Refine
Year of publication
Document Type
- Conference Proceeding (29)
- Article (3)
- Report (3)
- Working Paper (2)
- Part of a Book (1)
- Preprint (1)
Has Fulltext
- yes (39)
Is part of the Bibliography
- no (39)
Keywords
- Semantik (5)
- Text Mining (5)
- Concreteness (4)
- Information Retrieval (4)
- Computerlinguistik (3)
- Distributional Semantics (3)
- German (3)
- Klassifikation (3)
- Machine Learning (3)
- Open Access (3)
Institute
- Fakultät III - Medien, Information und Design (39) (remove)
Integrating distributional and lexical information for semantic classification of words using MRMF
(2016)
Semantic classification of words using distributional features is usually based on the semantic similarity of words. We show on two different datasets that a trained classifier using the distributional features directly gives better results. We use Support Vector Machines (SVM) and Multirelational Matrix Factorization (MRMF) to train classifiers. Both give similar results. However, MRMF, that was not used for semantic classification with distributional features before, can easily be extended with more matrices containing more information from different sources on the same problem. We demonstrate the effectiveness of the novel approach by including information from WordNet. Thus we show, that MRMF provides an interesting approach for building semantic classifiers that (1) gives better results than unsupervised approaches based on vector similarity, (2) gives similar results as other supervised methods and (3) can naturally be extended with other sources of information in order to improve the results.
The CogALex-V Shared Task provides two datasets that consists of pairs of words along with a classification of their semantic relation. The dataset for the first task distinguishes only between related and unrelated, while the second data set distinguishes several types of semantic relations. A number of recent papers propose to construct a feature vector that represents a pair of words by applying a pairwise simple operation to all elements of the feature vector. Subsequently, the pairs can be classified by training any classification algorithm on these vectors. In the present paper we apply this method to the provided datasets. We see that the results are not better than from the given simple baseline. We conclude that the results of the investigated method are strongly depended on the type of data to which it is applied.
In distributional semantics words are represented by aggregated context features. The similarity of words can be computed by comparing their feature vectors. Thus, we can predict whether two words are synonymous or similar with respect to some other semantic relation. We will show on six different datasets of pairs of similar and non-similar words that a supervised learning algorithm on feature vectors representing pairs of words outperforms cosine similarity between vectors representing single words. We compared different methods to construct a feature vector representing a pair of words. We show that simple methods like pairwise addition or multiplication give better results than a recently proposed method that combines different types of features. The semantic relation we consider is relatedness of terms in thesauri for intellectual document classification. Thus our findings can directly be applied for the maintenance and extension of such thesauri. To the best of our knowledge this relation was not considered before in the field of distributional semantics.
For indexing archived documents the Dutch Parliament uses a specialized thesaurus. For good results for full text retrieval and automatic classification it turns out to be important to add more synonyms to the existing thesaurus terms. In the present work we investigate the possibilities to find synonyms for terms of the parliaments thesaurus automatically. We propose to use distributional similarity (DS). In an experiment with pairs of synonyms and non-synonyms we train and test a classifier using distributional similarity and string similarity. Using ten-fold cross validation we were able to classify 75% of the pairs of a set of 6000 word pairs correctly.
NOA is a search engine for scientific images from open access publications based on full text indexing of all text referring to the images and filtering for disciplines and image type. Images will be annotated with Wikipedia categories for better discoverability and for uploading to WikiCommons. Currently we have indexed approximately 2,7 Million images from over 710 000 scientific papers from all fields of science.
Scientific papers from all disciplines contain many abbreviations and acronyms. In many cases these acronyms are ambiguous. We present a method to choose the contextual correct definition of an acronym that does not require training for each acronym and thus can be applied to a large number of different acronyms with only few instances. We constructed a set of 19,954 examples of 4,365 ambiguous acronyms from image captions in scientific papers along with their contextually correct definition from different domains. We learn word embeddings for all words in the corpus and compare the averaged context vector of the words in the expansion of an acronym with the weighted average vector of the words in the context of the acronym. We show that this method clearly outperforms (classical) cosine similarity. Furthermore, we show that word embeddings learned from a 1 billion word corpus of scientific exts outperform word embeddings learned from much larger general corpora.
