Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
16504 papers
This is a small hand-annotated partial treebank of Modern Tibetan, primarily in CoNLL-U format. Some texts were POS-tagged by machine, and then dependency relations between verbs and their arguments were added by hand. Other texts include only dependency relations and relevant POS-tags. A number of the texts have English translations which have been manually aligned to the Tibetan text. This work was created as part of the AHRC-funded project <em>Lexicography in Motion</em> (PI Ulrich Pagel, 2017-2021).
[Context:] Causal relations (e.g., If A, then B) are prevalent in functional\nrequirements. For various applications of AI4RE, e.g., the automatic derivation\nof suitable test cases from requirements, automatically extracting such causal\nstatements are a basic necessity. [Problem:] We lack an approach that is able\nto extract causal relations from natural language requirements in fine-grained\nform. Specifically, existing approaches do not consider the combinatorics\nbetween causes and effects. They also do not allow to split causes and effects\ninto more granular text fragments (e.g., variable and condition), making the\nextracted relations unsuitable for automatic test case derivation. [Objective &\nContributions:] We address this research gap and make the following\ncontributions: First, we present the Causality Treebank, which is the first\ncorpus of fully labeled binary parse trees representing the composition of\n1,571 causal requirements. Second, we propose a fine-grained causality\nextractor based on Recursive Neural Tensor Networks. Our approach is capable of\nrecovering the composition of causal statements written in natural language and\nachieves a F1 score of 74 % in the evaluation on the Causality Treebank. Third,\nwe disclose our open data sets as well as our code to foster the discourse on\nthe automatic extraction of causality in the RE community.\n
Older adults may be better able to modulate their emotional experiences than younger adults, and thus may recover more quickly from negative stressors. Additionally, older adults may be more likely to experience co-occurrence of negative and positive emotions in the setting of negative stressors, which may facilitate emotion recovery. To date, few studies have investigated the nature of age group differences in spontaneous emotional responses to a standardized stressor. The current study utilizes a laboratory mood manipulation to determine age group differences in emotion recovery in negative and positive affects, as well as age group differences in the co-occurrence of negative and positive affect. Older adults reported greater reactivity in one and greater recovery in two negative affect scales than younger adults; however, these differences did not remain significant when controlling for overall arousal ratings of the mood induction. There were no age group differences in reactivity or recovery of positive affects. Both younger and older adults returned to baseline in negative affects by the end of the recovery period despite age group differences in affect responses and arousal ratings. Older adults reported greater co-occurrence of negative and positive emotions in response to the mood induction as compared to younger adults. Overall, these results provide support for age group similarities in reactivity and recovery in discrete affects, and age group differences in mixed emotion states. Greater co-occurrence appears to reflect greater baseline endorsement of positive affect in older as compared to younger adults. Thus, higher baseline positive affect may create greater opportunities for older adults to experience mixed emotion states, which may in turn serve as an adaptive resource for older adults.
Recent work has shown that monolingual masked language models learn to represent data-driven notions of language variation which can be used for domain-targeted training data selection. Dataset genre labels are already frequently available, yet remain largely unexplored in cross-lingual setups. We harness this genre metadata as a weak supervision signal for targeted data selection in zero-shot dependency parsing. Specifically, we project treebank-level genre information to the finer-grained sentence level, with the goal to amplify information implicitly stored in unsupervised contextualized representations. We demonstrate that genre is recoverable from multilingual contextual embeddings and that it provides an effective signal for training data selection in cross-lingual, zero-shot scenarios. For 12 low-resource language treebanks, six of which are test-only, our genre-specific methods significantly outperform competitive baselines as well as recent embedding-based methods for data selection. Moreover, genre-based data selection provides new state-of-the-art results for three of these target languages.
Media language is a prototype of the public consent for the media to be defined through compromise as a fourth position in the paradigm of power as a philosophical category, whose explications before the media are legislative, executive, judicial. The linguistic norm and the cognitive-rhetorical characteristic of the media discourse are the prototype of the metaphor of the "fourth power". The formation of the information-language culture and the preservation of the language norm is the high social responsibility of the media discourse. The media is a prototype of public consciousness, a “picture” of national identity – a unit of political and socio-economic information and cultural “taste” (a sample of art and its list).
Recent work has shown that monolingual masked language models learn to represent data-driven notions of language variation which can be used for domain-targeted training data selection. Dataset genre labels are already frequently available, yet remain largely unexplored in cross-lingual setups. We harness this genre metadata as a weak supervision signal for targeted data selection in zero-shot dependency parsing. Specifically, we project treebank-level genre information to the finer-grained sentence level, with the goal to amplify information implicitly stored in unsupervised contextualized representations. We demonstrate that genre is recoverable from multilingual contextual embeddings and that it provides an effective signal for training data selection in cross-lingual, zero-shot scenarios. For 12 low-resource language treebanks, six of which are test-only, our genre-specific methods significantly outperform competitive baselines as well as recent embedding-based methods for data selection. Moreover, genre-based data selection provides new state-of-the-art results for three of these target languages.
The present study investigates the effect of background luminance on the self-reported valence ratings of auditory stimuli, as suggested by some earlier work. A secondary aim was to better characterise the effect of auditory valence on pupillary responses, on which the literature is inconsistent. Participants were randomly presented with sounds of different valence categories (negative, neutral, and positive) obtained from the IADS-E database. At the same time, the background luminance of the computer screen (in blue hue) was manipulated across three levels (i.e., low, medium, and high), with pupillometry confirming the expected strong effect of luminance on pupil size. Participants were asked to rate the valence of the presented sound under these different luminance levels. On a behavioural level, we found trend-level evidence for a small effect of background luminance on the self-reported valence rating, with generally more positive ratings as background luminance increases. Turning to valence effects on pupil size, irrespective of background luminance, interestingly, we observed that pupils were smallest in the positive valence and the largest in negative valence condition, with neutral sounds in between. In sum, the present findings therefore provide some evidence concerning the relationship between luminance perception (and hence pupil size) and self-reported valence of auditory stimuli, indicating a possible cross-modal interaction of auditory valence processing with completely task-irrelevant visual background luminance. The present experiment furthermore contributes new data on the relationship between valence and pupil size for auditory stimuli.
