and Liu (2004a), Moilanen and Stephen (2007)), but the effect of different alternative polarity lexicons is not explicitly investi- gated. Conversely, research efforts that focus on constructing a general purpose polarity lexicon (e.g., Takamura et al. (2005), Adreevskaia and Bergler (2006), Esuli and Sebastiani (2006), Rao and Ravichandran (2009)) generally evaluate the lexicon in isolation from any potentially relevant NLP task, and it is unclear how the new lexicon might affect end-to-end performance of a concrete NLP application.
fine-grained annotation system provides reliable results. Remus und Hänig (2011) present the Polarity Composition Model (PCM), which is a two-level structure, where evaluations are ana- lyzed on the word- and phrase-level. The authors draw the conclusion that evaluations are mani- fested through negations, positive and negative reinforces and the formal construct within the phrase; the world-level polarity analysis is car- ried out with recourse to the German-language polarity dictionary SentiWS (see also Remus et al., 2010). Lately, Fink et al. (2011) published their fine-grained annotation approach for senti- ment analysis. Thereby, they identify sentiment- bearing sentences to spot sentiment targets and their valence.
On the other hand, there are approaches that exploit machine learning techniques to induce a sentiment extraction model from training data, either in a fully supervised or weakly supervised fashion. Fully supervised systems that train on manually annotated data are commonly used to extract as- pects and subjective phrases (Klinger and Cimiano, 2013a; Klinger and Cimiano, 2013b; Li et al., 2010) or in order to classify the polarity or subjectivity of text (T¨ackstr¨om and McDonald, 2011; Sayeed et al., 2012; Shi and Li, 2011; Pang and Lee, 2004; Wiebe, 2000). In contrast to these fully supervised systems, Turney (2002) for instance proposed a system that is in this sense weakly supervised in that it relies on the two seed words “excellent” and “poor” and textual similarity to induce other “similar” adjectives that express a positive or negative sentiment, respectively. Completely unsupervised approaches have also been applied to the task (Titov and McDonald, 2008).
With the in-depth study of sentiment analysis research, finer-grainedopinion mining, which aims to detect opinions on different review fea- tures as opposed to the whole review level, has been receiving more and more attention in the sentiment analysis research community re- cently. Most of existing approaches rely mainly on the template extraction to identify the ex- plicit relatedness between product feature and opinion terms, which is insufficient to detect the implicit review features and mine the hid- den sentiment association in reviews, which satisfies (1) the review features are not appear explicit in the review sentences; (2) it can be deduced by the opinion words in its context. From an information theoretic point of view, this paper proposed an iterative reinforcement framework based on the improved information bottleneck algorithm to address such problem. More specifically, the approach clusters prod- uct features and opinion words simultaneously and iteratively by fusing both their semantic in- formation and co-occurrence information. The experimental results demonstrate that our ap- proach outperforms the template extraction based approaches.
Several researchers have attempted to determine whether a term is a marker of subjective content and what is its senti- ment orientation (e.g. positive, negative, neutral) (Hatzi- vassiloglou and McKeown, 1997; Turney, 2002). How- ever, the sentiment is conveyed not only by single words or phrases but rather by their combinations or contexts. Therefore, some researchers examine whether a given text has a factual nature or expresses an opinion by means of subjective patterns (Riloff and Wiebe, 2003; Riloff, et al., 2006; Popescu and Etzioni, 2005; Wilson, et al., 2005). Taking advantages of above approaches into account, we introduce a two-step learning method: the first step is to acquire sentiment knowledge (i.e. lexical sentiment ori- entation and negation words), and the second is to train a Na¨ıve Bayes (NB) classifier by subjective patterns which are generated on top of dependency structure with lexical sentiment knowledge.
position transported in a speech as a separate point on scale. The tool WordFish developed by Slapin and Proksch [Slapin and Proksch, 2008] is perfectly suitable for such a task, given that one input document contains exactly one position. In order to scale the positions towards all topics covered in a speech separately, we need to split the speech into documents containing a single topic. This again asks for topic classification. The most straightforward way is to classify units of a speech, i.e. sentences or paragraphs, with its topic. This is an easy task with supervised models using training data. However, we are not aware of an appropriate training set, and hand-coding such a resource set is time-consuming and costly. As an alternative, we propose a bootstrapping approach to create a training data set. Starting from a small set of party manifestos hand-coded on sentence level with their political topic by the Comparative Party Manifesto Project (cf. [Volkens et al., 2015]), we build a system to increase the training data with more, unlabeled manifestos. For this system, we first train state-of-the art classifiers on those manifestos. In a second step, we combine their predictions for sentences of unla- beled manifestos with knowledge about the structure of those documents and about typical sequences topics. A Markov Logic framework assess these features in mutual dependence to find the optimal global topic classifications for all sentences of an unlabeled manifesto. The total set of hand-coded and automatically labeled manifestos is then used in a state-of-the art machine learning model, which is then applied to the speeches to detect its topics.
