Korpuslinguistik
Refine
Year of publication
Document Type
- Conference Proceeding (161) (remove)
Keywords
- Korpus <Linguistik> (144)
- Deutsch (33)
- Annotation (28)
- Gesprochene Sprache (21)
- Forschungsdaten (17)
- corpus linguistics (13)
- Computerlinguistik (12)
- Corpus linguistics (12)
- Datenmanagement (11)
- Corpus technology (10)
Publicationstate
- Veröffentlichungsversion (127)
- Zweitveröffentlichung (14)
- Postprint (3)
Reviewstate
- Peer-Review (85)
- (Verlags)-Lektorat (41)
- Peer-review (2)
- Review-Status-unbekannt (1)
Publisher
- European Language Resources Association (ELRA) (22)
- European Language Resources Association (17)
- Institut für Deutsche Sprache (17)
- Leibniz-Institut für Deutsche Sprache (9)
- CLARIN (8)
- Linköping University Electronic Press (8)
- Nisaba (4)
- University of Birmingham (4)
- Association for Computational Linguistics (3)
- Extreme Markup Languages Conference (3)
We present a collection of (currently) about 5.500 commands directed to voice-controlled virtual assistants (VAs) by sixteen initial users of a VA system in their homes. The collection comprises recordings captured by the VA itself and with a conditional voice recorder (CVR) selectively capturing recordings including the VA-directed commands plus some surrounding context. Next to a description of the collection, we present initial findings on the patterns of use of the VA systems during the first weeks after installation, including usage timing, the development of usage frequency, distributions of sentence structures across commands, and (the development of) command success rates. We discuss the advantages and disadvantages of the applied collection-specific recording approach and describe potential research questions that can be investigated in the future, based on the collection, as well as the merit of combining quantitative corpus linguistic approaches with qualitative in-depth analyses of single cases.
Corpus-based identification and disambiguation of reading indicators for German nominalizations
(2010)
Corpus data is often structurally and lexically ambiguous; corpus extraction methodologies thus must be made aware of ambiguities. Therefore, given an extraction task, all relevant ambiguities must be identified. To resolve these ambiguities, contextual data responsible for one or another reading is to be considered. In the context of our present work, German -ung-nominalizations and their sortal readings are under examination. A number of these nominalizations may be read as an event or a result, depending on the semantic group they belong to. Here, we concentrate on nominalizations of verbs of saying (henceforth: "verba dicendi"), identify their context partners and their influence on the sortal reading of the nominalizations in question. We present a tool which calculates the sortal reading of such nominalizations and thus may improve not only corpus extraction, but also e.g. machine translation. Lastly, we describe successful attempts to identify the correct sortal reading, conclusions and future work.
Projektvorstellung – Redewiedergabe. Eine literatur- und sprachwissenschaftliche Korpusanalyse
(2018)
Das laufende DFG-Projekt „Redewiedergabe“ stellt einen Anwendungsfall quantitativer Sprach-und Literaturwissenschaft dar und beschäftigt sich mit dem Phänomen „Redewiedergabe“ auf der Grundlage großer Datenmengen. Zu diesem Zweck wird zum einen ein Korpus manuell mit Redewiedergabeformen annotiert, zum anderen werden Verfahren zur automatischen Erkennung des Phänomens entwickelt. Ziel ist es, Forschungsfragen nach der Entwicklung von Redewiedergabe vor allem im 19. Jahrhundert zu beantworten.
This paper reports on recent developments within the European Reference Corpus EuReCo, an open initiative that aims at providing and using virtual and dynamically definable comparable corpora based on existing national, reference or other large corpora. Given the well-known shortcomings of other types of multilingual corpora such as parallel/translation corpora (shining-through effects, over-normalization, simplification, etc.) or web-based comparable corpora (covering only web material), EuReCo provides a unique linguistic resource offering new perspectives for fine-grained contrastive research on authentic cross-linguistic data, applications in translation studies and foreign language teaching and learning.
This paper describes the application of probabilistic part of speech taggers to the Dzongkha language. A tag set containing 66 tags is designed, which is based on the Penn Treebank. A training corpus of 40,247 tokens is utilized to train the model. Using the lexicon extracted from the training corpus and lexicon from the available word list, we used two statistical taggers for comparison reasons. The best result achieved was 93.1% accuracy in a 10-fold cross validation on the training set. The winning tagger was thereafter applied to annotate a 570,247 token corpus.
We present recognizers for four very different types of speech, thought and writing representation (STWR) for German texts. The implementation is based on deep learning with two different customized contextual embeddings, namely FLAIR embeddings and BERT embeddings. This paper gives an evaluation of our recognizers with a particular focus on the differences in performance we observed between those two embeddings. FLAIR performed best for direct STWR (F1=0.85), BERT for indirect (F1=0.76) and free indirect (F1=0.59) STWR. For reported STWR, the comparison was inconclusive, but BERT gave the best average results and best individual model (F1=0.60). Our best recognizers, our customized language embeddings and most of our test and training data are freely available and can be found via www.redewiedergabe.de or at github.com/redewiedergabe.
The QUEST (QUality ESTablished) project aims at ensuring the reusability of audio-visual datasets (Wamprechtshammer et al., 2022) by devising quality criteria and curating processes. RefCo (Reference Corpora) is an initiative within QUEST in collaboration with DoReCo (Documentation Reference Corpus, Paschen et al. (2020)) focusing on language documentation projects. Previously, Aznar and Seifart (2020) introduced a set of quality criteria dedicated to documenting fieldwork corpora. Based on these criteria, we establish a semi-automatic review process for existing and work-in-progress corpora, in particular for language documentation. The goal is to improve the quality of a corpus by increasing its reusability. A central part of this process is a template for machine-readable corpus documentation and automatic data verification based on this documentation. In addition to the documentation and automatic verification, the process involves a human review and potentially results in a RefCo certification of the corpus. For each of these steps, we provide guidelines and manuals. We describe the evaluation process in detail, highlight the current limits for automatic evaluation and how the manual review is organized accordingly.
Metadata provides important information relevant both to finding and understanding corpus data. Meaningful linguistic data requires both reasonable annotations and documentation of these annotations. This documentation is part of the metadata of a dataset. While corpus documentation has often been provided in the form of accompanying publications, machinereadable metadata, both containing the bibliographic information and documenting the corpus data, has many advantages. Metadata standards allow for the development of common tools and interfaces. In this paper I want to add a new perspective from an archive’s point of view and look at the metadata provided for four learner corpora and discuss the suitability of established standards for machine-readable metadata. I am are aware that there is ongoing work towards metadata standards for learner corpora. However, I would like to keep the discussion going and add another point of view: increasing findability and reusability of learner corpora in an archiving context.
This paper describes the TEI-based ISO standard 24624:2016 ‘Transcription of spoken language’ and other formats used within CLARIN for spoken language resources. It assesses the current state of support for the standard and the interoperability between these formats and with rele- vant tools and services. The main idea behind the paper is that a digital infrastructure providing language resources and services to researchers should also allow the combined use of resources and/or services from different contexts. This requires syntactic and semantic interoperability. We propose a solution based on the ISO/TEI format and describe the necessary steps for this format to work as an exchange format with basic semantic interoperability for spoken language resources across the CLARIN infrastructure and beyond.