Statistical learning (SL) refers to the ability to extract regularities in the environment and has been well-documented to play a key role in speech segmentation and language acquisition. Whether SL requires top-down attention is an unresolved question. The current study examined whether SL can occur outside the focus of attention. Participants either focused or diverted their attention away from a heard nonsense language. Visual attention was taxed by requiring tracking of multiple randomly moving dots. Linguistic attention was taxed through a self-paced reading task. SL was assessed with an explicit familiarity rating task, and an implicit reaction-time (RT) based memory task. Explicit learning was only reduced when linguistic resources were taxed, but unimpaired when visual resources were taxed. On our implicit measure of SL, learning was unimpaired when attention was taxed. These results suggest implicit aspects of SL to be more robust to an attention diversion as compared to explicit aspects of SL. Findings suggest L2 learners may engage in demanding visual tasks, as well as offer insight into the neurocognitive underpinnings of SL.
Abstract Depression is characterized by disturbed emotional processing, with increased amygdala reactivity during perception of negative emotional stimuli. Here, we assessed whether intermittent theta-burst stimulation (iTBS) modulates amygdala activity during an emotional picture anticipation paradigm using functional magnetic resonance imaging (fMRI) in a patient sample with depression. Patients were randomized to either active (n=21) or sham (n=21) iTBS over the dorsomedial prefrontal cortex (DMPFC), delivered twice daily for ten days at target intensity. Depression symptom assessment and fMRI scanning took place just before treatment start and once again four weeks later. During fMRI scanning, picture stimuli of negative and positive valence were presented, indicated by a prior red or green screen, respectively. Behavioral valence ratings of the picture stimuli were conducted outside of the scanner. Amygdala activation during perception of negative picture stimuli was reduced after active, but not sham, iTBS (left amygdala: F(1,25)= 7.22, p=.013, right amygdala: F(1,25)=8.65, p=.007). Baseline amygdala reactivity was not correlated with depressive symptom levels. Behavioral valence ratings remained unchanged after active iTBS. The findings suggest a treatment effect of active iTBS in an emotional processing network in depression, spanning the DMPFC and amygdala. Keywords: dorsomedial prefrontal cortex, emotion, iTBS, transcranial magnetic stimulation
This article presents a method for automatic assignment of syntactic dependency relations to the corpus of American Norwegian speech (CANS). Different machine learning techniques and corpora are used. Finally, an accuracy measure is computed and compared with a relatively new treebank for spoken Norwegian.
DEFI is a prototype computer tool aimed at ranking (from most to least relevant) the French translations of an English lexical item in context. This paper deals with the strategies used by DEFI to recognize multi-word units (mwus) in running text. Any lexical unit included in the lexical database used in the project (a merge of the Oxford/Hachette and Robert/Collins English-to-French dictionaries) and longer than a single word is submitted to a surface parser, and the same process is applied to the user ’s text. A program written in Prolog assesses the quality of the match between the parsed user’s text and candidate mwus retrieved from the project’s lexical database. The matcher is able to account for some of the distortions undergone by the mwu, e.g. movement of a constituent as a result of relativization or passivization.
A linguistic norm (literary norm) is the rules for the use of speech means in a certain period of the development of the literary language, i.e. rules of pronunciation, word use, use of traditionally established grammatical, stylistic and other linguistic means adopted in social and linguistic practice. This is a uniform, exemplary, generally recognized use of language elements (words, phrases, sentences). Linguistic norms are a historical phenomenon. Changes in literary norms are due to the constant development of the language. What was the norm in the last century and even 15-20 years ago today can become a deviation from it. The norms help the literary language to maintain its integrity and comprehensibility. They protect the literary language from the flow of dialectal speech, social and professional jargon, and vernacular. This allows the literary language to fulfill its main function - cultural.
Space-valence metaphors (e.g., bad is down) are embedded within cognitive and emotional processing (e.g., negative stimuli at a lower space capture visual attention more than those at an upper space). Previous studies have revealed that motor action to vertical direction affects the emotional valence rating of stimuli in a metaphor-congruent manner only when the action was introduced after the stimuli presentation. In the present study, we hypothesized that motor action before the stimuli presentation does not affect valence rating while it may affect visual selective attention. In Experiment 1 (participants: 28 university students; mean age=19.50 years), we partially replicated the previous result with repeated ANOVA and t -tests; manual action introduced before the stimuli presentation does not affect the valence rating. Then, in Experiment 2 (participants: 28 university students; mean age=19.57 years), we employed a modified version of the dot-probe task as a measure of visual selective attention to emotional stimuli, where participants' vertical or horizontal manual action was introduced before the presentation of a pair of emotional words. The results of the t -tests revealed that an upward manual action promoting selective attention to negative words, which was incongruent with the space-valence metaphorical correspondence. These results suggest that even though manual action does not affect the evaluative process of emotional stimuli prospectively, upward manual action introduced before stimuli presentation can promote visual attention to the subsequent negative stimuli in a way that is incongruent with the space-valence metaphor.
Constituency parsing is generally evaluated superficially, particularly in a multiple language setting, with only F-scores being re-ported. As new state-of-the-art chart-based parsers have resulted in a transition from traditional PCFG-based grammars to span-based approaches (Stern et al., 2017; Gaddy et al.,2018), we do not have a good understanding of how such fundamentally different approaches interact with various treebanks as results show improvements across treebanks (Kitaev and Klein, 2018), but it is unclear what influence annotation schemes have on various treebank performance (Kitaev et al., 2019). In particular, a span-based parser’s capability of creating novel rules is an unknown factor. We perform an analysis of how span-based parsing performs across 11 treebanks in order to examine the overall behavior of this parsing approach and the effect of the treebanks’ specific annotations on results. We find that the parser tends to prefer flatter trees, but the approach works well because it is robust enough to adapt to differences in annotation schemes across treebanks and languages.
Image beauty assessment is an important subject of computer vision. Therefore, building a model to mimic the image beauty assessment becomes an important task. To better imitate the behaviours of the human visual system (HVS), a complete survey about images of different categories should be implemented. This work focuses on image beauty assessment. In this study, the pairwise evaluation method was used, which is based on the Bradley-Terry model. We believe that this method is more accurate than other image rating methods within an image group. Additionally, Convolution neural network (CNN), which is fit for image quality assessment, is used in this work. The first part of this study is a survey about the image beauty comparison of different images. The Bradley-Terry model is used for the calculated scores, which are the target of CNN model. The second part of this work focuses on the results of the image beauty prediction, including landscape images, architecture images and portrait images. The models are pretrained by the AVA dataset to improve the performance later. Then, the CNN model is trained with the surveyed images and corresponding scores. Furthermore, this work compares the results of four CNN base networks, i.e., Alex net, VGG net, Squeeze net and LSiM net, as discussed in literature. In the end, the model is evaluated by the accuracy in pairs, correlation coefficient and relative error calculated by survey results. Satisfactory results are achieved by our proposed methods with about 70 percent accuracy in pairs. Our work sheds more light on the novel image beauty assessment method. While more studies should be conducted, this method is a promising step.
