This documentation concerns the SIMPLE part (semantic layer) of the Dutch PAROLE lexicon. For extensive documentation on the Dutch PAROLE lexicon, we refer to the INL's website: www.inl.nl

Contents of this report:

The Dutch SIMPLE lexicon dd. 28 April 2000 contains 10,472 semantic units (Usems): 7326 noun Usems, 2114 verb Usems and 1032 adjective Usems. The 10,472 Usems are distributed over 3710 lemmata (head words): 2797 nouns, 559 verbs and 354 adjectives. 681 lemmata cover one or more Base Concepts (see below).

For each part of speech, starting point for the lemmalist to be provided with Usems, were the English base concepts (BC) selected by the Linguistic Specification Group from EuroWordnet lexicon (see general SIMPLE documentation). For each BC, a set of related Dutch equivalents (near-synonyms) was chosen, keeping the concept in mind (to avoid mere translation of the English word). For a number of BC's, no appropriate Dutch equivalent could be found. The whole set of Dutch equivalents was checked on occurrence in the Dutch PAROLE lexicon, so as to be able to connect the semantic descriptions to syntactic descriptions in the PAROLE lexicon. For nouns, occurrence in the Dutch EuroWordnet lexicon could also be checked (by cooperation of Piek Vossen), which resulted into a finetuned list of Dutch equivalents. For verbs and adjectives, this comparison was not possible due to pragmatic reasons. Per BC, a 'prototypical' Dutch equivalent was selected. There were three reasons for selecting more than one prototypical equivalent per BC: (1) 'real' synonyms from which a choice would be too arbitrary, (2) no single lemma covering the BC could be found, (3) for nouns: the preferred prototypical lemma was at the time not in the Dutch EuroWordnet lexicon. Totally 681 SIMPLE lemmata cover one or more BC's: 371 nouns, 172 verbs and 138 adjectives. See appendix 1 for the noun, verb and adjective lemmata covering one or more BC's.

The lemmalists were extended on the basis of an automatically lemmatized type-frequency list derived from the Dutch PAROLE Reference corpus (the corpus itself is not yet lemmatized; cf. 1.5). This list was compared with the PAROLE lemma list. The list of matching entries was ranked from high to low frequency. In the range of lower frequencies, priority was given to lemmata corresponding with BC's.

BC-meanings were covered by Dutch Usems as precisely as possible, in view of the multilingual links foreseen.

Other meanings per lemma were selected by consultation of several medium-sized dictionaries of Dutch, and if necessary with other reference works (a.o. Wordnet). A criterion generally applied was that meanings shared by at least two of the dictionaries were selected as Usems for SIMPLE, based on the assumption that these meanings can be considered 'standard Dutch'. However, meanings that were considered outdated or obsolete (relics from older dictionaries) were not included. Domain specific meanings were included due to their importance for language technology (the domain field in SIMPLE).

For reasons of the rather limited size of the PAROLE lexicon (ca. 20,000 lemmata) and the SIMPLE lexicon ( 3,710 lemmata), target Usems in the qualia roles (formal, agentive, constitutive and telic) were determined on the basis of their suitability as target Usem, rather than their occurrence in the PAROLE or SIMPLE lexicon. The tables in 2.3.3. show that at the lemma level, target Usems are covered by the PAROLE lexicon between 76% and 85%, and by the (much smaller) SIMPLE lexicon between 36% and 50%.

The standard templates delivered by the Specification Group were loaded into a database. For each Usem, the lexicographers fill in a database template form. The template forms are automatically converted into the SIMPLE SGML format, by software developed for the purpose. For further details about tools, see section 1.4. A lexicographer's manual (written in Dutch), which was of course based on the guidelines but specified some aspects and some working procedures, was used in order to enhance quality and consistency among the lexicographers.

Apart from the obligatory template fields, the lexicographers compiled the recommended 'type hierarchy information' ('template_supertype', 'unification_path'), 'polysemous class' and 'qualia roles' (formal, agentive, constitutive and telic), for reasons of their relevance as establishing relationships between different (groups of) lemmata irrespective of Part of Speech, and more specifically between specific meanings of lemmata. It is just this feature that is missing, or only partially or not systematically treated in dictionaries. Particular these relationships will be used in one of our institutional projects on the longer term (see 1.5.).

