Abstract: This paper explores the problem of sockpuppet detection in deceptive opinion spam using authorship attribution and verification approaches. Two methods are explored. The first is a feature subsampling scheme that uses the KL-Divergence on stylistic language models of an author to find discriminative features. The second is a transduction scheme, spy induction that leverages the diversity of authors in the unlabeled test set by sending a set of spies (positive samples) from training set to retrieve hidden samples in the unlabeled test set using nearest and farthest neighbors. Experiments using ground truth sockpuppet data show the effectiveness of the proposed schemes.

Abstract: Generic sentiment and emotion lexicons are widely used for the fine-grained analysis of human affects from text. In order to accurately detect affect, there is a need for domain intelligence, that enables understanding of the perceived interpretation of the same words in varied contexts. Recent work has focused on automatically inducing the polarity of given terms in changing contexts. We propose an unsupervised approach for the construction of domain-specific affect lexicons along these lines. The algorithm is seeded with existing standard lexicons and expanded based on context-relevant word associations. Experiments show that our lexicon provides better coverage than standard lexicons on both short as well as long texts, and corresponds well with human-annotated affect values. Our framework outperforms the state-of-the-art generic and domain-specific approaches with a precision of over 70% for the emotion detection task on the SemEval 2007 Affect Corpus.

Abstract: In this paper, we introduce a trie-structured Bayesian model for unsupervised morphological segmentation. In the model, we adopt prior information from different sources. We use neural word embeddings to discover words that are morphologically derived from each other and thereby that are semantically similar. We use letter successor variety counts obtained from tries that are built by neural word embeddings. Our results show that using different information sources such as neural word embeddings and letter successor variety as prior information improves morphological segmentation in a Bayesian model. Our model outperforms the recent and oldest unsupervised models on morphological segmentation for Turkish and gives promising results on English and German.

Abstract: The consumption of Linked Data has dramatically increased with the increasing momentum towards semantic web. Linked data is essentially a very simplistic format for representation of knowledge in that all the knowledge is represented as triples which can be linked using one or more components from the triple. To date, most of the efforts has been towards either creating linked data by mining the web or making it available for users as a source of knowledgebase for knowledge engineering applications. In recent times there has been a growing need for these applications to interact with users in a natural language which required the transformation of the linked data knowledge into a natural language. The aim of the RealText project described in this paper, is to build a scalable framework to transform Linked Data into natural language by generating lexicalization patterns for triples. A lexicalization pattern is a syntactical pattern that will transform a given triple into a syntactically correct natural language sentence. Using DBpedia as the Linked Data resource, we have generated 283 accurate lexicalization patterns for a sample set of 25 ontology classes. We performed human evaluation on a test sub-sample with an inter-rater agreement of 0.86 and 0.80 for readability and accuracy respectively. This results showed that the lexicalization patterns generated language that are accurate, readable and emanates qualities of a human produced language.

Abstract: Discriminative language models (DLMs) have been widely
used for reranking competing hypotheses produced by an Automatic
Speech Recognition (ASR) system. While existing DLMs suer from
limited generalization power, we propose a novel DLM based on a dis-
criminatively trained Restricted Boltzmann Machine (RBM). The hid-
den layer of the RBM improves generalization and allows for employing
additional prior knowledge, including pre-trained parameters and entity-
related prior. Our approach outperforms the single-layer-perceptron (SLP)
reranking model, and fusing our approach with SLP achieves up to 1.3%
absolute Word Error Rate (WER) reduction and a relative 180% im-
provement in terms of WER reduction over the SLP reranker. In par-
ticular, it shows that proposed prior informed RBM reranker achieves
largest ASR error reduction (3.1% absolute WER) on content words.

Abstract: Microblogging has become a popular method to disseminate information quickly, but also for many other dialogue acts such as expression opinion and advertising. As the volumes have risen, the task of filtering messages for wanted information has become increasingly important.

In this work we examine the potential of natural language processing and machine learning to filter short messages for those that state items of news. We propose an approach that makes use of information carried at a deeper level than message's lexical surface, and show that this can be used effectively improve precision in filtering Twitter messages. Our method outperforms a baseline unigram ``bag-of-words'' approach to selecting news-event Tweets, yielding a 4.8% drop in false detection.

Abstract: In recent years, sentiment analysis has become a hot topic
in natural language processing. Although sentiment analysis research in
English is rather mature, Chinese sentiment analysis has just set sail,
as the limited amount of sentiment resources in Chinese severely limits
its development. In this paper, we present a method for the construc-
tion of a Chinese sentiment resource. We utilize both English sentiment
resources and the Chinese knowledge base NTU Multi-lingual Corpus.
In particular, we first propose a resource based on SentiWordNet and a
second version based on SenticNet.

Abstract: The detection of aggressive behavior in online discussion communities is of great interest, due to the large number of users, especially of young age, who are frequently exposed to such behaviors in social networks. Verbal aggression in online chat-rooms and fora is a popular form of cyberbullying among youngsters either in group-wide discussions or pairwise conversations. Research on cyberbullying prevention focuses on the detection of potentially harmful messages and develops intelligent systems which identify automatically potential risks, such as insults and threats. Text mining techniques are among the most promising tools used so far in the field in order to detect the expression of aggressive sentiments in short texts, which correspond to the comments, reviews, tweets and in general the textual digital fingerprints of an aggressive user. This article presents a novel approach which operates via sentiment analysis at the level of each comment, but considers the thread of messages (i.e., user discussions) as a whole to embed the notion of the context. The suggested approach is able to detect aggressive, inappropriate or antisocial behavior, relatively to the wider discussion thread context and sentiment. Key aspects of the approach are the monitoring and analysis of the most recently published comments, and the application of text classification techniques for detecting whether an aggressive action actually emerges in a discussion thread. Thorough experimental validation of the suggested approach in a previously used benchamrk demonstrates its applicability and advantages compared to other approaches.

Abstract: Opinion target extraction is a crucial task of opinion mining, aiming to extract occurrences of the different entities of a corpus that are subjects of an opinion. In order to produce a readable and comprehensible opinion summary, which is the main application of opinion target extraction, these occurrences are consolidated at the entity level in a second task. In this paper we argue that combining the two tasks, i.e. extracting opinion targets using entities as labels instead of binary labels, yields better results for opinion target extraction. We compare the binary approach and the multi-class approach on available datasets in English and French, and conduct several investigation experiments to explain the promising results. Our experiment show that an entity-based labelling not only improves opinion extraction in a single domain setting, but also let us combine training data from different domains to improve the extraction, a result that has never been achieved on target-based training data.

Abstract: Prior question generation systems typically create questions from sentences in a text. In contrast, the work presented here creates questions from a text passage in a holistic approach to natural language understanding and generation. Several NLP techniques including topic modeling are combined in an ensemble approach to identify important concepts, which then are used to create questions. Evaluation of the generated questions revealed that they are of high linguistic quality and are also important, conceptual questions, compared to questions generated by sentence-level question generation systems.

Abstract: In this paper we proposed a supervised domain adaptation technique for opinion expression extraction task. The technique generates low dimensional projections that can improve model performance in the target domain by aligning features with true label sequences. We test our methods on product reviews and observed significant improvement over baseline methods.

