Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
18625 papers
For a resource poor language like Hindi, it becomes very difficult to bracket a noun sequence using approaches which are only based on corpus or lexical database. For semantic knowledge, power of both type of resources is needed to be combined. Therefore, affinity in between two nouns is preferred to be measured using backoff association which is the combination of lexical and conceptual association. Also, syntax is important for this task. But syntactic rules do not work for the compound nouns which is a special case of noun sequences and it may also occur as the sub-sequence. Using hybrid approach, accuracy of 86.33% has been obtained.
In the article the author makes an attempt to classify radiodiscourse in according to the existing classifications of discourses. Being a type of oral discourse of mass media it possesses its certain characteristics such as unprepared nature of official speech, tendencies to deterioration of a linguistic norm and an intimate manner of speaking. Moreover, clearly defined roles, specified place and context of communication, chronotope, certain typical discursive formulae allow to define radiodiscourse as an institutional discourse. However, extralinguistic conditions in which radiocommunication takes place bring about such peculiarities of radiodiscourse as distance and mediation, mass nature of audience, omnitude, high information rate, direct connection to time, in particular irreversibility, linearity, continuity. Though radiodiscourse has universal characteristics of both oral mediaand institutional discourses, the named peculiarities make it possible to distinguish it in discourse diversity, where it has a certain quite independent position.
Our knowledge of objects reflects the statistics of the visual environment. From our experiences in the world, we store information about categories of objects and the features that define them. One important statistical property of objects is the co-occurrence of their constituent features. For example, the round shape of an apple co-occurs frequently with the color red, but not the color blue. Here we examine the neural mechanisms that encode such feature co-occurrence statistics at the interface of perception and memory. In an fMRI experiment, subjects viewed images of colored objects while performing an unrelated scrambled-object detection task. The stimuli included exemplars from three different categories: apples, leaves, and roses. To create stimuli that sampled a range of co-occurrence statistics, each exemplar image had its color systematically manipulated to be red, pink, yellow, green, or blue (Fig.1A). We quantified co-occurrence frequencies of color-object combinations (e.g., “yellow apple”) in a large lexical corpus. A separate norming study demonstrated that this metric was strongly correlated with subjective ratings of color-object typicality. Importantly, the co-occurrence of object and color information is independent of the frequencies of each feature alone. We tested the hypothesis that feature co-occurrence information is encoded in semantic memory regions and automatically retrieved during object perception. Using representational similarity analysis, we identified regions where response patterns were similar for category exemplars with similar co-occurrence statistics (Fig.2B). We expected that the angular gyrus would encode combinatorial information given its proposed role in semantic integration. Indeed, we found that this region and the anterior fusiform cortex encode high-level feature co-occurrence statistics, while early visual cortex, lateral-occipital complex, and inferior-temporal cortex did not. These results suggest that regions at the interface of vision and semantic memory encode combinatorial information that underlies real-world knowledge of objects and is independent of coding for individual features. Meeting abstract presented at VSS 2015
Moral choices are the result of an interplay between emotion and cognition [1,2]: emotions usually elicit deontological choices while cognition leads people to make utilitarian choices.Accordingly, patients with emotional deficits give more utilitarian responses compared to healthy controls [3][4][5][6][7].Recently, studies have shown that this response pattern is present also in individuals with alexithymia [8,9].Alexithymia describes a subclinical condition characterized by affect dysregulation and it comprises an affective dimension, the level of subjective emotional experience, and a cognitive dimension, referring to the cognitive control of emotions [10].The aim of the present study was to investigate the impact of alexithymia on "moral decision making", taking these two dimensions into account.To this end, we recruited 22 alexithymic participants and 22 non-alexithymic controls, and asked them to perform a moral decision-making task, which consisted in choosing between two solutions of a dilemma and in rating the valence of emotions and arousal elicited by the decision [11].While participants performed the task, their skin conductance was measured.No differences emerged between alexithymic participants and controls in the percentage of utilitarian answers and decision time.However, alexithymic participants rated as significantly less unpleasant the emotions associated with the decision.Moreover, valence and arousal ratings significantly correlated with the affective component of alexithymia.No significant differences in skin conductance between the two groups were observed.In conclusion, alexithymia seems to influence emotions rising from the moral decision but not the decision itself.In addition, the affective components of alexithymia seems to play an important role in the evaluation of the affective reactions to moral decisions.