Background: This manuscript evaluates patient and provider perspectives on the core components of a Behavioral Health Home (BHH) implemented in an urban, safety-net health system. The BHH integrated primary care and wellness services (e.g., on-site Nurse Practitioner and Care Manager, wellness groups and tools, population health management) into an existing outpatient clinic for people with serious mental illness (SMI). Methods: As the qualitative component of a Hybrid Type I effectiveness-implementation study, semi-structured interviews were conducted with providers and patients 6 months after program implementation, and responses were analyzed using thematic analysis. Valence coding (i.e., positive vs. negative acceptability) was also used to rate interviewees' transcriptions with respect to their feedback of the appropriateness, acceptability, and feasibility/sustainability of 9 well-described and desirable Integrated Behavioral Health Core components (seven from prior literature and two additional components developed for this intervention). Themes from the thematic analysis were then mapped and organized by each of the 9 components and the degree to which these themes explain valence ratings by component. Results: Responses about the team-based approach and universal screening for health conditions had the most positive valence across appropriateness, acceptability, and feasibility/sustainability by both providers and patients. Areas of especially high mismatch between perceived provider appropriateness and measures of acceptability and feasibility/sustainability included population health management and use of evidence-based clinical models to improve physical wellness where patient engagement in specific activities and tools varied. Social and peer support was highly valued by patients while incorporating patient voice was also found to be challenging. Conclusions: Findings reveal component-specific challenges regarding the acceptability, feasibility, and sustainability of specific components. These findings may partly explain mixed results from BHH models studied thus far in the peer-reviewed literature and may help provide concrete data for providers to improve BHH program implementation in clinical settings. Plain language abstract: Many people with serious mental illness also have medical problems, which are made worse by lack of access to primary care. The Behavioral Health Home (BHH) model seeks to address this by adding primary care access into existing interdisciplinary mental health clinics. As these models are implemented with increasing frequency nationwide and a growing body of research continues to assess their health impacts, it is crucial to examine patient and provider experiences of BHH implementation to understand how implementation factors may contribute to clinical effectiveness. This study examines provider and patient perspectives of acceptability, appropriateness, and feasibility/sustainability of BHH model components at 6-7 months after program implementation at an urban, safety-net health system. The team-based approach of the BHH was perceived to be highly acceptable and appropriate. Although providers found certain BHH components to be highly appropriate in theory (e.g., population-level health management), their acceptability of these approaches as implemented in practice was not as high, and their feedback provides suggestions for model improvements at this and other health systems. Similarly, social and peer support was found to be highly appropriate by both providers and patients, but in practice, at months 6-7, the BHH studied had not yet developed a process of engaging patients in ongoing program operations that was highly acceptable by providers and patients alike. We provide these data on each specific BHH model component, which will be useful to improving implementation in clinical settings of BHH programs that share some or all of these program components.
Background: Whether affective states acutely predict the hypothalamic–pituitary–adrenal (HPA) axis activities and whether energy balance-related behaviors moderate the affect–HPA axis relationship in obese youths are not well-understood. Methods: 87 mostly obese (94.3% obese) minority adolescents (mean: 16.3 ± 1.2 years old; 56.8% Latino and 43.2% African American) participated in a randomized crossover trial in an observation laboratory, where they received either high-sugar/low-fiber (HSLF) or low-sugar/high-fiber (LSHF) meals first and then crossed over in the next visit 2–4 weeks later. During each visit, they rated five affective states and provided a saliva sample every 30 min for the first 5 h and wore a waist-worn accelerometer. The association between the affect ratings and cortisol levels in the subsequent 30 min and the moderation effect of energy balance-related behavior were examined using multilevel models. Results: Within-person negative affect (β = 0.02, p = 0.0343) and feeling of panic (β = 0.007, p = 0.004) were acutely related to the subsequent cortisol level only during the HSLF condition. The time spent in moderate-to-vigorous physical activity did not moderate the acute relationship between affect and the subsequent cortisol level. Conclusions: Negative affect could be acutely related to heightened HPA axis activities in youths, but only when they were exposed to meals with high sugar and low fiber content. These results suggest that the meals’ sugar and fiber content may modulate HPA axis reactivity to negative affect in youths.