Up to now, all noun Usems are compiled as for obligatory fields and the recommended fields mentioned. For the verb and adjective Usems, the recommended and obligatory fields are compiled, but the obligatory fields 'predicative representation', 'selectional restrictions' and link with syntax have not yet been finished (this is reported on in the last bimonthly). This is mainly due to the complexity of these fields, the time we needed to become familiar with the matter and its representation in SGML, and the problems we met (cf. 2.1, 2.5). However, now work is going on without major problems and completion is guaranteed by permanent staff with thorough knowledge of the matter working on it. Additionally, we found a way to continue contracts with temporary staff (see last bi-monthly). Apart from the 7326 noun Usems, ca. 450 verb Usems and ca. 150 adjective Usems have these fields finished now.

Appendices 1- 4 show more detailed information on the current lexicon contents.

The Dutch sample of 100 entries has the following characteristics.

As a starting point for the composition of the entrylist, we adopted the distribution figures applied in SIMPLE: 70% noun, 20% verb and 10% adjective Usems. The sample contains 71 noun entries with 263 Usems, 20 verbs entries with 90 Usems and 10 adjective entries with 43 Usems. In addition to the 20 verb entries, 9 Usems of 6 verbs that have a master link with noun Usems in the sample, have been included, which results in a total of 99 verb usems. The sample contains a total of 405 Usems.

The noun and adjective entries are the most frequent ones from the corresponding lemmalists (cf. 1.1.). Due to the short term of preparation of the review, the selection of the verbs is less elegant: we selected 2 x 10 verbs from 2 working files used for the compilation of the predicative fields (cf. 1.2). In order to demonstrate the non-master/master link between noun and verb Usems, the 9 Usems of the corresponding 6 verbs were added. These noun/verb Usem pairs concern: begin_1 ~ beginnen_4; gebruik_1 ~ gebruiken_1; leven_1, _2, _6, _7, _8 ~ leven_1, _6; onderzoek_1 ~ onderzoeken_1; werk_1,_4,_5,_6,_7 ~ werken_1,_2; wil_1,_2,_3 ~ willen_1,_3.

See appendices 2B-4B for the number of Usems per template type, per domain and semantic class for the Usems in this sample, separately for each Part of Speech.

Software tools used by INL for SIMPLE are:

• Database load tool
• Data entry tool
• Report tools
• Conversion to SGML tool

This tool (written in Perl) converts the standard 'skeleton' templates provided by the Linguistic Specification Group into an SQL command file, mainly consisting of 'insert'-statements. The tool searches for the relevant attribute-value pairs and builds the appropriate insert statement for it. The attribute-value pairs are easy to recognise as they are mostly in in the format: attribute:value.

When the command file is executed against the SIMPLE database, the skeleton template is loaded.

As an example we take the template for artwork. For this example we assume that it only consists of:

Usem: 1

Template_Type: [Artwork]

Unification_path: [Concrete_entity | ArtifactAgentive | Telic]

Applying the tool results in the following insert statement:

This tool is a data entry form similar to the standard templates, built with Uniface, a visual development environment available for a variety of platforms and databases systems. Although we dot not use the latest release, our release is still perfectly suited for the SIMPLE tasks.

The main functionality for SIMPLE was to enable the lexicographer to add, modify or delete information about Usems. Nearly all this functionality could be generated with Uniface, so we only had to write some additional code for e.g. checking purposes.

Report tools (all written in Perl) are used to collect information from the SIMPLE database which is not easy to collect by means of an SQL query. A typical approach is to obtain a set of data from the database and then applying the appropriate report tool.

As an example we look at the report tool for obtaining the number of Usems per domain. One Usem can have more than one value for domain (separated by ,). We need the separate values, so we have to split domain. As this cannot be done by SQL, we use a report tool. E.g the data set obtained from the database (one row represents one usem):

POLITICS_AND_GOVERNMENT, HISTORY, MONARCHY

FINANCE, ECONOMICS

FINANCE, SOCIOLOGY, GENERAL

POLITICS_AND_GOVERNMENT, GENERAL

The result after applying the tool:

The software for conversion to SGML consists of a large suite of Perl and C++ programs converting the database contents to SGML and establishing the link with the PAROLE lexicon. The conversion implies formal error correction (due to manual work in the database template forms) and the creation of additional SGML objects. The latter is automized, the former as far as possible.