L’etiquetage morpho-syntaxique est une tâche fondamentale du Traitement Automatique de la Langue, sur laquelle reposent souvent des traitements plus complexes tels que l’extraction d’information ou la traduction automatique. L’etiquetage en domaine de specialite est limite par la disponibilite d’outils et de corpus annotes specifiques au domaine. Dans cet article, nous presentons le developpement d’un corpus clinique du francais annote morpho-syntaxiquement a l’aide d’un jeu d’etiquettes issus des guides d’annotation French Treebank et Multitag. L’analyse de ce corpus nous permet de caracteriser le domaine clinique et de degager les points cles pour l’adaptation d’outils d’analyse morpho-syntaxique a ce domaine. Nous montrons egalement les limites d’un outil entraine sur un corpus journalistique applique au domaine clinique. En perspective de ce travail, nous envisageons une application du corpus clinique annote pour ameliorer l’etiquetage morpho-syntaxique des documents cliniques en francais.
Existing models of female sexual dysfunction (FSD) are broad and do not provide information about how to improve existing interventions. The purpose of the current study is to extend the empirical application of a disgust model of FSD (de Jong, van Overveld, & Borg, 2013) to a population of young women reporting difficulties with sexual desire and/or arousal and related distress. Sixty college-aged females participated in the study and were placed into two groups based upon their reports of sexual functioning and sexual distress: a control group (i.e., no sexual difficulties or distress) and a clinical group (i.e., difficulties with sexual desire and/or arousal and accompanying distress). Participants were attached to physiological equipment and shown images displaying neutral, positive, disgusting, and erotic content. It was hypothesized that the clinical group would show more evidence of disgust (via affective and autonomic responses) than the control group. Consistent with hypotheses, no group differences were found in any of the affective or autonomic measures during presentation of the neutral, positive, or disgust images. Group differences during presentation of the erotic images (i.e., in facial EMG, heart rate, and self-report affective ratings) and follow-up analyses provided preliminary evidence for generalizing the disgust model of female sexual dysfunction beyond disorders of sexual pain, at least among some women. Exploratory analyses implicated a relationship between a history of sexual victimization and self-report disgust ratings of erotic images. Future research should further explore these relationships in order to shed more light on how disgust-based mechanisms impact the onset and maintenance of female sexual dysfunction.
We examine the nature of phonological and semantic similarity\nin early language learning. We consider how the use of this\ninformation might change over the course of development. To\nthis end, we represent the lexicon as either a phonological or\nsemantic network and model the growth of this network. Constructing\nnormative vocabularies from the Communicative Development\nInventory norms, we utilize a preferential attachment\ngrowth algorithm. We predict and quantify the words\nwhich will be learned next, comparing the two network representations.\nWe consider the effect of age, total vocabulary size\nand language ability as measured through CDI percentile. Our\nfindings suggest that the semantic representation does not outperform\nthe baseline bag-of-words model, whereas the phonological\nrepresentation conditionally does. More generally, we\nshow that the network representation influences the ability of\na model to capture vocabulary growth. We further offer a\nmethod of analysis for testing representational assumptions in\nnetwork models.
International audience
Abstract This article is devoted to the functioning of Russian language in the modern world and raises a few issues: What are the regions of Russian diasporas nowadays? Under what conditions can Russian language survive in the foreign countries? How does Russian language change in its mother country and outside? It is also under discussion how language norms transform in modern Russian. It is stated that liberalization in lexical and register choice is opposed to the requirements of grammatical accuracy. Grammar is still considered as a marker of good education and culture of a person.
The ability to integrate contextual information is important for the comprehension of emotional and social situations. While some studies have shown that emotional processes and social cognition are impaired in people with hypomanic personality trait, no results have been reported concerning the neurophysiological processes mediating the processing of emotional information during the integration of contextual social information in this population. We therefore chose to conduct an ERP study dealing with the integration of emotional information in a population with hypomanic personality trait. Healthy participants were evaluated using the Hypomanic Personality Scale (HPS), and ERPs were recorded during a linguistic task in which participants silently read sentence pairs describing short social situations. The first sentence implicitly conveyed the positive or negative emotional state of a character. The second sentence was emotionally congruent or incongruent with the first sentence. We)
The article is devoted to the conceptual analysis of Orthodontic Terminology. According to the basic positions of Cognitive Linguistics the concepts are nominated on the basis of the ideas about them and develop in the course of practical human activity. The category of Space is the main one for all sections of Orthodontics. On the basis of this fact, the main spatual (dimentional) components were observed, which are the main components of the conceptual framework and are more often find their linguistic objectification: the place, the form, the size. These components are very important in the Orthodontics consepts formation, because of the need in accurate diagnosis and treatment methods. The article investigates the traditional ways of term formation: morphological, lexical-semantic, syntactic and mixed.