This paper presents an algorithm and an implementation for efficient tokenization of texts of space-delimited languages based on a deterministic finite state automaton. Two representations of the underlying data structure are presented and a model implementation for German is compared with state-of-the-art approaches. The presented solution is faster than other tools while maintaining comparable quality.
We present the use of count-based and predictive language models for exploring language use in the German Reference Corpus DeReKo. For collocation analysis along the syntagmatic axis we employ traditional association measures based on co-occurrence counts as well as predictive association measures derived from the output weights of skipgram word embeddings. For inspecting the semantic neighbourhood of words along the paradigmatic axis we visualize the high dimensional word embeddings in two dimensions using t-stochastic neighbourhood embeddings. Together, these visualizations provide a complementary, explorative approach to analysing very large corpora in addition to corpus querying. Moreover, we discuss count-based and predictive models w.r.t. scalability and maintainability in very large corpora.
This paper describes the TEI-based ISO standard 2462:2016 “Transcription of spoken language” and other formats used within CLARIN for spoken language resources. It assesses the current state of support for the standard and the interoperability between these formats and with relevant tools and services. The main idea behind the paper is that a digital infrastructure providing language resources and services to researchers should also allow the combined use of resources and/or services from different contexts. This requires syntactic and semantic interoperability. We propose a solution based on the ISO/TEI format and describe the necessary steps for this format to work as an exchange format with basic semantic interoperability for spoken language resources across the CLARIN infrastructure and beyond.
Signposts for CLARIN
(2021)
An implementation of CMDI-based signposts and its use is presented in this paper. Arnold, Fisseni et al. (2020) present signposts as a solution to challenges in long-term preservation of corpora. Though applicable to digital resources in general, we focus on corpora, especially those that are continuously extended or subject to modification, e.g., due to legal injunctions, but also may overlap with respect to constituents, and may be subject to migrations to new data formats. We describe the contribution signposts can make to the CLARIN infrastructure, notably virtual collections, and document the design for the CMDI profile.
This paper presents the QUEST project and describes concepts and tools that are being developed within its framework. The goal of the project is to establish quality criteria and curation criteria for annotated audiovisual language data. Building on existing resources developed by the participating institutions earlier, QUEST also develops tools that could be used to facilitate and verify adherence to these criteria. An important focus of the project is making these tools accessible for researchers without substantial technical background and helping them produce high-quality data. The main tools we intend to provide are a questionnaire and automatic quality assurance for depositors of language resources, both developed as web applications. They are accompanied by a knowledge base, which will contain recommendations and descriptions of best practices established in the course of the project. Conceptually, we consider three main data maturity levels in order to decide on a suitable level of strictness of the quality assurance. This division has been introduced to avoid that a set of ideal quality criteria prevent researchers from depositing or even assessing their (legacy) data. The tools described in the paper are work in progress and are expected to be released by the end of the QUEST project in 2022.
CMDI Explorer
(2021)
We present CMDI Explorer, a tool that empowers users to easily explore the contents of complex CMDI records and to process selected parts of them with little effort. The tool allows users, for instance, to analyse virtual collections represented by CMDI records, and to send collection items to other CLARIN services such as the Switchboard for subsequent processing. CMDI Explorer hence adds functionality that many users felt was lacking from the CLARIN tool space.
Towards comprehensive definitions of data quality for audiovisual annotated language resources
(2021)
Though digital infrastructures such as CLARIN have been successfully established and now provide large collections of digital resources, the lack of widely accepted standards for data quality and documentation still makes re-use of research data a difficult endeavour, especially for more complex resource types. The article gives a detailed overview over relevant characteristics of audiovisual annotated language resources and reviews possible approaches to data quality in terms of their suitability for the current context. Conclusively, various strategies are suggested in order to arrive at comprehensive and adequate definitions of data quality for this specific resource type and possibly for digital language resources in general.
The article focuses on determining responsible parties and the division of potential liability arising from sharing language data (LD) containing personal data (PD). A key issue here is to identify who has to make sure and guarantee the GDPR compliance. The authors aim to answer 1) whether an individual researcher is a controller and 2) whether sharing LD results in joint controllership or separate controllership (whether the data's transferee becomes the controller, the joint controller or the processor). The article also analyses the legal relations of parties involved in data sharing and potential liability. The final section outlines data sharing in the CLARIN context. The analysis serves as a preliminary analytical background for redesigning the CLARIN contractual framework for sharing data.
N-grams are of utmost importance for modern linguistics and language technology. The legal status of n-grams, however, raises many practical questions. Traditionally, text snippets are considered copyrightable if they meet the originality criterion, but no clear indicators as to the minimum length of original snippets exist; moreover, the solutions adopted in some EU Member States (the paper cites German and French law as examples) are considerably different. Furthermore, recent developments in EU law (the CJEU's Pelham decision and the new right of press publishers) also provide interesting arguments in this debate. The paper presents the existing approaches to the legal protection of n-grams and tries to formulate some clear guidelines as to the length of n-grams that can be freely used and shared.
Making research data publicly available for evaluation or reuse is a fundamental part of good scientific practice. However, regulations such as copyright law can prevent this practice and thereby hamper scientific progress. In Germany, text-based research disciplines have for a long time been mostly unable to publish corpora made from material outside of the public domain, effectively excluding contemporary works. While there are approaches to obfuscate text material in a way that it is no longer covered by the original copyright, many use cases still require the raw textual context for evaluation or follow-up research. Recent changes in copyright now permit text and data mining on copyrighted works. However, questions regarding reusability and sharing of such corpora at a later time are still not answered to a satisfying degree. We propose a workflow that allows interested third parties to access customized excerpts of protected corpora in accordance with current German copyright law and the soon to be implemented guidelines of the Digital Single Market directive. Our prototype is a very lightweight web interface that builds on commonly used repository software and web standards.
In this paper, we present our experiences and decisions in dealing with challenges in developing, maintaining and operating online research software tools in the field of linguistics. In particular, we highlight reproducibility, dependability, and security as important aspects of quality management – taking into account the special circumstances in which research software
is usually created.
Ungoliant: An optimized pipeline for the generation of a very large-scale multilingual web corpus
(2021)
Since the introduction of large language models in Natural Language Processing, large raw corpora have played a crucial role in Computational Linguistics. However, most of these large raw corpora are either available only for English or not available to the general public due to copyright issues. Nevertheless, there are some examples of freely available multilingual corpora for training Deep Learning NLP models, such as the OSCAR and Paracrawl corpora. However, they have quality issues, especially for low-resource languages. Moreover, recreating or updating these corpora is very complex. In this work, we try to reproduce and improve the goclassy pipeline used to create the OSCAR corpus. We propose a new pipeline that is faster, modular, parameterizable, and well documented. We use it to create a corpus similar to OSCAR but larger and based on recent data. Also, unlike OSCAR, the metadata information is at the document level. We release our pipeline under an open source license and publish the corpus under a research-only license.