<h3>Introduction</h3><br> BOLT Egyptian Arabic SMS/Chat Parallel Training Data was developed by LDC and consists of approximately 723,000 tokens of Egyptian Arabic SMS/Chat data collected for the DARPA BOLT program along with their corresponding English translations. <br> The DARPA <a href="https://www.ldc.upenn.edu/collaborations/current-projects/bolt"> BOLT</a> (Broad Operational Language Translation) program developed machine translation and information retrieval for less formal genres, focusing particularly on user-generated content. LDC supported the BOLT program by collecting informal data sources -- discussion forums, text messaging and chat -- in Chinese, Egyptian Arabic and English. The collected data was translated and annotated for various tasks including word alignment, treebanking, propbanking and co-reference. <br> <h3>Data</h3><br> The source date in this release was collected using two methods: new collection via LDC's collection platform, and donation of SMS or chat archives from BOLT collection participants. All data were reviewed manually to exclude any messages/conversations that were not in the target language or that had sensitive content, such as personal identifying information. <br> Data was manually selected for translation. Messages/conversations were arranged in chronological order, segmented into sentence units (all or portions of message threads depending on their length), and assigned to translation vendors. Translators followed LDC's BOLT translation guidelines. <br> Source and translation files are presented in UTF-8 encoded XML format. <br> <h3>Sponsorship</h3><br> This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR0011-11-C-0145. The content does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred. <br> <h3>Samples</h3><br> Please view this <a href="desc/addenda/LDC2021T15.arz.xml">Egyptian Arabic sample (XML)</a> and <a href="desc/addenda/LDC2021T15.eng.xml">English sample (XML)</a>. <br> <h3>Updates</h3><br> None at this time. </br> Portions © 2021 Trustees of the University of Pennsylvania
The article presents an attempt of analyzing and summarizing the experience the author acquired over a number of years and which is connected with the problems of the Russian (L1, native) language acquisition by schoolchildren in Latvia. The author aims to describe the speech portrait of Russian-speaking schoolchildren in Latvia. The material presented in the article is based on the results of the author’s long-term research. It includes: 1) Russian language textbooks (grades 4–9) development, testing and monitoring, 2007–2019; 2) project work which has been and still is focused on the analysis of Russian speech quality among university students and schoolchildren, 2008–2019; 3) research into the quality of writing skills of university students in Latvia (University of Latvia and RISEBA), 2013–2019; 4) the author’s scholarly and methodological dialogue with the teachers of Russian in Latvia for more than twenty years. The author mainly focuses on several problems. Firstly, it is the analysis of the oral speech of schoolchildren and of the Russian verbal environment whose developing potential is not high enough for the successful development of a child’s language personality. The natural Russian verbal environment (media language, everyday speech) often offers low-quality material. Secondly, the author carries out a special study of the quality of schoolchildren’s writing skills as, according to scholars, it is written language, due to its specificity, that most vividly indicates the development of a language personality. Taking into account the results of the research, the author points out that there are many serious problems in the verbal behavior of most Russian-speaking schoolchildren. One of them is departures from linguistic norms at all linguistic system levels. The nature of these departures depends directly on the model of bilingual education that is implemented at schools children attend. The author pays attention to the problem areas of the schoolchildren’s written language: 1) breaking of the spelling and punctuation rules of Russian; 2) confusion in speech of words and collocations belonging to different styles, lack of sense of style; 3) quality of written texts created by schoolchildren: problems with text structure; schoolchildren more often use explicit “I-communication” (reflective statements, texts with egocentric perspective), have a rather limited vocabulary; schoolchildren’s essays contain very few intertextual inclusions, as a result, their texts seem very poor semantically; schoolchildren have no desire to express emotions in their writing (rational/pragmatic description of a problem, situation prevails). The author concludes that the development of a diaspora schoolchild’s language personality in L1 (the Russian language) is not successful enough.
Indonesia is one of the largest coffee exporting countries in the United States market after Brazil, Colombia, Vietnam, and Guatemala. It is still unable to shift the export of coffee commodities from these four countries. This research aims to analyze the competitiveness and performance of coffee exports in the United States market using data analysis methods such as Revealed Symmetric Comparative Advantage (RSCA) and Constant Market Share (CMS). Research is classified as quantitative research that utilizes secondary data, an annual time series data, namely 2010-2019. The data source is exported data for Indonesian coffee commodity digit 6 with HS 090111 (Coffee, not roasted, not decaffeinated) obtained from the International Trade Center (ITC). This study's value results indicate that RSCA Indonesia is 0.87, where the RSCA is> 0. This shows that Indonesia still has competitiveness, although it is lower than Brazil0.95, Colombia, 0.96, and Guatemala, 0.97, and Indonesia is still superior to Vietnam, which is equivalent. 0.79. Meanwhile, the CMS value states that the Indonesian coffee commodity is less desirable in the United States market with an average commodity composition effect value of -0.00006. However, an increase in demand for Indonesian coffee commodities with an average market distribution effect value of 0.00002 and commodity Indonesian coffee has a competitive edge. Strong in the US market with an average competitiveness affect rating of 0.00001.
Words in the natural language have forms and meanings, and there might not always be a one-to-one match between them. This property of the language causes words to have more than one meaning; as a result, a text processing system faces challenges to determine the precise meaning of the target word in a sentence. Using lexical resources or lexical databases, such as WordNet, might be a help, but due to their manual development, they become outdated by passage of time and language change. Moreover, the lexical resources might be domain dependent which are unusable for open domain natural language processing tasks. These drawbacks are a strong motivation to use unsupervised machine learning approaches to induce word senses from the natural data. To reach the goal, the clustering approach can be utilized such that each cluster resembles a sense. In this paper, we study the performance of a word sense induction model by using three variables: a) the target language: in our experiments, we run the induction process on Persian and English; b) the type of the clustering algorithm: both parametric clustering algorithms, including hierarchical and partitioning, and non-parametric clustering algorithms, including probabilistic and density-based, are utilized to induce senses; c) the context of the target words to capture the information in vectors created for clustering: for the input of the clustering algorithms, the vectors are created either based on the whole sentence in which the target word is located; or based on the limited surrounding words of the target word. We evaluate the clustering performance externally. Moreover, we introduce a normalized, joint evaluation metric to compare the models. The experimental results for both Persian and English test data showed that the window-based partitioningK-means algorithm obtained the best performance.
This article discusses one of the forms of machine translation, the Instagram translation feature called “see translation”. The research is focused on the translation techniques applied by the machine in translating Banyumas batik motifs from Indonesian to English found in @batikantodjamil and @batk_rd. This topic is worth discussing since machine translation is now getting more developed and is projected to replace human translator. However, in some cases, for example in dealing with culturally-bound terms, machine translation cannot perform contextual knowledge as well as the human translator. this mini research was conducted by applying qualitative research with purposive sampling technique in which the researchers obtain the data by selecting two batik center Instagram accounts containing batik motif names in the captions. The result shows that there are three translation techniques applied by the Instagram translation features, namely literal, borrowing, and particularization. The most dominant technique to use is borrowing technique, and it shows a tendency that such cultural terms in the source language do not have one-to-one correspondence in the target language. In other words, the touch of human translator is very important in the post-editing process of translation by machine to make the translation more acceptable. However, if it is impossible to involve human translator, the Instagram administrator should enrich the machine with more contextual linguistic database to provide the users with better translation results.
