Refine
Year of publication
Document Type
- Preprint (25) (remove)
Has Fulltext
- yes (25)
Is part of the Bibliography
- no (25)
Keywords
- Deutsch (5)
- Satzanalyse (3)
- Acquisition (1)
- Arabisch (1)
- Automatische Sprachanalyse (1)
- Englisch (1)
- Fehlererkennung (1)
- Fehlertoleranz (1)
- German (1)
- Ilmenau <2000> (1)
Institute
- Extern (25) (remove)
The purpose of this paper is to describe recent developments in the morphological, syntactic, and semantic annotation of the TüBa-D/Z treebank of German. The TüBa-D/Z annotation scheme is derived from the Verbmobil treebank of spoken German [4, 10], but has been extended along various dimensions to accommodate the characteristics of written texts. TüBa-D/Z uses as its data source the "die tageszeitung" (taz) newspaper corpus. The Verbmobil treebank annotation scheme distinguishes four levels of syntactic constituency: the lexical level, the phrasal level, the level of topological fields, and the clausal level. The primary ordering principle of a clause is the inventory of topological fields, which characterize the word order regularities among different clause types of German, and which are widely accepted among descriptive linguists of German [3, 6]. The TüBa-D/Z annotation relies on a context-free backbone (i.e. proper trees without crossing branches) of phrase structure combined with edge labels that specify the grammatical function of the phrase in question. The syntactic annotation scheme of the TüBa-D/Z is described in more detail in [12, 11]. TüBa-D/Z currently comprises approximately 15 000 sentences, with approximately 7 000 sentences being in the correction phase. The latter will be released along with an updated version of the existing treebank before the end of this year. The treebank is available in an XML format, in the NEGRA export format [1] and in the Penn treebank bracketing format. The XML format contains all types of information as described above, the NEGRA export format contains all sentenceinternal information while the Penn treebank format includes only those layers of information that can be expressed as pure tree structures. Over the course of the last year, more fine grained linguistic annotations have been added along the following dimensions: 1. the basic Stuttgart-Tübingen tagset, STTS, [9] labels have been enriched by relevant features of inflectional morphology, 2. named entity information has been encoded as part of the syntactic annotation, and 3. a set of anaphoric and coreference relations has been added to link referentially dependent noun phrases. In the following sections, we will describe each of these innovations in turn and will demonstrate how the additional annotations can be incorporated into one comprehensive annotation scheme.
Part-of-Speech tagging is generally performed by Markov models, based on bigram or trigram models. While Markov models have a strong concentration on the left context of a word, many languages require the inclusion of right context for correct disambiguation. We show for German that the best results are reached by a combination of left and right context. If only left context is available, then changing the direction of analysis and going from right to left improves the results. In a version of MBT (Daelemans et al., 1996) with default parameter settings, the inclusion of the right context improved POS tagging accuracy from 94.00% to 96.08%, thus corroborating our hypothesis. The version with optimized parameters reaches 96.73%.
The definition of similarity between sentences is formulated on the levels of words, POS tags, and chunks (Abney 91; Abney 96). The evaluation of this approach shows that while precision and recall based on the PARSEVAL measures (Black et al. 91) do not reach state of the art Parsers yet (F1=87.19 on syntactic constituents, F1=77.78 including functionargument structure), the parser shows a very reliable performance where function-argument structure is concerned (F1=96.52). The lower F-scores are very often due to unattached constituents.
The problem of vocalization, or diacritization, is essential to many tasks in Arabic NLP. Arabic is generally written without the short vowels, which leads to one written form having several pronunciations with each pronunciation carrying its own meaning(s). In the experiments reported here, we define vocalization as a classification problem in which we decide for each character in the unvocalized word whether it is followed by a short vowel. We investigate the importance of different types of context. Our results show that the combination of using memory-based learning with only a word internal context leads to a word error rate of 6.64%. If a lexical context is added, the results deteriorate slowly.