CLARIN contractual framework for sharing language data: the perspective of personal data protection
(2020)
The article analyses the responsibility for ensuring compliance with the General Data Protection Regulation (GDPR) in research settings. As a general rule, organisations are considered the data controller (responsible party for the GDPR compliance). Research constitutes a unique setting influenced by academic freedom. This raises the question of whether academics could be considered the controller as well. However, there are some court cases and policy documents on this issue. It is not settled yet. The analysis serves a preliminary analytical background for redesigning CLARIN contractual framework for sharing data.
N-grams are of utmost importance for modern linguistics and language theory. The legal status of n-grams, however, raises many practical questions. Traditionally, text snippets are considered copyrightable if they meet the originality criterion, but no clear indicators as to the minimum length of original snippets exist; moreover, the solutions adopted in some EU Member States (the paper cites German and French law as examples) are considerably different. Furthermore, recent developments in EU law (the CJEU's Pelham decision and the new right of newspaper publishers) also provide interesting arguments in this debate. The proposed paper presents the existing approaches to the legal protection of n-grams and tries to formulate some clear guidelines as to the length of n-grams that can be freely used and shared.
The CMDI Explorer
(2020)
We present the CMDI Explorer, a tool that empowers users to easily explore the contents of complex CMDI records and to process selected parts of them with little effort. The tool allows users, for instance, to analyse virtual collections represented by CMDI records, and to send collection items to other CLARIN services such as the Switchboard for subsequent processing. The CMDI Explorer hence adds functionality that many users felt was lacking from the CLARIN tool space.
Signposts for CLARIN
(2020)
An implementation of CMDI-based signposts and its use is presented in this paper. Arnold et al. 2020 present Signposts as a solution to challenges in long-term preservation of corpora, especially corpora that are continuously extended and subject to modification, e.g., due to legal injunctions, but also may overlap with respect to constituents, and may be subject to migrations to new data formats. We describe the contribution Signposts can make to the CLARIN infrastructure and document the design for the CMDI profile.
Towards Comprehensive Definitions of Data Quality for Audiovisual Annotated Language Resources
(2020)
Though digital infrastructures such as CLARIN have been successfully established and now provide large collections of digital resources, the lack of widely accepted standards for data quality and documentation still makes re-use of research data a difficult endeavour, especially for more complex resource types. The article gives a detailed overview over relevant characteristics of audiovisual annotated language resources and reviews possible approaches to data quality in terms of their suitability for the current context. Conclusively, various strategies are suggested in order to arrive at comprehensive and adequate definitions of data quality for this particular resource type.
This paper presents the QUEST project and describes concepts and tools that are being developed within its framework. The goal of the project is to establish quality criteria and curation criteria for annotated audiovisual language data. Building on existing resources developed by the participating institutions earlier, QUEST develops tools that could be used to facilitate and verify adherence to these criteria. An important focus of the project is making these tools accessible for researchers without substantial technical background and helping them produce high-quality data. The main tools we intend to provide are the depositors’ questionnaire and automatic quality assurance, both developed as web applications. They are accompanied by a Knowledge base, which will contain recommendations and descriptions of best practices established in the course of the project. Conceptually, we split linguistic data into three resource classes (data deposits, collections and corpora). The class of a resource defines the strictness of the quality assurance it should undergo. This division is introduced so that too strict quality criteria do not prevent researchers from depositing their data.
Song lyrics can be considered as a text genre that has features of both written and spoken discourse, and potentially provides extensive linguistic and cultural information to scientists from various disciplines. However, pop songs play a rather subordinate role in empirical language research so far - most likely due to the absence of scientifically valid and sustainable resources. The present paper introduces a multiply annotated corpus of German lyrics as a publicly available basis for multidisciplinary research. The resource contains three types of data for the investigation and evaluation of quite distinct phenomena: TEI-compliant song lyrics as primary data, linguistically and literary motivated annotations, and extralinguistic metadata. It promotes empirically/statistically grounded analyses of genre-specific features, systemic-structural correlations and tendencies in the texts of contemporary pop music. The corpus has been stratified into thematic and author-specific archives; the paper presents some basic descriptive statistics, as well as the public online frontend with its built-in evaluation forms and live visualisations.
In this article, we describe a user support solution for the digital humanities. As a case study, we show the development of the CLARIN-D Helpdesk from 2013 into the current support solution that has been extended for several other CLARIN-related software and projects and the DARIAH-ERIC. Furthermore, we describe a way towards a common support platform for CLARIAH-DE, which is currently in the final phase. We hope to further expand the help desk in the following years in order to act as a hub for user support and a central knowledge resource for the digital humanities not only in the German, but also in the European area and perhaps at some point worldwide.
The sentiment polarity of an expression (whether it is perceived as positive, negative or neutral) can be influenced by a number of phenomena, foremost among them negation. Apart from closed-class negation words like no, not or without, negation can also be caused by so-called polarity shifters. These are content words, such as verbs, nouns or adjectives, that shift polarities in their opposite direction, e. g. abandoned in “abandoned hope” or alleviate in “alleviate pain”. Many polarity shifters can affect both positive and negative polar expressions, shifting them towards the opposing polarity. However, other shifters are restricted to a single shifting direction. Recoup shifts negative to positive in “recoup your losses”, but does not affect the positive polarity of fortune in “recoup a fortune”. Existing polarity shifter lexica only specify whether a word can, in general, cause shifting, but they do not specify when this is limited to one shifting direction. To address this issue we introduce a supervised classifier that determines the shifting direction of shifters. This classifier uses both resource-driven features, such as WordNet relations, and data-driven features like in-context polarity conflicts. Using this classifier we enhance the largest available polarity shifter lexicon.
We present a fine-grained NER annotations scheme with 30 labels and apply it to German data. Building on the OntoNotes 5.0 NER inventory, our scheme is adapted for a corpus of transcripts of biographic interviews by adding categories for AGE and LAN(guage) and also adding label classes for various numeric and temporal expressions. Applying the scheme to the spoken data as well as a collection of teaser tweets from newspaper sites, we can confirm its generality for both domains, also achieving good inter-annotator agreement. We also show empirically how our inventory relates to the well-established 4-category NER inventory by re-annotating a subset of the GermEval 2014 NER coarse-grained dataset with our fine label inventory. Finally, we use a BERT-based system to establish some baselines for NER tagging on our two new datasets. Global results in in-domain testing are quite high on the two datasets, near what was achieved for the coarse inventory on the CoNLLL2003 data. Cross-domain testing produces much lower results due to the severe domain differences.
