MonaSearch – A Tool for Querying Linguistic Treebanks
Contents
1. cat y PX Tree 13 Viele meinen da Perot mit seinem Befehlston auf dem Capitol gegen eine Tree 14 So erkl rt etwa Edward Brandon von dem Unternehmen National City in Ohic Tree 15 Ich glaube kaum da mit seinem naja etwas undiplomatischen Stil im Tree 17 Und ein anderer Manager vermutet da sich ein Dogmatiker wie Perot ir Tree 18 Ein ehemaliger Gesch ftsf hrer der heute in f nf Konzernen im Aufsichtsgr rl P Submit save Figure 2 The MonaSearch GUI compose and result window 5 Performance In order to give an impression of the performance of MonaSearch we compare query times for a series of queries for three query tools TIGERSearch 2000 fsq 2003 and MonaSearch These tools were chosen because they can be used to query any treebank and already have rather high expressiveness We tested their forces on the TIGER treebank and on T Ba D Z The queries are ordered by complexity starting with simple ones 1 An NP dominating an S dominating a PP 2 An NP dominating an S dominating a PP and an NP which do not dominate each other 3 Sentences where the verb precedes the subject 4 An NP not dominating an S which dominates a PP 5 A PP dominating an NP which is part of a chain of embedded PPs see introduction Table 2 shows that MonaSearch performs quite well TIGERSearch may be faster on simple queries but its deficit in expressive power makes it unsuitable2. for querying more advanced linguistic phenomena In fsq such phenomena can be queried but MonaSearch is faster and even more powerful Furthermore fsq requires a certain skill in composing a query Processing a query with quantifier depth 4 takes 23 seconds when the query is composed optimally But it can also take 46 minutes when the query is composed in a less skilled fashion MonaSearch does not suffer from this problem The processing time of logically equivalent queries is largely independent of the particular phrasing of the queries This makes MonaSearch significantly easier to use 38 Query TIGERSearch fsq MonaSearch 5 5 5 5 23 13 5 15 10 9 5 5 23 13 5 15 10 15 16 23 13 5 15 10 _ 23 13 5 15 10 15 10 Query times in seconds For each query tool the left column contains the query time on the TIGER treebank the right column the one on the T Ba D Z A dash indicates that the query language of the tool is not powerful enough to express the query System Pentium D 3 2GHz 1GB Ram OS Linux openSUSE 10 3 Table 2 Query times of MonaSearch in comparison to other tools Since all trees are evaluated separately the evaluation time of a query is linear in the number of trees in the treebank Effectively a huge treebank with one million trees can be queried in about five minutes i e quite fast 6 Related Work Alternative query tools and their performance have just been discussed in the previ ous section
3. automata tech niques and applications 2007 URL http tata gforge inria fr Stephan Kepser Finite Structure Query A tool for querying syntactically annot ated corpora In Ann Copestake and Jan Haji editors Proceedings EACL 2003 pages 179 186 2003 Stephan Kepser Querying linguistic treebanks with monadic second order logic in linear time Journal of Logic Language and Information 13 457 470 2004 Stephan Kepser Using MONA for querying linguistic treebanks In Chris Brew Lee Feng Chien and Katrin Kirchhoff editors Proceedings HLT EMNLP 20035 pages 555 563 2005 40 Nils Klarlund and Anders M ller MONA Version 1 4 User Manual BRICS Department of Computer Science University of Aarhus January 2001 URL http www brics dk mona Wolfgang Lezius Tigersearch ein suchwerkzeug fiir baumbanken In Stephan Busemann editor Proceedings der 6 Konferenz zur Verarbeitung natiirlicher Sprache KONVENS 2002 Saarbr cken 2000 Heike Telljohann Erhard W Hinrichs and Sandra K bler The T Ba D Z treebank annotating German with a context free backbone In Proceedings of the Fourth International Conference on Language Resources and Evaluation LREC 2004 pages 2229 2232 2004
4. bottom up tree automata For a general introduction to tree automata we refer the reader to 2007 The connection between MSO and tree automata is very strong A set of trees is definable by an MSO formula if and only if there exists a tree automaton accepting this set This equivalence is constructive there is an algorithm that constructs an automaton from a given MSO formula Hence the general evaluation strategy of MonaSearch can be summarized as follows 1 convert the query from the user into a tree automaton 2 run the auto maton on each tree in the tree bank In addition some auxiliary steps are necessary to smooth computation This strategy has the advantage that checking whether an automaton accepts a tree is very fast viz linear in the size of the tree 3 MONA The largest and most demanding subpart of this process is the development of a tree automata toolkit a toolkit that compiles formulas into tree automata by performing standard operations on tree automata such as union intersection negation and determinization The most notable and successful is MONA Klarlund and Moller More precisely over graphs with bounded tree width 32 2001 Although developed for hardware verification MONA can be used to query treebanks The language of MONA is pure monadic second order logic of two successors Note that there is no way to extend this language This has important consequences Firstly we are restricted to using b