The PAROLE lexicon is distributed by ELRA, and, for researchers in the Netherlands and Belgium only, by our institute. The PAROLE lexicon (without the SIMPLE part) is furthermore used in two (inter)national projects: the Dutch-Flemish project Corpus Gesproken Nederlands (Corpus of Spoken Dutch, comparable with BNC) and the Dutch project ToKeN2000, which aims at a sophisticated knowledge retrieval system, including modules concerning automatic language generation and spoken answers to questions by users. The latter project will also use the SIMPLE part of PAROLE.

The SIMPLE data will be used in the institutional project Integrated Language Database of 8th-21st Century Dutch, a long term project approved by the Dutch and Flemish governments. This project aims at creating a database in which data from linguistically annoted texts, electronic dictionaries and linguistic files will be linked in a meaningful way, in order to function as an instrument for research into the Dutch language and culture throughout the centuries. Especially the SIMPLE data establishing relationships between different (meanings of) lemmata are interesting for this project (cf. 1.2) and will be further developed in this framework. Furthermore, the PAROLE POS tagset will, with some extensions for the historical periods, probably be used for POS tagging of the texts in the database.

A current activity of our institute is to make the PAROLE corpus accessible over the Internet in a way similar to (but more modern than) the three INL corpora already operational (see www.inl.nl). Work is going on automatic lemmatization, POS tagging according to the PAROLE tagset and global syntactic tagging. A retrieval system is being developed which will give access to this corpus on linguistic parameters.

As explained in section 1.2., compilation of the fields 'predicative representation', 'selectional restrictions' and link with syntax have not yet been finished for verbs and adjectives, but completion is guaranteed by permament staff with thorough knowledge of the matter working on it. Additionally, we found a way to continue contracts with temporary staff (see last bi-monthly).

Permanent staff will be concerned with checks on quality and consistency.

General criteria for assigning readings to syntactic descriptions.

Syntax is linked to semantics by way of connecting the positions of the complementation frames of the Parole lexicon entries with the arguments of the Simple predicates. So the starting point for the link between Usem and Usyn are the syntactic descriptions (if any) of a lemma in the PAROLE lexicon. These syntactic descriptions are corpus based. In order to establish the connection we first had to decide about the semantic correspondence of each syntactic complement frame of the lexicon entry with the specific argument structure of the related Usem. Secondly, we had to decide which of the complements of a frame could actually be semantically linked to an argument. These decisions were made by interpretation of the relevant predicate and then compare it with the example phrases belonging to the syntactic complementation frames in question in order to see whether a correspondence might be established.

In the Dutch PAROLE lexicon, the basic fact of an adjective determining a noun is considered to belong to the grammar. In the syntactic frames, as a consequence, there is no position available for the noun in question. So, in our lexicon, adjectival complementation as found in: the man is angry with me, is described in the grammatically correct phrasing: The ‘with me angry’ man, where the entry ‘angry’ has a one-place frame with the prepositional complement PP(with) on the first and only position (P0).

For that reason, up to now, we described adjectival predication as one-place predicates:

angry_1(<arg0>),

where <arg0>=PP(with) with selectional restriction [Living Entity]

This implies that we don’t have the possibility to describe what kind of noun (for example: only living entity) the adjective selects for. Moreover, subject clauses which have the same P0 position in the frame, like:

it is good for me to see you (read: ‘to see you’ is good for me)

it is easy for me to do that (read: ‘to do that’ is easy for me)

cannot be linked to a semantic argument either.

This fact causes a considerable loss of semantic information and we plan to change the one-place predicates into two-place predicates:

angry_1(<arg0>)(<arg1>),

where <arg0>=NP-[Living_Entity] and <arg1>=PP(with) [Living_Entity]

By consequence we can only link <arg1> to a syntactic position, and have leave <arg0> unlinked, which is allowed in Simple.

As said above, the selection of syntactic complements in the Parole lexicon has been based on their freqency in our corpora. Sometimes, however, we considered a certain complementation pattern as non-prototypical from a semantic point of view. So in these cases, the predicate differs from the corpus based complementation patterns. By consequence, the ID shows that the non-prototypical complementation pattern is possible, but the respective positions have not been linked to an argument. Cf: boek_4, (book_4) with gloss ‘deel van een meerdelig boekwerk’ (part of a multipartite work). We choose for a predicate with only one (prototypical) argument: the name of the book, like the book Genesis. The two possible syntactic complements PP(van), naming the author of a book and PP(over), naming the topic the book is about, are not considered prototypical for that specific noun Usem and is therefore not linked to the only available argument.