The paper presents a discussion on the main linguistic phenomena of user-generated texts found in web and social media, and proposes a set of annotation guidelines for their treatment within the Universal Dependencies (UD) framework. Given on the one hand the increasing number of treebanks featuring user-generated content, and its somewhat inconsistent treatment in these resources on the other, the aim of this paper is twofold: (1) to provide a short, though comprehensive, overview of such treebanks - based on available literature - along with their main features and a comparative analysis of their annotation criteria, and (2) to propose a set of tentative UD-based annotation guidelines, to promote consistent treatment of the particular phenomena found in these types of texts. The main goal of this paper is to provide a common framework for those teams interested in developing similar resources in UD, thus enabling cross-linguistic consistency, which is a principle that has always been in the spirit of UD.
We present a new resource for German causal language, with annotations in context for verbs, nouns and adpositions. Our dataset includes 4,390 annotated instances for more than 150 different triggers. The annotation scheme distinguishes three different types of causal events (CONSEQUENCE, MOTIVATION, PURPOSE). We also provide annotations for semantic roles, i.e. of the cause and effect for the causal event as well as the actor and affected party, if present. In the paper, we present inter-annotator agreement scores for our dataset and discuss problems for annotating causal language. Finally, we present experiments where we frame causal annotation as a sequence labelling problem and report baseline results for the prediciton of causal arguments and for predicting different types of causation.
Interoperability in an Infrastructure Enabling Multidisciplinary Research: The case of CLARIN
(2020)
CLARIN is a European Research Infrastructure providing access to language resources and technologies for researchers in the humanities and social sciences. It supports the use and study of language data in general and aims to increase the potential for comparative research of cultural and societal phenomena across the boundaries of languages and disciplines, all in line with the European agenda for Open Science. Data infrastructures such as CLARIN have recently embarked on the emerging frameworks for the federation of infrastructural services, such as the European Open Science Cloud and the integration of services resulting from multidisciplinary collaboration in federated services for the wider domain of the social sciences and humanities (SSH). In this paper we describe the interoperability requirements that arise through the existing ambitions and the emerging frameworks. The interoperability theme will be addressed at several levels, including organisation and ecosystem, design of workflow services, data curation, performance measurement and collaboration. For each level, some concrete outcomes are described.
This paper presents experiments on sentence boundary detection in transcripts of spoken dialogues. Segmenting spoken language into sentence-like units is a challenging task, due to disfluencies, ungrammatical or fragmented structures and the lack of punctuation. In addition, one of the main bottlenecks for many NLP applications for spoken language is the small size of the training data, as the transcription and annotation of spoken language is by far more time-consuming and labour-intensive than processing written language. We therefore investigate the benefits of data expansion and transfer learning and test different ML architectures for this task. Our results show that data expansion is not straightforward and even data from the same domain does not always improve results. They also highlight the importance of modelling, i.e. of finding the best architecture and data representation for the task at hand. For the detection of boundaries in spoken language transcripts, we achieve a substantial improvement when framing the boundary detection problem as a sentence pair classification task, as compared to a sequence tagging approach.
Beyond Citations: Corpus-based Methods for Detecting the Impact of Research Outcomes on Society
(2020)
This paper proposes, implements and evaluates a novel, corpus-based approach for identifying categories indicative of the impact of research via a deductive (top-down, from theory to data) and an inductive (bottom-up, from data to theory) approach. The resulting categorization schemes differ in substance. Research outcomes are typically assessed by using bibliometric methods, such as citation counts and patterns, or alternative metrics, such as references to research in the media. Shortcomings with these methods are their inability to identify impact of research beyond academia (bibliometrics) and considering text-based impact indicators beyond those that capture attention (altmetrics). We address these limitations by leveraging a mixed-methods approach for eliciting impact categories from experts, project personnel (deductive) and texts (inductive). Using these categories, we label a corpus of project reports per category schema, and apply supervised machine learning to infer these categories from project reports. The classification results show that we can predict deductively and inductively derived impact categories with 76.39% and 78.81% accuracy (F1-score), respectively. Our approach can complement solutions from bibliometrics and scientometrics for assessing the impact of research and studying the scope and types of advancements transferred from academia to society.
The newest generation of speech technology caused a huge increase of audio-visual data nowadays being enhanced with orthographic transcripts such as in automatic subtitling in online platforms. Research data centers and archives contain a range of new and historical data, which are currently only partially transcribed and therefore only partially accessible for systematic querying. Automatic Speech Recognition (ASR) is one option of making that data accessible. This paper tests the usability of a state-of-the-art ASR-System on a historical (from the 1960s), but regionally balanced corpus of spoken German, and a relatively new corpus (from 2012) recorded in a narrow area. We observed a regional bias of the ASR-System with higher recognition scores for the north of Germany vs. lower scores for the south. A detailed analysis of the narrow region data revealed – despite relatively high ASR-confidence – some specific word errors due to a lack of regional adaptation. These findings need to be considered in decisions on further data processing and the curation of corpora, e.g. correcting transcripts or transcribing from scratch. Such geography-dependent analyses can also have the potential for ASR-development to make targeted data selection for training/adaptation and to increase the sensitivity towards varieties of pluricentric languages.
As a part of the ZuMult-project, we are currently modelling a backend architecture that should provide query access to corpora from the Archive of Spoken German (AGD) at the Leibniz-Institute for the German Language (IDS). We are exploring how to reuse existing search engine frameworks providing full text indices and allowing to query corpora by one of the corpus query languages (QLs) established and actively used in the corpus research community. For this purpose, we tested MTAS - an open source Lucene-based search engine for querying on text with multilevel annotations. We applied MTAS on three oral corpora stored in the TEI-based ISO standard for transcriptions of spoken language (ISO 24624:2016). These corpora differ from the corpus data that MTAS was developed for, because they include interactions with two and more speakers and are enriched, inter alia, with timeline-based annotations. In this contribution, we report our test results and address issues that arise when search frameworks originally developed for querying written corpora are being transferred into the field of spoken language.
We evaluate a graph-based dependency parser on DeReKo, a large corpus of contemporary German. The dependency parser is trained on the German dataset from the SPMRL 2014 Shared Task which contains text from the news domain, whereas DeReKo also covers other domains including fiction, science, and technology. To avoid the need for costly manual annotation of the corpus, we use the parser’s probability estimates for unlabeled and labeled attachment as main evaluation criterion. We show that these probability estimates are highly correlated with the actual attachment scores on a manually annotated test set. On this basis, we compare estimated parsing scores for the individual domains in DeReKo, and show that the scores decrease with increasing distance of a domain to the training corpus.
This paper addresses long-term archival for large corpora. Three aspects specific to language resources are focused, namely (1) the removal of resources for legal reasons, (2) versioning of (unchanged) objects in constantly growing resources, especially where objects can be part of multiple releases but also part of different collections, and (3) the conversion of data to new formats for digital preservation. It is motivated why language resources may have to be changed, and why formats may need to be converted. As a solution, the use of an intermediate proxy object called a signpost is suggested. The approach will be exemplified with respect to the corpora of the Leibniz Institute for the German Language in Mannheim, namely the German Reference Corpus (DeReKo) and the Archive for Spoken German (AGD).