5. 29 MonaSearch A Tool for Querying Linguistic Treebanks Hendrik Maryns and Stephan Kepser Collaborative Research Centre 441 University of T bingen Germany hendrik kepser sfs uni tuebingen de Abstract MonaSearch is a new powerful query tool for linguistic treebanks The query language of MonaSearch is monadic second order logic an extension of first order logic capable of expressing probably all linguistically interest ing queries In order to process queries efficiently they are compiled into tree automata A treebank is queried by checking whether the automaton repres enting the query accepts the tree for each tree Experiments show that even complex queries can be executed very efficiently The tree automaton toolkit MONA is used for the computation of the automata 1 Introduction A frequent critique to most available query tools is their poor expressive power most of them only incorporate a subset of first order logic as a query language Furthermore often the semantics of the query language is not clearly defined it is only defined by the functionality of the program The importance of the expressive power of the query language is a consequence of the sizes of the available treebanks nowadays tree banks of several tens of thousands of trees are no exception It is clearly impossible to browse these treebanks manually searching for linguistic phenomena But a query tool that does not permit the user to specify the sou
6. The aim of this section is to explain how the current approach differs from the one in 2005 explores a different strategy for using MONA to query linguistic treebanks This different strategy proved unsuc cessful The reason for this being that trees were encoded as MONA automata featuring all linguistic labels of the treebank As a result the automata which en coded the trees got so big they could not be processed any more The current approach does not try to code trees as automata The current approach avoids large label sets on trees by equipping each tree that is to be checked by an automaton with only those labels which are relevant to the current query as described in step 4 1 in section 4 2 By avoiding these pitfalls the method presented in this paper is successful 7 Conclusion and Future Work In this paper we presented MonaSearch a powerful tool for querying linguistic treebanks The two main features of MonaSearch are the very high expressive power of the query language and the high performance of the query engine The query language is monadic second order logic This is to our knowledge the most expressive query language offered in any query tool useful for linguists Mon 39 aSearch is on the other hand also the fastest query system available for advanced queries Thanks to MonaSearch s speed it is for the first time within reach to query automatically annotated treebanks with several million trees with non trivial
7. a simple tree it ignores the root node and starts looking at its left child This left child will necessarily be the first root node of the original tree no further tweaking is needed 4 3 User interface MonaSearch offers MSO as a query language The usual logical connectives are provided along with first and second order quantification For the atomic formu las those relevant for relations of nodes in a tree are chosen along with relations that express relations between nodes and node sets Furthermore linguistic labels of nodes can be queried First order variables are supposed to range over nodes whereas second order variables the particular property of MSO range over sets of nodes The user can compose these formulas stepwise in a bottom up fashion in the graphical user interface which acts like a scrap book of formulas that can be ed ited connected using the logical connectives and submitted for querying Base formulas can be entered from menus We illustrate the query language by an example of a query In most main sen tences in German as in English the subject precedes the inflected verb In yes no questions this order is reversed To find trees where the verb precedes the subject in the T Ba D Z treebank the following query may be used x y z cat x SIMPX cat y VXFIN A fet z ONAx lt IT yAx lt it zAy lt z There exist a node with category SIMPX a node with category VXFIN and a node with func
8. ach node in a tree processable by the automaton is a bit vector The index of the vector states which variable is stored at that place The value 0 or 1 indicates if the node is in the set of this variable Since linguistic labels are treated as set variables each binary tree must undergo this translation process We impose some arbitrary but fixed order on the labels The length of the bit vector is the number of labels in the query The index states 36 which particular label is referred to Now consider a node in the binary tree and its original set L of labels The new label is a bit vector If the index refers to a label in L the bit is set to 1 otherwise it is set to 0 As a consequence the bit vectors in the translated tree only refer to labels occurring in the query After this translation step the input tree is finally ready for processing The MONA library contains functions that load the precompiled tree automaton from file and apply it to a translated tree returning true or false depending on whether the tree automaton accepts the tree or not It remains to solve one small technicality Remember that in the binarization process a virtual root node was introduced to connect disconnected parts of the tree Care has to be taken to ignore this node when evaluating the query in order not to skew the results Here fortune has smiled upon us the particular form of tree automaton MONA uses happens to do just that when evaluating
