SRCMF tutorial
Contents
1. clitic headpos PROadv amp tokenarity clitic 1 Most clitics are captured by the first option in the query which matches personal pronouns with Obj Cmpl or Rfc function or the negative particle The second option following the OR operator simply matches all words tagged as PROadv en or i only Finally we used the tokenarity function to ensure that the node has only one child i e that the Obj Ng Cmpl Rfc dependant consists only of a single word 3 Exporting the results as a Key node in context KNIC concordance The SRCMF team has developed an alternative export format for TIGERSearch the KNIC concordance Rainsford and Heiden 2014 This presents the results of a syntactic query in a tabular format with one particular structure the key node occupying the central column in the table It is based on the Key Word in Context concordances generated by a number of corpus query programs e g TXM for the Base de Francais m di val Philologic TIGERSearch also provides a number of other export options which we will not cover here Further details may be found in the TIGERSearch manual ch 4 89 In this chapter we will first present the knicmaker tool 3 1 before describing the use of concordances in greater detail 3 2 3 1 Creating KNIC concordances KNIC concordances present the results of a syntactic query in a tabular format with one particular structure the key node2. 1b 20_1263221221 22 Li rois pense que par folie Si 1b 60 1263462997 12 reigne Ja nul verroient en la fa 1b 69_ 1263223807 37 Bien sai qu il me dorroit la mort 6 beroul_pb 3 1b 85 1263224824 55 chose qui n est voire Sire 7 beroul pb 4 1b 92 1263475749 96 Si a l on fait de mon seignor gt Sire J 1 i 73 3 S Figure 12 Configure TSV import gt The following settings are required for the import to function o Character set Unicode UTF 8 o Separated by Tab ONLY o Merge delimiters OFF o Text delimiter NONE empty box If your spreadsheet software will not load the file directly Cancel Help gt open the file first in a text editor making sure the encoding is set to unicode UTF 8 gt copy paste the table into the spreadsheet program If the concordance does not appear correctly check the following likely problems gt If accented characters do not appear correctly gt check the character set is UTF 8 gt If some rows do not seem to have the correct number of columns gt check that Text Delimiter is set to nothing the default is usually double quote which will cause an error where the text contains double quotes merge delimiters is OFF and TAB is the only separator selected If zeros appear rather than punctuation unlikely use the Fields section of the import window to set every column type to Text rather than Standard 3 2 Desi
3. occupying the central column in the table The concordance is generated by the knicmaker tool see 3 1 2 which post processes an XML file exported from TIGERSearch containing structures which match the user s query 7 Itis not perfect it will fail to match clitics with modifiers i e where a pronominal clitic is modfied by a dislocated argument or a discontinuous relative clause and it will also match the rare cases of stressed personal pronouns used with object or complement function 3 1 1 Specifying the key node the pivot identifier The knicmaker identifies the keynode in the query from the node identifier pivot In order for a TIGERSearch query to be compatible with the knicmaker it must contain the node identifier pivot For example the following query is correct in TIGERSearch it finds all subjects headed by the word rois but cannot be used to produce a concordance cat SjPer gt L word rois In order to make the query compatible with the knicmaker we must use the node identifier pivot to identify the key node In this case there are two possibilities e The whole of the subject is the key node pivot cat SjPer gt L word rois e Only the word rois is the key node cat SjPer gt L pivot word rois 3 1 2 Exporting the results from TIGERSearch Firstly you need to design and execute a TIGERSearch query containing the pivot node identifiers as
4. 1 2 2 Browsing the corpus TIGERGraph Viewer Corpus Query Help O yvainku tsbin G gt hm F 8 J wainku tsbin Explore corpus Textual mode Graphical mode J Documentation tie Summary view ti Detailed view 9 Edge labels edges amp secondary edges Nonterminal features cat annotationFile annotationUri coord dom headpos nodom type Terminal features word When opening a corpus for the first time it s often helpful simply to browse the first few annotated sentence to get an idea of how it has been annotated Once the Yvain corpus is open click on the Explore Corpus button in the toolbar see figure 2 This will open the Graph Viewer 1 2 2 1 Changing the graph viewer display options We can change the node features which are visible in the graph viewer by editing the Display options Display Options General settings Corpus dependent settings Maximum width of terminal node Display secondary edges Displayed non terminal feature Displayed terminal features Display virtual root node word C editionld _ editionNs ditionUri edge label _ editionuri Liq Number of context sentences _ form Hide feature value OK Cancel Reset Default Figure 3 Display options gt Click Options gt Display Options to open the Display Options dialogue box see fig 3 gt Select word and pos only in the list
5. VROOT _ gt nt Lr Lo Co cluster Ats E EJ me Lo fr sd m ModA Gpcoo GiPep lslustay K w m Reino GpCoo GiPen Reino Moda Em Que je fui plus petiz de lui Et ses chevax miaudres del Ji mien CONcoo PROper VERcjg ADVgen ADJqua PRE PROper CONcoo DETpos NOMcom ADjqua PRE DETdef PRE DETdef PROpos Figure 10 Complex coordination The representation of this complex coordination structure in TIGERSearch is given in figure 10 The two complex conjuncts are coordinated Each argument in both clauses has a dependency relation to the finite verb fui The complex conjuncts are represented by two nodes labelled GpCoo one linked to je and plus petiz de lui and the other to ses chevax and miaudres del mien by secondary edges labelled cluster The GpCoo nodes are children of a Coo node which as in the previous example is outside the dependency structure The coordinating conjunction et is a child of the second GpCoo node 1 3 3 Further reading This section has covered the basic structures of the SRCME as found in TIGERSearch For more information on the SRCMEF dependency grammar model see Pr vost and Stein REF For a linguistic justification of the coordination analysis see Mazziotta REF For more information on the syntactic tagset and the analysis of specific structures please see the SRCME annotation guide http www srcmf orgf iches index html in French 1 4 The TIGERSearch query lan