The translation of corporate publicity material,(TCPM for short), as a pragmatic text, calls for guidance of relevant theo-ry since someword-for-wordtranslation is far from adequate to achieve its purpose. International exchange and communicationin the Southern areas of Jiangsu Province are developing fast and of important value. Therefore, the present study attempts to con-duct the research on in the lights ofFunctional EquivalenceandSkopos Theory. In the proposed translation strategy, the studygives due attention to the TCPM's unique aspect, which has both vocative and informative function. Thus the translator shall taketarget language and its culture norms into consideration so as to make the translated text readable. To achieve the desired effect ofthe translated text, the translator shall free themselves from theFormal Equivalenceand rigid translation, and take efforts toadapt to target reader's cultural background and language norm, thus improving translsted texts at lexical, syntactic, and textuallevels.
Theory of Mind (ToM) has repeatedly been defined as the ability to understand that others believe their own things based on their own subjective interpretations and experiences, and that their thoughts are determined independently from your own. In this study, we wanted to see if individual differences in ToM are capable of causing different perceptions of an individual's interactions with human like robotics and highlight whether or not individual differences in ToM account for different levels of how individuals experience what is called the "Uncanny Valley phenomenon" and to see whether or not having a fully developed theory of mind is essential to the perception of the interaction. This was assessed by inquiring whether or not individuals with Autism Spectrum Disorder (ASD) perceive robotics and artificially intelligent technology in the same ways that typically developed individuals do; we focused on the growing use of social robotics in ASD therapies. Studies have indicated that differences of ToM exist between individuals with ASD and those who are typically developed. Comparably, we were also curious to see if differences in empathy levels also accounted for differences in ToM and thus a difference in the perceptions of human like robotics. A robotic image rating survey was administered to a group of University of central Florida students, as well as 2 surveys - the Autism Spectrum Quotient (ASQ) and the Basic Empathy Scale (BES), which helped optimize a measurement for theory of mind. Although the results of this study did not support the claim that individuals with ASD do not experience the uncanny valley differently than typically developed individuals, there were significant enough results to conclude that different levels of empathy may account for individual differences in the uncanny valley. People with low empathy seemed to have experienced less of an uncanny valley feeling, while people with higher recorded empathy showed to experience more of an uncanny valley sensitivity.
The PROIEL Treebank is a dependency treebank with morphosyntactic and information-structure annotation. It includes texts in several ancient Indo-European languages and is freely available under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License.
Recent research reveals an age-related increase in positive autobiographical memory retrieval using a number of positivity measures, including valence ratings and positive word use. It is currently unclear whether the positivity shift in each of these measures co-occurs, or if age uniquely influences multiple components of autobiographical memory retrieval. The current study examined the correspondence between valence ratings and emotional word use in young and older adults' autobiographical memories. Positive word use in narratives was associated with valence ratings only in young adults' narratives. Older adults' narratives contained a consistent level of positive word use regardless of valence rating, suggesting that positive words and concepts may be chronically accessible to older adults during memory retrieval, regardless of subjective valence. Although a relation between negative word use in narratives and negative valence ratings was apparent in both young and older adults, it was stronger in older adults' narratives. These findings confirm that older adults do vary their word use in accordance with subjective valence, but they do so in a way that is different from young adults. The results also point to a potential dissociation between age-related changes in subjective valence and in positive word use.