9. ary trees we employ the First Child Next Sib ling encoding a rather standard technique An arbitrary node x in the original tree is transformed into x in the binary tree as follows e If x has any children call its leftmost child y then y will become the left child of x e If x has any right siblings call the leftmost one z then z will become the right child of x For example the left tree in fig 1 stemming from the T Ba D Z treebank is trans formed into the tree on the right For real trees the original tree can easily be recre ated from this binary form in fact that is what the formula translation described later is implicitly doing The simplifications mentioned before however are not retrievable from the binary form and are therefore not available for querying Since we refer to trees using an identifier we do not need to do this back translation ex plicitly we simply use the original tree for further processing Note how the disconnected punctuation node at the lower right corner in the original tree becomes the right child of the SIMPX node in the transformed tree it is treated as if it was its sibling in the original tree As can be seen in the figure edge labels are moved down to the node below it thus a node in the translated tree carries a set of labels This is necessary either way since even in the original tree the leaves carry multiple labels We d like to mention that it is only a coincide
10. ght linguistic phenomenon quite precisely is not too helpful either If the user can The work presented here is part of the project A2 in the Collaborative Research Centre 441 Linguistic Data Structures at the University of Tiibingen funded by a grant of the German Re search Council DFG SFB 441 We would like to thank Michael Jakl Uwe M nnich and Fang Wei for interesting and fruitful discussions We are particularly grateful to Michael for pointing us to the function mgCheck in the MONA gta library We also thank the anonymous reviewers for their remarks 30 only approximate the phenomenon he seeks answer sets will be very big often containing several hundreds to thousands of trees Weeding through answer sets of this size is still cumbersome and not really fruitful If the task is to gain small answer sets then query languages have to be powerful As a query language offering the necessary expressive power we present Mon adic Second Order logic henceforth MSO describes its power and how it can be exploited to query tree banks in linear time This paper describes how these ideas were successfully implemented to obtain the most powerful query engine for treebanks to date MSO is an extension of first order predicate logic by set variables These vari ables can be quantified over representing finite sets of nodes in a tree MSO offers a firm expressive power while at the same time the automaton model dis cussed below a
11. inary trees only secondly it cannot handle node labels The main part of MONA is a compiler that compiles formulas in an idiosyn cratic logical language into a specific variant of tree automata The input is a file containing the formulas The default output is some information on whether the formula is satisfiable or not it is however possible to let it output a representation of a compiled automaton into a file Additionally MONA offers a small library with functions that allow one to load those precompiled automata from file and to check if an automaton accepts a binary coded tree The strategy for employing MONA to query treebanks consists of two parts Since we don t want to burden the user with learning the complex MONA lan guage a query from the user in MSO is translated into a MONA formula and this formula is compiled into a tree automaton by the main program The library func tions are then used to check each tree in the treebank for acceptance The trees from the treebank have to be transformed in a suitable way described below such that they can be used as input to MONA s particular type of automata 4 MonaSearch In this section we ll go into some more detail about the steps involved in a query and the preparatory tasks and explain how they are handled in MonaSearch 4 1 Precompilation of the Treebank Linguistic treebanks are usually provided in some type of data exchange format This format is typically unsuitable fo
