- 1 http://www.perseus.tufts.edu/
- 2 http://www.ime.usp.br/~tycho/
1In this survey, we focus on Medieval French and English corpora to highlight the discrepancy between the resources in both languages. We are nevertheless aware of ongoing work, especially for Romance corpora (Latin Dependency Treebank (Bamman and Gregory 2006, Perseus Project1, Medieval Portuguese2, etc.).
2For medieval English, important corpus resources have been built in various projects. Today, they constitute the reference in the domain of syntactically annotated medieval corpora and have become an important base for theoretical research in diachronic syntax.
3Two corpora, the York-Toronto-Helsinki Parsed Corpus of Old English Prose (YCOE, Taylor et al. 2003) and the Penn-Helsinki Parsed Corpus of Middle English (PPCME2, Kroch and Taylor 2000) have been annotated either at the University of Pennsylvania (UPenn, project directed by A. Kroch) or using the probabilistic parsing method developed there. They are publicly available (but the PPCME2 is not free of charge) and distributed with CorpusSearch, a search tool conceived for these (and only these) corpora. The probabilistic parsing method, although leading to a robust automatic tool, implies important cost (pre-tagging, manual correction) for building the training data, and for post-processing (manual correction) the output, at least if precision is an issue.
- 3 http://www-users.york.ac.uk/~lang18/Documentation/syn-ppcme2-lite.htm
- 4 See http://www.ling.upenn.edu/mideng for further examples and documentation
4The syntactic model uses a « limited tree representation in the form of labelled parentheses ».3 The structure in (1) shows an annotated sentence from the PPCME2.4
An annotated sentence from the PPCME2
5In the past few years, two independent projects have built the two largest text corpora for Medieval French, and made them available for research: the Base de Français Médiéval (BFM), initiated in 1989 and since then augmented and enhanced by C. Marchello-Nizia and her team (ENS-LSH Lyon, see Guillot & Marchello-Nizia & Lavrentiev 2007), and the Nouveau Corpus d’Amsterdam (NCA, University of Stuttgart, see Stein et al. 2006). In section 2 of this article, we will discuss the annotation of examples taken from the NCA.
- 5 Alexander-von-Humboldt-Stiftung, Transcoop grant III-DEU1112357: « Computergestützte Analyse und E (...)
6The Nouveau Corpus d’Amsterdam (NCA, Stein et al. 2006) is the first edition of the Old French data collected in the 1980s in the research project of Anthonij Dees for his Atlas des formes linguistiques des textes littéraires de l’ancien français (Dees 1987). In a joint project5 with P. Kunstmann (Ottawa University) the 3,3 million word corpus has been edited in a POS-tagged, lemmatised and XML-formatted version (Stein et al. 2006). The 300 texts and text extracts were selected according to Dees’s principles in order to cover the regional varieties of Old French (see Kunstmann & Stein (2007a) for a general description of the NCA and Stein (2007) for technical information on POS-tagging, lemmatisation and XML markup).
7A syntactic reference corpus for Medieval French is a desideratum not only for French corpus linguistics, but for comparative typology and diachronic syntax, where corpus-based approaches and the consideration of older periods have become increasingly important.
8The publication of the corpora for Medieval English has had a striking impact on morphological and syntactic research, directly or indirectly connected to these corpora, e.g. Kroch & Taylor (1997), Pintzuk et al. (2000), Kroch (2001), Trips (2002), Pintzuk & Taylor (2006), Haeberli & Ingham, (2007), Trips (2007), Stein & Trips (2008). Comparable publications on French diachronic syntax were either lacking an empirical base entirely or based on a small number of texts and could therefore, although of high quality, not discuss important typological questions on an empirical base of comparable size; we only cite Roberts (1993), Marchello-Nizia (1996), Vance (1997), Esperling (2001), Kaiser (2002), Becker (2005) for many others.