Abstract: The extraction of Semantic Relationship (RS) is an important task allowing the identification of relevant semantic information in the annotated textual resources. Besides, extracting SR between Named Entities (NE) is a process, which consists in guessing the significant semantic links related to them. This process is very useful to enhance the NLP-application performance, such as Question Answering systems. In this paper, we propose a rule-based method to extract and annotate SR between Arabic NEs (ANE) using an annotated Arabic Wikipedia corpus. In fact, our proposed method is composed of two main cascades regrouping respectively analysis and synthesis transducers. The analysis transducer cascade is dedicated to extract five SR types, which are synonymy, meronymy, accessibility, functional and proximity. However, synthesis one is devoted to normalize the SR and NE annotation using the TEI (Text Encoding Initiative) recommendation. Furthermore, the established transducer cascades are implemented and generated using the CasSys tool available under Unitex linguistic platform. Finally, the obtained results showed by the calculated measure values are encouraging.

Abstract: After more than a decade of the dominance of the statistical paradigm in NLP, a new wave of R&D has reverted to the primacy of rule-based approaches. This is particu-larly true for those applications where high quality results are a must. The ultimate examples of such applications are the analysis, generation and machine translation of patent claims. It is well known, however, that high quality rule-based NLP require rich knowledge resources (world models, grammar rules and lexicons) which nowa-days are painstakingly handcrafted from scratch for every new application, language or language pair.
The paper suggests a set of universal computational formalisms, which facilitate rule and lexicon acquisition and allows migrating from one rule-based application to another within one language or cross-linguistically reusing processing algorithms and, partially, linguistic knowledge thus saving a great deal of development effort. The suggested rule-based approach still absorbs the benefits of the statistics-oriented methodologies that speeded up the development of the new resources and systems.
The basic idea in our approach to the computational formalisms is two-fold. On the one hand, they are based on grammars that in a uniform way deal with atomic informational structures and then manipulate these structures by means of a few well-defined operations that build new more complex structures. On the other hand, within the frame of this formalism both the atomic, and complex structures, are application oriented and motivated by processing considerations. We consider a computational formalism to be well-defined if its semantics (mind, not the semantics of the language described by the formalism) is also well defined that we achieved by characterizing them algebraically.
We describe several such formalisms that take care of the main traditional process-ing steps in the rule-based NLP, - tagging, syntactic and shallow semantic parsing, (that results in a number of atomic predicate-argument structures), transfer (for multi-lingual applications) and generation, the latter includes a number of formalisms for building trees of atomic structures, linearization and grammaticalization. The formal-isms for different steps of processing can be applied in a cascade mode and are de-fined over a specially designed tagset, syncretically coding morphosyntactic and se-mantic information and.
All formalisms are implemented in a set of compilers (modules of developer’s environment), by means of which a developer can acquire or edit processing rules that are immediately propagated into a corresponding NLP program, which is language-independent. The formalisms described have been successfully used in several unilin-gual and multilingual systems, which are illustrated in the paper.

Abstract: In this paper, we propose domain adaptation in word sense disambiguation (WSD) using word embeddings.
The validity of the word embeddings from a huge corpus, e.g., Wikipedia, for WSD had already been shown, but
their validity in a domain adaptation framework has not been discussed before.
In addition, if they are valid, the difference in effects according to the genre of the corpora is still unknown.
Therefore, we investigate the performances of domain adaptation in WSD using the word embeddings from the source, target, and general corpora and
examine (1) whether the word embeddings are valid for domain adaptation of WSD and
(2) they are, the effects in accordance with the genre of the corpora.
The experiments using Japanese corpora revealed that the accuracy of WSD was highest when we used the word embeddings obtained from the target corpus.

Abstract: In the medical domain, text simplification is both a desirable and challenging natural language processing task . Indeed, first, medical texts can be difficult to understand for patient, because of the presence of specialized medical terms. Replacing these difficult terms with easier words can lead to improve patient’s understanding both is as straightforward as one can think because of polysemous, elisions and fuzzy semantics. In this paper, we present a lexical network based method to simplify health information in French language. We deal with semantic difficulty by replacement difficult term with supposedly easier synonyms or by using semantically related term with the help of a french lexical semantic network. We extract semantic and lexical information present in the network. In this paper, we present such a method for text simplification along with its qualitative evaluation.

Abstract: So far, the Stanford Arabic statistic parser is considered as the best parsing tool in terms of performance compared to other parsers. This performance is not stable and may vary depending on the given corpus. A more detailed method to evaluate this parser may help the users to address the causes of a performance loss. We propose, for this reason, to evaluate the Stanford Parser using the verification of the satisfaction of the syntactic constraints (called, properties) based on the analysis results of the corpus. We may obtain these properties from a reference Arabic property grammar. By the way, we enriched the simple representation of the parsing result with syntactic properties. This allows to explicit several implicit information that are the relations between syntactic units. Therefore, we had both a detailed method for the evaluation of parsers and a more syntactically informative representation for the analysis. We obtained widely detailed and encouraging results.

Abstract: Requirements form a specific class of arguments. They are designed to specify the features of systems.
Even for a simple system, several thousands of requirements produced by different authors are often needed.
It is then frequent to observe overlap and incoherence problems.
This has several consequences: significant
waste of time to identify incoherences, difficulty to update requirements,
and, most importantly, risks of misconceptions by manufacturers and users leading to incorrect
realizations and exposure to health or ecological risks.

It is very difficult and costly to manually identify incoherent requirements in large specifications.
In this paper, we propose a method to construct a corpus of various types of incoherent requirements and a categorization method that leads to the definition of patterns to detect incoherence.
We focus in this contribution on incoherences (1) which can be detected solely from linguistic factors and (2) which concern pairs of requirements. These represent about
40\% of the different types of incoherences; the other types often require extensive domain knowledge and reasoning.

This contribution opens new perspectives (1) on incoherence analysis in texts and (2) on mining incoherent requirements and arguments more generally.

Abstract: The meaning and the argument structure of particle verbs are determined by the combination of a verb and a particle. In Hungarian, verbal particles (preverbs) can occupy various positions in the sentence: they can be preverbal (detached from the verb), immediately preverbal or postverbal. The syntax of these particles is discussed in a wide range of theoretical literature. This paper presents a performance-based analysis, using corpus-driven method to reveal the distribution patterns of verbal particles in more than 21.5 million sentences. In order to obtain these data, we improved the POS-tagging of verbal particles and developed a semi-automatic method to decide which verb the detached particle belongs to. The distribution patterns gave us better insight into the phonological and pragmatic factors that may determine the position of verbal particles.

Abstract: Event detection in Twitter is an attractive and hard task. Existing methods mainly consider words co-occurrence or topic distribution of the tweets to detect the event. Few of them consider the time-series information in the text stream. In this paper, for event detection in twitter, we propose a novel multi-view clustering model which can consider both topic information and time-series information. First, we build a topic similarity matrix and a time-series similarity matrix by using the topic model and the wavelet analysis, respectively. Then, the multi-view clustering algorithm are used to group keywords. Each cluster of keywords is finally represented as an event. The experiments show that our method achieves better performance than other related work.

Abstract: In this paper, we explore the utility of sentiment analysis and text classification of voice of the customer (VOC) for improving churn prediction, which is a task to detect customers who are about to quit. Our work is motivated by the observation that the increase of customer satisfaction will reduce churn and the customer satisfaction can be reflected in some degree by applying NLP techniques on VOC, the unstructured textual information which captures a view of customer’s attitude and feedbacks. To the best of our knowledge, this is the first work that introduces text classification of VOC to churn prediction task. Experiments show that adding VOC analysis into a conventional churn prediction model results in a significant increase in predictive performance.

Abstract: Text classification is among the most broadly used machine learning tools in computational linguistic. Web information retrieval is one of the most important sectors that took advantage from this technique. Applications range from page classification, used by search engines, to URL classification used for focus crawling and on-line time-sensitive applications.
Due to the pressing need for the highest possible accuracy, a supervised learning approach is always preferred when an adequately large set of training examples is available.
Nonetheless, since building such an accurate and representative training set often becomes impractical when the number of classes increases over a few units, alternative unsupervised or semi-supervised approaches have come out.
The use of standard web directories as a source of examples can be prone to undesired effects due, for example, to the presence of maliciously misclassified web pages. In addition, this option is subjected to the existence of all the desired classes in the directory hierarchy.

