Refine
Document Type
Language
- English (27) (remove)
Has Fulltext
- yes (27)
Is part of the Bibliography
- no (27)
Keywords
- Computerlinguistik (27) (remove)
Institute
In this paper, we present our Named Entity Recognition (NER) system for German – NERU (Named Entity Rules), which heavily relies on handcrafted rules as well as information gained from a cascade of existing external NER tools. The system combines large gazetteer lists, information obtained by comparison of different automatic translations and POS taggers. With NERU, we were able to achieve a score of 73.26% on the development set provided by the GermEval 2014 Named Entity Recognition Shared Task for German.
This paper presents a Named Entity Recognition system for German based on Conditional Random Fields. The model also includes language-independant features and features computed form large coverage lexical resources. Along side the results themselves, we show that by adding linguistic resources to a probabilistic model, the results improve significantly.
In the latest decades, machine learning approaches have been intensively experimented for natural language processing. Most of the time, systems rely on using statistics within the system, by analyzing texts at the token level and, for labelling tasks, categorizing each among possible classes. One may notice that previous symbolic approaches (e.g. transducers) where designed to delimit pieces of text. Our research team developped mXS, a system that aims at combining both approaches. It locates boundaries of entities by using sequential pattern mining and machine learning. This system, intially developped for French, has been adapted to German.
In this paper, we investigate a semi- supervised learning approach based on neu- ral networks for nested named entity recog- nition on the GermEval 2014 dataset. The dataset consists of triples of a word, a named entity associated with that word in the first-level and one in the second-level. Additionally, the tag distribution is highly skewed, that is, the number of occurrences of certain types of tags is too small. Hence, we present a unified neural network archi- tecture to deal with named entities in both levels simultaneously and to improve gen- eralization performance on the classes that have a small number of labelled examples.
In this paper we present Nessy (Named Entity Searching System) and its application to German in the context of the GermEval 2014 Named Entity Recognition Shared Task (Benikova et al., 2014a). We tackle the challenge by using a combination of machine learning (Naive Bayes classification) and rule-based methods. Altogether, Nessy achieves an F-score of 58.78% on the final test set.
This paper presents the BECREATIVE Named Entity Recognition system and its participation at the GermEval 2014 Named Entity Recognition Shared Task (Benikova et al., 2014a). BECREATIVE uses a hybrid approach of two commonly used procedural methods, namely list-based lookups and machine learning (Naive Bayes Classification), which centers around the classifier. BECREATIVE currently reaches an F-score of 37.34 on the strict evaluation setting applied on the development set provided by GermEval.
This paper describes the DRIM Named Entity Recognizer (DRIM), developed for the GermEval 2014 Named Entity (NE) Recognition Shared Task. The shared task did not pose any restrictions regarding the type of named entity recognition (NER) system submissions and usage of external data, which still resulted in a very challenging task. We employ Linear Support Vector Classification (Linear SVC) in the implementation of SckiKit, with variety of features, gazetteers and further contextual information of the target words. As there is only one level of embedding in the dataset, two separate classifiers are trained for the outer and inner spans. The system was developed and tested on the dataset provided by the GermEval 2014 NER Shared Task. The overall strict (fine-grained) score is 70.94% on the development set, and 69.33% on the final test set which is quite promising for the German language.
This paper describes our classification and rule-based attempt at nested Named Entity Recognition for German. We explain how both approaches interact with each other and the resources we used to achieve our results. Finally, we evaluate the overall performance of our system which achieves an F-score of 52.65% on the development set and 52.11% on the final test set of the GermEval 2014 Shared Task.
MoSTNER is a German NER system based on machine learning with log-linear models and morphology-aware features. We use morphological analysis with Morphisto for generating features, moreover we use German Wikipedia as a gazetteer and perform punctuation-aware and morphology-aware page title matching. We use four types of factor graphs where NER labels are single variables or split into prefix (BILOU) and type (PER, LOC, etc.) variables. Our system supports nested NER (two levels), for training we use SampleRank, for prediction Iterated Conditional Modes, the implementation is based on Python and Factorie.
Collobert et al. (2011) showed that deep neural network architectures achieve state- of-the-art performance in many fundamental NLP tasks, including Named Entity Recognition (NER). However, results were only reported for English. This paper reports on experiments for German Named Entity Recognition, using the data from the GermEval 2014 shared task on NER. Our system achieves an F1 -measure of 75.09% according to the official metric.
Modular Classifier Ensemble Architecture for Named Entity Recognition on Low Resource Systems
(2014)
This paper presents the best performing Named Entity Recognition system in the GermEval 2014 Shared Task. Our approach combines semi-automatically created lexical resources with an ensemble of binary classifiers which extract the most likely tag sequence. Out-of-vocabulary words are tackled with semantic generalization extracted from a large corpus and an ensemble of part-of-speech taggers, one of which is unsupervised. Unknown candidate sequences are resolved using a look-up with the Wikipedia API.