6. n word rois returns identical results to word rois with the difference that each matched node is also assigned the label n 1 4 3 Node relations There exist two core node relations in a TIGERSearch graph e precedence expressed by the dot operator e dominance expressed by the greater than operator gt 1 4 3 1 Precedence relations Precedence relations are used to query the order of terminal nodes i e words in the query For example if we wish to find all tokens of li rois the king we need to use the following query word 1i word rois Here two nodes are defined using two separate feature constraints The relation between these two n nodes is defined as direct precedence This will match all tokens of li directly preceding rois There exist a number of variants of the precedence operator of which the most useful is precedes at any distance For example if we wish to find tokens of negated verbs we could try to use the following query pos ADVneg pos VERcjg The query identifies all cases of negative ne pos tag ADVneg preceding a finite verb pos tag VERcjg Precedence at any distance ensure that cases in which clitic pronouns intervene are counted along with cases in which ne is directly before the verb However this query also returns a lot of noise since there is at present nothing to restrict the words to the same clause 1 4 3 2 Dominance r
7. which is the object of the sentence Note that words appearing in the pivot headed structure column are also found in the two context columns The original sentence may be read across the columns left context pivot right context Appendix Node features and edge labels in the SRCMF corpus Features of terminal nodes word word form as found in the base edition e g rois pos part of speech tag using the Cattex schema developed by the Base de Fran ais M di val See Guillot et al 2013 e form shows whether the word is in a prose text or a verse text and its position in the line Possible values o prose prose text o vers verse text word within the line o vers_debut verse text word at the beginning of the line o vers fin verse text word at the end of the line q shows whether the word occurs in direct discourse or not N b property is not implemented for the Yvain text o y word occurs in direct discourse o n word does not occur in direct discourse editionId xml id identifier of the word used in the source edition editionNs namespace of the source edition editionUri URI for the word in the SRCMF corpus editionNs editionId Features of non terminal nodes cat syntactic function of the node For the tagset please refer dom list of all syntactic functions which depend on the node ordered alphabetically and separated by _ For example a verb with a subject a
8. Displayed Non terminal features You may also wish to increase the number of preceding following sentences displayed below the graph this can be adjusted with the Number of context sentences option 2 1 2 2 A sample sentence The first sentence in Yvain is rather long so we ll start with the second gt Click the Next gt button in the navigation panel at the bottom of the Graph Viewer to advance to the second sentence D R odA CReIN RelN Li rojs fu a Carduel en Gales DETdef NOMcam VERcjg PRE WNOMproa PRE WNOMpro Figure 4 Yvain second sentence You should see the sentence tree in figure 4 for the following Old French sentence 1 Li rois fu a Carduel en Gales The king was at Carduel in Wales The words appear along the bottom of the graph with part of speech tags underneath e g rois is tagged as NOMcom common noun Syntactic functions are given in the white ovals within the graph itself e g the group Li rois is tagged as SjPer personal subject 1 2 2 3 Nodes and edges To use the correct terminology we can describe each TIGERSearch graph in terms of NODES and EDGES There are two types of nodes in this sample graph TERMINAL NODES these always appear along the bottom line of the graph and contain the words in the text Terminal nodes as the name suggests have no child nodes e NON TERMINAL NODES these are represented by the white ovals Non terminal nodes must h
9. described above Once you have run the query gt Select Query gt Export Matches from the toolbar above gt In the export window which pops up o set Export Format to XML o set Export to file to the output file of your choice o set Export includes to Whole corpus gt Click Submit This will create a Tiger XML file compatible with the knicmaker Export format XML M XML Export to file don t refer to schema horne tomr beroul_sj_vo xml Search Export header T v Export graph structure Export includes All matching corpus graphs vi Export match info Current matching corpus graph From matching corpus graph 1 to 1442 Select matching corpus graphs XML piped through XSLT All non matching corpus graphs Whole corpus Submit N Cancel Done Figure 11 TigerSearch export settings 3 1 2 Installing the knicmaker The knicmaker is tested on Windows and Linux systems and requires a Java runtime environment to be installed on your computer e Download the knicmaker from http sourceforge net projects knicconcordances files software e Unzip the ZIP archive to a convenient location on your computer 3 1 3 Configuring the knicmaker The knicmaker has no user interface It must be configured using the text file knicmaker properties before it is run It outputs any error success messages to the file knicmak
10. table 1 contains square brackets in the pivot x Mes messire Yvains pas ne fuit Qui de lui siudre ne se faint But Lord Yvain does not flee he does not hesitate to follow him lit who does not hesitate J These are used in all concordances when the structure in the pivot column is discontinuous The annotated subject in this sentence is messire Yvains qui de lui siudre ne se faint In between the first part of the subject messire Yvains and the relative which modifies the subject qui de lui siudre ne se faint who does not hesitate to follow him is the negated main verb of the sentence pas ne fuit does not flee The words pas ne fuit which separate the two parts of the subject are included in the pivot column surrounded by square brackets This means that e the pivot column contains all parts of discontinuous pivots e reading the concordance from left to right will always give the original sentence Asterisks and hashes are dummy words in the SRCMF and may be ignored see 1 3 2 Slashes indicate divisions between sentences in the syntactic annotation and are only present in the context outside sentence columns of the concordance 3 2 2 Pivot and block concordance The pivot and block concordance is designed to highlight the position of certain structures called blocks e g the subject with respect to a pivot e g the verb The resulting tables are complex