- 6 This project is part of the German SFB 538 « Mehrsprachigkeit », cf. http://www.uni-hamburg.de/fac (...)
9For French and other Romance languages, J. Meisel’s project « Mehrsprachigkeit als Ursache und Folge von Sprachwandel: historische Syntax romanischer Sprachen »6 has collected and analysed syntactic data for selected phenomena in the diachronic development of French and other Romance languages. Although the project has made an important contribution to this domain and part of the research was corpus-based, neither elaboration nor annotation of corpora were intended: a small number of texts was analysed syntactically, and the sentences were stored in a database in order to extract examples for specific syntactic properties (see Kaiser & Meisel 1991, Rinke 2003, Goldbach 2006).
- 7 GTRC project 412-2004-1002, University of Ottawa, speaker: France Martineau.
10In the Canadian GTRC project Les voies du français7, Kroch’s probabilistic parsing methodology is currently being applied to Old French texts. Martineau et al. (2007) describe the intended results for French. Syntactic structures produced in this project will be similar to the Middle English example (1) cited in section 1.1.
11Didier Bourigault has built Syntex, a partial dependency parser for Modern French. It analyses noun phrases and verb phrases. However, Bourigault et al. (2005) have shown that the choice of a syntactic representation does not imply the choice of a particular method of automatic analysis. They also show that annotated dependencies can easily be translated to constituent structures by conversion of dependency types into labelled edges, so that existing search tools like TigerSearch (Lezius 2002) can be used for corpus queries.
12A similarity between our approach of manual pre-annotation and Syntex is that our syntactic annotation also builds on POS tagging (Syntex uses POS annotated texts tagged by the French TreeTagger, cf. Stein & Schmid 1995). Therefore Syntex is theoretically an interesting tool – at least for partial syntactic pre-annotation. However, the technology is not at our disposal, since it was partly developed with private funding (société Synomia) and is not freely available for research purposes.
13To sum up, neither automatic parsing method meets our goal of producing publicly available and reusable resources and tools: A. Kroch’s probabilistic parser is trained and can be used only within the GTRC project, and Syntex is unavailable for research for commercial reasons.
14With respect to the grammar model used for annotation, our proposal is in line with most of the treebank approaches: we avoid decisions in certain cases of ambiguity and adopt as few theoretical assumptions as possible, with a preference for flat syntactical structures.
15Our grammar model is situated between the two extremes applied in existing treebanks: constituency-style annotation, as in the Old and Middle English Corpora, and dependency-style annotation, as in Prague Dependency Treebanks (Hajicová 2002): on the one hand, in our annotation, certain phrases will be either automatically chunked or manually annotated as constituents (e.g. NP, AP, PP). On the other hand, grammatical functions will not be defined by constituency, but annotated directly by named functions or by dependencies. A similar hybrid annotation is used for the Paris 7 French Treebank (Abeillé 2003) and explained in more detail in Abeillé & Barrier (2004).
16The short sentence in (2) will exemplify some of these principles:
(2) tresqu en la mer cunquist la tere altaigne
‘as far as to the sea (he) conquered the high land’
17In our analysis of (2),
-
a) noun phrases (here: the direct object la tere altaigne) are not annotated as a DP, but an NP, since functional categories are not heads in our annotation;
-
- 8 In the terminology used here, « A governs B » and « B depends on A » are equivalent and express th (...)
b) this object NP it is not embedded in a VP, but depends on (or is governed by8) cunquist;
-
c) the empty subject category is not annotated: pro drop sentences will be marked by attributes at sentence level (since defining the empty subject position by annotating an empty category would be impossible without adhering to disputable theoretical assumptions);
-
d) the adverbial phrase tresqu en la mer is not adjoined to (governed by) cunquist, but depends on the phrase node.