Concreteness of words has been studied extensively in psycholinguistic literature. A number of datasets have been created with average values for perceived concreteness of words. We show that we can train a regression model on these data, using word embeddings and morphological features, that can predict these concreteness values with high accuracy. We evaluate the model on 7 publicly available datasets. Only for a few small subsets of these datasets prediction of concreteness values are found in the literature. Our results clearly outperform the reported results for these datasets.
Concreteness of words has been measured and used in psycholinguistics already for decades. Recently, it is also used in retrieval and NLP tasks. For English a number of well known datasets has been established with average values for perceived concreteness.
We give an overview of available datasets for German, their correlation and evaluate prediction algorithms for concreteness of German words. We show that these algorithms achieve similar results as for English datasets. Moreover, we show for all datasets there are no significant differences between a prediction model based on a regression model using word embeddings as features and a prediction algorithm based on word similarity according to the same embeddings.
Image captions in scientific papers usually are complementary to the images. Consequently, the captions contain many terms that do not refer to concepts visible in the image. We conjecture that it is possible to distinguish between these two types of terms in an image caption by analysing the text only. To examine this, we evaluated different features. The dataset we used to compute tf.idf values, word embeddings and concreteness values contains over 700 000 scientific papers with over 4,6 million images. The evaluation was done with a manually annotated subset of 329 images. Additionally, we trained a support vector machine to predict whether a term is a likely visible or not. We show that concreteness of terms is a very important feature to identify terms in captions and context that refer to concepts visible in images.
Regional knowledge map is a tool recently demanded by some actors in an institutional level to help regional policy and innovation in a territory. Besides, knowledge maps facilitate the interaction between the actors of a territory and the collective learning. This paper reports the work in progress of a research project which objective is to define a methodology to efficiently design territorial knowledge maps, by extracting information of big volumes of data contained in diverse sources of information related to a region. Knowledge maps facilitate management of the intellectual capital in organisations. This paper investigates the value to apply this tool to a territorial region to manage the structures, infrastructures and the resources to enable regional innovation and regional development. Their design involves the identification of information sources that are required to find which knowledge is located in a territory, which actors are involved in innovation, and which is the context to develop this innovation (structures, infrastructures, resources and social capital). This paper summarizes the theoretical background and framework for the design of a methodology for the construction of knowledge maps, and gives an overview of the main challenges for the design of regional knowledge maps.
Discovery and efficient reuse of technology pictures using Wikimedia infrastructures. A proposal
(2016)
Multimedia objects, especially images and figures, are essential for the visualization and interpretation of research findings. The distribution and reuse of these scientific objects is significantly improved under open access conditions, for instance in Wikipedia articles, in research literature, as well as in education and knowledge dissemination, where licensing of images often represents a serious barrier.
Whereas scientific publications are retrievable through library portals or other online search services due to standardized indices there is no targeted retrieval and access to the accompanying images and figures yet. Consequently there is a great demand to develop standardized indexing methods for these multimedia open access objects in order to improve the accessibility to this material.
With our proposal, we hope to serve a broad audience which looks up a scientific or technical term in a web search portal first. Until now, this audience has little chance to find an openly accessible and reusable image narrowly matching their search term on first try - frustratingly so, even if there is in fact such an image included in some open access article.
In the present paper we sketch an automated procedure to compare different versions of a contract. The contract texts used for this purpose are structurally differently composed PDF files that are converted into structured XML files by identifying and classifying text boxes. A classifier trained on manually annotated contracts achieves an accuracy of 87% on this task. We align contract versions and classify aligned text fragments into different similarity classes that enhance the manual comparison of changes in document versions. The main challenges are to deal with OCR errors and different layout of identical or similar texts. We demonstrate the procedure using some freely available contracts from the City of Hamburg written in German. The methods, however, are language agnostic and can be applied to other contracts as well.