For binary polarity classification, the polarity that appears more in the text is assigned to the text. For example, if a sentence has more positive words than the negative ones, it is classified as a positive sentence. For fine-grained classifi- cation, the frequency of the opinionated words is also taken into consideration. For example, if a sentence contains 4 positive and 1 negative words, the sentiment strength of the sentence will be +3. This value can also be normalized to have a better and comparative representation of sentence sentiments. These approaches, however, are not accurate; because all positive or negative words are considered the same. For example, the sentences “The song was good.” and “It was an excellent song.” will be assigned the same sentiment, because both “good” and “excellent” have the same label in the dictionary. To overcome this problem, we need a dictionary with more detailed information about opinionated words such that in addition to the polarity of the words, each word is assigned a sentiment degree. Table 2 presents sample words appeared in an opinion dictionary with fine-grained sentiment scores.
Opinionanalysis deals with public opin- ions and trends, but subjective language is highly ambiguous. In this paper, we follow a simple data-driven technique to learn fine-grained opinions. We select an intersection set of Wall Street Jour- nal documents that is included both in the Penn Discourse Tree Bank (PDTB) and in the Multi-Perspective Question Answer- ing (MPQA) corpus. This is done in or- der to explore the usefulness of discourse- level structure to facilitate the extraction of fine-grainedopinion expressions. Here we perform shallow parsing of MPQA ex- pressions with connective based discourse structure, and then also with Named Enti- ties (NE) and some syntax features using conditional random fields; the latter fea- ture set is basically a collection of NEs and a bundle of features that is proved to be useful in a shallow discourse parsing task. We found that both of the feature-sets are useful to improve our baseline at different levels of this fine-grainedopinion expres- sion mining task.
As a final extrinsic evaluation of the system, we evaluated the usefulness of its output in a number of applications. Although there have been several publications detailing the extraction of MPQA-style opinion expressions, as far as we are aware there has been no attempt to use them in an application. In contrast, we show that the opinion expressions as defined by the MPQA corpus may be used to derive machine learning features that are useful in two practical opinion mining tasks; the addition of these features leads to statistically significant improvements in all scenarios we evaluated. First, we develop a system for the extraction of evaluations of product attributes from product reviews (Hu and Liu 2004a, 2004b; Popescu and Etzioni 2005; Titov and McDonald 2008), and we show that the features derived from opinion expressions lead to significant improvement. Secondly, we show that fine-grainedopinion structural information can even be used to build features that improve a coarse-grained sentiment task: document polarity classification of reviews (Pang, Lee, and Vaithyanathan 2002; Pang and Lee 2004).
Recent years have seen a surge of interest in the automatic processing of subjective language. The technologies emerging from this research have ob- vious practical uses, either as stand-alone appli- cations or supporting other NLP tools such as information retrieval or question answering sys- tems. While early efforts in subjectivity analysis focused on coarse-grained tasks such as retriev- ing the subjective documents from a collection, most recent work on this topic has focused on fine- grained tasks such as determining the attitude of a particular person on a particular topic. The devel- opment and evaluation of such systems has been made possible by the release of manually anno- tated resources using fairly fine-grained represen- tations to describe the structure of subjectivity in language, for instance the MPQA corpus (Wiebe et al., 2005).