11. with a large number of columns and are intended as the basis for more detailed analysis 3 2 2 1 Basic structure of the pivot and block concordance The pivot and block concordance has the following basic structure sentence ID e left context outside sentence e left context within sentence e pre pivot blocks e pivot e post pivot blocks e right context within sentence e right context outside sentence As with the basic concordance TIGERSearch queries must define a pivot node Additionally users may define any number of other nodes in the query as blocks using a node identifier of the form blockn i e block1 block2 block3 For example the following query will generate a pivot and block concordance to show the position of the subject block1 with respect to the verb in main clauses pivot snt cat Snt gt D block1 cat Sj amp snt gt L pivot The key section of the resulting concordance will take the following form Left context Block Pivot Block Right context Li rois fu a Carduel en Gales Apr s mangier Cil chevalier s atropelerent La ou dames parmi ces sales les apelerent Ou dameiseles ou puceles Or est Amors tornee a fable Por ce que cil qui rien n en santent Dient qu il aiment Table 2 Simplified pivot and block concordance block is subject pivot is finite verb Where the subject is pre verbal it appears in the block column to the le
12. To express feature values observe that exact values strings are given between double quote marks while regular expressions are given between forward slashes When combining two feature constraints two operators are available e amp AND both feature value pairs match e g word roi amp pos NOMcom e OR at least one of the feature value pairs match e g cat SjPer cat SjImp will find tokens of both personal and impersonal subjects It is also possible to negate a feature value pair using the operator does not match For example to match instances of the word es other than as a common grammatical form definite 2 This tutorial will not cover regular expression syntax as it is neither specific to TIGERSearch nor to the SRCMF corpus This does not imply that they are not useful A good online tutorial may be found at http www regular expressions info or see the TIGERSearch manual sec 3 5 determiner and personal pronoun word les amp pos DETdef amp pos PROper 1 4 2 2 Node identifiers A node identifier is a user ddined name attributed to a particular node in the query It is introduced by the hash symbol n n word rois The node identifier alone i e n has no feature constraints If entered as a query it will match every node in every graph in the corpus To combine a node identifier with a feature constraint the colon is used as a separator The query
13. Using the Syntactic Reference Corpus of Medieval French in TIGERSearch SRCMF T M Rainsford September 2015 Introduction This tutorial provides a step by step guide to working with the Syntactic Reference Corpus of Medieval French SRCMF using the TIGERSearch query engine preferred by the project It is intended for those who are familiar with syntactic analysis and the concept of treebank corpora The document is divided into four main chapters 1 Getting started Covers installation of the software 1 1 the basic structure of the syntactic annotation in the SRCMEF 1 2 and the basics of the TIGERSearch query language 1 3 2 Writing successful queries Tips for writing effective queries to identify structures in the the SRCMF 3 Exporting the results as a KNIC concordance How to export the results of your searches in a tabular format This is a user guide intended to help new users start using the corpus and is not a comprehensive reference manual Those looking for a reference manual are referred to the following resources e The TIGERSearch manual K nig Lezius and Voormann 2003 online at http www ims uni stuttgart de forschung ressourcen werkzeuge TIGERSearch manual html e The SRCMF annotation manual in French online at http www srcmf orgfiches index html The Cattex annotation manual part of speech tagset in French Guillot Pr vost and Lavrentiev 2013 online at http bfm ens lyon fr spip
14. aightforward queries Use the precedes at any distance operator in combination with a definition of the dependencies e Post verbal subjects verbword pos VERcjg subjectword amp subject cat Sj gt L subjectword amp verb gt L verbword amp verb gt D subject e Pre nominal adjectives adjword pos ADJ nounword pos NOM amp noun gt L nounword amp adj cat ModA gt L adjword amp noun gt D adj 2 1 4 How do find all xs that don t dominate a y the dom feature E g verbs without subjects nouns without determiners etc It is impossible in a TIGERSearch query to require the non existence of a particular kind of node see 1 4 3 3 However the SRCMF includes a feature to make the most common requests of this type possible If y can be dined solely by its syntactic function i e the cat feature use the dom feature with a regular expression The dom feature is present on each structure node and lists the functions of all its dependants and relators in alphabetical order separated by underscores For example if a verb has a subject object and two adjuncts the dom feature will have the value Circ_Circ_Obj_SjPer e Clauses without a subject headpos VERcjg amp dom Sj e Unmodified nouns headpos NOM amp dom ModA Single part clausal negation i e ne only no reinforcing particle such a
15. al form For example we may wish to create a pivot and block concordance showing the position of the subject and non pronominal object relative to the main clause verb In this case it is vital that the exported concordance labels each of the blocks with its syntactic function at Snt gt D block1 cat Sj gt D block2 cat 0bj amp headpos PROper gt L pivot amp sni amp sni snt c Block Block Pivot Block Li un SjPer recontoient noveles Obj que Obj j voldroies tu SjPer divers chanz Obj chantoit chascuns SjPer Li uns SjPer lautre a l Obj assaut espee Table 5 Simplified pivot and block concordance blocks are subject and object pivot is finite verb By inserting the cat feature of each block into the concordance a very clear overview of the word order in each sentence is obtained The resulting table can also be sorted by the columns containing the cat feature in order to group sentences with a similar word order together 3 2 4 Single word pivot concordance The final type of concordance can only take a word node as its pivot and contains a minimum of 7 columns based on the following structure e sentence ID e left context outside sentence e left context within sentence e pivot single word e structure of which pivot is head e right context within sentence e right context outside sentence The single word pivot conco
16. ave at least one child node Each node has a number of FEATURES For instance terminal nodes have the features word word form and pos part of speech which we are currently viewing and a number of others which are not currently being shown in the graph However in the SRCMF terminal and non terminal nodes do not have the same features for example syntactic function tags such as SjPer personal subject or Snt sentence are stored as the cat feature of non terminal nodes gt Hover your mouse pointer over any node in the graph to see a full list of features The nodes are connected by EDGEs These are represented by a line connecting the two nodes Edges in the SRCMEF are labelled this appears as a single letter L D in a grey box on the line 1 2 2 4 Secondary edges La dames les apelerent Ou dameiseles ou puceles 4DVgen PROrel PROrel NOMcom PROper VERcig CONcoo NOMcom CONcoo NOMcom Figure 5 Yvain third sentence secondary edges Let s briefly advance to the third sentence in the Yvain text by clicking the Next gt button You ll notice that this graph contains some green lines in the second half of the sentence see fig 5 which reads as follows 2 La ou dames les apelerent Ou dameiseles ou puceles Where ladies damsels or maidens called them The green lines are known as SECONDARY EDGES and link nodes together in a non hierarchical manner There are two main uses of secondar