The reuse of scientific raw data is a key demand of Open Science. In the project NOA we foster reuse of scientific images by collecting and uploading them to Wikimedia Commons. In this paper we present a text-based annotation method that proposes Wikipedia categories for open access images. The assigned categories can be used for image retrieval or to upload images to Wikimedia Commons. The annotation basically consists of two phases: extracting salient keywords and mapping these keywords to categories. The results are evaluated on a small record of open access images that were manually annotated.
For the analysis of contract texts, validated model texts, such as model clauses, can be used to identify used contract clauses. This paper investigates how the similarity between titles of model clauses and headings extracted from contracts can be computed, and which similarity measure is most suitable for this. For the calculation of the similarities between title pairs we tested various variants of string similarity and token based similarity. We also compare two additional semantic similarity measures based on word embeddings using pre-trained embeddings and word embeddings trained on contract texts. The identification of the model clause title can be used as a starting point for the mapping of clauses found in contracts to verified clauses.
In order to ensure validity in legal texts like contracts and case law, lawyers rely on standardised formulations that are written carefully but also represent a kind of code with a meaning and function known to all legal experts. Using directed (acyclic) graphs to represent standardized text fragments, we are able to capture variations concerning time specifications, slight rephrasings, names, places and also OCR errors. We show how we can find such text fragments by sentence clustering, pattern detection and clustering patterns. To test the proposed methods, we use two corpora of German contracts and court decisions, specially compiled for this purpose. However, the entire process for representing standardised text fragments is language-agnostic. We analyze and compare both corpora and give an quantitative and qualitative analysis of the text fragments found and present a number of examples from both corpora.
Legal documents often have a complex layout with many different headings, headers and footers, side notes, etc. For the further processing, it is important to extract these individual components correctly from a legally binding document, for example a signed PDF. A common approach to do so is to classify each (text) region of a page using its geometric and textual features. This approach works well, when the training and test data have a similar structure and when the documents of a collection to be analyzed have a rather uniform layout. We show that the use of global page properties can improve the accuracy of text element classification: we first classify each page into one of three layout types. After that, we can train a classifier for each of the three page types and thereby improve the accuracy on a manually annotated collection of 70 legal documents consisting of 20,938 text elements. When we split by page type, we achieve an improvement from 0.95 to 0.98 for single-column pages with left marginalia and from 0.95 to 0.96 for double-column pages. We developed our own feature-based method for page layout detection, which we benchmark against a standard implementation of a CNN image classifier. The approach presented here is based on corpus of freely available German contracts and general terms and conditions.
Both the corpus and all manual annotations are made freely available. The method is language agnostic.
Generalisierte Rechtsdokumente, bei denen für die individuellen Ausprägungen eines Vertrages die Positionen im Text bekannt sind, können eingesetzt werden, um erstens das Genehmigungsverfahren von Neuverträgen automatisiert zu unterstützen und zweitens als Vertragsgenerator neue Rechtsdokumente vorausgewählt zur Verfügung zu stellen. In diesem Beitrag wird, mithilfe von bekannten juristischen Texten gezeigt, wie formelhafte Textabschnitte identifiziert und häufige individuelle Ausprägungen klassifiziert werden können, um als Musterabschnitte eingesetzt zu werden. Es werden Einsatzbereiche vorgestellt und vorhandenes Potential für Legal Tech-Anwendungen aufgezeigt.
To learn a subject, the acquisition of the associated technical language is important.
Despite this widely accepted importance of learning the technical language, hardly any studies are published that describe the characteristics of most technical languages that students are supposed to learn. This might largely be due to the absence of specialized text corpora to study such languages at lexical, syntactical and textual level. In the present paper we describe a corpus of German physics text that can be used to study the language used in physics. A large and a small variant are compiled. The small version of the corpus consists of 5.3 Million words and is available on request.