The motivation for this paper is from Shishir et al. . They proposed a heuristic approach on the ordering of clean-up jobs of the workflow. A fine and course-grained genetic approach for the Data-intensive workflows was proposed to optimize the schedule. Applying both heuristic and genetic algorithms, they reduced the overall cost as well as the execution time of large Data-intensive workflows for Grid Resources. Arun et al. presented an algorithm for scheduling the workflow tasks to the resources taking into account disk-space constraints and attained a feasible solution for Grid environment. An Algorithm proposed by Chauhan et al.  yields schedule based on both the communication cost and computation cost related to tasks. Hence unlike computing field scheduling which is applicable only in case of computation intensive tasks, this new fully decentralized algorithm for Peer to Peer (P2P) grid gives good schedule for tasks; irrespective of the fact whether they are computation intensive or communication intensive in nature. Taura and Andrew  designed a heuristic algorithm that maps data- processing tasks onto heterogeneous resources. This algorithm achieves a good throughput of the whole data-processing pipeline, taking both parallelism (load balance) and communication volume (locality) into account. It performs well both under computation intensive and communication intensive conditions. A decentralized scheduling algorithm for Peer to Peer grid systems proposed by Piyush  optimizes the schedule compared to the conventional approaches. This algorithm takes accurate scheduling decisions depending on both computation cost and communication cost associated with DAG’s subtasks.Cloud computing is an emerging paradigm where traditional resource allocation approaches, inherited from cluster computing and grid computing systems, fail to provide efficient performance. The main reason is that most of cloud applications require availability of communication resources for information exchange between tasks, with databases, or end users. A non-linear programming model to minimize the data retrieval and execution cost of data-intensive workflows in Cloud was formulated by Suraj et al. . In this paper, a comparative study was made between Amazon Cloud Front’s ‘nearest’ single data source selection and the non-linear algorithm. The non-linear algorithm saved three-quarters of total cost. A new communication-aware model for cloud computing applications, called CA-DAG proposed by Kliazovich et al.  overcomes shortcomings of existing approaches using communication awareness. Using this model developing a new scheduling algorithm of improved efficiency can be cracked.
Gh Cloud technology provides the QOS to the user with less response time with heterogeneous systems. To improve the features of this environment, in this work we propose a new heuristic algorithm interference aware group scheduling (Priority Based on Deadline and Size) which increases the throughput of the overall system by providing the immediate response to the jobs based on the priority. Jobs with earliest deadline with short burst time will be executed by the resources first. Short jobs are benefited with this method. It also decreases the system overhead by preventing the task migration between the processors and system power is not wasted unnecessarily in context switching between the processes.
Movie reviews were one of the first research do- mains for sentiment analysis as they (i) have the properties of a short message, and (ii) are already manually annotated by the author, as the score generally reflects sentiment polarity. Popular fea- tures for score/sentiment prediction include POS tags, word n-grams, word lemmata, and various context features based on the distance from a topic word. The challenge with movie reviews is that only some of the words are relevant for sentiment analysis. In fact, often the review is just a short narrative of the movie plot. One way to approach the problem is to use a subjectivity classifier (Pang and Lee, 2004), which can be used to filter out ob- jective sentences from the reviews, thus allowing the classifier then to focus on the subjective sen- tences only.
In this paper, we presented a new concept of kernel trace analysis adapted to cloud computing and virtualized sys- tems that can help for the monitoring and tuning of such systems and the development of those technologies. This concept is independent of the kernel tracer and hyper- visor used. By creating a new view in Trace Compass, we showed that it was possible to display an overview of the full hierarchy of the virtualized systems running on a physical host, including VMs and containers. Finally, by adding a new dynamic filter feature to the FVS view, in addition to a permanent filter for any VS, we showed how it is possible to observe the real execution on the host of a virtual machine, one of its virtual CPUs, its processes and its containers.
Table 2 lists the evaluation results for the dif- ferent classifiers. To determine statistical signif- icance of the relative effectiveness of two clas- sifiers we applied a paired t-test at a signifi- cance level of p < 0.01. The classifiers ex- clusively using polarity features have compara- ble accuracy values. While the SVM is show- ing a bias towards classifying segments as nega- tive ML polarity shows the opposite trend. Al- though the accuracy of SVM is slightly higher the relative difference of the accuracy values is not statistically significant. Including neighbor- hood relations increases the effectiveness rela- tive to both non-structure based classifiers signifi- cantly. MLN neigborhood achieves an F-measure of 69.50% for positive segments and 68.52% for negative segments with an overall accuracy of 69.02%. It also significantly outperforms the majority baseline which achieves an accuracy of 51.60%. Contrary to our hypothesis, distinguish- ing between contrast and ncontrast rela- tions did not improve the effectiveness relative to MLN neigborhood. MLN contrast achieves a slightly lower accuracy than MLN neighborhood although the difference is not statistically signif- icant. These results suggest that the correlation of contrast relations and polarity changes is not significant. Furthermore, the number of con- trast relations in product reviews is too small to have a significant impact. Finally, employing a discourse parser as a component of a sentiment analysis poses the problem that misclassifications might as well be caused by erroneous decisions of the component. Figure 2 depicts the accuracy val- ues for the different classifiers on each of the ten cross-validation folds.
