Post PAROLESIMPLE lexical resources and initiatives in Sweden

1
Post PAROLE/SIMPLE lexical resources and
initiatives in Sweden

• Maria Toporowska Gronostaj
• Maria.Gronostaj_at_svenska.gu.se

New Horizons for Linguistic Resources in a Global
Context 7-8th July 2009, Barcelona
2
The main aims of my talk

• present a work in progress on building a free
  full-scale lexical resource, the Swedish
  FrameNet (SFN), being conducted by the
  Swedish Language Bank
• give an overview of lexical resources
  contributing to the development of SFN
• describe a core of SFN, the SALDO lexicon
• reflect on merging lexical data from different
  resources to acquire information on frames

3
The overall objectives of the SFN

• create a robust lexical resource aimed at LT
  applications with
• exhaustive morphological, syntactic and semantic
  description of lexical units, incl. information
  on frames and world knowledge relevant for
  word/text understanding
• produce it cost-effectively by merging data from
  free lexical resources and re-using free software
  tools
• ensure its content interoperability
• create an interactive text-lexicon block with
  morphological and semantic annotations on the fly

4
Content interoperability a challenge for SFN

• The contributing lexicons are heterogeneous in
  several respects
• have partly different types of content
• were developed for different purposes
• were or are developed by different groups of
  contributors
• language experts
• a collective effort of both language engineers
  and users of web-lexicons

5
Free lexical resources behind SFN (1)

• SALDO Swedish monolingual lexicon with semantic
  and morphological layers
• 76,750 entries 74,000 distinct semantic units
• The Swedish Associative Thesaurus by
• L. Lönngren (1992) reincarnated by L. Borin
• enhanced with a complete morphological
  description by L. Borin M. Forsberg
• People's Synonym Dictionary (web-lexicon)
• 80,000 Swedish synonym pairs
• synonymy graded from 5 to 0 by lexicon users
• collective effort of web-lexicon users
• language engineering Viggo Kann

6
Free lexical resources behind SFN (2)

• The People's Dictionary (Swedish/English)
• collective effort of web-lexicon users
• equivalents are graded by lexicon users
• language engineering Viggo Kann
• SemNet
• 52,800 hyperonymy/hyponymy relations
  automatically retrieved from the definitions of
  nouns and verbs in GLDB
• Parole/Simple lexicons
• 29,000 syntactic units (valency) and 8,500
  semantic units encoded with mandatory information

7
(No Transcript)
8
SALDO unusual semantic network

• Lexemes, arranged in a hierarchical network
  according to the principle of centrality,
  capturing semantic closeness between two lexemes
• Semantic relations are postulated for both open
  and closed classes and can go beyond a word class
• There are 51 primitive semantically unrelated
  concepts being the top nodes of the hierarchies
  capturing the centrality. These nodes are
  connected to an artificial top node PRIM to form
  a tree
• There are no synsets in the sense of Wordnet.
  Neither glosses of the lexemes, nor semantic
  relations, such as hyponymy, hyperonymy or qualia
  relations are explicitly specified there.

9
Semantic centrality in SALDO

• Each lexical unit is given
• an obligatory main descriptor, mother, which can
  be complemented by an optional determinative
  descriptor, father
• bröd (bread) mat mjöl (foodflour)
• brud (bride) gifta sig hon (get marriedshe)
• bröllop (wedding) gifta sig (get married)
• gifta sig (get married) par (pair)

10
Semantic relations in SALDO

• Mother descriptor is usually
• semantically more close to the key word,
• semantically and/or morphologically less complex
  than the key word
• more frequent
• stylistically more unmarked
• acquired earlier in the first and second language
  acquisition
• Father descriptors are used mainly to
  differentiate lexemes having the same mother.
• They are assigned to ca 50 of words

11
Associative sets, assets

• Keywords can function as mother- or father
  descriptors for other lexemes and thus form the
  basis of any number of derived relations,
  referred to as assets
• brud (bride) get married she
• kronbrud (crown bride) bride chastity
• brudbukett (bride bouquet) bouquet bride
• brudklänning (bride dress) dress bride
• brudkrona (bride crown) crown bride
• brudgum (bridegroom) get maried he
• no assets

12
Assets sharing mother relations build natural
semantic groupings

• sol (sun) lysa himmel (shine sky)
• comet, moon, star (shine sky)
• blinka (blink) lysa snabbt (shine quickly)
• ljus (candle) lysa brinna (shine burn)
• The lexemes sharing mother and father relations
  are closer related to each other, as compared to
  those having different father descriptors