Sometimes different syntactic complements (e.g. an N or a PP) might be linked to one semantic argument, but this is not possible because of the selectional restriction on the argument, eg:

nota_3(<arg0>)(<arg1>) de nota van de minister over het kunstbeleid (the note of the minister on art management)

where <arg0> = Role_ProtoAgent and <arg1> = Role_ProtoPatient

The PAROLE lexicon has three descriptions for nota:

(1) de nota Kunstbeleid (the note (called)Art management)
(2) de nota van de minister (the note of the minister)
(3) de nota over het kunstbeleid (the note on art management)

Only two syntactic frames (2 and 3) can be linked. The complement in (1) cannot be linked to <arg1> because of its role: Role_Adjunct instead of Role_ProtoPatient.

The general SIMPLE criterion for the selection of a domain value from the SIMPLE domain list is the topic of texts in which a Usem usually appears, or is most likely to appear. The most specific domain in the hierarchy was selected. If no suitable specific domain was available (e.g. a particular branch of sport), the immediate node was selected. Usems of very common usage have the value 'general'. A Usem can have one or more domain values. If a Usem is both domain specific (i.e. likely to appear in domain specific texts) and of very common usage, both specific domain values and the value 'general' are assigned. See appendix 3A for frequencies of Usems per domain value, separately for each Part of Speech (3.A.1-3.A.3).

The principles for the selection of a semantic class from the SIMPLE hierarchy were essentially the same as for domain: whenever possible the most specific one. Only one value is either assigned or selected from the list. A semantic class value can be specified by a distinctive feature. See appendix 4A for frequencies per semantic class, separately for each Part of Speech (4.A.1-4.A.3).

The selection of a template type was based on knowledge of the ontology and on the linguistic tests and the information per template type provided by the Specification Group. Furthermore, the suitability of the qualia roles and their values was used as a check on the correct template type: the selection was probably wrong if the qualia roles were not suitable for the particular meaning of an Usem. Irrespective of its status as core or recommended, the template type which was considered most suitable and most specific was selected.

After the final and complete set of template types was provided by the Specification Group, the lexicographers checked the selection of template type for noun Usems corresponding with Base Concepts, as at the time of compilation the set of template types was much smaller.

No language specific typing was applied.

No language specific template subtyping was applied.

Relations. In the standard templates, template-specific relations for the qualia roles are 'predefined', with an indication of their status (type-defining, optional). These were taken as starting point. Whenever a predefined relation could not be filled in sensefully, this is marked as such in the database and then the relation is not implemented in the SGML file. Relations were added if they were judged essential for the particular Usem. These cases are marked as such in the database and implemented in the SGML file.

Choice of target Usems. As said above (1.1), due to the limited size of both the SIMPLE lexicon and the PAROLE lexicon, it was considred not desirable to select only target Usems that are in those lexicons. For this reason, our first criterion for the selection of target Usems was suitability for the specific relation between the Usem and its target Usem.

The present state of the art of coverage of target Usems in the lexicons is as follows. Per Part of Speech, the target Usems were collected and compared with the complete entrylist of the PAROLE lexicon and the SIMPLE lexicon, respectively. Note that 'target lemmata' are lemmatized target Usems, that is a target lemma includes all target Usems with an identical word form.

NOUNS

VERBS

ADJECTIVES

Not applicable. This field has not been compiled.

Polysemic relations given in the standard templates provided by the Specification Group were judged for their suitability for a particular Usem. If not, the relation is marked as such in the database template field and not implemented in the SGML file.

Only polysemous relations that have a counterpart in another Usem of the lemma concerned, are implemented in the SGML file.

Synonymy and hyponmy is not (yet) encoded in our lexicon.

The first step on the predication path was to decide about a Usem’s Type of Link with another Part of Speech item. Our starting point in this has been the basic SIMPLE assumption (Del. 2.1, p34/35): "We assume that verbs and adjectives always have a master relation with the predicate". For that reason even denominal verbs and adjectives are considered to be Master (where their true (etymological) derivation is to be described in the derivation field of the template). Nouns therefore were the only Part of Speech to be decided upon. As a consequence of what has been said, noun Usems are only Master if they cannot be semantically related to an adjectival or verbal predicate. All other noun Usems are non-Master, independently of the factual etymological relation between the noun and the verb or the noun and the adjective.

Whenever a noun Usem could be semantically related to the predication of two different verbs the semantically most resembling was taken. In the case of:

gevoel~ gevoelen|voelen

feeling~feel|feel

we choose for ‘voelen’.