12. nce and precedence being the most relevant Their trans lation uses two auxiliary predicates on the binary tree which can be defined in the MONA language e dom x y expresses that x dominates y in the binary tree e right_branch x y expresses that y lies at the branch of right children start ing at x Note that both express the transitive closure of a first order relation parenthood and right child respectively and as such cannot be expressed in first order logic Table 1 shows the translation of the parent and precedence relations on the original trees into a formula on binary trees in the language of MONA see the user manual for the notation Note that the dominance relation could equally well be expressed as the transitive closure of the parent relation We include it since it turns out that a direct translation yields a simpler formula Finally all types of linguistic labels are treated as free set variables This yields a MONA formula 9 In step B the MONA formula is compiled into a tree automaton by a call to the MONA compiler This implies that MONA has to be installed on the system in order for MonaSearch to work Step 4 comprises the actual querying of the treebank The aim is to run the tree automaton on each binary coded tree Remember that MONA is not able to handle the preprocessed trees directly The process of compiling a formula into an automaton translates variables into bit vectors More precisely the label of e
13. nce of the chosen annotation scheme and source tree that all and only the leaf nodes of the binary tree are labelled with words This need not be the case in general Presently labels for the category of a node its function the word at a leaf its lemma or its morphological information occur however other labels can be treated the same way The preprocessing of the labels is still not enough to enable MONA to handle the binary trees The trees it expects contain bit vectors as labels which are closely 34 virtual root a SIMPX a KOORD KON gt IR Oder p S VXFIN MF l SIMPX E HD 2 E VAFIN ne a HD On C ON MF PRED ist gt S NE a ae HD ADVX Bremerhaven p amp a aa 3 ADJX A HD PRED HD HD HD HD nicht a ADJD Oder ist Bremerhaven nicht g nstiger HD KON VAFIN NE PTKNEG ADJD g nstiger Figure 1 An example sentence from T Ba D Z and its binary coding related to the translation process These bit vectors are however dependent on the specific query and are therefore generated while querying 4 2 Querying The detailed steps of a query are as follows 1 getting a query from the user 2 translating the query into a MONA formula on binary trees 3 compiling the MONA formula into a MONA tree automaton 4 for each tree from the precompiled treebank 1 preparing the
14. other as indicated This contrasts with a structure like in The dog buried the bone pp with his paws pp under a stone pp behind the tree pp in the afternoon where each PP is an independent adjunct of the verb Both examples can stepwise be extended to sequences of PPs of arbitrary length The query for an arbitrary 31 number of embedded PPs is not definable in first order logic but can be defined in MSO We can define the transitive closure of a PP dominating an NP as a new relation and state that the highest PP node and the most deeply embedded NP node stand in this new relation Another property of MSO which comes in nicely is its decidability over trees This assures that when a query gives no results this effectively means no tree in the tree bank satisfies the formula rather than that no tree was found In this paper we present MonaSearch a query tool which implements MSO as a query language using automata techniques This way we show that the theoretical considerations are indeed also practically valid We compare run times of a set of linguistically motivated queries to show that MonaSearch is not only powerful but also faster than its main competitors MonaSearch is publicly available at http tcl sfs uni tuebingen de Monasearch 2 Overview of the Querying Process What makes MSO usable as a query language is the existence of an automaton model for this logic The type of automata that correspond to MSO is
15. quer ies Future work has two directions Firstly usability will be enhanced by simplify ing the syntax of the query language extending the range of treebank input formats and enhancing the GUI An important component would be a graphical treebank browser Secondly it would be worth investigating whether the restriction to proper trees can be overcome As stated there exists a theoretical solution for this prob lem But the recoding of trees and formulas that is involved in this solution is very complicated Formulas may get very large It is therefore an open question whether these formulas can still be compiled into automata But the potential advantage of coping with arbitrary tree like structures should make it worthwhile to investigate the practicability of this solution References Gosse Bouma and Geert Kloosterman Mining syntactically annotated corpora with xquery In Proceedings of the Linguistic Annotation Workshop pages 17 24 Prague June 2007 Association for Computational Linguistics URL http www aclweb org anthology W W07 WO7 1503 Sabine Brants Stefanie Dipper Peter Eisenberg Silvia Hansen Esther K nig Wolfgang Lezius Christian Rohrer George Smith and Hans Uszkoreit TI GER Linguistic interpretation of a German corpus Research on Language and Computation 2 4 597 620 2004 Hubert Comon Max Dauchet Remi Gilleron Christof L ding Florent Jacque mard Denis Lugiez Sophie Tison and Marc Tommasi Tree