17. d word node may need to be defined e Pronominal objects cat Obj amp headpos PROper e Yvain heads the subject 5 cat Sj gt L word Yvain sz 2 1 2 How do I find xs that dominate a y and a z2 E g verbs with subject and object nouns with a relative clause infinitives introduced by a preposition These are straightforward queries Use feature constraints to define each node and dominance relations to specify the relations between them e Finite verbs with subject and object verb headpos VERcjg gt D cat Sj amp verb gt D cat Obj e Nouns governing a relative clause headpos NOM gt D cat ModA amp headpos VERcjg Infinitival clauses introduced by a preposition recall from 2 2 2 that the edge linking a governing node to a preposition is labelled R headpos VERinf gt R headpos PRE In some cases e g when interested in a particular lexical form we may need to query a property of the word node too For example if we are searching for all cases in which the word Yvain occurs as the object of a finite verb 5 Note that when Yvain is preceded by a title e g messire Yvain it is the title that heads the subject object cat Obj gt L objectword word Yvain sz amp verb headpos VERcjg gt D object 2 1 3 How do find all xs that precede a y E g verb subject inversion pre nominal adjectives These are relatively str
18. e written in the main right hand panel of the TIGERSearch main window Writing queries for the SRCMEF will be covered in more detail in section 3 gt Copy the following text in the query window word rois Corpus Query Help 0 yvainku tsbin CHA i 5 f3 bd 8 2 WO wainku tsbin 2 a Documentation tie Summary view ti Detailed view 9 Edge labels edges amp secondary edges Nonterminal features cat annotationFile annotationUri coord dom headpos gt nodom type Terminal features word Textual mode Graphical mode word rois ay Figure 6 Launch a search You ll see colour coding appear in the search window the feature name word is red and the value is blue gt Click the gt Search button in the bottom right of the screen or the play arrow in the toolbar see fig 6 1 2 3 2 Results graphs and subgraphs The Graph Viewer will now pop up but instead of showing the whole corpus it will only show those graphs which match the query in this case all those containing the word rois king s The matching node is highlighted in red Graphs 54 Previous i Next gt Subgraph 1 2 Subgraphs 56 First y 54 Last a Figure 7 Graphs and subgraphs The navigation panel below the graph contains some useful quantitative information about the query results see figure 7 in particular the number o
19. elations querying edges A dominance relation is used to query relations between nodes represented by EDGES in the TIGERSearch graph A simple application of this operator in the SRCMEF corpus is to associate a structure node with its word node cf 2 2 2 Suppose we wish to find every instance of the word rois as a subject The syntactic function is marked on the structure node but the lexical information 3 TIGERSearch can also calculate precedence between non terminal nodes based on the relative position of the leftmost terminal node dominated We do not recommend using this implicit shortcut in the SRCMF corpus since it is not sensitive to the SRCMF node pairs is marked on the word node cat SjPer gt L word rois Our feature constraints identify two nodes a structure node with subject function and a word node rois The dominance relation gt L specifies that an edge labelled L exists linking both nodes with the structure node dominating the word node i e we require these two nodes to form a node pair Try querying cat SjPer amp word rois to see why we need to specify the relationship between the two nodes Secondary edges can also be queried using the operator gt For example to find all tokens of words with a double function see 2 2 2 1 we may search for secondary edges labelled dupl gt dup1l Here we use an empty feature constraint to define the two
20. er log not to the screen or terminal Before running the knicmaker you should edit the file knicmaker properties gt InFile enter the full path and filename of the XML file you have just exported from TIGERSearch e g home tomr beroul_sj_vb xml gt OutFile enter the full path and filename of the output file you wish to generate gt ConcordanceType select the type of concordance you would like to generate using the keywords o simple for the basic concordance o word pivot for the single word pivot concordance blocks for the pivot and block concordance gt TFeatures enter a comma separated list of all the word node features you would like to have in the final concordance apart from word which is always shown Leave blank if 8 Mac users have reported problems with the jar We suggest that Mac users either i run the compiled jar in a Linux virtual machine or ii download and run the source code directly using the Groovy interpreter this latter option is reported to work no features required E g o TFeatures pos form show the pos and the form features o TFeatures i e blank show no features gt NTFeatures enter a comma separated list of all the structure node features that should be shown in the concordance Leave blank if no features required E g o NT Features cat headpos show the cat and headpos features o NTFeatures cat show the cat
21. f matching GRAPHS and SUBGRAPHS It is very important to understand what these numbers mean e Graphs 54 There are 54 graphs in the corpus which contain a match for this query i e 54 sentences contain the word rois e Subgraphs 56 There are 56 structures which match this query in the corpus i e there are 56 tokens of the word rois The difference between the two figures tells us that one or two sentences contain more than a single token of the word rois The first matching sentence is one such case Note the green Subgraph box at the right of the navigation panel Subgraph 1 2 There are two structures in this graph matching the query the first is highlighted By clicking on the green forward arrow the red highlight in the graph shifts to the second of the two matching subgraphs You can use the lt Previous and Next gt buttons in the navigation panel to view the other 53 matching graphs one by one 1 2 3 Further reading We have now covered the basics of the TIGERSearch interface More details can be found in the TIGERSearch manual available on the project homepage http www ims uni stuttgart de forschung ressourcen werkzeuge tigersearch html Some users may be interested in exploring TIGERSearch s graphical interface for writing queries full details may be found in the TIGERSearch manual 1 3 The SRCMF in TIGERSearch The SRCMF grammar model will not be pre