The amount of papers published yearly increases since decades. Libraries need to make these resources accessible and available with classification being an important aspect and part of this process. This paper analyzes prerequisites and possibilities of automatic classification of medical literature. We explain the selection, preprocessing and analysis of data consisting of catalogue datasets from the library of the Hanover Medical School, Lower Saxony, Germany. In the present study, 19,348 documents, represented by notations of library classification systems such as e.g. the Dewey Decimal Classification (DDC), were classified into 514 different classes from the National Library of Medicine (NLM) classification system. The algorithm used was k-nearest-neighbours (kNN). A correct classification rate of 55.7% could be achieved. To the best of our knowledge, this is not only the first research conducted towards the use of the NLM classification in automatic classification but also the first approach that exclusively considers already assigned notations from other
classification systems for this purpose.
Editorial for the 17th European Networked Knowledge Organization Systems Workshop (NKOS 2017)
(2017)
Knowledge Organization Systems (KOS), in the form of classification systems, thesauri, lexical databases, ontologies, and taxonomies, play a crucial role in digital information management and applications generally. Carrying semantics in a well-controlled and documented way, Knowledge Organization Systems serve a variety of important functions: tools for representation and indexing of information and documents, knowledge-based support to information searchers, semantic road maps to domains and disciplines, communication tool by providing conceptual framework, and conceptual basis for knowledge based systems, e.g. automated classification systems. New networked KOS (NKOS) services and applications are emerging, and we have reached a stage where many KOS standards exist and the integration of linked services is no longer just a future scenario. This editorial describes the workshop outline and overview of presented papers at the 17th European Networked Knowledge Organization Systems Workshop (NKOS 2017) which was held during the TPDL 2017 Conference in Thessaloniki, Greece.
Editorial for the 15th European Networked Knowledge Organization Systems Workshop (NKOS 2016)
(2016)
Knowledge Organization Systems (KOS), in the form of classification systems, thesauri, lexical databases, ontologies, and taxonomies, play a crucial role in digital information management and applications generally. Carrying semantics in a well-controlled and documented way, Knowledge Organisation Systems serve a variety of important functions: tools for representation and indexing of information and documents, knowledge-based support to information searchers, semantic road maps to domains and disciplines, communication tool by providing conceptual framework, and conceptual basis for knowledge based systems, e.g. automated classification systems. New networked KOS (NKOS) services and applications are emerging, and we have reached a stage where many KOS standards exist and the integration of linked services is no longer just a future scenario. This editorial describes the workshop outline and overview of presented papers at the 15th European Networked Knowledge Organization Systems Workshop (NKOS 2016) in Hannover, Germany.
This paper summarizes the results of a comprehensive statistical analysis on a corpus of open access articles and contained figures. It gives an insight into quantitative relationships between illustrations or types of illustrations, caption lengths, subjects, publishers, author affiliations, article citations and others.
We compare the effect of different text segmentation strategies on speech based passage retrieval of video. Passage retrieval has mainly been studied to improve document retrieval and to enable question answering. In these domains best results were obtained using passages defined by the paragraph structure of the source documents or by using arbitrary overlapping passages. For the retrieval of relevant passages in a video, using speech transcripts, no author defined segmentation is available. We compare retrieval results from 4 different types of segments based on the speech channel of the video: fixed length segments, a sliding window, semantically coherent segments and prosodic segments. We evaluated the methods on the corpus of the MediaEval 2011 Rich Speech Retrieval task. Our main conclusion is that the retrieval results highly depend on the right choice for the segment length. However, results using the segmentation into semantically coherent parts depend much less on the segment length. Especially, the quality of fixed length and sliding window segmentation drops fast when the segment length increases, while quality of the semantically coherent segments is much more stable. Thus, if coherent segments are defined, longer segments can be used and consequently less segments have to be considered at retrieval time.