This paper describes the GermEval 2014 Named Entity Recognition (NER) Shared Task workshop at KONVENS. It provides background information on the motivation of this task, the data-set, the evaluation method, and an overview of the participating systems, followed by a discussion of their results. In contrast to previous NER tasks, the GermEval 2014 edition uses an extended tagset to account for derivatives of names and tokens that contain name parts. Further, nested named entities had to be predicted, i.e. names that contain other names. The eleven participating teams employed a wide range of techniques in their systems. The most successful systems used state-of-the- art machine learning methods, combined with some knowledge-based features in hybrid systems.
Ironic speech act detection is indispensable for automatic opinion mining. This paper presents a pattern-based approach for the detection of ironic speech acts in German Web comments. The approach is based on a multilevel annotation model. Based on a gold standard corpus with labeled ironic sentences, multilevel patterns are deter- mined according to statistical and linguis- tic analysis. The extracted patterns serve to detect ironic speech acts in a Web com- ment test corpus. Automatic detection and inter-annotator results achieved by human annotators show that the detection of ironic sentences is a challenging task. However, we show that it is possible to automatically detect ironic sentences with relatively high precision up to 63%.
The workshops hosted at this iteration of KONVENS also reflect the interaction of, and common themes shared between, Computational Linguistics and Information Science: a focus on on evaluation, represented by shared tasks on Named Entity Recognition (GermEval) and on Sentiment Analysis (GESTALT); a growing interest in the processing of non-canonical text such as that found in social media (NLP4CMC) or patent documents (IPaMin); multi-disciplinary research which combines Information Science, Computer Aided Language Learning, Natural Language Processing, and E-Lexicography with the objective of creating language learning and training systems that provide intelligent feedback based on rich knowledge (ISCALPEL).
It is of interest to study sentence construction for children’s writing in order to understand grammatical errors and their influence on didactic decisions. For this purpose, this paper analyses sentence structures for various age groups of children’s writings in contrast to text taken from children’s and youth literature. While valency differs little between text type and age group, sentence embellishments show some differences. Both use of adjectives and adverbs increase with age and book levels. Furthermore books show a larger use thereof. This work presents one of the steps in a larger ongoing effort to understand children’s writing and reading competences at word and sentence level. The need to look at variable from non-variable features of sentence structures separately in order to find distinctive features has been an important finding.
We present WebNLP, a web-based tool that combines natural language processing (NLP) functionality from Python NLTK and text visualizations from Voyant in an integrated interface. Language data can be uploaded via the website. The results of the processed data are displayed as plain text, XML markup, or Voyant visualizations in the same website. WebNLP aims at facilitating the usage of NLP tools for users without technical skills and experience with command line interfaces. It also makes up for the shortcomings of the popular text analysis tool Voyant, which, up to this point, is lacking basic NLP features such as lemmatization or POS tagging.
This paper presents Atomic, an open-source platform-independent desktop application for multi-level corpus annotation. Atomic aims at providing the linguistic community with a user-friendly annotation tool and sustainable platform through its focus on extensibility, a generic data model, and compatibility with existing linguistic formats. It is implemented on top of the Eclipse Rich Client Platform, a pluggable Java-based framework for creating client applications. Atomic - as a set of plug-ins for this framework - integrates with the platform and allows other researchers to develop and integrate further extensions to the software as needed. The generic graph-based meta model Salt serves as Atomic’s domain model and allows for unlimited annotation levels and types. Salt is also used as an intermediate model in the Pepper framework for conversion of linguistic data, which is fully integrated into Atomic, making the latter compatible with a wide range of linguistic formats. Atomic provides tools for both less experienced and expert annotators: graphical, mouse-driven editors and a command-line data manipulation language for rapid annotation.
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.
German Perception Verbs: Automatic Classification of Prototypical and Multiple Non-literal Meanings
(2014)
This paper presents a token-based automatic classification of German perception verbs into literal vs. multiple non-literal senses. Based on a corpus-based dataset of German perception verbs and their systematic meaning shifts, we identify one verb of each of the four perception classes optical, acoustic, olfactory, haptic, and use Decision Trees relying on syntactic and semantic corpus-based features to classify the verb uses into 3-4 senses each. Our classifier reaches accuracies between 45.5% and 69.4%, in comparison to baselines between 27.5% and 39.0%. In three out of four cases analyzed our classifier’s accuracy is significantly higher than the according baseline.
In this work we consider the problem of social media text Part-of-Speech tagging as fundamental task for Natural Language Processing. We present improvements to a social media Markov model tagger, by adapting parameter estimation methods for unknown tokens. In addition, we propose to enrich the social media text corpus by a linear combination with a newspaper training corpus. Applying our tagger to a social media text corpus results in accuracies of around 94.8%, which comes close to accuracies for standardized texts.