16. r direct querying Plain text files are not suit able for efficient processing by their nature pro pose an efficient way of handling XML encoded treebanks but it requires the tree bank to be in a specific format E g the T Ba D Z and TIGER treebanks cannot be queried in that way MonaSearch does not suffer from this restriction In the case of MonaSearch there is a particular reason for a precompilation step As mentioned before MONA can only cope with proper binary trees When translating trees from the treebank into MONA trees we consider proper trees only Many treebanks contain more complex structures than proper trees At the current 33 stage of the development of MonaSearch we simplify these structures as follows 1 secondary relations are ignored 2 disconnected subparts are integrated into the tree by introducing a new virtual root and connecting the disconnected parts to this super root and 3 crossing edges are ignored by taking only the order of the children as seen by the parent into account In principle these restrictions can be released There exists a more general method that encodes tree like structures into proper binary trees as is illustrated in Kepser 2004 As stated above the precompilation of trees into formulas has to perform two tasks 1 the trees which are arbitrarily branching must be transformed into binary trees and 2 the linguistic labels have to be taken care of For the transformation into bin
17. ssures the evaluation time of an MSO query on a tree is linear in the size of the tree An example of its strength is the ability to express the transitive closure of any binary relation which is definable in the language Since MSO is an extension of first order logic any first order relation can be expressed in it and thus also the transitive closure of any first order relation Transitive closures appear quite often and very naturally in linguistics Most often the need for them has been met by introducing atomic formulas which dir ectly express the transitive relation For the ancestor relation for example an atomic formula x lt y is introduced next to the more basic parent relation x lt y The limitation of this approach is that those relations are hard coded in the logic There is no way to define new ones apart from extending the language In the case of a corpus query tool this would mean the extra relation has to be implemented by the programmer himself In MSO this can be done dynamically by the user As it is hard to foresee the need for such relations this feature can come in handy An example illustrating this which might be regarded as a natural query for MonaSearch is the following Consider sentences ending in a sequence of prepos itional phrases like The dog buried the bone pp behind the tree pp in the garden pp in front of the house pp at the end of the street where the PPs are all embedded by one an
18. tion ON the first dominates the other two and the second comes before the third ON codes the grammatical function subject in the T Ba D Z The tree in fig lis a sample match for this query 37 l MonaSearch ry x TE Results for E1 x E1 y amp amp gt x y cat x PP cat y APPR _ o x File Query Base Complex Info File Treebank tiger1000 utf export Tree 2 Konzernchefs lehnen den Milliard r als US Pr sidenten ab a Tree 3 Texaner gibt nur vage Auskunft ber seine Wirtschaftspolitik Ely fct y Tree 4 Der texanische Milliard r Ross Perot hat das politische Establishment in Wash E1 y fct y OA ei yllammav einfl en Tree 5 Nach Meinungsumfragen liegt der parteilose Self mademan gut im Rennen ul y yal Tree 7 Ob diese F higkeiten aber ausreichen um die gr te Volkswirtschaft der Wel E1 y morph y 3 Sg Pres ind Tree 8 Und auch die Konzernchefs in den USA halten nicht viel von dem 62j hrigen El d tzt E1 y word y jetzt Tree 9 Zwar k nnen sich die meisten Topmanager durchaus einen Unternehmer als E1 x cat x KOUS Tree 10 Nach einer Umfrage des Wirtschaftsmagazins Fortune unter den Bossen vor Free TEN SAE HEREBY Tree 12 Als gr te Schw che des Texaners nennen die Befragten seinen Mangel an x x x x E1 x cat x PPER E1 x cat x DM E1 x E1 y amp amp gt x y cat x NF cat y PX E1 x El y amp amp gt xy cat x NF
19. tree for the query 2 running the automaton on the translated tree noting whether it is ac cepted or not 5 presenting the results to the user Steps I and 5 are handled in a graphical user interface For step 2 one has to keep in mind that the user formulates his query on the original treebank but the trees that are used for querying are the binary coded trees Hence it is necessary to translate the original query in such a way into a MONA formula that the original query is true for an original tree t if and only if the translated formula is true on the binary recoding r of t This step also has the advantage that it makes us independent of the idiosyn cratic MONA language We can offer the user an easily understandable semantic 35 Relation Formula Translation parenthood x lt y ex1 z x 0 z amp right_branch z y precedence x lt y ex1 z x z dom z 0 x amp dom z 1 y dominance x lt ty ext z x 0 z amp dom z y Table 1 MONA translation of the more involved base relations ally clear language that is almost the pure MSO logic It also makes extending the query language with typical linguistic relations such as c command relatively easy For most connectives the MONA counterpart can be taken over directly This is the case for all boolean connectives quantification and some of the atomic rela tions The main problem lies with the relations that express something about the shape of the tree domina
Download Pdf Manuals
Related Search