In syntax, the trend nowadays is towards lexicalized grammar formalisms. It is now widely accepted that dividing words into wordclasses may serve as a laborsaving mechanism - but at the same time, it discards all detailed information on the idiosyncratic behavior of words. And that is exactly the type of information that may be necessary in order to parse a sentence. For learning approaches, however, lexicalized grammars represent a challenge for the very reason that they include so much detailed and specific information, which is difficult to learn. This paper will present an algorithm for learning a link grammar of German. The problem of data sparseness is tackled by using all the available information from partial parses as well as from an existing grammar fragment and a tagger. This is a report about work in progress so there are no representative results available yet.
This paper presents a comparative study of probabilistic treebank parsing of German, using the Negra and TüBa-D/Z treebanks. Experiments with the Stanford parser, which uses a factored PCFG and dependency model, show that, contrary to previous claims for other parsers, lexicalization of PCFG models boosts parsing performance for both treebanks. The experiments also show that there is a big difference in parsing performance, when trained on the Negra and on the TüBa-D/Z treebanks. Parser performance for the models trained on TüBa-D/Z are comparable to parsing results for English with the Stanford parser, when trained on the Penn treebank. This comparison at least suggests that German is not harder to parse than its West-Germanic neighbor language English.
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.
In the last decade, the Penn treebank has become the standard data set for evaluating parsers. The fact that most parsers are solely evaluated on this specific data set leaves the question unanswered how much these results depend on the annotation scheme of the treebank. In this paper, we will investigate the influence which different decisions in the annotation schemes of treebanks have on parsing. The investigation uses the comparison of similar treebanks of German, NEGRA and TüBa-D/Z, which are subsequently modified to allow a comparison of the differences. The results show that deleted unary nodes and a flat phrase structure have a negative influence on parsing quality while a flat clause structure has a positive influence.
Chunk parsing has focused on the recognition of partial constituent structures at the level of individual chunks. Little attention has been paid to the question of how such partial analyses can be combined into larger structures for complete utterances. Such larger structures are not only desirable for a deeper syntactic analysis. They also constitute a necessary prerequisite for assigning function-argument structure. The present paper offers a similaritybased algorithm for assigning functional labels such as subject, object, head, complement, etc. to complete syntactic structures on the basis of prechunked input. The evaluation of the algorithm has concentrated on measuring the quality of functional labels. It was performed on a German and an English treebank using two different annotation schemes at the level of function argument structure. The results of 89.73% correct functional labels for German and 90.40%for English validate the general approach.
In this paper, we investigate the role of sub-optimality in training data for part-of-speech tagging. In particular, we examine to what extent the size of the training corpus and certain types of errors in it affect the performance of the tagger. We distinguish four types of errors: If a word is assigned a wrong tag, this tag can belong to the ambiguity class of the word (i.e. to the set of possible tags for that word) or not; furthermore, the major syntactic category (e.g. "N" or "V") can be correctly assigned (e.g. if a finite verb is classified as an infinitive) or not (e.g. if a verb is classified as a noun). We empirically explore the decrease of performance that each of these error types causes for different sizes of the training set. Our results show that those types of errors that are easier to eliminate have a particularly negative effect on the performance. Thus, it is worthwhile concentrating on the elimination of these types of errors, especially if the training corpus is large.