Long short-term memory (LSTM) applications need fast yet compact models. Neural network compression approaches, e.g., the grow-and-prune paradigm, have proved to be promising for cutting down network complexity by skipping insignificant weights. However, current compression strategies remain mostly hardware-agnostic and network complexity reduction does not always translate to execution efficiency. In this work, we propose a hardware-guided symbiotic training methodology for compact, accurate, yet execution-efficient inference models. It is based on our observation that hardware may introduce substantial non-monotonic behavior, which we call the latency hysteresis effect, when evaluating network size versus inference latency. This observation raises question about the mainstream smaller-dimension-is-better compression strategy, which often leads to a sub-optimal model architecture. Leveraging the hardware-impacted hysteresis effect and sparsity, we enable a symbiosis of model compactness and accuracy with execution efficiency, thus reducing LSTM latency while increasing its accuracy. We have evaluated our approach on language modeling and speech recognition applications. Relative to the traditional stacked LSTM architecture obtained for the Penn Treebank dataset, we reduce the number of parameters by 18.0× (30.5×) and measured run-time latency by up to 2.4× (5.2×) on Nvidia GPUs (Intel Xeon CPUs) without any accuracy degradation. For the DeepSpeech2 architecture obtained for the AN4 dataset, we reduce the model size by 7.0× (19.4×), word error rate from 12.9% to 9.9% (10.4%), and measured run-time latency by up to 1.7× (2.4×) on Nvidia GPUs (Intel Xeon CPUs). Our method consistently outperforms prior art for both applications, with compact, accurate, and execution-efficient inference models.
BACKGROUND: Studies on food cue reactivity have documented that altered responses to high-calorie food are associated with bulimic symptomatology, however, alterations in sexual motivations and behaviors are also associated clinical features in this population, which justify their inclusion as a research target. Here, we study responses to erotic cues-alongside food, neutral and aversive cues-to gain an understanding of specificity to food versus a generalized sensitivity to primary reinforcers. METHODS: We recorded peripheral psychophysiological indices -the startle reflex, zygomaticus, and corrugator responses-and self-reported emotional responses (valence, arousal, and dominance) in 75 women completing the Bulimia Test-Revised (BULIT-R). Multiple regression analysis tested whether BULIT-R symptoms were predicted by self-reported and psychophysiological responses to food versus neutral and erotic versus neutral images. RESULTS: The results showed that individuals with higher bulimic symptoms were characterized by potentiated eye blink startle response during binge food (vs. neutral images) and more positive valence ratings during erotic (vs. neutral) cues. CONCLUSIONS: The results highlight the negative emotional reactivity of individuals with elevated bulimic symptoms toward food cues, which could be related to the risk of progression to full bulimia nervosa and thereby addressed in prevention efforts. Results also point to the potential role of reactivity to erotic content, at least on a subjective level. Theoretical models of eating disorders should widen their conceptual scope to consider reactivity to a broader spectrum of primary reinforcers, which would have implications for cue exposure-based treatments. We examined appetitive and aversive cue responses in college women to investigate how bulimic symptoms relate to primary reinforcers such as food and erotic images. We recorded peripheral psychophysiological indices (the startle reflex, zygomaticus, and corrugator responses) and self-reported emotional responses (valence, arousal, and dominance) in 75 college women that were presented with the Spanish version of the Bulimia Test-Revised. The results showed that bulimic symptoms increase both psychophysiological defensiveness toward food cues and subjective pleasure toward erotic cues. The findings suggest a generalized sensitivity to primary reinforcers in the presence of bulimic symptoms, and emphasize the relevance of adopting a wider framework in research and treatment on bulimia nervosa.
Recent impressive improvements in NLP, largely based on the success of contextual neural language models, have been mostly demonstrated on at most a couple dozen high-resource languages. Building language models and, more generally, NLP systems for non-standardized and low-resource languages remains a challenging task. In this work, we focus on North-African colloquial dialectal Arabic written using an extension of the Latin script, called NArabizi, found mostly on social media and messaging communication. In this low-resource scenario with data displaying a high level of variability, we compare the downstream performance of a character-based language model on part-of-speech tagging and dependency parsing to that of monolingual and multilingual models. We show that a character-based model trained on only 99k sentences of NArabizi and fined-tuned on a small treebank of this language leads to performance close to those obtained with the same architecture pre-trained on large multilingual and monolingual models. Confirming these results a on much larger data set of noisy French user-generated content, we argue that such character-based language models can be an asset for NLP in low-resource and high language variability set-tings.
This paper examines the use of Italian digital language, which is often evaluated in negative terms. Considering the fact that internet communication occupies an important place in the life of modern man, the study of the features of digital language has been the subject of much research. For those born in the digital age (it. nativi digitali), digital has become the norm to the extent that it is difficult to imagine life without multimedia interaction through modern means of communication (Bralić 145). Digital text is different from traditional written text and the rapid obsolescence of new media is changing the habits of digital language users. Italian, which has existed exclusively in the traditional written form for centuries, and has received full spoken use in the last seventy years (largely thanks to television), faces today a new revolutionary phase of development in which the majority of Italians in everyday life use written digital language. In this way, the digital age marked a return to the Italian written language. However, the language of forums and social networks is an informal language (e-Italian), quite different from the former, exceptionally formal, written Italian. The aim of this paper is to study and explain the linguistic features of the Italian language in Internet communication. The focus is on the language of blogs, forums, and social networks written in Italian over the last three years, from the beginning of 2018 to the end of 2020. The question is whether everything that deviates from the norm in the language is wrong or if, on the contrary, demonstrates the stability and ability of the language to adapt to new media and thus new conditions. The major changes on social networks are the result of the transition from the elite use of the network to the “mass network” (Gheno 2017, 103). The changes are also heading towards the direction that has yet to be identified. Thus, we notice that the use of certain language features on social networks such as abbreviations, acronyms and other similar phenomena was a way of distinction, but also a necessity dictated by technical limitations such as restricted space for writing messages and the high cost of network connection. Therefore, it comes to no surprise that in recent years we have witnessed a writing normalization directed towards approximating some kind of linguistic norm. Finally, after having removed the space and time limitations and as a result of the possibility of spell checking that is suggested by smart devices while writing, even the so-called “language play and use of creative forms of writing” has become practically a waste of time. The fact that we are in the normalization phase can also be seen thanks to other novelties on social networks. One of them is caused by the policy of some platforms that is aimed at using one’s own name and abandoning the nickname, leading to an interesting social effect demonstrating that haters do not necessarily hide behind nicknames. Moreover, there is a tendency to give more importance to the interlocutor who signs with his own name, as contrasted with those who use a nickname. It becomes normal again to introduce yourself by your real name and surname, without leaving the impression of a person that is hidden behind a mask or nickname. The use of language on social networks has changed thoroughly over time and continues to change even today, both in Italian and in other languages. It is highly probable that over time users will pay more attention to the impression they leave online and, thus, be more careful when it comes to the language, they use by respecting the prescribed language norms. In addition to dealing with language dilemmas, it is necessary to establish the right habits that will allow us to live a comfortable life online and accept the fact that we have become like mini public figures who are responsible for what they say. We should also keep in mind that, on social networks, the most emphasized part of our online personality is presented primarily by words.