We studied the dynamics of lexical decisions by asking participants to categorize lexical and nonlexical stimuli and recording their mouse movements toward response buttons during the choice. In a previous report we revealed greater trajectory curvature and attraction to competitors for Low Frequency words and Pseudowords. This analysis did not clarify whether the trajectory curvature in the two conditions was due to a continuous dynamic competition between the response alternatives or if a discrete revision process (a "change of mind") took place during the choice from an initially selected response to the opposite one. To disentangle these two possibilities, here we analyse the velocity and acceleration profiles of mouse movements during the choice. Pseudowords' peak movement velocity occurred with 100ms delay with respect to words and Letters Strings. Acceleration profile for High and Low Frequency words and Letters Strings exhibited a butterfly plot with one acceleration peak at 4)
Temporal information in a signal can be partitioned into temporal envelope (E) and fine structure (FS). Fine structure is important for lexical tone perception for normal-hearing (NH) listeners, and listeners with sensorineural hearing loss (SNHL) have an impaired ability to use FS in lexical tone perception due to the reduced frequency resolution. The present study was aimed to assess which of the acoustic aspects (E or FS) played a more important role in lexical tone perception in subjects with auditory neuropathy spectrum disorder (ANSD) and to determine whether it was the deficit in temporal resolution or frequency resolution that might lead to more detrimental effects on FS processing in pitch perception. Fifty-eight native Mandarin Chinese-speaking subjects (27 with ANSD, 16 with SNHL, and 15 with NH) were assessed for (1) their ability to recognize lexical tones using acoustic E or FS cues with the “auditory chimera” technique, (2) temporal resolution as measured with temporal )
International audience
Graph-based dependency parsing algorithms commonly employ features up to third order in an attempt to capture richer syntactic relations. However, each level and each feature combination must be defined manually. Besides that, input features are usually represented as huge, sparse binary vectors, offering limited generalization. In this work, \nwe present a deep architecture for dependency parsing based on a convolutional neural network. It can examine the whole sentence structure before scoring each head/modifier \ncandidate pair, and uses dense embeddings as input. Our model is still under ongoing work, achieving 91.6% unlabeled attachment score in the Penn Treebank.
A lexicostatistical classification is proposed for 20 languages and dialects of the Lezgian group of the North Caucasian family, based on meticulously compiled 110-item wordlists, published as part of the Global Lexicostatistical Database project. The lexical data have been subsequently analyzed with the aid of the principal phylogenetic methods, both distance-based and character-based: Starling neighbor joining (StarlingNJ), Neighbor joining (NJ), Unweighted pair group method with arithmetic mean (UPGMA), Bayesian Markov chain Monte Carlo (MCMC), Unweighted maximum parsimony (UMP). Cognation indexes within the input matrix were marked by two different algorithms: traditional etymological approach and phonetic similarity, i.e., the automatic method of consonant classes (Levenshtein distances). Due to certain reasons (first of all, high lexicographic quality of the wordlists and a consensus about the Lezgian phylogeny among Caucasologists), the Lezgian database is a perfect testing are)
Memory is essential to many cognitive tasks including language. Apart from empirical studies of memory effects on language acquisition and use, there lack sufficient evolutionary explorations on whether a high level of memory capacity is prerequisite for language and whether language origin could influence memory capacity. In line with evolutionary theories that natural selection refined language-related cognitive abilities, we advocated a coevolution scenario between language and memory capacity, which incorporated the genetic transmission of individual memory capacity, cultural transmission of idiolects, and natural and cultural selections on individual reproduction and language teaching. To illustrate the coevolution dynamics, we adopted a multi-agent computational model simulating the emergence of lexical items and simple syntax through iterated communications. Simulations showed that: along with the origin of a communal language, an initially-low memory capacity for acquired ling)
Multilingual digital resources with Bulgarian languageThe paper presents in brief Bulgarian language resources as a part of multilingual digital resources developed in the frame of some international projects, among them parallel annotated and aligned corpora, comparable corpora, morpho-syntactic specifications for corpora annotation and dictionaries encoding, lexicons, lexical databases, and electronic dictionaries.
Automatic speech recognition (ASR) technology has matured over the past few decades and has made significant impacts in a variety of fields, from assistive technologies to commercial products. However, ASR system development is a resource intensive activity and requires language resources in the form of text annotated audio recordings and pronunciation dictionaries. Unfortunately, many languages found in the developing world fall into the resource-scarce category and due to this resource scarcity the deployment of ASR systems in the developing world is severely inhibited. One approach to assist with resource-scarce ASR system development, is to select “useful” training samples which could reduce the resources needed to collect new corpora. In this work, we propose a new data selection framework which can be used to design a speech recognition corpus. We show for limited data sets, independent of language and bandwidth, the most effective strategy for data selection is frequency-matched selection and that the widely-used maximum entropy methods generally produced the least promising results. In our model, the frequency-matched selection method corresponds to a logarithmic relationship between accuracy and corpus size; we also investigated other model relationships, and found that a hyperbolic relationship (as suggested from simple asymptotic arguments in learning theory) may lead to somewhat better performance under certain conditions.