22. feature gt ContextSize Size of the context to show in the concordance in number of words Ignores sentence boundaries Remember to save the changes to the knicmaker properties file 3 1 4 Launching the knicmaker Once the knicmaker properties file has been edited launch the knicmaker gt on Windows by double clicking knicmaker exe gt on Linux o Open a terminal o Change the working directory to where you installed the knicmaker using the cd command e g cd home tmr knicmaker o Launch knicmaker jar using the following command java jar knicmaker jar Check the knicmaker log file for any errors during execution e g mistyped file names 3 1 5 Opening the output file The knicmaker exports the concordance as a unicode UTF 8 tab separated text file This simple format is easily loaded in LibreOffice OpenOffice gt open the file gt the spreadsheet software should open a configuration window such as the following from LibreOffice Calc v 3 4 4 Text Import concordance csv Import Character set Unicode UTF 8 Language Default English UK From row 1 z Separator options Fixed width Separated by Tab Comma Other l Semicolon Space Merge delimiters Text delimiter Other options Quoted field as text Detect special numbers Fields Column type Standard Standard b 5_1263220951 97 com el l a devanci Tb 18 1263221009 34 nule foiz mais Je vos di bien
23. ft of the pivot Where it is post verbal it appears in the block column to the right of the pivot Pivot and block concordances may contain braces in the block column e g cil chevalier s These mark words which intervene between the block and the following column in this case the pivot The reflexive pronoun s is here not part of the subject but intervenes between the subject and the verb As with the basic concordance text word order is maintained if the concordance is read from right to left 3 2 2 2 Why so many columns The pivot and block concordance shows only one result per pivot Continuing to work with the same example if a single verb pivot has multiple subjects which is quite possible in cases of coordination each subject occupies a separate column Pivot Block Block Block Block Fu Didonez et Sagremors Et Kex et messire Gauvains Table 3 Coordinated subjects as blocks However due to the way the number of columns is calculated some may be empty throughout the concordance These may be deleted in the spreadsheet software if you wish 9 Those used to TIGERSearch will note that the concordance combines multiple matches within a single sentence where the pivot variable is the same 3 2 2 3 Discontinuous single structures vs two structures matching block Discontinuous structures are always in a single column Left context Block Pivot Block Righ
24. gn of KNIC concordances Three variants of the KNIC concordance are currently implemented in the knicmaker e basic concordance e pivot and block concordance e single word pivot concordance 3 2 1 Basic concordance The basic concordance has a minimum of six columns e sentence ID e left context outside sentence e left context within sentence e pivot e right context within sentence e right context outside sentence The pivot can be any node in the graph either a word node or a structure node For example we may wish to create a concordance of all the cases in which the word Yvain occurs within the main clause subject cat Snt amp pivot gt gt D pivot cat SjPer word Yvain sz The pivot node denotes the subject of the clause Below is a selection of the results from the concordance ID LeftCxInsideSnt pivot RightCxInsideSnt YvainKu_pb 79_ b 56 Bea 20 20 1 YvainKu_01 _1309961775 91 Et sii fu messire Yvains Et avoec ax Qualogrenanz Uns chevaliers mout avenanz Qui YvainKu_pb 90_ b 27 74Bea 20 20 1 YvainKu_10 _1321888863 94 Yvains respondre ne li puet YvainKu_pb 91_ b 32 66Bea 20 20 1 YvainKu_12 _1322227968 63 Mes messire Yvains pas ne fuit Qui de lui siudre ne se faint Table 1 Basic concordance subject contains Yvain Note that the pivot may be one or more words The third example in
25. guage This section provides a brief introduction to the syntax of the TIGERSearch query language for the SRCMF 1 4 1 Basics A TIGERSearch query defines a particular subgraph that we wish to search for The subgraph may contain a single or multiple nodes and makes use of a range of wildcards The two most important aspects of a TIGERSearch query are therefore 1 Node definitions 2 Relations between nodes 1 4 2 Node definitions Node definitions consist of two parts e Anode identifier e A feature constraint Either part may be omitted in a query 1 4 2 1 Feature constraints Feature constraints consist of one or more feature value pairs enclosed in square brackets for example word rois cat SjPer word roi amp pos NOMcom As we saw earlier the query word rois will find all instances of the word rois Put technically it matches all nodes whose word feature is rois This matches the word node of a node pair as the word feature is on the terminal node The query cat SjPer matches all nodes whose cat feature is SjPer i e find all subjects This matches the structure node of the node pair since this has the cat feature Multiple properties can be combined in a node definition The query word roi amp pos NOMcom matches all nodes whose word feature matches the regular expression roi and whose pos feature is NOMcom i e all nouns beginning roi
26. inite verb dominates the structure node for the negation negword pos ADVneg verbword pos VERcjg amp verbstructure gt L verbword amp negstructure gt L negword amp verbstructure gt D negstructure Three node relations are added to the query two which identify the structure nodes verbstructure and negstructure from their word nodes and one which defines the relationship between these two structure nodes The four node relations are conjoined by the ampersand amp Note the importance of node identifiers which enable us to refer to the same node multiple times within a query 2 3 4 Further reading This section has covered the basic syntax of TIGERSearch queries For further information refer to the TIGERSearch manual ch 3 and in particular the quick reference ch 3 12 4 Note also that this is not the simplest way to search for negated verbs 2 Writing successful queries This chapter is intended for users with a basic knowledge of e the structure of the SRCMEF corpus in TIGERSearch see 1 3 e TIGERSearch query syntax see 1 4 It aims to explain and to provide models for common types of query 2 1 General structures 2 1 1 How do find all words x with syntactic function y E g pronominal objects the word Yvain as subject These are very straightforward queries Use feature constraints to identify the node In some cases both a structure node an