Subject inversion in French is usually considered to be optional (Le Bidois 1952; Kayne & Pollock 1978) and more costly than variants with preverbal subject. As the result of verb movement (Hulk & Pollock 2001), it is claimed to demand higher processing cost (Holmes & O’Regan 1981). However, some studies suggest that subject inversion in relative clauses may even be favoured by certain semantic or heaviness constraints (Fuchs 2006; Marandin 2011). In this paper, we take an empirical approach to this question. In our corpus study using the French Treebank described<br> in Abeillé et al. (2019), we found that subject inversion in object relatives can be as frequent as cases without inversion. We also found that inversion is preferred with longer subjects and shorter and non-agentive verbs. This pattern was confirmed in an acceptability judgement experiment as well as in a self-paced reading experiment. Thus, object relatives with and without inversion are not merely stylistic variants (i.e. two equivalent syntactic ways of expressing one meaning), but are more or less preferred depending on their properties. Our results are compatible with semantic accounts of relative clause processing (Mak et al. 2006; Traxler et al. 2002).
Language Identification in textual documents is the process of automatically detecting the language contained in a document based on its content. The present Language Identification techniques presume that a document contains text in one of the fixed set of languages, however, this presumption is incorrect when dealing with multilingual document which includes content in more than one possible language. Due to the unavailability of large standard corpora for Hindi-English mixed lingual language processing tasks we propose the language lexicons, a novel kind of lexical database that supports several multilingual language processing tasks. These lexicons are built by learning classifiers over transliterated Hindi and English vocabulary. The designed lexicons possess richer quantitative characteristic than its primary source of collection which is revealed using the visualization techniques.
Statistical machine translation (SMT) approaches extract translation knowledge automatically from parallel corpora. They additionally take advantage of monolingual text for target-side language modelling. Syntax-based SMT approaches also incorporate knowledge of source and/or target syntax by taking advantage of monolingual grammars induced from treebanks, and semantics-based SMT approaches use knowledge of source and/or target semantics in various forms. However, there has been very little research on incorporating the considerable monolingual knowledge encoded in deep, hand-built grammars into statistical machine translation. Since deep grammars can produce semantic representations, such an approach could be used for realization as well as MT. In this thesis I present a hybrid approach combining some of the knowledge in a deep hand-built grammar, the English Resource Grammar (ERG), with a statistical machine translation approach. The ERG is used to parse the source sentences to obtain Dependency Minimal Recursion Semantics (DMRS) representations. DMRS representations are subsequently transformed to a form more appropriate for SMT, giving a parallel corpus with transformed DMRS on the source side and aligned strings on the target side. The SMT approach is based on hierarchical phrase-based translation (Hiero). I adapt the Hiero synchronous context-free grammar (SCFG) to comprise graph-to-string rules. DMRS graph-to-string SCFG is extracted from the parallel corpus and used in decoding to transform an input DMRS graph into a target string either for machine translation or for realization. I demonstrate the potential of the approach for large-scale machine translation by evaluating it on the WMT15 English-German translation task. Although the approach does not improve on a state-of-the-art Hiero implementation, a manual investigation reveals some strengths and future directions for improvement. In addition to machine translation, I apply the approach to the MRS realization task. The approach produces realizations of high quality, but its main strength lies in its robustness. Unlike the established MRS realization approach using the ERG, the approach proposed in this thesis is able to realize representations that do not correspond perfectly to ERG semantic output, which will naturally occur in practical realization tasks. I demonstrate this in three contexts, by realizing representations derived from sentence compression, from robust parsing, and from the transfer-phase of an existing MT system. In summary, the main contributions of this thesis are a novel architecture combining a statistical machine translation approach with a deep hand-built grammar and a demonstration of its practical usefulness as a large-scale machine translation system and a robust realization alternative to the established MRS realization approach.
This paper uses the COVID-19 pandemic as an extreme case to test whether reference points affect how citizens use performance information on effectiveness, cost, and equality. Drawing on the evaluability hypothesis, the paper argues that citizens are more likely to make decisions based on performance information on equality and disregard performance information on effectiveness and costs when no reference points are available to aid interpretation. The paper uses a pre-registered between-subject conjoint survey experiment on 2,025 Danish citizens to test expectations. Respondents were randomly drawn to rate either one fictive government strategy to combat the Coronavirus—with no opportunity to compare performance information between strategies—or two strategies—with the opportunity to compare performance information between strategies. The strategies varied on effectiveness (mortality rate), costs (overall economic costs) and equality (distribution of the economic costs and access to testing). Results show that when respondents are presented with one strategy, only performance information on equality affects ratings. Strategies with lower fatality and lower economic costs are thus not rated higher than strategies with higher fatality and higher economic costs holding other factors constant. In contrast, when respondents are presented with two strategies, performance information on mortality rate and economic cost plays a significant role for citizens’ ratings. Even during a high-information high salience crisis such as COVID-19, citizens are thus more likely to make decisions based on performance information on equality than effectiveness and cost when no ‘yardstick’ is available. Results imply that performance information on effectiveness and cost risk being drown out by other information easier to interpret if not presented with relevant reference points.
Dependency distance minimization (DDm) is a well-established principle of word order. It has been predicted theoretically that DDm implies compression, namely the minimization of word lengths. This is a second order prediction because it links a principle with another principle, rather than a principle and a manifestation as in a first order prediction. Here we test that second order prediction with a parallel collection of treebanks controlling for annotation style with Universal Dependencies and Surface-Syntactic Universal Dependencies. To test it, we use a recently introduced score that has many mathematical and statistical advantages with respect to the widely used sum of dependency distances. We find that the prediction is confirmed by the new score when word lengths are measured in phonemes, independently of the annotation style, but not when word lengths are measured in syllables. In contrast, one of the most widely used scores, i.e. the sum of dependency distances, fails to confirm that prediction, showing the weakness of raw dependency distances for research on word order. Finally, our findings expand the theory of natural communication by linking two distinct levels of organization, namely syntax (word order) and word internal structure.