Propaganda aims at influencing people’s mindset with the purpose of advancing a spe- cific agenda. Previous work has addressed propaganda detection at the document level, typically labelling all articles from a propa- gandistic news outlet as propaganda. Such noisy gold labels inevitably affect the quality of any learning system trained on them. A further issue with most existing systems is the lack of explainability. To overcome these lim- itations, we propose a novel task: performing fine-grainedanalysis of texts by detecting all fragments that contain propaganda techniques as well as their type. In particular, we cre- ate a corpus of news articles manually anno- tated at the fragment level with eighteen pro- paganda techniques and we propose a suit- able evaluation measure. We further design a novel multi-granularity neural network, and we show that it outperforms several strong BERT-based baselines.
MT-based approaches. The workflow of these approaches is as follows. Starting with a text in the target language to be labeled with coreference, it first must be machine-translated to the source language. A coreference resolver for the source language is then applied on the translated text and, finally, the newly established coreference links are projected back to the target language. Flexibility of this approach lies in the fact that it can be ap- plied in both train and test time, and no linguis- tic processing tools for the target language are re- quired. To the best of our knowledge, this ap- proach has been applied to coreference only twice, by Rahman and Ng (2012) on projection from En- glish to Spanish and Italian, and by Ogrodniczuk (2013) on projection from English to Polish. Corpus-based approaches. In these ap- proaches, a human-translated parallel corpus of the two languages is available and the projection mechanism is applied within this corpus. Coref- erence annotation in the source-language side of the corpus may be both labeled by humans or a coreference system. The target-language side of the corpus then serves as a training dataset for a coreference resolver. This approach thus must be applied in train time and, moreover, it requires a coreference resolver trainable on the target-language data. As a consequence, linguistic processing tools should be available for the target language as most of the resolvers depend on some amount of additional linguistic information. On the other hand, human translation and gold coref- erence annotation, if available, should increase the quality of the projected coreference. This approach has been used to create a coreference resolver by multiple authors, e.g. de Souza and Or˘asan (2011), Martins (2015), Wallin and Nugues (2017), and Nov´ak et al. (2017). However, since the present work employs the corpus-based approach on gold annotations of coreference, we offer more details on works of Postolache et al.
Bar et al  used features from pre-trained CNN  for style classification; however they used a small dataset. Karayev et al  explored hand crafted features vs deep learning features for style classification; however their model extracted CNN features and used various classification algorithms on top of it thus not being end-to-end. Saleh et al  created the Wikiart dataset and used features ranging from low-level to high-level semantic features with metric learning approaches for the task of style, artist and genre classification. The closest to our work was done by Tan et al.  where they fine-tuned a VGG network on the same Wikiart dataset  to improve classification scores. However, our models not only improve the classification scores significantly from all the above works but also work quite well on the task of fine-grained classification where we consider 194 artists labels instead of 23.
4.1.1 Baseline Models. We implemented our baseline (denoted by “Baseline”) inspired by Paraskevopoulos et al.  work. To do so, for each of our cities, we first created a grid structure of squared areas with a side length of 1 km. For each of these defined squared areas, we created a document by concatenating the text of the tweets associated with each area. We then indexed these documents. As a preprocessing step, usernames and hashtags were preserved as tokens, all hyperlinks were removed from tweets, and re-tweets were preserved in the dataset. Then, we retrieved the most content-based similar document (Top-1) for each non-geo- tagged tweet. As the model returns the Top-1 tweet, the longitude and latitude coordinates of the tweet are returned as the predicted location instead of the squared area associated to the post. We investigated several retrieval models to maximise the performance of our baseline. Five different retrieval models were evaluated: Divergence From Randomness (dfr), Language Model with Dirichlet Smoothing (lmd), IDF (idf), TF-IDF (tf idf) and BM25 (bm25) using the Apache Lucene 1 implementation. The difference between our baseline and the work by  are two-fold. First, we removed stop- words  and applied Porter stemming. 2 Second, we also did not consider the time dimension, as described in Section 2.