Distributional semantics tries to characterize the meaning of words by the contexts in which they occur. Similarity of words hence can be derived from the similarity of contexts. Contexts of a word are usually vectors of words appearing near to that word in a corpus. It was observed in previous research that similarity measures for the context vectors of two words depend on the frequency of these words. In the present paper we investigate this dependency in more detail for one similarity measure, the Jensen-Shannon divergence. We give an empirical model of this dependency and propose the deviation of the observed Jensen-Shannon divergence from the divergence expected on the basis of the frequencies of the words as an alternative similarity measure. We show that this new similarity measure is superior to both the Jensen-Shannon divergence and the cosine similarity in a task, in which pairs of words, taken from Wordnet, have to be classified as being synonyms or not.
Lemmatization is a central task in many NLP applications. Despite this importance, the number of (freely) available and easy to use tools for German is very limited. To fill this gap, we developed a simple lemmatizer that can be trained on any lemmatized corpus. For a full form word the tagger tries to find the sequence of morphemes that is most likely to generate that word. From this sequence of tags we can easily derive the stem, the lemma and the part of speech (PoS) of the word. We show (i) that the quality of this approach is comparable to state of the art methods and (ii) that we can improve the results of Part-of-Speech (PoS) tagging when we include the morphological analysis of each word.
We compare the effect of different segmentation strategies for passage retrieval of user generated internet video. We consider retrieval of passages for rather abstract and complex queries that go beyond finding a certain object or constellation of objects in the visual channel. Hence the retrieval methods have to rely heavily on the recognized speech. Passage retrieval has mainly been studied to improve document retrieval and to enable question answering. In these domains best results were obtained using passages defined by the paragraph structure of the source documents or by using arbitrary overlapping passages. For the retrieval of relevant passages in a video no author defined paragraph structure is available. We compare retrieval results from 5 different types of segments: segments defined by shot boundaries, prosodic segments, fixed length segments, a sliding window and semantically coherent segments based on speech transcripts. We evaluated the methods on the corpus of the MediaEval 2011 Rich Speech Retrieval task. Our main conclusions are (1) that fixed length and coherent segments are clearly superior to segments based on speaker turns or shot boundaries; (2) that the retrieval results highly depend on the right choice for the segment length; and (3) that results using the segmentation into semantically coherent parts depend much less on the segment length. Especially, the quality of fixed length and sliding window segmentation drops fast when the segment length increases, while quality of the semantically coherent segments is much more stable. Thus, if coherent segments are defined, longer segments can be used and consequently fewer segments have to be considered at retrieval time.
The dependency of word similarity in vector space models on the frequency of words has been noted in a few studies, but has received very little attention. We study the influence of word frequency in a set of 10 000 randomly selected word pairs for a number of different combinations of feature weighting schemes and similarity measures. We find that the similarity of word pairs for all methods, except for the one using singular value decomposition to reduce the dimensionality of the feature space, is determined to a large extent by the frequency of the words. In a binary classification task of pairs of synonyms and unrelated words we find that for all similarity measures the results can be improved when we correct for the frequency bias.
This paper describes the approach of the Hochschule Hannover to the SemEval 2013 Task Evaluating Phrasal Semantics. In order to compare a single word with a two word phrase we compute various distributional similarities, among which a new similarity measure, based on Jensen-Shannon Divergence with a correction for frequency effects. The classification is done by a support vector machine that uses all similarities as features. The approach turned out to be the most successful one in the task.
This paper presents a possibility to extend the formalism of linear indexed grammars. The extension is based on the use of tuples of pushdowns instead of one pushdown to store indices during a derivation. If a restriction on the accessibility of the pushdowns is used, it can be shown that the resulting formalisms give rise to a hierarchy of languages that is equivalent with a hierarchy defined by Weir. For this equivalence, that was already known for a slightly different formalism, this paper gives a new proof. Since all languages of Weir's hierarchy are known to be mildly context sensitive, the proposed extensions of LIGs become comparable with extensions of tree adjoining grammars and head grammars.