Taking as input a textual description of each class and a set of URLs, in this paper we propose a new framework to automatically build a representative training set able to reasonably approximate the classification accuracy obtained by means of a manually-curated training set.
Our approach leverages on the observation that a not negligible fraction of website names is the result of the juxtaposition of few keywords. Yet, the entire URL can often be converted into a meaningful text snippet. When this happens, we can label the URL by measuring its degree of similarity with each class description. The text contained in the pages corresponding to labelled URLs can be used as a training set for any subsequent classification task (not necessarily on the web).

Experiments on a set of 20 thousand web pages belonging to 9 categories have shown that our auto-labelling framework is able to attain an approximation factor over 88% of the accuracy of a pure supervised classification trained with manually-curated examples.

Abstract: In this work, we introduce a supervised model for learning textual similarity, which can identify and score the alternative versions of a given query text. By combining dependency graph similarity and coverage features with lexical similarity measures using neural networks, we show that most relevant documents to a given text can be more accurately ranked and scored than if the similarity measures were used in isolation. Additionally, we introduce an approximate dependency subgraph alignment approach allowing node gaps and mismatch, where a certain word in one dependency graph cannot be mapped to any word in the other graph. We apply our model into two different applications, namely re-ranking for improving document retrieval precision on a new dataset, and automatic short answer scoring on a standard dataset. Experimental results indicate that our approach is easily adaptable to different tasks and languages, and works well for long texts as well as short texts.

Abstract: Personality Recognition is an emerging task in Natural Language Processing due to its potential applications. However, there is still limited literature on the topic, and those models which address this task rely on handcrafted resources; therefore, they are restricted by the domain of the problem and by the availability of resources. We propose a Convolutional Neural Network architecture trained using pre-trained word embeddings that is capable of learning the best features for the task at hand without any external dependence. The results show the potential of this approximation. The proposed model achieves comparable results with state-of-the-art models and is able to predict the personality traits of authors regardless of the social network and the availability of resources.

Abstract: The presence of figurative language represents a big challenge for sentiment analysis. In this work, we address the task of assigning sentiment polarity to Twitter texts when figurative language is employed, with a special focus on the presence of ironic devices.
We introduce a pipeline model which aims to assign a polarity value exploiting, on the one hand, irony-aware features, which rely on the outcome of a state-of-the-art irony detection model, on the other hand a wide range of affective features that cover different facets of affect exploiting information from various sentiment and emotion lexical resources for English available to the community, possibly referring to different psychological models of affect.
The proposed method has been evaluated on a set of tweets especially rich in figurative language devices proposed as a benchmark in the shared task
on ``Sentiment Analysis of Figurative Language" at SemEval-2015. Experiments and results of feature ablation show the usefulness of irony-aware features and the impact of using different affective lexicons for the task.

Abstract: Miss-translation or dropping of proper noun affects quality of machine translation output. In this paper, we propose an automatic method to build a location dependent dictionary for speech translation system. The method consists of two parts: location dependent word extraction part and word classification part. The first part extract the word by using micro blog date based on Akaike's information criteria. And the second parts classifies the words using machine learning using crawled corpus as training data.

Abstract: Timeline construction task has become popular as a way of multi-document summarization. Dealing with such a problem, it is essential to anchor each event to an appropriate time expression in a document. In this paper, we present a supervised machine learning model, two-stage event-time anchoring model. In the first stage, our system estimates event-time relations using local features. In the second stage, the system re-estimates them using the result of first stage and global features. Our experimental results show that the proposed method surpasses the state-of-the-art system by 3.5 F-score points in the TimeLine shared task of SemEval 2015.

Abstract: This paper presents a comparative study of several different approaches to speech recognition for the Tatar language. All the compared systems use a corpus-based approach, so recent results in speech and text corpora creation are also shown. The recognition systems differ in acoustic modelling algorithms, basic acoustic units, and language modelling techniques. The DNN-based system shows the best recognition result obtained on the test part of speech corpus.

Abstract: Classifying stance expressed in a text toward specific target is an emerging problem in opinion mining. A major difference between stance detection and traditional aspect-level sentiment classification is that the target of the stance might not be explicitly mentioned in text. In this paper, we show that the stance polarity of a text is not merely dependent on the content but is also highly determined by the concerned target. To this end, We propose a neural network based model, which incorporate target-specific information into stance classification using a novel attention mechanism. The proposed attention mechanism can focus on critical parts of a text. We evaluate our model on the SemEval 2016 Task 6 Twitter Stance Detection corpus achieving satisfactory performance. Our model achieves significant and consistent improvements on this task as compared with baselines.

Abstract: Arabic WordNet (AWN) is a lexical database, freely available, and useful resource to Natural Language Processing (NLP) research and applications (Information Retrieval, Machine Translation\ldots). This project is built following the methods developed for Princeton WordNet (PWN) and EuroWordNet (EWN). However, this database needs more than what it gets to improve NLP applications. Compared with others wordnets, AWN has a very poor content in both, quantity and quality levels. In this paper, we concentrate on the quality plan, especially on the antonym relations. Therefore, we propose a pattern-based approach to extend these relations, using Arabic Corpus and a corpus analysis tool. Our proposed method relies on two steps: patterns definition and automatic antonym pair extraction. The evaluation of our approach has given good results.

Abstract: Building a meta-model of dialogue act taxonomies, such as those of DAMSL, DiAML or the HCRC dialogue structure, would allow for the modelization of taxonomies themselves through the characterization of their labels using primitive features. Doing so enables the re-exploitation of annotated data for automatic dialogue act recognition tasks across taxonomies, i.e. it gives us the means to make a classifier learn from data annotated according to taxonomies different from the target taxonomy. We propose a meta-model covering several well-known taxonomies of dialogue acts, and we demonstrate its usefulness for the task of cross-taxonomy dialogue act recognition.

Abstract: The main challenge of plagiarism is how to detect the semantic relationship between the suspect text document and the source text document. Nowadays, the combinations of Natural Language Processing NLP and deep learning based approaches have a booming in the field of text analysis, including: text classification, machine translation, text similarity detection, etc.

In this context, we proposed a deep learning based method to detect Arabic plagiarism composed by the following phases: First, we started with a preprocessing phase by extracting the relevant information from text document. Then, word2vec algorithm was used to generate word vectors representation which they would be combined subsequently to generate a sentence vectors representation. Finally, we used a Convolutional Neural Networks CNN to improve the ability to capture statistical regularities in the context of sentences which then makes it possible to facilitate the similarity measurement operation between the representations of source and suspicious sentences.
The evaluation of our proposed approach gave us a promising result in term of precision.

Abstract: Semantic search is an advanced topic in Information Retrieval (IR) and has attracted increasing attention in recent years. The growing availability of the structured semantic data offers opportunities for semantic search engines, which can support more expressive queries that address complex information needs. However, due to the fact that many new concepts (mined from the Web or learned through crowd-sourcing) are continuously integrated into knowledge bases, those search engines face the challenging issues performance of scalability. In this paper, we present a parallel method, called $gSparql$, which utilizes the massive computation power of General Purpose GPUs to accelerate the performance of query processing and inference. Our method is based on the backward-chaining approach which makes inferences at query time. The experimental results show that gSparql outperforms the state-of-the-art algorithm and efficiently answers structured queries on large datasets.