13
SALDO world knowledge (1)

• SALDO an intrinsic network capturing the world
  knowledge underlying lexical-semantic relations
• The network relations are based on the notion of
  centrality by the depth of an entry, its
  distance down from the PRIM root node
• The deeper an entry lies in the tree, the less
  central it is
• PRIM
• one
• unit
• two
• pair
• get married
• bride
• The average depth of entries in SALDO is 5, 7

14
SALDO world knowledge (2)

• SALDO is supportive in recognizing entailments by
  pointing out the mother to a key word, which
  promotes word text understanding
• It provides explicit information on distribution
  of the associative sets among lexemes (e.g. bride
  bridegroom)
• It includes named entities as entries
• Bulgakov författare rysk (writer Russian)

15
Approaches towards frames acquisition in SFN

• Merging relevant lexical data from available free
  lexical resources
• Cross-language transfer of lexical units with
  information on the frames and frame elements from
  FN to SFN
• Automatic acquisition of frames from corpora
  using a software tool, FrameNet Labeler system
  for Swedish text

16
Merging lexical data with SALDO involves

• interlinking the morphological units from the
  component lexicons (based on lemmas form, part
  of speech and inflectional patterns, whenever
  possible)
• augmenting the SALDOs lexical units with the
  semantic content from SemNet, SIMPLE, Peoples
  Synonym Dictionary and English equivalents from
  the Peoples Dictionary (Swedish/English)
• adding syntactic information from the PAROLE
  lexicon to SALDO

17
Frame acquistion supported by PAROLE/SIMPLE

• V gifta sig (to marry/get married)
• PAROLE
• Sub. (Anim.) V (refl.) PrepObj (Anim.) med
  (with)
• Sub. (Plural) (Anim.) V (refl.)
• SIMPLE
• Semantic type V Cooperative activity
• Selection restrictions Human V Human, Human V
• HumanVCooperative activityHuman gt Partner(s)
• In FN the Partner role is a core FE in the
  frames
• Collaboration, Forming Relationship, Personal
  relations
• Due to the semantic syntactic data in the P/S
  lexicon, the frame Forming Relationship is
  selected for the verb marry

18
Automatic acquistion of frames and FEs

• a software tool FrameNet Labeler for Swedish
  text
• elaborated by R. Johansson, P. Nugues
• trained on semantically annotated corpus,
  produced by a cross-language transfer
• 75 accuracy in classification of FEs

19
Populating the frames in SFNwith lexical units

• re-using the lexical data retrieved from corpora
  by the FrameNet labeler
• cross-language transfer of lexical units from FN
  to SFN
• semantic mining and refining lexical data in the
  SIMPLE lexicon
• enhancing the repository of lexical units with
  synonyms, hyponyms and siblings

20
Conclusions (1)

• Lexicons can be re-purposed and re-used for the
  task of SFN creation
• Content integration and interoperability seems to
  be feasible to achieve
• SFN can be augmented with
• synsets to compensate for the lack of glosses,
  (data from Peoples Synonym Dictionary)
• hyperonymy/hyponymy relationer from SemNet
• world knowledge from the SALDO lexicon
• Creation of a text-lexicon block with SALDO
  annotations on the fly is in progress

21
Conclusions (2)

• Desirable further extensions of SFN
• valency information
• explicit semantic typing of lexical units
• multi-word expressions
• broader coverage of different domains
• creation of text-lexicon block with semantic role
  annotations
• SFN will make a Swedish contribution to
  BLARK/CLARIN available under Creative Commons
  Attribute-Share Alike Licence and LGPL 3.0

22
References

• Borin L., Forsberg M. 2009. All in the Family A
  comparison of SALDO and WordNet. Proceedings of
  the 17th Nordic Conference of Computational
  Linguistics NODALIDA 2009. Odense.
• Johansson, R. Nugues, P. 2007. Construction of a
  FrameNet labeler for Swedish Text. NODALIDA 2007.
  Helsinki.
• Kann, V. , Rosell, M. 2005. Free Construction of
  a Swedish Dictionary of Synonyms. NODALIDA 2005.
  Joensuu.
• Lönngren, L. 1989. Svensk associationslexikon.
  Del /-IV Institutionen för lingvistik. Uppsala
  universitet. Rapport UCDL-R-89-1.
• Lönngren, L. 1998. A Swedish associative
  thesaurus. In Euralex 98 proceedings, Vol.2. pp
  467-474.
• SALDO http//spraakbanken.gu.se/sal/