Psychologists have used experimental methods to study language for more than a century. However, only with the recent availability of large-scale linguistic databases has a more complete picture begun to emerge of how language is actually used, and what information is available as input to language acquisition. Analyses of such "big data" have resulted in reappraisals of key assumptions about the nature of language. As an example, we focus on corpus-based research that has shed new light on the arbitrariness of the sign: the longstanding assumption that the relationship between the sound of a word and its meaning is arbitrary. The results reveal a systematic relationship between the sound of a word and its meaning, which is stronger for early acquired words. Moreover, the analyses further uncover a systematic relationship between words and their lexical categories-nouns and verbs sound differently from each other-affecting how we learn new words and use them in sentences. Together, these results point to a division of labor between arbitrariness and systematicity in sound-meaning mappings. We conclude by arguing in favor of including "big data" analyses into the language scientist's methodological toolbox.
We investigate the relation between general affective meaning and the use of particular phonological segments in poems, presenting a novel quantitative measure to assess the basic affective tone of a text based on foregrounded phonological units and their iconic affective properties. The novel method is applied to the volume of German poems “verteidigung der wölfe” (defense of the wolves) by Hans Magnus Enzensberger, who categorized these 57 poems as friendly, sad, or spiteful. Our approach examines the relation between the phonological inventory of the texts to both the author’s affective categorization and readers’ perception of the poems—assessed by a survey study. Categorical comparisons of basic affective tone reveal significant differences between the 3 groups of poems in accordance with the labels given by the author as well as with the affective rating scores given by readers. Using multiple regression, we show our sublexical measures of basic affective tone to account for a considerable part of variance (9.5%−20%) of ratings on different emotion scales. We interpret this finding as evidence that the iconic properties of foregrounded phonological units contribute significantly to the poems’ emotional perception—potentially reflecting an intentional use of phonology by the author. Our approach represents a first independent statistical quantification of the basic affective tone of texts. (PsycINFO Database Record (c) 2015 APA, all rights reserved)
We investigated music and language processing in a group of early bilinguals who spoke a tone language and a non-tone language (Cantonese and Dutch). We assessed online speech-music processing interactions, that is, interactions that occur when speech and music are processed simultaneously in songs, with a speeded classification task. In this task, participants judged sung pseudowords either musically (based on the direction of the musical interval) or phonologically (based on the identity of the sung vowel). We also assessed longer-term effects of linguistic experience on musical ability, that is, the influence of extensive prior experience with language when processing music. These effects were assessed with a task in which participants had to learn to identify musical intervals and with four pitch-perception tasks. Our hypothesis was that due to their experience in two different languages using lexical versus intonational tone, the early Cantonese-Dutch bilinguals would outperform )
Recent years have seen an increased interest in and availability of many different kinds of corpora. These range from small, but carefully annotated treebanks to large parallel corpora and very lar...
Although there has been extensive research on the processing of the emotional meaning of music, little is known about other aspects of listeners’ experience of music. The present study investigated the neural correlates of the iconic meaning of music. Event-related potentials (ERP) were recorded while a group of 20 music majors and a group of 20 non-music majors performed a lexical decision task in the context of implicit musical iconic meaning priming. ERP analysis revealed a significant N400 effect of congruency in time window 260-510 ms following the onset of the target word only in the group of music majors. Time-course analysis using 50 ms windows indicated significant N400 effects both within the time window 410-460 ms and 460-510 ms for music majors, whereas only a partial N400 effect during time window 410-460 ms was observed for non-music majors. There was also a trend for the N400 effects in the music major group to be stronger than those in the non-major group in the sub-wi)
The paper describes an ongoing experiment consisting in the attempt to quantify word-order proper- ties of three Indo-European languages (Czech, English and German). The statistics are collected from the syntacti- cally annotated treebanks available for all three languages. The treebanks are searched by means of a universal query tool PML-TQ. The search concentrates on the mutual or- der of a verb and its complements (subject, object(s)) and the statistics are calculated for all permutations of the three elements. The results for all three languages are compared and a measure expressing the degree of word order free- dom is suggested in the final section of the paper. This study constitutes a motivation for formal modeling of natural language processing methods.