A common way to express sentiment about some product is by comparing it to a different product. The anchor for the comparison is a comparative predicate like “better”. In this work we concentrate on the annotation of multiword predicates like “more powerful”. In the single-token-based approaches which are mostly used for the automatic detection of comparisons, one of the words has to be selected as the comparative predicate. In our first experiment, we investigate the influence of this decision on the classification performance of a machine learning system and show that annotating the modifier gives better results. In the annotation conventions adopted in standard datasets for sentiment analysis, the modified adjective is annotated as the aspect of the comparison. We discuss problems with this type of annotation and propose the introduction of an additional argument type which solves the problems. In our second experiment we show that there is only a small drop in performance when adding this new argument type.
Enterprises express the concepts of their electronic business-to-business (B2B) communication in individual ontology-like schemas. Collaborations require merging schemas’ common concepts into Business Entities (BEs) in a Canonical Data Model (CDM). Although consistent, automatic schema merging is state of the art, the task of labeling the BEs with descriptive, yet short and unique names, remains. Our approach first derives a heuristically ranked list of candidate labels for each BE locally from the names and descriptions of the underlying concepts. Second, we use constraint satisfaction to assign a semantically unique name to each BE that optimally distinguishes it from the other BEs. Our system’s labels outperform previous work in their description of BE content and in their discrimination between similar BEs. In a task-based evaluation, business experts estimate that our approach can save about 12% of B2B integration effort compared to previous work and about 49% in total.
Researchers in the (digital) humanities have identified a large potential in the use of automatic text analysis capabilities in their text studies, scaling up the amount of material that can be explored and allowing for new types of questions. To support the scholars’ specific research questions, it is important to embed the text analysis capabilities in an appropriate navigation and visualization framework. However, study-specific tailoring may make it hard to migrate analytical components across projects and compare results. To overcome this issue, we present in this paper the first version of TEANLIS (text analysis for literary scholars), a flexible framework designed to include text analysis capabilities for literary scholars.
Verb Polarity Frames: a New Resource and its Application in Target-specific Polarity Classification
(2014)
We discuss target-specific polarity classification for German news texts. Novel, verb-specific features are used in a Simple Logistic Regression model. The polar perspective a verb casts on its grammatical roles is exploited. Also, an additional, largely neglected polarity class is examined: controversial texts. We found that the straightforward definition of ’controversial’ is problematic. More or less balanced polarities in a text are a poor indicator of controversy. Instead, non-polar wording helps more than polarity aggregation. However, our novel features proved useful for the remaining polarity classes.
We study the influence of information structure on the salience of subjective expressions for human readers. Using an online survey tool, we conducted an experiment in which we asked users to rate main and relative clauses that contained either a single positive or negative or a neutral adjective. The statistical analysis of the data shows that subjective expressions are more prominent in main clauses where they are asserted than in relative clauses where they are presupposed. A corpus study suggests that speakers are sensitive to this differential salience in their production of subjective expressions.
The 2014 issue of KONVENS is even more a forum for exchange: its main topic is the interaction between Computational Linguistics and Information Science, and the synergies such interaction, cooperation and integrated views can produce. This topic at the crossroads of different research traditions which deal with natural language as a container of knowledge, and with methods to extract and manage knowledge that is linguistically represented is close to the heart of many researchers at the Institut für Informationswissenschaft und Sprachtechnologie of Universität Hildesheim: it has long been one of the a institute’s research topics, and it has received even more attention over the last few years. The main conference papers deal with this topic from different points of view, involving flat as well as deep representations, automatic methods targeting annotation and hybrid symbolic and statistical processing, as well as new Machine Learning-based approaches, but also the creation of language resources for both machines and humans, and methods for testing the latter to optimize their human-machine interaction properties. In line with the general topic, KONVENS-2014 focuses on areas of research which involve this cooperation of information science and computational linguistics: for example learning-based approaches, (cross-lingual) Information Retrieval, Sentiment Analysis, paraphrasing or dictionary and corpus creation, management and usability. The workshops hosted at this iteration of KONVENS also reflect the interaction of, and common themes shared between, Computational Linguistics and Information Science: a focus on on evaluation, represented by shared tasks on Named Entity Recognition (GermEval) and on Sentiment Analysis (GESTALT); a growing interest in the processing of non-canonical text such as that found in social media (NLP4CMC) or patent documents (IPaMin); multi-disciplinary research which combines Information Science, Computer Aided Language Learning, Natural Language Processing, and E-Lexicography with the objective of creating language learning and training systems that provide intelligent feedback based on rich knowledge (ISCALPEL).