Abstract: The number of word forms in agglutinating languages is theoretically infinite, which introduces sparsity in many natural language processing tasks. Part-of-speech tagging is one of these tasks that often suffers from sparsity. In this paper, we present an unsupervised model based on Bayesian Hidden Markov Models for jointly part-of-speech tagging and stemming for agglutinating languages. We use stemming in order to reduce the sparsity and emit the stems rather than words. Our results show that joint part-of-speech tagging and stemming improves both stemming and tagging scores. We present results for Turkish as an agglutinating language and English as a morphologically less complex language.

Abstract: Micro-blogging services have emerged as a powerful, real-time, way to disseminate information on the web. A small fraction of the colossal volume of posts overall are relavant. We propose Curator, a micro-blogging recommendation system that ranks micro-blogs appearing on a user's timeline according to her context. Curator learns user's time variant preferences from the text of the micro-blogs the user interacts with. Furthermore, Curator infers the user's home location and the micro-blog's subject location with the help of textual features. Precisely, we use a set of machine learning and natural language processing techniques to analyze the user's context dynamically from the micro-blogs and rank them accordingly. Curator’s extensive performance evaluation on a publicly available dataset show that it outperforms the competitive state-of-the-art by up to 154% on NDCG@5 and 105% on NDCG@25. The results also show that location is a salient feature in Curator.

Abstract: The automatic detection of uncertain statements can benefit
NLP tasks such as deception detection and information extraction. Furthermore,
it can enable new analyses in social sciences such as business
where the quantification of uncertainty or risk plays a significant role. In
spite of this, the task has only been performed on biomedical scientific
texts and Wikipedia articles. Hence, we approach the automatic detection
of uncertain statements as a binary sentence classification task on
the transcripts of spoken language in the financial domain. We present
a new dataset and { besides using features proven on the encyclopedic
domain such as bag-of-words, part-of-speech tags, and dictionaries { develop
rule-based features tailored to our task. We provide a systematic
analysis showing which type of features perform best in the financial domain
as opposed to the already explored encyclopedic domain.

Abstract: NER is an important task in NLP, often used as a basis for further treatments. A new challenge has emerged in the last few years: structured named entity recognition, where not only named entities must be identified but also their hierarchical components. In this article, we describe a cascading CRFs approach to address this challenge. It reaches the state of the art while remaining very simple on a structured NER challenge. We then offer an error analysis of our system based on a detailed, yet simple, error classification.

Abstract: Online reviews often contain user's specific opinions on aspects (features) of items. These opinions are very useful to merchants and customers, but manually extracting them is time-consuming. Several topic models have been proposed to simultaneously extract item aspects, user opinions on the aspects, and to detect sentiment of the opinions. However, existing models tend to find poor aspect-opinion associations when limited examples of the required word co-occurrences are available. Existing models often also assign incorrect sentiment to words. In this paper, we propose a Latent embedding structured Opinion mining Topic model, called the LOT, to simultaneously discover relevant aspect-level specific opinions and to assign accurate sentiment to words. Our model does this in a way that works well with small or large numbers of reviews. Experimental results for topic coherence, document sentiment classification, and a human evaluation all show that our proposed model achieves significant improvements over several state-of-the-art baselines.

Abstract: Speech is the most natural form of communication for human beings, and among others, it provides information about the speaker’s emotional state. The current study focuses on automatic speech emotion recognition based on classic and innovated machine learning approaches. Specifically, individual Gaussian mixture models (GMM) trained for each emotion, a universal background GMM model (UBM-GMM) adapted to each emotion using maximum posteriori (MAP) adaptation, and an approach based on i-vector paradigm, widely used in speaker recognition and language identification, and adapted to emotion recognition are used. When using individual GMMs, a novel technique based on multiple classifiers and late fusion is also applied. In this case, a 90.9% recognition rate is been obtained. When the state of the art, i-vector paradigm based method, along with probabilistic linear discriminant analysis (PLDA) model is used, a 91.4% average rate for speaker-independent Japanese speech emotion recognition is achieved, which is a very promising result and superior to similar studies. In addition to the Japanese emotion recognition, pair-wise recognition for seven emotions in German language has also been conducted. In this experiment, an 89.2% rate has been achieved.

Abstract: We will demonstrate several morphological analyzers of languages for which morphological analysis is very difficult, and/or that are under-resourced. It will cover at least French, German, Khmer, Lao, Lithuanian, Portuguese, Quechua, Spanish and Russian. These morphological analyzers all run on the collaborative platform lingwarium.org that supports the ARIANE-H lingware development environment. Some will also be presented as stand-alone Windows applications.

Abstract: We extend the Morphological Chain model (Narasimhan et al., 2015) by expanding the candidate space used in the model recursively to cover more split points for agglutinating languages. The results show that we improve the scores with 10% by having a F-measure of 71% in Turkish, which is higher than the current state-of-art. It shows that candidate generation plays an important role in such a log-linear model with contrastive estimation.

Abstract: We introduce a model for attitude prediction that takes the reader and the writer perspectives into account
and enables a joint reception of the attitudinal dispositions involved.
For instance, a proponent of the reader might turn out to be a villain, or some moral values of his might be negatively affected.
A formal model is specified that induces in a compositional, bottom-up manner informative relation tuples which indicate perspectives on attitudes.
This enables the reader to focus on interesting cases,
since they are directly accessible from the parts of the relation tuple.

Abstract: Extractive text summarization aims at selecting a small subset of sentences so that the contents and meaning of the original document are preserved in the best possible way. In this paper we describe an unsupervised approach to extractive summarization. It combines hierarchical topic modeling (TM) with the Minimal Description Length (MDL) principle and applies them to Chinese language. Our summarizer strives to extract information that provides the best description
of text topics in terms of MDL. This model is applied to the NLPCC
2015 Shared Task of Weibo-Oriented Chinese News Summarization [1],
where Chinese texts from news articles were summarized with the goal
of creating short meaningful messages for Weibo [2]. The experimental
results disclose superiority of our approach over other summarizers from
the NLPCC 2015 competition.

Abstract: In order to extract the real-time information referring to a specific place from SNS text such as tweets, it is necessary to analyze the temporal semantics of the mentions.
To solve this problem, we created corpus with multiple annotations using crowdsourcing for more than 10,000 tweets.
We constructed an automatic analysis model based on multiple neural networks and compared their characteristics.

Abstract: Words that participate in the sentiment (positive or negative) classification decision are known as significant words for sentiment classification. Identification of such significant words as features from the corpus reduces the amount of irrelevant information in the feature set under supervised sentiment classification settings. In this paper, we conceptually study and compare various types of feature building methods, viz., unigrams, TFIDF, Relief, Delta-TFIDF, χ2 test and Welch’s t-test for sentiment analysis task. Unigrams and TFIDF are the classic ways of feature building from the corpus. Relief, Delta-TFIDF and χ2 test have recently attracted much attention for their potential use as feature selection methods in sentiment analysis. On the contrary, t-test is the least explored for the identification of significant words from the corpus as features.

We show the effectiveness of significance tests over other feature building methods for three types of sentiment analysis tasks, viz., in-domain, cross-domain and cross-lingual. Delta-TFIDF, χ2 test and Welch’s t-test compute the significance of the word for classification in the corpus, whereas unigrams, TFIDF and Relief do not observe the significance of the word for classification. Furthermore, significance tests can be divided into two categories, bag-of-words-based test and distribution-based test. Bag-of-words-based test observes the total count of the word in different classes to find significance of the word, while distribution-based test observes the distribution of the word. In this paper, we substantiate that the distribution-based Welch’s t-test is more accurate than bag-of-words-based χ2 test and Delta-TFIDF in identification of significant words from the corpus.