18Compared to modern language corpora, the Old French resources are of limited size. Still, in its current format, the NCA has about 300.000 <s>-boundaries corresponding to either sentences in prose texts or verses. The BFM is of comparable size and expected to grow in the future. The time for manual annotation per sentence depends on many parameters (e.g. depth of annotation and performance of the tool for manual annotation (TMA)) and is therefore hard to estimate. But it is quite obvious that partial automatic annotation is necessary to avoid the project continuing for decades. Annotation practice will show later to what extent such pre-chunking can be helpful for manual annotation. The tool NotaBene (Mazziotta 2006) will have the functionality to deal with pre-annotated structures, and in some cases, e.g. complex phrases, the pre-annotation might speed up the manual annotation process, whereas in others, the annotators might simply delete the chunks altogether.
- 9 If probabilistic parsers are ruled out for methodological reasons, XSLT (« Extensible Style Sheet (...)
- 10 TWIC is the query engine used for online access to version 2 of the Nouveau Corpus d’Amsterdam.
19In this contribution we are not going to discuss the different methods for syntactic pre-annotation9; rather, we will focus on concrete syntactic questions. That is why we chose to make first tests using our own tool, a Perl tool based on the TWIC (Tagged Words in Context) search tool originally designed for querying part-of-speech-tagged and XML formatted text corpora.10 TWIC’s « chunker » function can annotate chunks based on POS tag definitions. A first version handles frequent types of AP, NP, and PP phrases, as well as some verb and determiner clusters (i.e. unstructured sequences of verb and determiner types). It produces chunks like the ones shown in (6).
20The chunker function allows the user to assign syntactic markup codes to these queries by formulating rules like the one in (3), which states that markup for a noun phrase (NP) will be introduced for each sequence of determiner, adjective and noun. The options on the left side of the rule tell the chunker to include the name of the rule (option « r ») and the chunked string (« s ») in the markup and to apply the rule only if the constituents agree in number, gender and case (« a »).
(3) NP,det-a-n,rsa -> pos=DET pos=ADJ pos=NOM
21TWIC does not compile a grammar from these rules, but simply processes them one by one, in the order in which they are specified. A rule can therefore build on the syntactic markup inserted by previous rules, as in (4), where the attribute « xml » on the right side refers to a previously inserted category, but the markup resulting from subsequent rules is of course not available at processing time.
(4) NP,det-a-np -> pos=DET pos=ADJ xml=NP
22Other limitations of this method come from the rule input: the chunker relies on the part of speech annotation, which means that the rules have to account for errors and particuliarities like for example the underspecified category « PROCON » (originally « 600 » in A. Dees’s annotation scheme) which is used for pronouns, conjunctions and some adverbs.
23In the current experimental phase, we apply about thirty rules for complex verb forms (VCOMP), adjective phrases (AP), noun phrases (NP) and prepositional phrases (PP). Our goal is to speed up the manual tagging of the corpus by pre-annotating frequent chunks in a reliable way, favouring precision over recall since manual corrections would be very time-consuming. We therefore abandoned rules which produced too many errors, like « NP, det-n-a », because many adjectives in postposition do not actually depend on the preceding noun:
(5) sun braz teneit *[le chef enclin]
24The following table lists the rules ranked by their frequency of application in a sample corpus (La chanson de Roland: about 30.000 words, 2558 sentences).
Table 1: Rules ranked by frequency of application
25The 30 rules of our experimental « grammar » provide markup for more than 6.000 annotated phrases, which is a good start for the manual annotation of dependencies at a higher level, since the most common noun and prepositional phrases will have been annotated, and the annotator will be able to focus on the more complex tasks.
26Example (6) shows the output of the chunker for sentence (2) discussed in section 2.1. and its preceding context. The XML elements inserted by the chunker have been prefixed by « CHUNK- » to distinguish them from existing elements. Some default attributes (« rule » to indicate the chunking rule) have been eliminated here for better readability. The chunker has inserted the attributes which allow to build the dependency structure: « id » is a unique identifier for each phrase, and « gov » indicates the id of the governing phrase and must be filled by a phrase id (the default being the top node, i.e. the sentence id). The annotator corrects the « gov » values where necessary and thus builds the dependency structure for the whole sentence, either manually or, ideally, assisted by the annotation tool. The chunker also inserts the attribute « func », for the grammatical function of the phrase, and can suggest a default value based on the phrase type (« mod(ifier) » for AP and PP) or on the feature information (« suj(et) » for subject case, « obj(et) » for oblique case).