Since 2013 representatives of several French and German CMC corpus projects have developed three customizations of the TEI-P5 standard for text encoding in order to adapt the encoding schema and models provided by the TEI to the structural peculiarities of CMC discourse. Based on the three schema versions, a 4th version has been created which takes into account the experiences from encoding our corpora and which is specifically designed for the submission of a feature request to the TEI council. On our poster we would present the structure of this schema and its relations (commonalities and differences) to the previous schemas.
Text corpora come in many different shapes and sizes and carry heterogeneous annotations, depending on their purpose and design. The true benefit of corpora is rooted in their annotation and the method by which this data is encoded is an important factor in their interoperability. We have accumulated a large collection of multilingual and parallel corpora and encoded it in a unified format which is compatible with a broad range of NLP tools and corpus linguistic applications. In this paper, we present our corpus collection and describe a data model and the extensions to the popular CoNLL-U format that enable us to encode it.
Common Crawl is a considerably large, heterogeneous multilingual corpus comprised of crawled documents from the internet, surpassing 20TB of data and distributed as a set of more than 50 thousand plain text files where each contains many documents written in a wide variety of languages. Even though each document has a metadata block associated to it, this data lacks any information about the language in which each document is written, making it extremely difficult to use Common Crawl for monolingual applications. We propose a general, highly parallel, multithreaded pipeline to clean and classify Common Crawl by language; we specifically design it so that it runs efficiently on medium to low resource infrastructures where I/O speeds are the main constraint. We develop the pipeline so that it can be easily reapplied to any kind of heterogeneous corpus and so that it can be parameterised to a wide range of infrastructures. We also distribute a 6.3TB version of Common Crawl, filtered, classified by language, shuffled at line level in order to avoid copyright issues, and ready to be used for NLP applications.
Nearly all of the very large corpora of English are “static”, which allows a wide range of one-time, pre-processed data, such as collocates. The challenge comes with large “dynamic” corpora, which are updated regularly, and where preprocessing is much more difficult. This paper provides an overview of the NOW corpus (News on the Web), which is currently 8.2 billion words in size, and which grows by about 170 million words each month. We discuss the architecture of NOW, and provide many examples that show how data from NOW can (uniquely) be extracted to look at a wide range of ongoing changes in English.
As the Web ought to be considered as a series of sources rather than as a source in itself, a problem facing corpus construction resides in meta-information and categorization. In addition, we need focused data to shed light on particular subfields of the digital public sphere. Blogs are relevant to that end, especially if the resulting web texts can be extracted along with metadata and made available in coherent and clearly describable collections.
This paper reports on the latest developments of the European Reference Corpus EuReCo and the German Reference Corpus in relation to three of the most important CMLC topics: interoperability, collaboration on corpus infrastructure building, and legal issues. Concerning interoperability, we present new ways to access DeReKo via KorAP on the API and on the plugin level. In addition we report about advancements in the EuReCo- and ICC-initiatives with the provision of comparable corpora, and about recent problems with license acquisitions and our solution approaches using an indemnification clause and model licenses that include scientific exploitation.
Distributional models of word use constitute an indispensable tool in corpus based lexicological research for discovering paradigmatic relations and syntagmatic patterns (Belica et al. 2010). Recently, word embeddings (Mikolov et al. 2013) have revived the field by allowing to construct and analyze distributional models on very large corpora. This is accomplished by reducing the very high dimensionality of word cooccurrence contexts, the size of the vocabulary, to few dimensions, such as 100-200. However, word use and meaning can vary widely along dimensions such as domain, register, and time, and word embeddings tend to represent only the most prevalent meaning. In this paper we thus construct domain specific word embeddings to allow for systematically analyzing variations in word use. Moreover, we also demonstrate how to reconstruct domain specific co-occurrence contexts from the dense word embeddings.
In this paper, we describe MLSA, a publicly available multi-layered reference corpus for German-language sentiment analysis. The construction of the corpus is based on the manual annotation of 270 German-language sentences considering three different layers of granularity. The sentence-layer annotation, as the most coarse-grained annotation, focuses on aspects of objectivity, subjectivity and the overall polarity of the respective sentences. Layer 2 is concerned with polarity on the word- and phrase-level, annotating both subjective and factual language. The annotations on Layer 3 focus on the expression-level, denoting frames of private states such as objective and direct speech events. These three layers and their respective annotations are intended to be fully independent of each other. At the same time, exploring for and discovering interactions that may exist between different layers should also be possible. The reliability of the respective annotations was assessed using the average pairwise agreement and Fleiss’ multi-rater measures. We believe that MLSA is a beneficial resource for sentiment analysis research, algorithms and applications that focus on the German language.
This paper analyses reply relations in computer-mediated communication (CMC), which occur between post units in CMC interactions and which describe references between posts. We take a look at existing practices in the description and annotation of such relations in chat, wiki talk, and blog corpora. We distinguish technical reply structures, indentation structures, and interpretative reply relations, which include reply relations induced by linguistic markers. We sort out the different levels of description and annotation that are involved and propose a solution for their combined representation within the TEI annotation framework.
In the NLP literature, adapting a parser to new text with properties different from the training data is commonly referred to as domain adaptation. In practice, however, the differences between texts from different sources often reflect a mixture of domain and genre properties, and it is by no means clear what impact each of those has on statistical parsing. In this paper, we investigate how differences between articles in a newspaper corpus relate to the concepts of genre and domain and how they influence parsing performance of a transition-based dependency parser. We do this by applying various similarity measures for data point selection and testing their adequacy for creating genre-aware parsing models.
This presentation introduces a new collaborative project: the International Comparable Corpus (ICC) (https://korpus.cz/icc), to be compiled from European national, standard(ised) languages, using the protocols for text categories and their quantities of texts in the International Corpus of English (ICE).
New exceptions for Text and Data Mining and their possible impact on the CLARIN infrastructure
(2018)
The proposed paper discusses new exceptions for Text and Data Mining that have recently been adopted in some EU Member States, and probably will soon be adopted also at the EU level. These exceptions are of great significance for language scientists, as they exempt those who compile corpora from the obligation to obtain authorisation from rightholders. However, corpora compiled on the basis of such exceptions cannot be freely shared, which in a long run may have serious consequences for Open Science and the functioning of research infrastructure such as CLARIN ERIC.
This paper presents Release 2.0 of the SALSA corpus, a German resource for lexical semantics. The new corpus release provides new annotations for German nouns, complementing the existing annotations of German verbs in Release 1.0. The corpus now includes around 24,000 sentences with more than 36,000 annotated instances. It was designed with an eye towards NLP applications such as semantic role labeling but will also be a useful resource for linguistic studies in lexical semantics.
We present a testsuite for POS tagging German web data. Our testsuite provides the original raw text as well as the gold tokenisations and is annotated for parts-of-speech. The testsuite includes a new dataset for German tweets, with a current size of 3,940 tokens. To increase the size of the data, we harmonised the annotations in already existing web corpora, based on the Stuttgart-Tübingen Tag Set. The current version of the corpus has an overall size of 48,344 tokens of web data, around half of it from Twitter. We also present experiments, showing how different experimental setups (training set size, additional out-of-domain training data, self-training) influence the accuracy of the taggers. All resources and models will be made publicly available to the research community.