Abstract: The objective of Entity Linking is to connect an entity mention in a text to a known entity in a knowledge base. The general approach for this task is to generate, for a given mention, a set of candidate entities from the base and determine, in a second step, the best one. This paper focuses on this last step and proposes a method based on learning a function that discriminates an entity from its most ambiguous ones. Our contribution lies in the strategy to learn efficiently such a model while keeping it compatible with large knowledge bases. We propose three strategies with different efficiency/performance tradeoff, that are experimentally validated on six datasets of the TAC evaluation campaigns by using Freebase and DBpedia as reference knowledge bases.

Abstract: In this paper, we introduce a neural network based sequence learning approach for the task of Arabic dialect classification. Character models based on recurrent neural networks with Long Short-Term Memory (LSTM) are suggested to classify short texts, such as tweets, written in different Arabic dialects. The LSTM-based character models can handle long-term dependencies in character sequences and do not require a set of linguistic rules at word-level, which is especially useful for the rich morphology of the Arabic language and the lack of strict orthographic rules for dialects. On the Tunisian Election Twitter dataset, our system achieves a promising average accuracy of 92.2\% for distinguishing Modern Standard Arabic from Tunisian dialect. On the Multidialectal Parallel Corpus of Arabic, the proposed character models can distinguish six classes, Modern Standard Arabic and five Arabic dialects, with an average accuracy of 63.4\%. They clearly outperform a standard word-level approach based on statistical n-grams as well as several other existing systems.

Abstract: Performance appraisal (PA) is an important HR process to periodically measure and evaluate every employee's performance vis-a-vis the goals established by the organization. A PA process involves purposeful multi-step multi-modal communication between employees, their supervisors and their peers, such as self-appraisal, supervisor assessment and peer feedback. Analysis of the structured data and text produced in PA is crucial for measuring the quality of appraisals and tracking actual improvements. In this paper, we apply text mining techniques to produce insights from PA text. First, we perform sentence classification to identify strengths, weaknesses and suggestions of improvements found in the supervisor assessments and then use clustering to discover broad categories among them. Next we use multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Finally, we propose a short-text summarization technique to produce a summary of peer feedback comments for a given employee and compare it with manual summaries. All techniques are illustrated using a real-life dataset of supervisor assessment and peer feedback text produced during the PA of 4528 employees in a large multi-national IT company.

Abstract: This paper proposes a service-oriented architecture, called as HANS, to facilitate Chinese language processing. Base on this unified framework, fundamental NLP tasks, namely word segmentation, part-of-speech tagging, named entity recognition, chunking, paring, and semantic role labeling can be seamlessly integrated to enhance functionalities for Chinese language processing. We take basic Chinese word segmentation task for example to demonstrate the effects. Evaluated benchmarks are originated from the SIGHAN 2005 bakeoff and the NLPCC 2016 shared task. We implement publicly released toolkits including Stanford CoreNLP, Fudan NLP and CKIP as services in our HANS framework for performance comparison. Experimental results confirm the feasibility of the proposed architecture. Findings are also discussed to sketch potential developments in the future.

Abstract: The automatic extraction of relations between medical entities found in related texts is considered to be a very important task, due to the multitude of applications that it can support, from question answering systems to the development of medical ontologies. Many different methodologies have been presented and applied to this task over the years. Of particular interest are hybrid approaches, in which different techniques are combined in order to improve the individual performance of either one of them. In this study, we extend a previously established hybrid framework for medical relation extrac-tion, which we modify by enhancing the pattern-based part of the framework and by applying a more sophisticated weighting method. Most notably, we replace the use of regular expressions with finite state automata for the pattern-building part, while the fusion part is replaced by a weighting strategy that is based on the operational capabilities of the Random Forests algorithm. The experimental results indicate the superiority of the proposed approach against the aforementioned well-established hybrid methodology and other state-of-the-art approaches.

Abstract: In these last years, many works have been published in the video indexing and retrieval field. However, few are the methods that have been designed to Arabic video. This paper’s aim is to achieve a new approach for Arabic news video indexing which based on embedded text as the information source and Knowledge extraction techniques to provide a conceptual description of video content. Firstly, we applied a low level processing in order to detect and recognize the video texts. Then, we extract the conceptual information including name of person, Organization and location of event using local grammars which have been implemented with the linguistic platform NooJ. Our proposed approach was tested on a large collection of Arabic TV news and experimental results were satisfactory

Abstract: In this paper we describe information extraction from web pages of scientific conferences. We enrich already known features with our new features specific for this domain and show their importance in the process of extracting information. Moreover, we investigate various data representation models, e.g., based on single tokens or sequences, in order to find the best configuration for the task in question and set up a new baseline over publicly available corpus.

Abstract: Improving the quality of Machine Translation (MT) systems is an important task not only for researcher but it is a substantial need for a translating companies to create translations in quicker and cheaper way. Combination of more machine translation system outputs is a common technique to get better translation quality because the benefits of the different systems could be utilized. The main question is to find the best method for the combination. In this paper, we used the Quality Estimation (QE) technique to combine a phrase-based and a hierarchical-based machine translation systems. The composite system was tested on several language combinations. The QE module was used to compare the outputs of the different MT systems and gave the best one as the result translation of the composite system. Results of the composite system gained better translation quality than the results for separated systems.

Abstract: The paper evaluates neural machine translation systems and phrase-based machine translation systems for highly inflected and small languages. It analyses two translation scenarios: 1) when translating broad domain data from a morphologically rich language into a morphologically rich language or English (and vice versa), and 2) when translating narrow domain data and there are limited amounts of training data available for training machine translation systems. The paper reports on experiments for English (Germanic), Estonian (Finno-Ugric), Latvian (Baltic), and Russian (Slavic) languages. The scenarios are evaluated using automatic and manual – system comparative and error analysis-based – evaluation methods. The paper also analyses the aspects where neural machine translation systems are superior to statistical (phrase-based) machine translation systems and vice versa.

Abstract: The spread of English has led to the emergence of new English varieties worldwide. Existing quantitative approaches made use of several linguistic criteria, particularly morphological and syntactical features, in investigating variations across English varieties. Taking an alternative lexical perspective to the classification of World Englishes, this paper adopts a corpus-based approach in investigating the lexical frequency across 20 regional English varieties. Specifically, the lexical items in focus include culture-bound terms and words that have undergone semantic shift. The English varieties are categorized following a series of filtering rules and normalization techniques, and a hierarchical cluster of the varieties is subsequently formalized. Our findings generally corroborate
with Kachru's Three Circle of English model, with subtle differences to the Inner Circle-Outer Circle groupings. The taxonomy of the English varieties additionally reveals geographical and cultural correlations.

Abstract: In this paper, we present a simple and elegant algorithm to extract and visualize various concept relationships present in sections of a textbook. This can be easily extended to develop visualizations of entire chapters or textbooks, thereby opening up opportunities for developing a range of visual applications for e-learning and education in general. Our algorithm creates visualizations by mining relationships between concepts present in a text by applying the idea of transitive closure rather than merely counting co-occurrences of terms. It does not require any thesaurus or ontology of concepts We applied the algorithm to two textbooks - Theory of Computation and Machine Learning - to extract and visualize concept relationships from their sections. Our findings show that the algorithm is not only capable of capturing deep-set relationships between concepts which could not have been found by using a term co-occurrence approach, but also capable of word-sense disambiguation.