(6) pre-annotation of: set anz tuz pleins ad estét en espaigne / tresqu en la mer cunquist la tere altaigne
- 11 Our goal is not to give a complete description of the functionality of the TMA. Note that the tool (...)
27We have shown that automatic pre-annotation can complement manual annotation, and that the extent to which this technique will be applied depends on several factors (depth of annotation, precision of the chunker rules, skills of human annotators, etc.). A tool for manual annotation (TMA) should therefore be as flexible as possible. We therefore conclude our reflexions on manual annotation by stating some of the requirements which a TMA appropriate for a large-scale annotation project should meet11:
281 System
-
a) The TMA must be fast with regard to the user interaction (response to input, display of modifications, key and mouse buffering, etc.). The manual annotator should never have to wait for a process to finish.
-
b) As system-independent as possible (run on Linux, Mac and Windows machines and be integrated as well as possible into their surface with regard to standard keystrokes, etc.).
292 User Interface
-
a) Keyboard input is usually safer and faster than mouse input. The TMA should permit both types of input for the selection of text and the assignment of categories.
-
b) The possibility of defining shortcuts for keyboard and mouse actions is essential to speed up the manual annotation.
-
c) The user should be able to simplify the tree representation (fold and unfold substructures, e.g. embedded phrases) so that complex sentences will be easier to read.
-
d) The interface should be in English or provide methods for localisation.
303 Syntactic Annotation
-
a) The TMA should accept corpora with different degrees of previous syntactic annotation (manual or automatic) and thus allow increases in annotation depth.
-
b) Deletion and correction of previous annotation must be allowed for. Optionally, the TMA could keep track of deletions and corrections (by creating comments or writing them to a log file).
-
c) Dependencies between words and categories must be assignable as stated above. Default assignments (e.g. to the sentence node) should be possible.
-
d) The names of the inserted information (categories, functions, etc.) and their assignment to XML codes (elements, attributes, values) must be configurable.
-
e) Discontinuous elements (e.g. auxiliary and main verb separated by an NP) must be selectable.
-
f) The user should be able to insert comments at category level (within an XML element, e.g. to comment on a particular choice) and freely in the text (XML comment, e.g. with regard to a particular structure).
314 Input/Output
-
a) Support for standard character encoding (ISO Latin and Unicode).
-
b) Regular backups should be configurable.
-
c) The output of discontinuous elements should be configurable: either linear output (in word order, the annotation indicating the dependencies) or clustered (the annotation indicating the original position).
325 Options (useful, not mandatory):
-
a) Save preferences for different users or different corpora.
-
b) Provide configurable input/output formats or interfaces with scripts for input/output (e.g. export for query tools like TigerSearch).
-
c) Control input for free text (XML #CDATA) to avoid typing errors.
-
d) Definition of the annotation depth, so that the tool displays if the current annotation is complete or not.
33In contrast to other syntactic annotation projects, our goal is to annotate the Nouveau Corpus d’Amsterdam (NCA) manually, assisted by automatic pre-annotation (chunking). We have presented various projects for medieval corpora and discussed some examples for the interaction between pre-annotation, manual annotation, and the respective tools for these tasks.
34Further development of the project will depend on public funding and on cooperation with similar annotation initiatives for medieval French corpora, most of all the Base de Français Médiéval (BFM, Lyon). A joint German-French research proposal has therefore been submitted to ensure that NCA and BFM will be annotated according to the same principles, so that the future users will get maximum benefit from this research programme.