Knowledge in textual form is always presented as visually and hierarchically structured units of text, which is particularly true in the case of academic texts. One research hypothesis of the ongoing project Knowledge ordering in texts - text structure and structure visualisations as sources of natural ontologies1 is that the textual structure of academic texts effectively mirrors essential parts of the knowledge structure that is built up in the text. The structuring of a modern dissertation thesis (e.g. in the form of an automatically generated table of contents - toes), for example, represents a compromise between requirements of the text type and the methodological and conceptual structure of its subject-matter. The aim of the project is to examine how visual-hierarchical structuring systems are constructed, how knowledge structures are encoded in them, and how they can be exploited to automatically derive ontological knowledge for navigation, archiving, or search tasks. The idea to extract domain concepts and semantic relations mainly from the structural and linguistic information gathered from tables of contents represents a novel approach to ontology learning.
Extending the possibilities for collaborative work with TEI/XML through the usage of a wiki system
(2013)
This paper presents and discusses an integrated project-specific working environment for editing TEI/XML-files and linking entities of interest to a dedicated wiki system. This working environment has been specifically tailored to the workflow in our interdisciplinary digital humanities project GeoBib. It addresses some challenges that arose while working with person-related data and geographical references in a growing collection of TEI/XML-files. While our current solution provides some essential benefits, we also discuss several critical issues and challenges that remain.
The present paper reports the first results of the compilation and annotation of a blog corpus for German. The main aim of the project is the representation of the blog discourse structure and relations between its elements (blog posts, comments) and participants (bloggers, commentators). The data included in the corpus were manually collected from the scientific blog portal SciLogs. The feature catalogue for the corpus annotation includes three types of information which is directly or indirectly provided in the blog or can be construed by means of statistical analysis or computational tools. At this point, only directly available information (e.g. title of the blog post, name of the blogger etc.) has been annotated. We believe, our blog corpus can be of interest for the general study of blog structure or related research questions as well as for the development of NLP methods and techniques (e.g. for authorship detection).
The present paper reports the first results of the compilation and annotation of a blog corpus for German. The main aim of the project is the representation of the blog discourse structure and relations between its elements (blog posts, comments) and participants (bloggers, commentators). The data included in the corpus were manually collected from the scientific blog portal SciLogs. The feature catalogue for the corpus annotation includes three types of information which is directly or indirectly provided in the blog or can be construed by means of statistical analysis or computational tools. At this point, only directly available information (e.g., title of the blog post, name of the blogger etc.) has been annotated. We believe, our blog corpus can be of interest for the general study of blog structure or related research questions as well as for the development of NLP methods and techniques (e.g. for authorship detection).
In this paper we present an approach to faceted search in large language resource repositories. This kind of search which enables users to browse through the repository by choosing their personal sequence of facets heavily relies on the availability of descriptive metadata for the objects in the repository. This approach therefore informs the collection of a minimal set of metatdata for language resources. The work described in this paper has been funded by the EC within the ESFRI infrastructure project CLARIN.
The motivation for this article is to describe a methodology for interrelating and analyzing language and theory-specific corpus data from various languages. As an example phenomeon we use information structure (IS, see [3]) in treebanks from three languages: Spanish, Korean and Japanese. Korean and Japanese are typologically close, while both are typologically different from Spanish. Therefore, the problem of annotating IS is that there are diverging language-specific formal linguistic means for the realization of IS-functions (like “topicalization / contrast”) on various levels like prosody, morphology and word-order. Hence, it is necessary to describe the relations between language-specific formal means and functional views on IS, and how to operationalize these relations for corpus analysis.
This paper presents an extension to the Stuttgart-Tübingen TagSet, the standard part-of-speech tag set for German, for the annotation of spoken language. The additional tags deal with hesitations, backchannel signals, interruptions, onomatopoeia and uninterpretable material. They allow one to capture phenomena specific to spoken language while, at the same time, preserving inter-operability with already existing corpora of written language.
In the NLP literature, adapting a parser to new text with properties different from the training data is commonly referred to as domain adaptation. In practice, however, the differences between texts from different sources often reflect a mixture of domain and genre properties, and it is by no means clear what impact each of those has on statistical parsing. In this paper, we investigate how differences between articles in a newspaper corpus relate to the concepts of genre and domain and how they influence parsing performance of a transition-based dependency parser. We do this by applying various similarity measures for data point selection and testing their adequacy for creating genre-aware parsing models.
We present the annotation of information structure in the MULI project. To learn more about the information structuring means in prosody, syntax and discourse, theory- independent features were defined for each level. We describe the features and illustrate them on an example sentence. To investigate the interplay of features, the representation has to allow for inspecting all three layers at the same time. This is realised by a stand-off XML mark-up with the word as the basic unit. The theory-neutral XML stand-off annotation allows integrating this resource with other linguistic resources such as the Tiger Treebank for German or the Penn treebank for English.
The metadata management system for speech corpora “memasysco” has been developed at the Institut für Deutsche Sprache (IDS) and is applied for the first time to document the speech corpus “German Today”. memasysco is based on a data model for the documentation of speech corpora and contains two generic XML schemas that drive data capture, XML native database storage, dynamic publishing, and information retrieval. The development of memasysco’s information architecture was mainly based on the ISLE MetaData Initiative (IMDI) guidelines for publishing metadata of linguistic resources. However, since we also have to support the corpus management process in research projects at the IDS, we need a finer atomic granularity for some documentation components as well as more restrictive categories to ensure data integrity. The XML metadata of different speech corpus projects are centrally validated and natively stored in an Oracle XML database. The extension of the system to the management of annotations of audio and video signals (e.g. orthographic and phonetic transcriptions) is planned for the near future.
We present SPLICR, the Web-based Sustainability Platform for Linguistic Corpora and Resources. The system is aimed at people who work in Linguistics or Computational Linguistics: a comprehensive database of metadata records can be explored in order to find language resources that could be appropriate for one’s specific research needs. SPLICR also provides an interface that enables users to query and to visualise corpora. The project in which the system is being developed aims at sustainably archiving the ca. 60 language resources that have been constructed in three collaborative research centres. Our project has two primary goals: (a) To process and to archive sustainably the resources so that they are still available to the research community in five, ten, or even 20 years time. (b) To enable researchers to query the resources both on the level of their metadata as well as on the level of linguistic annota-tions. In more general terms, our goal is to enable solutions that leverage the interoperability, reusability, and sustainability of heterogeneous collections of language resources.