Abstract: Online communities have gained considerable importance in recent years due to the increasing number of people connected to the Internet. Moderating user content in online communities is mainly performed manually, and reducing the workload through automatic methods is of great financial interest for community maintainers. Often, the industry uses basic approaches such as bad words filtering and regular expression matching to assist the moderators. In this article, we consider the task of automatically determining if a message is abusive. This task is complex since messages are written in a non-standardized way, including spelling errors, abbreviations, community-specific codes...
First, we evaluate the system that we propose using standard features of online messages. Then, we evaluate the impact of the addition of pre-processing strategies, as well as original specific features developed for the community of an online in-browser strategy game.
We finally propose to analyze the usefulness of this wide range of features using feature selection. This work can lead to two possible applications: 1) automatically flag potentially abusive messages to draw the moderator's attention on a narrow subset of messages; and 2) fully automate the moderation process by deciding whether a message is abusive without any human intervention.

Abstract: A classifier-based pattern selection approach for relation instance extraction is proposed in this paper. The classifier-based pattern selection approach proposes to employ a binary classifier that filters patterns that extracts incorrect entities for a given relation, from pattern set obtained using global estimates such as high frequency. The proposed approach is evaluated using two large independent datasets. The results presented in this paper shows that the classifier-based approach provides a significant improvement in the task of relation extraction against standard methods of relation extraction, employing pattern sets based on high frequency. The higher performance is achieved through filtering out patterns that extract incorrect entities, which in turn improves the precision of applied patterns, resulting in significant improvement in the task of relation extraction.

Abstract: Adverse drug effects (ADEs) are known to be one of the leading causes of post-therapeutic death. Thus, their
identification constitutes an important challenge as the effects of ADEs are often underreported.
However, the recent popularity of different social media sources have make it a promising source for ADE extraction. In this paper we have explored different linguistic and graph topological features to
automatically classify short sentences or tweets into ADEs or Non-ADEs. We have further represented the ADE
knowledge base into an bipartite network structure of drugs and their side effects to model drug-side effect
relationships. The proposed model can also be used to discover implicit ADEs that are
not represented in the source data. We have evaluated our proposed models with two openly available ADE dataset.
Our evaluation results shows that the proposed model have surpasses the performance of the existing baseline systems.

Abstract: The goal of the present work is to extract Co-reference chains from a document. Co-reference chains show cohesiveness of the document. The cohesiveness in the document is marked by cohesive markers such as Reference, Substitution, Ellipsis, Conjunction and Lexical cohesion. In this work we will take up Pronominal, Reflexives, R-expressions and form Co-reference chains for each of the above markers. The Co-reference chains are very essential in building sophisticated natural language processing applications such as information extraction, profile generator, entity specific text summarization etc. It is also needed in machine translation and information retrieval task. Though pronominal resolution is dealt in few Indian languages such as Tamil, Hindi, Bengali, Malayalam etc, extraction of co-reference chain in Indian languages is not attempted. We extract co-reference chains from Tamil language text. We have evaluated the system with real time data and results are encouraging.

Abstract: Instructions for physical exercises leave many details underspecified that are taken for granted and inferred by the intended reader. For certain applications, such as generating virtual action visualizations from such textual instructions, advanced text processing is needed, requiring interpretation of both implicit and explicit information. This work presents an ontology that can support the semantic analysis of such instructions in order to support the identification of matching action constructs. The proposed ontology lays down a hierarchical structure following the human body structure along with various type of movement restrictions. This facilitates flexible yet adequate representations.

Abstract: Usage of online textual media is steadily increasing; news stories, blog posts, scientific articles, are all freely available and have been employed extensively in multiple research areas, e.g. automatic text summarization, information retrieval, information extraction. Meanwhile, online debate forums have recently become popular, but have remained largely unexplored. In this paper, we work on an annotation task for a new dataset, \emph{online debate data}. We present the judgements of annotators on similar-meaning sentences, which give a low-averaged baseline for Cohen's kappa and Krippendorff's alpha. Through the application of a semantic similarity approach during calculation of inter-annotator agreement, we increase kappa and alpha to 36\% and 50\%, respectively. Moreover, we also implement an extractive summarization system for online debates. Key features to extract salient sentences from online debates are sentence position, debate title words, cosine similarity of the debate title words, and cosine similarity of the debate sentences. Our ROUGE results reveal that system performance also increases after the semantic similarity approach is applied.

Abstract: Transition-based and graph-based parsing models are two of the most dominant approaches in dependency parsing. Both models are known to achieve state-of-the-art accuracy. While each method has its own set of advantages, they are also limited by the models they use for learning and inference. Although their approaches couldn't be more different from each other, both transition and graph-based models use single classifier based linear models to predict arcs or decisions for a given instance. In this paper we deal with problems arising with both these approaches during learning. We propose a neural network based classifier voting approach to dependency parsing using multiple classifiers as component systems in an ensemble and a neural network algorithm as an oracle. We show significant improvements over the best component systems for both transition-based and graph-based dependency parsing. We also investigate different weighting schemes for voting among individual classifiers in the ensemble. All our experiments were conducted on Hindi and Telugu language data but the approach is language-independent.

Abstract: Word Sense Disambiguation (WSD) is the field that seeks to determine the correct sense of a word in a given context. In this paper, we present a WSD method based on random walks over a dependency tree, whose nodes are word-senses from the WordNet. Besides, our method incorporates prior knowledge about the frequency of use of the word-senses. We observed that our results outperform several graph-based WSD methods in All-Word task of SensEval-2 and SensEval-3, including the baseline, where the nouns and verbs part-of-speech show the better improvement in their F-measure scores.

Abstract: Query expansion approaches based on terms correlations such as association rules (ARs) between terms have shown significant improvement in the performance of the information retrieval task. The highly sized set of generated ARs is considered as a real hamper to select only most interesting ones for query expansion. In this respect, we propose a new learning automatic query expansion approach using ARs between terms. The main idea of our proposal is to rank candidate ARs in order to select the most relevant rules to be used in the query expansion process. In particular, we propose an association rules ranking approach based on a pairwise learning model to generate relevant expansion terms. Experimental results on Robust TREC and CLEF test collections highlight that the retrieval performance can be improved when ARs ranking method is used.

Abstract: In the last few years, Recurrent Neural Networks (RNNs) have proved effective on several NLP tasks.
Despite such great success, their ability to model sequence labeling is still limited.
This lead research toward solutions where RNNs are combined with models which already proved effective in this domain, such as CRFs.
In this work we propose a solution far simpler but very effective: an evolution of the simple Jordan RNN, where labels are reinjected as input into the network, and converted into embeddings, in the same way as words.
We compare this RNN variant to all the other RNN models, Elman and Jordan RNN, LSTM and GRU, on two well-known tasks of Spoken Language Understanding (SLU).
Thanks to label embeddings and their combination at the hidden layer, the proposed variant, which uses more parameters than Elman and Jordan RNNs, but far fewer than LSTM and GRU, is more effective than other RNNs, but also outperforms sophisticated CRF models.

Abstract: Humans use a variety of modifiers while expressing their emotions. This paper aims to understand the influence of different modifiers on specific emotion categories, and to compare side-by-side their impact in an automatic manner. To do so, we propose a novel data analysis method that not only quantifies how much emotional statements change under each modifier, but also models how emotions shift and how their confidence changes. This method is based on comparing the distributions of emotion labels for modified and non-modified occurrences of emotional terms within labeled data. We apply this analysis to study six types of modifiers (negation, intensification, conditionality, tense, interrogation, and modality) within a large corpus of tweets with emotional hashtags. Our study sheds light on how to model negation relations between studied emotions, reveals the impact of previously under-studied modifiers, and suggests how to detect more precise emotional statements.