Whenever a noun Usem could be semantically related to a verbal as well as to an adjectival predicate, the verb was given priority, cf:

associatie ~ associeren/associatief

association~ associate/associative

Whenever a single noun Usem could be semantically related to two (or more) adjectives, the most meaningful was choosen: (the suffix –achtig (‘-like’) can be added to nearly every noun)

monster ~ monsterlijk/monsterachtig (we choose ‘monsterlijk’)

mos ~ mossig/mosachtig (we choose ‘mossig’)

paniek ~ panisch/paniekerig/paniekachtig (we choose ‘panisch’)

Sometimes a noun might be semantically related to an adjective, but this adjective is in fact the noun itself in an adjectival function, cf:

moslim_(noun), moslim_(adj) in: een moslim soldaat/aanval (the ‘real’ adjective: ‘moslims’ is not used)

model_(noun), model_(adj) in: een model vader (an exemplary father)

We considered the nouns model and moslim to be a Master.

Whenever a noun Usem could only be semantically related to another noun, we considered that noun Usem to be a Master, cf:

mensheid_2 ~ mens (manhood ~man)

ministerie~minister (ministry~minister)

The Guidelines propose to maximalize predication. We somehow restricted this principle by the criterion of prototypicality. For example, in many cases locative arguments are possible, and they are often part of the PAROLE descriptions of nouns. We, however, have only included them in the Simple predication if they were prototypical for the Usem in question. Take for example ‘artikel’ (article), where we consider the locative argument to be prototypical, because ‘artikel_2’, the item published in a paper, is different from, e.g. ‘artikel_3’, an item which only shows up in dictionaries.

As the semantic roles mentioned in the Simple Guidelines are merely syntactically defined, we were in need of a more semantically driven notion in order to decide upon the semantic role of verbal arguments. For example, syntactic subjects vary considerably where their semantic role is concerned. They can have semantic role ProtoAgent, ProtoPatient or Underspecified. So we introduced the notion of ‘Control’ to decide whether a syntactic subject had role ProtoAgent (cf. Oppentocht 1999). Our working definition of this role is: "someone having control over the act expressed in the verbal or nominal predicate". According to the Guidelines the direct object is always attributed the ProtoPatient-role. We missed the possibility to discriminate between whether the object is ‘undergoing the event’ or ‘benefitting from the event’. Moreover syntactic subjects may also have RoleProtoPatient, cf.

the action starts today, the boat is sinking, the storm died down, my cousin hates me

So in addition to the Guidelines definition: "for the direct object and strongly bound prepositional complements", we supplied "and for those subjects which experience or undergo the event expressed by the nominal or verbal predicate".

As a consequence Role_Underspecified is often attributed to subjects which neither control nor experience or undergo the action expressed in the verbal or nominal predicate: cf

the storm caused great damages

his anger showed how much he was affected by this

In addition to the closed list of Semantic Roles, we would be very happy with the semantic roles Beneficiary and Goal.

Problems are reported on in the respective sections.

See appendices 1 and 2A-4A for information concerning the complete dataset. See for information about the delivered 100 entries appendices 2B-4B.

Oppentocht, L. (1999), Lexical Semantic Classification of Dutch verbs. Towards constructing NLP and human-friendly definitions. Ph.D. dissertation, University of Leiden, The Netherlands.

Appendix 1: List of lemmata covering one or more base concepts, for nouns (1.1), verbs (1.2) and adjectives (1.3), respectively.

Appendix 2.A: Number of Usems per template type in the complete dataset, for nouns (2.A.1), verbs (2.A.2) and adjectives (2.A.3), respectively.

Appendix 2.B: Number of Usems per template type in the sample of 100 entries, for nouns (2.B.1), verbs (2.B.2) and adjectives (2.B.3), respectively.

Appendix 3.A: Number of Usems per Domain in the complete dataset, for nouns (3.A.1), verbs (3.A.2) and adjectives (3.A.3), respectively.

Appendix 3.B: Number of Usems per Domain in the sample of 100 entries, for nouns (3.B.1), verbs (3.B.2) and adjectives (3.B.3), respectively.

Appendix 4.A: Number of Usems per Semantic Class in the complete dataset, for nouns (4.A.1), verbs (4.A.2) and adjectives (4.A.3), respectively.

Appendix 4.B: Number of Usems per Semantic Class in the sample of 100 entries, for nouns (4.B.1), verbs (4.B.2) and adjectives (4.B.3), respectively.