Making 1:n explorable: a search interface for the ZAS database of clause-embedding predicates
(2017)
We introduce a recently published corpus-based database of German clause-embedding predicates and present an innovative web application for exploring it. The application displays the predicates and the corpus examples for these predicates in two separate tables that can be browsed and searched in real time. While familiar web interface paradigms make it easy for users to get started, the data presentation and the interactive advanced search components for the two tables are designed to accommodate remarkably complex query needs without the need for resorting to a dedicated query language or a more specialized tool. The 1:n relationship between predicates and their examples is exploited in the two tables in that, e.g. the predicate table also shows, for each predicate and each example attribute, all values that occur in the examples for this predicate. An easy-to-use visual query builder for arbitrary Boolean combinations of search criteria can optionally be displayed to pre-filter the underlying data presented in both tables. Several options for altering quantifier scope can be activated with simple checkboxes and considerably widen the space of searchable constellations.
CoMParS is a resource under construction in the context of the long-term project German Grammar in European Comparison (GDE) at the IDS Mannheim. The principal goal of GDE is to create a novel contrastive grammar of German against the background of other European languages. Alongside German, which is the central focus, the core languages for comparison are English, French, Hungarian and Polish, representing different typological classes. Unlike traditional contrastive grammars available for German, which usually cover language pairs and are based on formal grammatical categories, the new GDE grammar is developed in the spirit of functionalist typology. This implies that, instead of formal criteria, cognitively motivated functional domains in terms of Givón (1984) are used as tertia comparationis. The purpose of CoMParS is to document the empirical basis of the theoretical assumptions of GDE-V and to illustrate the otherwise rather abstract content of grammar books by as many as possible naturally occurring and adequately presented multilingual examples, including information on their use in specific contexts and registers. These examples come from existing parallel corpora, and our presentation will focus on the legal aspects and consequences of this choice of language data.
Our paper describes an experiment aimed to assessment of lexical coverage in web corpora in comparison with the traditional ones for two closely related Slavic languages from the lexicographers’ perspective. The preliminary results show that web corpora should not be considered ― inferior, but rather ― different.
Unlike traditional text corpora collected from trustworthy sources, the content of web based corpora has to be filtered. This study briefly discusses the impact of web spam on corpus usability and emphasizes the importance of removing computer generated text from web corpora.
The paper also presents a keyword comparison of an unfiltered corpus with the same collection of texts cleaned by a supervised classifier trained using FastText. The classifier was able to recognize 71% of web spam documents similar to the training set but lacked both precision and recall when applied to short texts from another data set.
Complex linguistic phenomena, such as Clitic Climbing in Bosnian, Croatian and Serbian, are often described intuitively, only from the perspective of the main tendency. In this paper, we argue that web corpora currently offer the best source of empirical material for studying Clitic Climbing in BCS. They thus allow the most accurate description of this phenomenon, as less frequent constructions can be tracked only in big, well-annotated data sources. We compare the properties of web corpora for BCS with traditional sources and give examples of studies on CC based on web corpora. Furthermore, we discuss problems related to web corpora and suggest some improvements for the future.
Creating CorCenCC (Corpws Cenedlaethol Cymraeg Cyfoes - The National Corpus of Contemporary Welsh)
(2017)
CorCenCC is an interdisciplinary and multiinstitutional project that is creating a large-scale, open-source corpus of contemporary Welsh. CorCenCC will be the first ever large-scale corpus to represent spoken, written and electronicallymediated Welsh (compiling an initial data set of 10 million Welsh words), with a functional design informed, from the outset, by representatives of all anticipated academic and community user groups.
Corpus researchers, along with many other disciplines in science are being put under continual pressure to show accountability and reproducibility in their work. This is unsurprisingly difficult when the researcher is faced with a wide array of methods and tools through which to do their work; simply tracking the operations done can be problematic, especially when toolchains are often configured by the developers, but left largely as a black box to the user. Here we present a scheme for encoding this ‘meta data’ inside the corpus files themselves in a structured data format, along with a proof-of-concept tool to record the operations performed on a file.
The Manatee corpus management system on which the Sketch Engine is built is efficient, but unable to harness the power of today’s multiprocessor machines. We describe a new, compatible implementation of Manatee which we develop in the Go language and report on the performance gains that we obtained.
This article describes a series of ongoing efforts at the Stanford Literary Lab to manage a large collection of literary corpora (~40 billion words). This work is marked by a tension between two competing requirements – the corpora need to be merged together into higher-order collections that can be analyzed as units; but, at the same time, it’s also necessary to preserve granular access to the original metadata and relational organization of each individual corpus. We describe a set of data management practices that try to accommodate both of these requirements – Apache Spark is used to index data as Parquet tables on an HPC cluster at Stanford. Crucially, the approach distinguishes between what we call “canonical” and “combined” corpora, a variation on the well-established notion of a “virtual corpus” (Kupietz et al., 2014; Jakubíek et al., 2014; van Uytvanck, 2010).
This paper outlines the broad research context and rationale for a new international comparable corpus (ICC). The ICC is to be largely modelled on the text categories and their quantities the International Corpus of English with only a few changes. The corpus will initially begin with nine European languages but others may join in due course. The paper reports on those and other agreements made at the inaugural planning meeting in Prague on 22-23 June 2017. It also sets out the project’s goals for its first two years.
Many (modernist) works of literature can be understood by their associativeness, be it constructed or “free”. This network-like character of (modernist) literature has often been addressed by terms like “free association”, connotation”, “context” or “intertext”. This paper proposes an experimental and exemplary approach to intraconnect a literary corpus of the Austrian writer Ilse Aichinger with semantic web-technologies to enable interactive explorations of word-associations.
The paper presents best practices and results from projects dedicated to the creation of corpora of computer-mediated communication and social media interactions (CMC) from four different countries. Even though there are still many open issues related to building and annotating corpora of this type, there already exists a range of tested solutions which may serve as a starting point for a comprehensive discussion on how future standards for CMC corpora could (and should) be shaped like.
Feedback utterances are among the most frequent in dialogue. Feedback is also a crucial aspect of all linguistic theories that take social interaction involving language into account. However, determining communicative functions is a notoriously difficult task both for human interpreters and systems. It involves an interpretative process that integrates various sources of information. Existing work on communicative function classification comes from either dialogue act tagging where it is generally coarse grained concerning the feed- back phenomena or it is token-based and does not address the variety of forms that feed- back utterances can take. This paper introduces an annotation framework, the dataset and the related annotation campaign (involving 7 raters to annotate nearly 6000 utterances). We present its evaluation not merely in terms of inter-rater agreement but also in terms of usability of the resulting reference dataset both from a linguistic research perspective and from a more applicative viewpoint.
We investigate how the granularity of POS tags influences POS tagging, and furthermore, how POS tagging performance relates to parsing results. For this, we use the standard “pipeline” approach, in which a parser builds its output on previously tagged input. The experiments are performed on two German treebanks, using three POS tagsets of different granularity, and six different POS taggers, together with the Berkeley parser. Our findings show that less granularity of the POS tagset leads to better tagging results. However, both too coarse-grained and too fine-grained distinctions on POS level decrease parsing performance.