Prepositional phrase (PP) attachment is one of the major sources for errors in traditional statistical parsers. The reason for that lies in the type of information necessary for resolving structural ambiguities. For parsing, it is assumed that distributional information of parts-of-speech and phrases is sufficient for disambiguation. For PP attachment, in contrast, lexical information is needed. The problem of PP attachment has sparked much interest ever since Hindle and Rooth (1993) formulated the problem in a way that can be easily handled by machine learning approaches: In their approach, PP attachment is reduced to the decision between noun and verb attachment; and the relevant information is reduced to the two possible attachment sites (the noun and the verb) and the preposition of the PP. Brill and Resnik (1994) extended the feature set to the now standard 4-tupel also containing the noun inside the PP. Among many publications on the problem of PP attachment, Volk (2001; 2002) describes the only system for German. He uses a combination of supervised and unsupervised methods. The supervised method is based on the back-off model by Collins and Brooks (1995), the unsupervised part consists of heuristics such as ”If there is a support verb construction present, choose verb attachment”. Volk trains his back-off model on the Negra treebank (Skut et al., 1998) and extracts frequencies for the heuristics from the ”Computerzeitung”. The latter also serves as test data set. Consequently, it is difficult to compare Volk’s results to other results for German, including the results presented here, since not only he uses a combination of supervised and unsupervised learning, but he also performs domain adaptation. Most of the researchers working on PP attachment seem to be satisfied with a PP attachment system; we have found hardly any work on integrating the results of such approaches into actual parsers. The only exceptions are Mehl et al. (1998) and Foth and Menzel (2006), both working with German data. Mehl et al. report a slight improvement of PP attachment from 475 correct PPs out of 681 PPs for the original parser to 481 PPs. Foth and Menzel report an improvement of overall accuracy from 90.7% to 92.2%. Both integrate statistical attachment preferences into a parser. First, we will investigate whether dependency parsing, which generally uses lexical information, shows the same performance on PP attachment as an independent PP attachment classifier does. Then we will investigate an approach that allows the integration of PP attachment information into the output of a parser without having to modify the parser: The results of an independent PP attachment classifier are integrated into the parse of a dependency parser for German in a postprocessing step.
Maschinelles Lernen wird häufig zur effzienten Annotation großer Datenmengen eingesetzt. Die Forschung zu maschinellen Lernverfahren beschränkt sich i.a. darauf unterschiedliche Lernverfahren zu vergelichen oder die optimale größe der Trainingsdaten zu bestimmen. Bisher wurde jedoch nicht untersucht, in wie weit sich linguistisches Wissen bei der Aufgabendefinition positiv auswirken kann. Dies soll hier anhand des Lernens von Base-Nominalphrasen mit drei unterschiedlichen Definitionen untersucht werden. Die Definitionen unterscheiden sich im Grad der linguistisch motivierten Erweiterungen, die zu einer eher praktisch motivierten ersten Definition hinzu kamen. Die Untersuchungen ergaben, dass sich die Anzahl der falsch klasssifizierten Wörter um ein Drittel reduzieren lässt.
This paper reports on the SYN-RA (SYNtax-based Reference Annotation) project, an on-going project of annotating German newspaper texts with referential relations. The project has developed an inventory of anaphoric and coreference relations for German in the context of a unified, XML-based annotation scheme for combining morphological, syntactic, semantic, and anaphoric information. The paper discusses how this unified annotation scheme relates to other formats currently discussed in the literature, in particular the annotation graph model of Bird and Liberman (2001) and the pie-in-thesky scheme for semantic annotation.
Chunk parsing has focused on the recognition of partial constituent structures at the level of individual chunks. Little attention has been paid to the question of how such partial analyses can be combined into larger structures for complete utterances. The TüSBL parser extends current chunk parsing techniques by a tree-construction component that extends partial chunk parses to complete tree structures including recursive phrase structure as well as function-argument structure. TüSBLs tree construction algorithm relies on techniques from memory-based learning that allow similarity-based classification of a given input structure relative to a pre-stored set of tree instances from a fully annotated treebank. A quantitative evaluation of TüSBL has been conducted using a semi-automatically constructed treebank of German that consists of appr. 67,000 fully annotated sentences. The basic PARSEVAL measures were used although they were developed for parsers that have as their main goal a complete analysis that spans the entire input.This runs counter to the basic philosophy underlying TüSBL, which has as its main goal robustness of partially analyzed structures.
This paper provides an overview of current research on a hybrid and robust parsing architecture for the morphological, syntactic and semantic annotation of German text corpora. The novel contribution of this research lies not in the individual parsing modules, each of which relies on state-of-the-art algorithms and techniques. Rather what is new about the present approach is the combination of these modules into a single architecture. This combination provides a means to significantly optimize the performance of each component, resulting in an increased accuracy of annotation.