27. n object and no other arguments will have the dom value Obj_SjPer headpos Part of speech tag of the word node see pos for finite verbs type Simplified headpos gives the part of speech of the head node as o VFin finite verb o VInf verb infinitive o Par verb participle o nV not a verb coord value is y if the node is a conjunct or part of a complex conjunct annotationUri URI of the structural annotation in the SRCMEF corpus annotationFile used in corpus development nodom redundant Edge labels L lexical links two members of a node pair D dependency marks a dependency relation between two structure nodes R relator links a governing node to prepositions and conjunctions P part links the coordination node Coo to nodes representing complex conjuncts GpCoo Secondary edge labels e cluster links the node representing a complex conjunct GpCoo to the structures which form the conjunct e coord links the coordination node Coo to its conjunctions simple coordination e dupl links a structure node to its true word node in cases of word with two functions References K nig Esther Wolfgang Lezius and Helger Voormann 2003 TIGERSearch User s Manual Stuttgart IMS University of Stuttgart http www ims uni stuttgart de forschung ressourcen werkzeuge TIGERSearch manual html Rainsford T M and Serge Heiden 2014 Key Node in Context KNIC Concorda
28. n of the word relative to its parent The two nodes are linked by an edge labelled L for lexical form The dependency relations forming the hierarchical syntactic structure are marked between the non terminal nodes of the node pair and are labelled D Prepositions and conjunctions are analysed as relators and the relation between these and their governing node is labelled R As a general rule the non terminal node of a node pair contains structural information e g dependencies and syntactic function while the terminal node contains only word level features For this reason we will henceforth adopt the term STRUCTURE NODE to denote the non terminal node of a node pair and the term WORD NODE to denote the terminal node 1 Indeed these features are incompatible with the standard CoNLL format used by dependency parsers me doie estre a mal escrite PROper VERcjg VERinf PRE NOMcom VERppe Figure 9 SRCMF corpus in TIGERSearch Figure 9 shows the TIGERSearch graph for sentence 3 The node pairs representing the first three words are highlighted in red 1 3 2 1 Representing nodes with two functions The word qui in 3 has two separate functions The representation of this structure in the TIGERSearch graph is highlighted in blue e The word qui is split into two node pairs The first node pair links the word node qui with the first of its two functions ordered alphabetically e The second node
29. n terms of syntactic functions the cat property negation is annotated using two separate labels e Ng when it is dependant on a verbal head e ModA when it is dependant on a non verbal head However this is rare in the SRCMF as negation is attached to the verb wherever possible In the part of speech tagging negative ne nen non are tagged as ADVneg Note that in cases of pronominal enclisis nel nes the tag will be ADVneg PROper Consequently the following query will find all negated xs x should be replaced with a feature constraint x gt D headpos ADVneg 2 2 2 How do identify subordinate clauses The SRCMF has no specific tag for subordinate clauses however it does for main clauses Snt The easiest way to identify subordinate clauses is therefore to identify all finite verbs that do not head a sentence cat Snt amp headpos VERcjg 2 2 3 How do identify auxiliary verbs compound verb forms The past participle of a compound verb is tagged as AuxA The two parts of a compound verb forms can be identified with the following query auxiliary headpos VER gt D participle cat AuxA amp headpos VERppe J Passives can be identified with a similar query replacing AuxA with AuxP 2 2 4 How do identify clitics The best query to match clitics is as follows verb headpos VERcjg gt D clitic cat Ng Obj Cmpl1 Rfc amp headpos PROper ADVneg
30. nces Improving Usability of an Old French Treebank SHS Web of Conferences 8 2707 2718 lt DOI 10 105 1 shsconf 20140801250 gt
31. nodes since we simply wish to find all instances of the secondary edge labelled dupl without any restrictions on the nodes that it links together There exist a number of variants of the dominance operator but few are essential for beginners See the TIGERSearch manual ch 3 1 4 3 3 Negated operators a word of caution A glance at the TIGERSearch manual reveals that it is possible to negate both dominance and precedence operators e g e 1 does not precede e gt does not dominate However these are confusing since they only negate the relationship not the existence of the nodes For example the following query does not find every null subject main clause try it cat Snt gt cat SjPer cat SjImp Instead it will return graphs containing a subject somewhere other than in the main clause There may well be a subject in the main clause too These operators are occasionally useful in very specific circumstances but should be avoided by beginners 1 4 3 4 Putting it all together Most queries are made up of a number of node definitions and node relations joined together with the amp AND or the OR operator Let us return for a moment to our unsatisfactory query to find negated verbs pos ADVneg pos VERcjg We need to add a requirement which states that these two nodes must be within the same clause In graphical terms we need to state that the structure node for the f
32. ow do find structures with only one x E g Verbs with only one complement nouns with only one preceding adjective These queries are often impossible to formulate in TIGERSearch However there are a couple of techniques which may arrive at something close to the desired result Firstly you may be able to use the dom property with a regular expression see 2 1 4 above e all verbs with only one complement object included dom Obj Cmpl amp dom Cmpl_ Cmpl amp dom Cmpl Obj Effectively we identify all dom strings containing Obj or Cmpl but then exclude those containing Cmpl twice or Obj and Cmpl However the query will also exclude verbs with two coordinated complements Secondly you can write a query to find all structures with at least one x and then generate a pivot and block concordance see below 2 1 3 This can then be sorted using spreadsheet software to separate occurrences of one x two xs three xs and so on e nouns with one preceding adjective postprocessing necessary using pivot and block concordance to identify which matches contain only one noun headpos NOM gt L pivot amp adj headpos ADJ gt L blockl amp noun gt D adj amp olockl pivot 2 2 Specific structures This section provides simple queries to identify common structures which are not directly annotated in the corpus 2 2 1 How do find all negated xs I