Abstract: In this paper we propose a computational method for determining
the syntactic similarity between languages. We investigate multiple
approaches and metrics, showing that the results are consistent
across methods. We report results on 15 languages belonging to various
language families. The analysis that we conduct is adaptable to any
languages, as far as resources are available.

Abstract: Languages borrow words from one another for various reasons. How the borrowing process takes place, how new words enter a target language is one of the key questions of historical linguistics. In this paper, we propose a multilingual method for word form production based on the orthography of the words. For borrowed words, we investigate the derivation from a source language into a target language. We also address the problem of genetic cognates derivation. We experiment with Romanian as a target language and we investigate borrowings from multiple source languages. The advantages of the proposed method are that it does not use any external knowledge, except for the training data, and it does not require the phonetic transcriptions of the input words.

Abstract: The results of an extractive automatic summarization task depends to a great extend on the nature of the processed texts (e.g., news, medicine, or literature). In fact, general-purpose methods usually need to be adhoc modified to improve their performance when dealing with a particular application context. However, this customization requires a lot of effort from domain experts and application developers, which makes it not always possible nor appropriate.

In this paper, we propose a multi-language approach to extractive summarization which adapts itself to different text domains in order to improve its performance. In a training step, our approach leverages the features of the text documents in order to classify them by using machine learning techniques. Then, once the text typology of each text is identified, it tunes the different parameters of the extraction mechanism solving an optimization problem for each of the text document classes. This classifier along with the learned optimizations associated with each document class allows our system to adapt to each of the input texts automatically. The proposed method has been applied in a real environment of a media company with promising results.

Abstract: Sentiment analysis is a fundamental natural language processing task that automatically analyzes raw textual data and infer from it seman-tic meaning. The inferred information focuses on the author’s attitude or opinion towards a written text. Although there is extensive research done on sentiment analysis on English language, there has been little work done that targets the morphologically rich structure of the Arabic language. In addition, most of the research done on Arabic either focus on introducing new datasets or new sentiment lexicons. We propose a supervised sentiment analysis approach for two tasks: posi-tive/negative classification and positive/negative/neutral classification. We focus on the morphological structure of the Arabic language by in-troducing filtering, segmentation and morphological processing specifi-cally for this language. We also manually create an emoticon senti-ment lexicon in order to stress the expressed emotions and improve on the sentiment analyzer.

Abstract: Some expressions can be ambiguous between idiomatic and literal
interpretation depending on the context they occur in (``sales hit
the roof'' vs ``hit the roof of the car''). Previous studies suggest
that idiomaticity is not a binary property, but rather a continuum
or the so-called ``scalar phenomenon'' ranging from completely
literal to highly idiomatic. In this paper we reported the results
of an experiment in which human annotators rank idiomatic
expressions in context on a scale from 1 (literal) to 4 (highly
idiomatic). Our experiment supports the hypothesis that idioms fall
on a continuum and that one might differentiate between highly
idiomatic, mildly idiomatic and weakly idiomatic expressions. In
addition, we measure the relative idiomaticity of 11 idiomatic
types and compute the correlation between the relative idiomaticity
of an expression and the performance of various automatic models for
idiom detection. We show that our model, based on the
distributional semantics ideas, not only outperforms the previous
models, but also positively correlates with the human judgements,
which suggests that we're moving in the right direction toward
automatic idiom detection.

Abstract: Clustering step in the mention-pair paradigm for coreference resolution, forms the chain of coreferent mentions from the mention pairs classified as coreferent. Clustering methods including best-first clustering considers each antecedent candidate individually, while selecting the antecedent for an anaphoric mention. Here we introduce an easy-to-implement modification to best-first clustering to improve coreference resolution on Indian classical music forums. This method considers the relation between the candidate antecedents along with the relation between the anaphoric mention and the candidate antecedent. We observe a modest but statistically significant improvement over the best-first clustering for this dataset.

Abstract: Vector based word representation models are typically developed from
very large corpora with the hope that the representations are reliable and that they cover many words. However, we often encounter words in real world applications that are not available in a single vector-based model. In this paper, we present a novel Neural Network (NN) based approach for obtaining representations for words that are missing in a target model from another model, called the source model, where representations for these words are available, effectively pooling together their vocabularies and the corresponding representations. Our experiments with three different types of pre-trained models (Word2vec, GloVe, and LSA) each providing representations for millions of unique words and yet having only a small percentage of words in common show that the representations obtained using our transformation approach can substantially and effectively extend the word coverage of existing models. The increase in the number of unique words covered by a model varies from few to several times depending on which model vocabulary is taken as a reference. The transformed word representations are well correlated (average correlation up to 0.801 for words in Simlex-999 dataset) with the native target model representations indicating that the transformed vectors can effectively be used as substitutes of native word representations. Furthermore, an extrinsic evaluation based on a word-to-word similarity task using the Simlex-999 dataset leads to results close to those obtained using native target model representations.

Abstract: The translation task of social media comments has attracted researchers in recent times because of the challenges to understand the nature of the comments and its representation and the need of its translation into a target language. In the present work, we attempt two approaches of translating the Facebook comments – one using a language identifier and other without using the language identifier.
We also attempt to handle some form of spelling variation of these comments towards
improving the translation quality with the help of state-of-the-art statistical machine translation techniques. Our approach employs n-best-list generation of the source language of the training dataset to address the spelling variation in the comments and also enrich the resource for translation. A small in-domain dataset could further boost the performance of the translation system. Our translation task focus on Hindi-English mixed comments collected from Facebook and our systems show
improvement of translation quality over the baseline system in terms of automatic evaluation scores.

Abstract: Nowadays, many NLP problems are tackled as supervised machine learning tasks. Consequently, the cost of the expertise needed to annotate the examples is a widespread issue. Active learning offers a framework to that issue, allowing to control the annotation cost while maximizing the classifier performance, but it relies on the key step of choosing which example will be proposed to the expert.
In this paper, we examine and propose such selection strategies in the specific case of Conditional Random Fields (CRF) which are largely used in NLP.
On the one hand, we propose a simple method to correct a bias of certain state-of-the-art selection techniques. On the other hand, we detail an original approach to select the examples, based on the respect of proportions in the datasets.
These contributions are validated over a large range of experiments implying several tasks and datasets, including named entity recognition, chunking, phonetization, word sense disambiguation.

Abstract: Existing pre computer era monolingual dictionaries are a voluminous loosely structured source of lexical and ontological information. Numerous attempts were made to extract WordNet or ontology relations from monolingual dictionaries with varying success, most based on morphosyntactic rules. Difficulty of information extraction task greatly depends on dictionary authors' discipline, which is rarely enough to allow effortless dictionary parsing.

Core of this work is an analysis of how can rule-based approach for relation extraction be improved and what relations can be extracted. To simplify this task syntactically similar definitions are clustered, clusters are analyzed and then one morphosyntactic rule per cluster is assigned. The author suggests to use mixed n-gram clustering features as a simple substitute for syntactic features.

The clustering is performed on Russian explanatory dictionary (ed. Kuznetsov). The results show that dictionary discipline in this dictionary is actually very strict with less than hundred definition styles adhered. This along with preliminary clustering allowed to develop rules for relation extraction with very high precision, not reported in similar works previously.

Abstract: In this paper we propose a novel approach to determine Textual Entailment (TE) relation between a pair of text snippets. It can be defined as a directional relationship between a pair of sentences, denoted by T – the entailing “Test” and H – the entailed “Hypothesis”. We can say that T entails H if the meaning of H can be relate with the meaning of T. Different machine translation along with summary evaluation metrics have been used as features for different machine learning classifiers to take the entailment decision in this study. We considered three machine translation evaluation metrics, namely BLEU, METEOR and TER and a summary evaluation metric namely ROUGE as similarity metrics for this task. Finally, we used polarity feature like negation in combination with the aforementioned features. We carried out the experiments on the datasets released in the shared tasks on textual entailment organized in RTE-1, RTE-2, RTE-3, RTE-4 and RTE-5. The best classification accuracies obtained by our system on the RTE-1, RTE-2, RTE-3, RTE-4 and RTE-5 dataset are 54%, 55%, 60%, 52% and 51% respectively.