Accurate dependency parsing requires large treebanks, which are only available for a few languages. We propose a method that takes advantage of shared structure across languages to build a mature parser using less training data. We propose a model for learning a shared “univer-sal ” parser that operates over an inter-lingual continuous representation of lan-guage, along with language-specific map-ping components. Compared with super-vised learning, our methods give a con-sistent 8-10 % improvement across several treebanks in low-resource simulations. 1
We compare two annotation styles, Prague dependencies and Universal\nStanford Dependencies, in their adequacy for parsing.\nWe specifically focus on comparing the adposition attachment style, used in these two\nformalisms, applied in multi-source cross-lingual delexicalized dependency\nparser transfer performed by parse tree combination.\nWe show that in our setting, converting the adposition annotation to Stanford style in the\nPrague style training treebanks leads to promising results.\nWe find that best results can be obtained by parsing the target sentences\nwith parsers trained on treebanks using both of the adposition annotation\nstyles in parallel,\nand combining all the resulting parse trees together\nafter having converted them to the Stanford adposition style (+0.39% UAS over Prague\nstyle baseline).\nThe score improvements are considerably more significant when using a smaller set of diverse source\ntreebanks (up to +2.24% UAS over the baseline).
The main goal of this research is to build a sentiment analysis system which automatically determines user opinions of the Stanford Sentiment Treebank in terms of three sentiments such as positive, negative, and neutral. Firstly, sentiment sentences are POS tagged and parsed to dependency structures. All nodes of the Treebank and their polarities are automatically extracted from the Treebank. We train two Support Vector Machines models. One is for a node level classification and the other is for a sentence level. We have tried various type of features such as word lexicons, POS tags, Sentiment lexicons, head-modifier relations, and sibling relations. Though we acquired 74.2% in accuracy on the test set for 3 class node level classification and 67.0% for 3 class sentence level classification, our experimental results for 2 class classification are comparable to those of the state of art system using the same corpus.
Developing an app version of a printed dictionary is a new challenge faced by lexicographers. Lexicographers involved in the app development process must consider fundamental lexicographic aspects as well as learn to understand technological and usage issues inherent to the new media. An inevitable question is how closely the content and layout can be made to match the printed dictionary while still offering ‘digital’ functionality such as linking, collapsed sections, audio, etc. Only a few reports discussing these issues have so far been published. The aim of our paper is to further advance the exchange of knowledge and experience by sharing our observations made during the development of a new app corresponding to the comprehensive printed dictionary, Svensk ordbok utgiven av Svenska Akademien (the Contemporary Dictionary of the Swedish Academy, 2009). The app is the result of close cooperation between the financer The Swedish Academy, lexicographers and system developers at the Department of Swedish, University of Gothenburg and Isolve AB, a Stockholm-based app development agency specializing in dictionary apps.
We propose a model of Tibetan syntactic parsing which is based on Tibetan syl lables instead of Tibetan words, change Tibetan syntactic Treebank use algorith m of labeling.
Presentations given at the workshop Language Comparison with Linguistic Databases (LanCLiD 2), held on 30 April 2015 at the Department of Linguistics of the Max Planck Institute for Evolutionary Anthropology (Leipzig, Germany).
We describe a well-performed semantic role labeling system that further extracts concepts (smaller semantic expressions) from unstructured natural language sentences language independently. A dual-layer semantic role labeling (SRL) system is built using Chinese Treebank and Propbank data. Contextual information is incorporated while labeling the predicate arguments to achieve better performance. Experimental results show that the proposed approach is superior to CoNLL 2009 best systems and comparable to the state of the art with the advantage that it requires no feature engineering process. Concepts are further extracted according to templates formulated by the labeled semantic roles to serve as features in other NLP tasks to provide semantically related cues and potentially help in related research problems. We also show that it is easy to generate a different language version of this system by actually building an English system which performs satisfactory.
In this paper, we tackle the problem of domain adaptation of statistical machine translation (SMT) by exploiting domain-specific data acquired by domain-focused crawling of text from the World Wide Web. We design and empirically evaluate a procedure for automatic acquisition of monolingual and parallel text and their exploitation for system training, tuning, and testing in a phrase-based SMT framework. We present a strategy for using such resources depending on their availability and quantity supported by results of a large-scale evaluation carried out for the domains of environment and labour legislation, two language pairs (English–French and English–Greek) and in both directions: into and from English. In general, machine translation systems trained and tuned on a general domain perform poorly on specific domains and we show that such systems can be adapted successfully by retuning model parameters using small amounts of parallel in-domain data, and may be further improved by using additional monolingual and parallel training data for adaptation of language and translation models. The average observed improvement in BLEU achieved is substantial at 15.30 points absolute.