Search is one of the key functionalities in digital platforms and applications such as an electronic dictionary, a search engine, and an e-commerce platform. While the search function in some languages is trivial, Khmer word search is challenging given its complex writing system. Multiple orders of characters and different spelling realizations of words impose a constraint on Khmer word search functionality. Additionally, spelling mistakes are common since robust spellcheckers are not commonly available across the input device platforms. These challenges hinder the use of Khmer language in search-embedded applications. Moreover, due to the absence of WordNet-like lexical databases for Khmer language, it is impossible to establish semantic relation between words, enabling semantic search. In this paper, we propose a set of robust solutions to the above challenges associated with Khmer word search. The proposed solutions include character order normalization, grapheme and phoneme-based spellcheckers, and Khmer word semantic model. The semantic model is based on the word embedding model that is trained on a 30-million-word corpus and is used to capture the semantic similarities between words.
In this research paper entitled "Linguistic Atlases and their role in building a database for Arabic terminology banks" we firstly talk about geographical linguistics, then the concept of Linguistic Atlases, mentioning the German experience, the French experience, and the experience of the German Orientalist Bragstrazer in the Arab countries. Secondly, we talk about Arabic terminology banks represented in the lexical database(Meaarby), the Saudi automatic bank for terms ( Bassem), the terminology database (Qimam) in Tunisia, and the Jordan Academy for Arabic language bank of terms. Finally we mention to the extent by which the Arabic terminology banks benefit from the linguistic Atlases.
Chinese is a discourse-oriented language. “Run-on” sentences (liushui ju) are a typical and prevalent form of discourse in Chinese. These sentences show the capacity of the Chinese language for organizing loose structures into an effective and coherent discourse. Despite their widespread use in Chinese, previous studies have only explored “run-on” sentences by using small-scale examples. In order to carry out a quantitative investigation of “run-on” sentences, we need to establish a corpus. The present study selects 500 “run-on” sentences and annotates them on the levels of discourse, syntax and semantics. We mainly adopt PDTB (Penn Discourse Treebank) styles in the discourse annotations but we also borrow some features from RST (rhetorical structure theory). We find that the distribution of the frequency of discourse relations in the data extracted from this corpus follows the power law. The preliminary results reveal that semantic leaps in “run-on” sentences are closely related to the use of the topic chain and the animacy and the span of discourse relations. This corpus can thus aid in carrying out further computational and cognitive studies of Chinese discourse.
Abstract It is commonly assumed that participles show a mixture of verbal and adjectival properties, but the issue of how this mixed nature can best be captured is anything but settled. Analyses range from the purely adjectival to the purely verbal with various shades in between. This lack of consensus is at least partly due to the fact that participles are used in a variety of ways and that an analysis which fits one of them is not necessarily equally plausible for the other. In an effort to overcome the resulting fragmentation this paper proposes an analysis that covers all uses of the participles, from the adnominal over the predicative to the free adjunct uses, including also the nominalized ones. To keep it feasible we focus on one language, namely Dutch. With the help of a treebank we first identify the uses of the Dutch participles and describe their properties in informal terms. In a second step we provide an analysis in terms of the notation of Head-driven Phrase Structure Grammar. A key property of the analysis is the differentiation between core uses and grammaticalized uses. The treatment of the latter is influenced by insights from Grammaticalization Theory.
Recent years have witnessed significant improvement in ASR systems to\nrecognize spoken utterances. However, it is still a challenging task for noisy\nand out-of-domain data, where substitution and deletion errors are prevalent in\nthe transcribed text. These errors significantly degrade the performance of\ndownstream tasks. In this work, we propose a BERT-style language model,\nreferred to as PhonemeBERT, that learns a joint language model with phoneme\nsequence and ASR transcript to learn phonetic-aware representations that are\nrobust to ASR errors. We show that PhonemeBERT can be used on downstream tasks\nusing phoneme sequences as additional features, and also in low-resource setup\nwhere we only have ASR-transcripts for the downstream tasks with no phoneme\ninformation available. We evaluate our approach extensively by generating noisy\ndata for three benchmark datasets - Stanford Sentiment Treebank, TREC and ATIS\nfor sentiment, question and intent classification tasks respectively. The\nresults of the proposed approach beats the state-of-the-art baselines\ncomprehensively on each dataset.\n
While the last few decades have seen impressive improvements in several areas in Natural Language Processing, asking a computer to make sense of the discourse of utterances in a text remains challenging. There are several different theories that aim to describe and analyse the coherent structure that a well-written text inhibits. These theories have varying degrees of applicability and feasibility for practical use. Presumably the most data-driven of these theories is the paradigm that comes with the Penn Discourse TreeBank, a corpus annotated for discourse relations containing over 1 million words. Any language other than English however, can be considered a low-resource language when it comes to discourse processing. This dissertation is about shallow discourse parsing (discourse parsing following the paradigm of the Penn Discourse TreeBank) for German. The limited availability of annotated data for German means the potential of modern, deep-learning based methods relying on such data is also limited. This dissertation explores to what extent machine-learning and more recent deep-learning based methods can be combined with traditional, linguistic feature engineering to improve performance for the discourse parsing task. A pivotal role is played by connective lexicons that exhaustively list the discourse connectives of a particular language along with some of their core properties. To facilitate training and evaluation of the methods proposed in this dissertation, an existing corpus (the Potsdam Commentary Corpus) has been extended and additional data has been annotated from scratch. The approach to end-to-end shallow discourse parsing for German adopts a pipeline architecture and either presents the first results or improves over state-of-the-art for German for the individual sub-tasks of the discourse parsing task, which are, in processing order, connective identification, argument extraction and sense classification. The end-to-end shallow discourse parser for German that has been developed for the purpose of this dissertation is open-source and available online. In the course of writing this dissertation, work has been carried out on several connective lexicons in different languages. Due to their central role and demonstrated usefulness for the methods proposed in this dissertation, strategies are discussed for creating or further developing such lexicons for a particular language, as well as suggestions on how to further increase their usefulness for shallow discourse parsing.
The article focuses on the linguistic personality of A.V. Suvorov – one of the most important historical figures famous for his military achievements, including his service and warfare in Slavic territories (military service in the Lublin region, participation in hostilities in Poland, the Battle of Brest, etc.), a prominent representative of the Russian cultural elite in the 18th century. The study employs methods of linguopersonology and historical lexicology. It aims to describe a historical language personality, viewed as a reflection of both an individual with a linguistic potential, a type of specific linguistic reflection, and of the era as a whole with its inherent linguistic features and tendencies. Processes in the language at a given time period are also manifested in the language of an individual. The study focuses on the borrowed lexis, reflecting the general instability of the Russian language system of the period, which resulted from the prevailing multilingualism of the Russian nation, a significant increase in foreign words in the Russian language, and a change in linguistic norms. Through the prism of xenolexis, the authors describe Suvorov’s language personality reflected in his letters to note his broad outlook, fluency in native and foreign languages, linguistic intuition, innovative use of units of native and foreign languages.