Abstract: In this paper, we model a transliteration system from Persian to English as grapheme-to-phoneme (G2P) and word lattice methods combined with statistical machine translation (SMT). Persian is an Indo-Iranian branch of the Indo-European family of languages be-longing to Arabic script-based languages. Our transliteration model is induced from a parallel corpus containing the Arabic script of a Persian book together with its romanized transcription scheme named Dabire. We manually aligned the sentence of this book in both scripts and used it as a parallel corpus. Our results indicate that the performance of the system is improved by adding grapheme-to-phoneme and word lattice methods for out-of-vocabulary handling task into the monotonic statistical machine transliteration sys-tem. In addition, the final performance on the test corpus shows that our system achieves comparable results with other state-of-the-art systems.

Abstract: Coordinations refer to phrases such as ``A and/but/or/... B".
The detection of coordinations remains a major problem due to the complexity of their components. Existing work normally classified the training data into two categories: correct and incorrect. This often caused the problem of data imbalance which inevitably damaged performances of the models they used.
We propose to fully exploit the differences between training data by
formulating the detection of coordinations as a ranking problem to remedy this problem. We develop a novel model based on the long short-term memory network. Experiments on Penn Treebank and Genia show the proposed method has outperformed previous work.

Abstract: Providing appealing brand names to newly launched products, newly formed companies, websites, campaigns or for renaming existing companies is highly important as it can play a crucial role in deciding its success or failure. In this work, we propose a computational method to generate appealing brand names based on the description of such entities. Further, we use quantitative scores for readability, pronounciability, memorability and uniqueness of a generated name to rank order them. A set of diverse appealing names is recommended to the user for brand naming task. These names also aid in ideation. Experimental results show that the names generated by our approach are more appealing than names which prior approaches and few recruited humans could come up.

Abstract: The effectiveness of character n-gram features for representing the stylistic properties of a text has been demonstrated in various independent Authorship Attribution (AA) studies. Moreover, it has been shown that some categories of character n-grams perform better than others both under single and cross-topic AA conditions. In this work, we present an improved algorithm for cross-topic AA. We demonstrate that the effectiveness of character n-grams representation can be significantly enhanced by performing simple pre-processing steps and appropriately tuning the number of features, especially in cross-topic conditions.

Abstract: The paper describes a named entity recognition system for Amharic, an under-resourced language, using a recurrent neural network, a bi-directional long short term memory model. Word vectors based on semantic information are built for all tokens using an unsupervised learning algorithm, word2vec, while a Conditional Random Fields classifier trained on language independent features predicts each token's named entity class. The predictions, features and word vectors are fed to the deep neural network which assigns labels to the words. When evaluated by 10-fold cross-validation, the Amharic named entity recogniser achieved good precision (86.0%), but worse recall (65.5%).

Abstract: We present the Contextual Specificity Similarity (CSS) measure, a new document similarity measure based on word embeddings and inverse document frequency. The idea behind the CSS measure is to score higher the documents that include words with close embeddings and frequency of usage. This paper provides a comparison with several methods of text classification, which will evince the accuracy and utility of CSS in k-nearest neighbour classification tasks for short texts.

Abstract: We present our work aiming at turning the linguistic material available in Grierson’s classical Linguistic Survey of India (LSI) into a digital language resource, a database suitable for a broad array of linguistic investigations of the languages of South Asia. While doing so, we develop state-of-the-art language technology for automatically extracting the relevant grammatical information from the text of the LSI, and interactive linguistic information visualization tools for better analysis and comparisons of languages based on their structural and functional features.

Abstract: To automatically identify sarcasm in text is a challenging task, since it can be difficult also for humans, in particular in very short texts with little explicit context, such as tweets (Twitter messages). The paper presents a comparison of three sets of tweets marked for sarcasm, two annotated manually and one annotated using the common strategy of relying on the authors correctly using hashtags to mark sarcasm. To evaluate the datasets, a state-of-the-art system for sarcasm detection in tweets was implemented. Experiments on the two manually annotated datasets show comparable results, while deviating considerably from results on the automatically annotated data, indicating that using hashtags is not a reliable approach for creating Twitter sarcasm corpora.

Abstract: Over the past few years, there has been a growing interest in applying Natural Language Processing (NLP) methods to social media texts due to the increased availability of social media platforms, such as social networks, forums and blogs, and the importance of the social media data to companies, governments and nonprofit organizations. Social media texts are usually very short texts which characterized by nonstandard spelling, inconsistent punctuation, abbreviations, emojis, and emoticons that pose special challenges for NLP.

Since social media allow direct contact with the target public, opinion and sentiment analysis of users' comments has received a lot of research attention. Most prior work on sentiment analysis addressed the problem of classifying comments as positive, negative or neutral with respect to the general attitude of the blog/post. In this paper, we focus on task of classifying the relevance of a comment to the content of its post, which requires a fuller understanding of the comment semantics. For example, the general attitude of the comment John, you are the best, and above the rest!!! is positive, but since the comment is not toward the content of the post, it is considered as irrelevant/negative in the relevance classification task.

Many text classification (TC) methods use the bag-of-words (BoW) representation. Due to the low frequency of occurrence of each word in short texts, their BoW representation of is very sparse. To overcome the sparsity problem, different dimensional reduction methods for semantic analysis, such as Latent Semantic Analysis (LSA), Word Embedding (Word2Vec), and Random Projection (RP) have been explored.

In this study, we explore four semantic vector representations for the relevance classification task. We investigate different types of large unlabeled data for learning the semantic vectors, namely comment texts, post texts and both post and comment texts. In addition, we examine whether expanding the input of the comment relevance classification task to include also the post text is beneficial for increasing the classification performance.

Abstract: The previous Named Entity Recognition (NER) models for Modern Standard Arabic (MSA) rely heavily on the use of features and gazetteers, which is time consuming. In this paper, we introduce a novel neural network architecture based on bidirectional Gated Recurrent Unit (GRU) combined with Conditional Random Fields (CRF). Our neural network uses minimal features: pretrained word representations learned from unannotated corpora and also character-level embeddings of words. This novel architecture allowed us to eliminate the need for most of handcrafted engineering features. We evaluate our system on a publicly available dataset where we were able to achieve comparable results to previous best-performing systems.

Abstract: Emotion detection in Arabic text is an emerging research area, but the efforts in this new field are hindered by the very limited availability of Arabic datasets annotated with emotions. In this paper, we review work that has been carried out in the area of emotion analysis in Arabic text. We then present an Arabic tweet dataset that we have built to serve this task. The efforts and methodologies followed to collect, clean, and annotate our dataset are described and preliminary experiments carried out on this dataset for emotion detection are presented. The results of these experiments are provided as a benchmark for future studies and comparisons with other emotion detection models. The best results over a set of eight emotions were obtained using a complement Naïve Bayes algorithm with an overall accuracy of 68.12 %.

Abstract: We propose a novel framework for multimodal sentiment
analysis and emotion recognition using convolutional
neural network-based feature extraction from text and visual
modality. We obtain a performance improvement of 10%
over the state of the art by combining visual, text and audio
features. To this end, we also discuss some of frequently
ignored major issues in this research field: role of speaker independent
models, importance of the modalities and generalizability.