This paper presents a short insight into a new project at the "Institute for the German Language” (IDS) (Mannheim). It gives an insight into some basic ideas for a corpus-based dictionary of spoken German, which will be developed and compiled by the new project "The Lexicon of spoken German” (Lexik des gesprochenen Deutsch, LeGeDe). The work is based on the "Research and Teaching Corpus of Spoken German” (Forschungs- und Lehrkorpus Gesprochenes Deutsch, FOLK), which is implemented in the "Database for Spoken German” (Datenbank für Gesprochenes Deutsch, DGD). Both resources, the database and the corpus, have been developed at the IDS.
We discovered several recurring errors in the current version of the Europarl Corpus originating both from the web site of the European Parliament and the corpus compilation based thereon. The most frequent error was incompletely extracted metadata leaving non-textual fragments within the textual parts of the corpus files. This is, on average, the case for every second speaker change. We not only cleaned the Europarl Corpus by correcting several kinds of errors, but also aligned the speakers’ contributions of all available languages and compiled every- thing into a new XML-structured corpus. This facilitates a more sophisticated selection of data, e.g. querying the corpus for speeches by speakers of a particular political group or in particular language combinations.
Designing a Bilingual Speech Corpus for French and German Language Learners: a Two-Step Process
(2014)
We present the design of a corpus of native and non-native speech for the language pair French-German, with a special emphasis on phonetic and prosodic aspects. To our knowledge there is no suitable corpus, in terms of size and coverage, currently available for the target language pair. To select the target L1-L2 interference phenomena we prepare a small preliminary corpus (corpus1), which is analyzed for coverage and cross-checked jointly by French and German experts. Based on this analysis, target phenomena on the phonetic and phonological level are selected on the basis of the expected degree of deviation from the native performance and the frequency of occurrence. 14 speakers performed both L2 (either French or German) and L1 material (either German or French). This allowed us to test, recordings duration, recordings material, the performance of our automatic aligner software. Then, we built corpus2 taking into account what we learned about corpus1. The aims are the same but we adapted speech material to avoid too long recording sessions. 100 speakers will be recorded. The corpus (corpus1 and corpus2) will be prepared as a searchable database, available for the scientific community after completion of the project.
The IFCASL corpus is a French-German bilingual phonetic learner corpus designed, recorded and annotated in a project on individualized feedback in computer-assisted spoken language learning. The motivation for setting up this corpus was that there is no phonetically annotated and segmented corpus for this language pair of comparable of size and coverage. In contrast to most learner corpora, the IFCASL corpus incorporate data for a language pair in both directions, i.e. in our case French learners of German, and German learners of French. In addition, the corpus is complemented by two sub-corpora of native speech by the same speakers. The corpus provides spoken data by about 100 speakers with comparable productions, annotated and segmented on the word and the phone level, with more than 50% manually corrected data. The paper reports on inter-annotator agreement and the optimization of the acoustic models for forced speech-text alignment in exercises for computer-assisted pronunciation training. Example studies based on the corpus data with a phonetic focus include topics such as the realization of /h/ and glottal stop, final devoicing of obstruents, vowel quantity and quality, pitch range, and tempo.
The paper presents best practices and results from projects in four countries dedicated to the creation of corpora of computer-mediated communication and social media interactions (CMC). Even though there are still many open issues related to building and annotating corpora of that type, there already exists a range of accessible solutions which have been tested in projects and which may serve as a starting point for a more precise discussion of how future standards for CMC corpora may (and should) be shaped like.
This paper is a contribution to the ongoing discussion on treebank annotation schemes and their impact on PCFG parsing results. We provide a thorough comparison of two German treebanks: the TIGER treebank and the TüBa-D/Z. We use simple statistics on sentence length and vocabulary size, and more refined methods such as perplexity and its correlation with PCFG parsing results, as well as a Principal Components Analysis. Finally we present a qualitative evaluation of a set of 100 sentences from the TüBa- D/Z, manually annotated in the TIGER as well as in the TüBa-D/Z annotation scheme, and show that even the existence of a parallel subcorpus does not support a straightforward and easy comparison of both annotation schemes.
The aim of this paper is to highlight the actual need for corpora that have been annotated based on acoustic information. The acoustic information should be coded in features or properties and is needed to inform further processing systems, i.e. to present a basis for a speech recognition system using linguistic information. Feature annotation of existing corpora in combination with segmental annotation can provide a powerful training material for speech recognition systems, but will as well challenge the further processing of features to segments and syllables. We present here the theoretical preliminaries for our multilingual feature extraction system, that we are currently working on.
This paper presents a thorough examination of the validity of three evaluation measures on parser output. We assess parser performance of an unlexicalised probabilistic parser trained on two German treebanks with different annotation schemes and evaluate parsing results using the PARSEVAL metric, the Leaf-Ancestor metric and a dependency-based evaluation. We reject the claim that the TüBa-D/Z annotation scheme is more adequate then the TIGER scheme for PCFG parsing and show that PARSEVAL should not be used to compare parser performance for parsers trained on treebanks with different annotation schemes. An analysis of specific error types indicates that the dependency-based evaluation is most appropriate to reflect parse quality.
Recent studies focussed on the question whether less-configurational languages like German are harder to parse than English, or whether the lower parsing scores are an artefact of treebank encoding schemes and data structures, as claimed by Kübler et al. (2006). This claim is based on the assumption that PARSEVAL metrics fully reflect parse quality across treebank encoding schemes. In this paper we present new experiments to test this claim. We use the PARSEVAL metric, the Leaf-Ancestor metric as well as a dependency-based evaluation, and present novel approaches measuring the effect of controlled error insertion on treebank trees and parser output. We also provide extensive past-parsing crosstreebank conversion. The results of the experiments show that, contrary to Kübler et al. (2006), the question whether or not German is harder to parse than English remains undecided.
How to Compare Treebanks
(2008)
Recent years have seen an increasing interest in developing standards for linguistic annotation, with a focus on the interoperability of the resources. This effort, however, requires a profound knowledge of the advantages and disadvantages of linguistic annotation schemes in order to avoid importing the flaws and weaknesses of existing encoding schemes into the new standards. This paper addresses the question how to compare syntactically annotated corpora and gain insights into the usefulness of specific design decisions. We present an exhaustive evaluation of two German treebanks with crucially different encoding schemes. We evaluate three different parsers trained on the two treebanks and compare results using EVALB, the Leaf-Ancestor metric, and a dependency-based evaluation. Furthermore, we present TePaCoC, a new testsuite for the evaluation of parsers on complex German grammatical constructions. The testsuite provides a well thought-out error classification, which enables us to compare parser output for parsers trained on treebanks with different encoding schemes and provides interesting insights into the impact of treebank annotation schemes on specific constructions like PP attachment or non-constituent coordination.
We present data-driven methods for the acquisition of LFG resources from two German treebanks. We discuss problems specific to semi-free word order languages as well as problems arising from the data structures determined by the design of the different treebanks. We compare two ways of encoding semi-free word order, as done in the two German treebanks, and argue that the design of the TiGer treebank is more adequate for the acquisition of LFG resources. Furthermore, we describe an architecture for LFG grammar acquisition for German, based on the two German treebanks, and compare our results with a hand-crafted German LFG grammar.