33. pair links the second of the two functions with a dummy word e A secondary edge labelled dupl links the structure node of the second node pair to its true word node the terminal node qui 1 3 2 2 Representing groups in coordination Let us return briefly to example 2 and figure 5 Here the subject is made up of three separate conjuncts dames ou dameiseles and ou puceles In the TIGERSearch representation e Each conjunct has a separate dependency SjPer on the finite verb apelerent e The three conjuncts are linked to a non terminal node labelled Coo by three secondary edges labelled coord e The node Coo is entirely outside the dependency hierarchy it has no parent node other than the automatically generated VROOT element and is paired with a dummy word node 1 3 2 3 Complex coordination Some cases of coordination involve more than a single conjunct Consider example 4 4 Que je fui plus petiz de lui et ses chevax miaudres del mien For I was smaller than he and his horse better than mine The SRCMF analysis of gapping constructions such as these is as follows The arguments of the gapped clause 1 e ses chevax his horse and miaudres del mien better than mine depend on the expressed verb of the first clause The arguments of the first clause i e je T and plus petiz de lui smaller than he and the arguments of the second clause constitute two COMPLEX CONJUNCTS lt
34. php article323 1 Getting started 1 1 Installation 1 1 1 Installing TIGERSearch You will need to install the TIGERSearch corpus query engine on your computer ZIP archives containing both Windows and Mac versions are available on www srcmf org in the Tools section of the website recommended e Alternatively you can download the program from the TIGERSearch homepage www ims uni stuttgart de forschung ressourcen werkzeuge tigersearch html This version does not include a Java Runtime Environment which you will have to install separately 1 1 2 Installing the SRCMF The SRCMEF corpus is available in the Access section of www srcmf org You have the choice of downloading the texts as a single corpus or individually Each text is available in three formats e aTIGERSearch binary tsbin recommended e aTIGER XML file tsxml this must be processed using the TIGERRegistry program before it can be used with TIGERSearch e a RDF file rdf not compatible with TIGERSearch Note that the full corpus and some of the individual texts are subject to licensing restrictions full details are provided on the website For the purpose of this tutorial we will be working with the Yvain text which is available without any licensing restrictions To download and install Yvain gt click on the tsbin link next to Yvain de Chr tien de Troyes in the list of texts and save the zip archive to your computer gt ex
35. rdance is designed to give as much information as possible about a single word It is designed to be used with multiple columns showing node features inserted For example a single word pivot concordance could be created around the word Yvain pivot word Y vain sz Below is a selection of the results from the concordance some columns are omitted Left context in Pivot Pivot Pivot Pivot Right context in sentence form headed headed sentence structure structure cat Et si i fu messire Yvains vers fin Yvains ModA Et avoec ax Qualogrenanz Uns chevaliers mout avenanz Qui Et toz les autres fors Yvain vers_ fin fors Yvain Le mangongie r le guileor Le desleal le tricheor ModA Le mangongier le guileor Le desleal le tricheor Dame je ai Yvain vers Yvain trov Le chevalier mialz esprove Del monde et le mialz antechi Obj trov Le chevalier mialz esprove Del monde Table 6 Single word pivot concordance pivot is Yvain sz The concordance is similar to the basic concordance but with the addition of a pivot headed structure column which shows the pivot with its dependents In the third example for instance the word Yvain heads the structure Yvain Le chevalier mialz esprov del monde et le mialz antechi Yvain the most distinguished knight in the world and the most noble
36. s far less noise even though it is still not perfect structures with a conjunction and a preposition before the determiner will slip through the net 2 1 5 How do find all non coordinated xs the coord feature E g conjoined objects gapping constructions non conjoined object If you only wish to specify whether or not x forms part of a coordinate structure use the coord feature coord has the value y when the node is part of a coordinate structure Find all objects in a coordinate structure cat Obj amp coord y Find all objects not in a coordinate structure cat Obj amp coord y If you wish to place more specific constraints on coordination you must dine the structure using node relations See sections 1 3 2 2 1 3 2 3 for an overview of how coordination is represented in the SRCMF Find all gapping constructions two complex conjuncts involving a subject and an object Note here the use of the siblings with precedence operator this ensures that i conjl and conj2 are separate nodes and ii conjl always refers to the first conjunct while conj2 always refers to the second conjunct coo cat Coo gt P conjl1 cat GpCoo amp conjl conj2 cat GpCoo amp conjl gt cluster cat Sj amp conj1 gt cluster cat 0bj amp conj2 gt cluster cat Sj amp conj2 gt cluster cat O0bj 2 1 6 H
37. s pas or mie headpos VERcjg amp dom NgPrt gt D cat Ng If y cannot be defined solely by its syntactic function the query is impossible For example while finding unmodfied nouns is straightforward ModA is a syntactic function finding nouns not modified by a determiner is impossible with a single query determiner is not a 6 The directly precedes operator is less useful especially since the SRCMF contains dummy words syntactic function Users are advised to think laterally and to formulate these corpus searches in a slightly different way For example while we cannot find all nouns without a determiner we can find all nominal structures which begin with an element that is not a determiner Note the operator gt I left corner dominance indicates the leftmost dominated terminal node headpos NOM gt 1 pos PRE DET This query produces a lot of noise since determiners are not always the first word in a nominal structure In particular prepositions and conjunctions will precede any determiner We should modify our query to check that if a nominal structure begins with a preposition or a conjunction the second word is not a determiner headpos NOM gt 1 pos PRE DET amp pos PRE amp pos CON n headpos NOM gt l w pos PRE pos CON amp w w2 pos PRE DET amp n gt w2 This query produce
38. sented in full in this tutorial Readers are instead referred both to Pr vost and Stein REF and to the SRCMF annotation guide http www srcmf orgfiches index html We assume moreover that readers are familiar with the basics of dependency grammar The purpose of this section is twofold firstly to highlight some of the less canonical aspects of the SRCMF grammar model which may be unfamiliar even to readers who have worked with dependency corpora before and secondly to document how the SRCMF grammar model is represented in TIGERSearch 1 3 1 Dependency in the SRCMF treebank 1 3 1 1 A dependency treebank The SRCMF is a dependency treebank the main clause finite verb is the head of the sentence exception non sentences with the exception of the head of the sentence each individual word depends on one other word in the sentence cuit ie je ne avoir chose dite i doie qui me estre mal escrite Figure 8 SRCMF dependency tree Figure 8 shows the following sentence represented using a dependency graph which closely reflects the grammar model of the SRCMF 3 Je ne cuit avoir chose dite qui me doie estre a mal escrite T do not think I have said anything that should be held against me 1 3 1 2 Head dependent directionality Since in a number of cases there is no clear consensus among linguists as to the correct directionality of head dependent relations it is worth summarizing