In this paper we investigate how concreteness and abstractness are represented in word embedding spaces. We use data for English and German, and show that concreteness and abstractness can be determined independently and turn out to be completely opposite directions in the embedding space. Various methods can be used to determine the direction of concreteness, always resulting in roughly the same vector. Though concreteness is a central aspect of the meaning of words and can be detected clearly in embedding spaces, it seems not as easy to subtract or add concreteness to words to obtain other words or word senses like e.g. can be done with a semantic property like gender.
Diese Studie untersucht Gruppen von Ortsnamen in Deutschland (in den Postleitregionen) nach vorhandenen Ähnlichkeiten. Als Messgröße wird ein Häufigkeitsvektor von Trigrammen in jeder Gruppe herangezogen. Mit der Anwendung des Average Linkage-Algorithmus auf die Messgröße werden Cluster aus räumlich zusammenhängenden Gebieten gebildet, obwohl das Verfahren keine Kenntnis über die Lage der Cluster zueinander besitzt. In den Clustern werden die zehn häufigsten n-Gramme ermittelt, um charakteristische Wortpartikel darzustellen. Die von den Clustern umschriebenen Gebiete lassen sich zwanglos durch historische oder linguistische Entwicklungen erklären. Das hier verwendete Verfahren setzt jedoch kein linguistisches, geographisches oder historisches Wissen voraus, ermöglicht aber die Gruppierung von Namen in eindeutiger Weise unter Berücksichtigung einer Vielzahl von Wortpartikeln in einem Schritt. Die Vorgehensweise ohne Vorwissen unterscheidet diese Studie von den meisten bisher angewendeten Untersuchungen.
Library of Congress Subject Headings (LCSH) are popular for indexing library records. We studied the possibility of assigning LCSH automatically by training classifiers for terms used frequently in a large collection of abstracts of the literature on hand and by extracting headings from those abstracts. The resulting classifiers reach an acceptable level of precision, but fail in terms of recall partly because we could only train classifiers for a small number of LCSH. Extraction, i.e., the matching of headings in the text, produces better recall but extremely low precision. We found that combining both methods leads to a significant improvement of recall and a slight improvement of F1 score with only a small decrease in precision.
Regional Innovation Systems describe the relations between actors, structures and infrastructures in a region in order to stimulate innovation and regional development. For these systems the collection and organization of information is crucial. In the present paper we investigate the possibilities to extract information from websites of companies. First we describe regional innovation systems and the information types that are necessary to create them. Then we discuss the possibilities of text mining and keyword extraction techniques to extract this information from company websites. Finally, we describe a small scale experiment in which keywords related to economic sectors and commodities are extracted from the websites of over 200 companies. This experiment shows what the main challenges are for information extraction from websites for regional innovation systems.
We present a simple method to find topics in user reviews that accompany ratings for products or services. Standard topic analysis will perform sub-optimal on such data since the word distributions in the documents are not only determined by the topics but by the sentiment as well. We reduce the influence of the sentiment on the topic selection by adding two explicit topics, representing positive and negative sentiment. We evaluate the proposed method on a set of over 15,000 hospital reviews. We show that the proposed method, Latent Semantic Analysis with explicit word features, finds topics with a much smaller bias for sentiments than other similar methods.
Automatic classification of scientific records using the German Subject Heading Authority File (SWD)
(2012)
The following paper deals with an automatic text classification method which does not require training documents. For this method the German Subject Heading Authority File (SWD), provided by the linked data service of the German National Library is used. Recently the SWD was enriched with notations of the Dewey Decimal Classification (DDC). In consequence it became possible to utilize the subject headings as textual representations for the notations of the DDC. Basically, we we derive the classification of a text from the classification of the words in the text given by the thesaurus. The method was tested by classifying 3826 OAI-Records from 7 different repositories. Mean reciprocal rank and recall were chosen as evaluation measure. Direct comparison to a machine learning method has shown that this method is definitely competitive. Thus we can conclude that the enriched version of the SWD provides high quality information with a broad coverage for classification of German scientific articles.