This thesis explores the use of Natural Language Processing (NLP) on the Akkadian language documented from 2400 BCE to 100 CE. The methods and tools proposed in this thesis aim to fill the gaps left in previous research in Computational Assyriology, contributing to the transformation of transliterated cuneiform tablets into richly annotated text corpora, as well as to the quantitative lexicographic analysis of cuneiform texts.\n\nThree contributions of this thesis address the task of transforming Akkadian from its basic Latinized representation, transliteration, into linguistically annotated text corpora. These include (I) neural network-based automatic phonological transcription of transliterated cuneiform text, which is essential for normalizing the diverse spelling variations encountered in the Akkadian writing system; (II) finite-state-based automatic morphological analysis of Akkadian that allows deconstructing word forms into morphological labels, lemmata and part-of-speech tags to improve the useability of Akkadian corpora for quantitative analysis; and (III) creation of a morphological gold standard, and a standardized Universal Dependencies approved morphological label set for Akkadian morphology as the byproduct of an Akkadian treebank. \n\nThree contributions address the previously unexplored quantitative analysis of Akkadian lexical semantics using word association measures and word embeddings in order to better understand the language in its own terms. One of these contributions is (IV) an algorithmic method for reducing the distortion caused by fully or partially duplicated sequences in Akkadian texts. This algorithm solves over-representation issues encountered in pointwise mutual information (PMI)-based collocation analysis, and according to preliminary results, also in PMI-based word embeddings. Two contributions (V and VI) are quantitative case studies that demonstrate the use of PMI and word embeddings in Akkadian lexicography, and compare the results with previous qualitative philological research.\n\nThe last contribution (VII) is a hybrid approach, where PMI is applied to social network analysis of the Neo-Assyrian pantheon in order to reinforce the statistical relevance between the actors. These "semantic" social networks are used to study the position of the Assyrian main god, Aššur, within the pantheon.\n\nIn addition to the contributions, this thesis presents the first survey of Computational Assyriology, which covers six decades of research on automatic artifact reconstruction, optical character recognition, linguistic annotation, and quantitative analysis of cuneiform texts.
The main idea of the present paper is to uncover the hidden potential of Karachay-Balkar language in Constrained Writing. Being well described and deeply analyzed in terms of major linguistics subdisciplines, Karachay-Balkar language (as well as many other minor languages) lacks academic attention towards creation of experimental forms, e.g., palindromes. The present paper aims to answer the question "How big is the potential of Karachay-Balkar language in creating palindromes of different kinds, and which methods are applicable in doing that?" In order to answer this question, we justify choosing the appropriate lexical databases, and demonstrate that two different approaches are effective: "heuristic method", based on some apriori knowledge about Karachay-Balkar language and our own creativity; and "algorithmic method", based on some set of rules executed on Python. As a result, we have generated many palindromes of different kinds (letter-palindromes, syllable-palindromes, educanto-palindromes, "magic squares"), and about 70 of them are demonstrated in the present paper.
Syntactic annotation of historical text, with no access to native-speaker intuitions, poses a number of problems to the annotator who is faced with the task of giving a single analysis of each sentence. This article reports on the experiences from annotating complementation structures in Old Church Slavonic and Old East Slavonic in the PROIEL and TOROT treebanks. Two case studies are examined: complement clauses in Old Church Slavonic and the history of Russian čьto ‘what, which, that’. In the first case the annotation scheme is shown to work well in terms of interannotator agreement and retrievability. However, the price is that a large number of examples with jako ‘that’ are analysed as complement clauses with a subjunction, even though many of these examples are in fact ambiguous and jako can equally well be interpreted as a quotative particle followed by direct speech. The second case study looks at a development from situation where čьto could be taken to be an interrogative pronoun in all subordinated clauses, to a situation where a subjunction and a relative pronoun analysis are also available. This leads to a large number of ambiguous occurrences. The solution in TOROT is to analyse unambiguous interrogative pronoun and subjunction examples at face value, while all of the remaining occurrences are analysed as relative clauses. This makes the annotator’s job manageable, but causes retrievability problems, since individual researchers will have to sift through the relative clause examples themselves.
This paper compares two influential theories of processing difficulty: Gibson (2000)'s Dependency Locality Theory (DLT) and Hale (2001)'s Surprisal Theory. While prior work has aimed to compare DLT and Surprisal Theory (see I compare estimated surprisal values from two models, an RNN and a Transformer neural network, as well as DLT integration cost from a hand-parsed treebank, to reading times from the Dundee Corpus. The results for integration cost corroborate those of Ultimately, I conclude that a broad-coverage model must integrate both theories in order to most accurately predict processing difficulty.
The article discusses means of expressing aggressiveness used in the newspaper “LDPR in Udmurtia” during election campaigns. Despite a large number of implicit means that politicians and journalists are equipped with, the newspaper demonstrates explicit means of speech aggressiveness. The style of the publications is a replica of speech manner of the LDPR leader. One of the major ways of expressing negative assessment of political opponents is conversational and colloquial vocabulary of deflated style. The authors of this media tend to use phraseology with non-normative connotation. Expressive means used by journalists tend to sound negative and aim to discredit the existing power. To demonstrate negative attitude to the government a variety of stylistic means are used including repetitions, gradation, rhetoric questions, rhetoric exclamations and addressing. In the newspaper of the opposition party ethical and linguistic norms are broken, negative information disclosing situation in the country tend to form readers’ sense of frustration.
This paper presents a full procedure for the development of a Part-of-Speech (POS) tagged corpus of Old Catalan. As an extremely low-resource language with rich inflection and frequent homographs, Old Catalan poses non-trivial problems in the development of a searchable constituency-based treebank. We demonstrate, however, that a semi-supervised method of incrementally building training data using both neural and memory-based taggers, together with the Pyrrha annotation tool is highly efficient and yields accurate results. We propose that this simple and effective method could easily be extended to other low-resource historical languages for which no NLP tools exist yet.