39. some of the key points of the linguistic analysis adopted in the SRCMF All clause level structures arguments adjuncts and non arguments such as parentheticals depend on the finite verb Subordinating conjunctions depend on the finite verb in the subordinate clause they are not heads cf qui me doie above Prepositions depend on the non finite verbal or non verbal head of the following structure they are not heads cf a mal above In compound verb tenses e g the perfect tense the passive the finite auxiliary verb is analysed as the head of the clause on which all other arguments and adjuncts depend cf avoir dite above The past participle is a childless dependant of the finite verb In constructions involving a modal verb pouvoir devoir savoir the finite modal verb is analysed as the head of the clause on which all other arguments and adjuncts depend The infinitive is a childless dependant of the finite verb cf the flat structure of doie estre escrite above e Determiners are treated like adjectives they depend on the nominal element which they determine 1 3 1 3 Departure from conventional dependency structure Two structural features of the SRCMEF call for particular comment as they do not occur in traditional dependency corpora 1 A single node may have more than one parent or indeed have more than one dependency relation with the same parent In the example above the word qui who has
40. t context Et cil come Vint plus tost c uns mautalentis alerions Fiers Vint plus tost par sanblant c uns come lions alerions Fiers par sanblant come lions Table 4 Discontinuous block This representation shows e a single discontinuous subject Cil come mautalentis Fiers par sanblant come lions in the block column e within the block column words that intervene between the two parts of the subject appear between square brackets even if these words appear later in other columns e the position of the block relative to the pivot is determined by its first word This case is distinguished from that in which there is more than one structure matching the definition of the block node in the sentence In this second case more than one block column is filled Compare the following result which shows the gapping construction discussed in 2 2 2 3 with table 4 Left context Block Pivot Block Right context Que je fui plus petiz miaudres del de lui Et ses mien chevax Table 5 Coordinated subjects two separate blocks This representation shows two separate subjects one preverbal je and one postverbal ses chevax The words plus petiz de lui intervene between the verb and the postverbal subject and are marked using braces 3 2 3 Adding features to the concordance It is also possible to create concordances which show node features other than lexic
41. tract the folder in the archive to the folder CorporaDir in your TIGERSearch directory That s it The corpus will be available to use when you start TIGERSearch 1 2 The TIGERSearch interface This section covers basic use of the TIGERSearch interface Users experienced with TIGERSearch may wish to skip to section 1 3 1 2 1 Opening a corpus To begin launch TIGERSearch gt Windows double click tigersearch exe in the bin folder of your TIGERSearch installation gt Mac launch runTS command in the lib folder of your TIGERSearch installation TIGERSearch no corpus loaded Corpus Query Help no corpus loaded g Textual mode Graphical mode CorporaRoot DemoCorpora gt TromsoCorpora abevilletest gt ROLAND abeville Query window camcorp_rev2 camcorp_rev3 List of corpora Corpus Chooser Please select a corpus Corpus information y Open GI Bookmarks Corpus abeville closed Search Figure 1 TIGERSearch main window Once launched you will see the TIGERSearch main window see figure 1 which contains three main panels e a list of installed corpora top left The Yvain text yvainku tsbin will be on this list e information about the currently open corpus bottom left e aquery window right Double click yvainku tsbin in the corpus list to open the Yvain text
42. two separate dependencies on the head of the clause doie should since it is both the subject of the clause and a subordinating conjunction 2 Coordination is not represented within the dependency structure Coordinated nodes are instead grouped together and this non hierarchical grouping complements hierarchical dependency relations Due to the more than one parent approach the SRCMEF is not obliged to split word forms containing an enclitic determiner e g del al des as or an enclitic pronoun e g nel nes jel which necessarily have a double function The Old French word division conventionally adopted in print edition is therefore observed throughout the corpus 1 3 2 Representing dependency in TIGERSearch node pairs structures and heads The distinction made between terminal and non terminal nodes in TIGERSearch see 1 2 2 3 reflects the fact that it was also designed for constituency based annotation in which words terminal nodes are grouped into constituents non terminal nodes While such a distinction is irrelevant in the SRCMF it has had to be followed in order to make the corpus compatible with TIGERSearch Consequently each word in the SRCMF dependency structure is represented by a NODE PAIR in the TIGERSearch graph e aterminal node with features denoting the word form word and the part of speech pos e anon terminal node with a feature cat denoting the syntactic functio
43. y edges in the SRCMF in coordinate structures and to mark words with a double syntactic function In this case you ll see that the three conjuncts which make up the subject of the clause are linked to a node labelled Coo by secondary edges We ll return to this in more detail in section 1 3 2 2 1 2 2 Viewing corpus information gt Close the Graph Viewer window to return to the TIGERSearch main window Now the Yvain text is open the left hand panels of the main window provide useful information about the annotation used in the corpus In the top left panel you l see a list of the elements that make up a TIGERSearch graph edges non terminal nodes and terminal nodes A complete list of features is provided for each type of node gt Click on the cat feature listed under Nonterminal features Clicking on some features such as the cat feature will bring up a complete list of all possible values of that feature in the lower left panel including a gloss in French In the SRCMF this information is provided for all features which form a closed class gt Click on Detailed view listed under Documentation in the top left panel The detailed view of the corpus information gives a easily readable list of all information provided in the corpus header including feature lists number of tokens and number of graphs in the corpus 1 2 3 A quick corpus query 1 2 3 1 Launching a query Corpus queries ar
Download Pdf Manuals
Related Search