ion language (SAL), used to generate hierarchical hyponymy and hypernymy relations. The paper also describes action-of, result-of, and synonymy relations between multiword units and single words, mostly where there is a morpho-syntactic and semantic relation between words of distinct parts-of-speech. The semantic relations were generated automatically, based on the linguistic information associated with each lexical entry in NooJ dictionaries. Local grammars were developed as a mechanism to read this linguistic information and generate the semantic relations, which have been used in paraphrasing and machine translation. Dictionaries and grammars can easily be adapted to distinct languages and are useful to various natural language processing monolingual or cross-language tasks.
Lexical Semantics
Semantic relations are crucial to understand and to structure the meaning of natural language. They are vital to communication overall, and highly employed in technical and specialized domains, where the most important content of texts is conveyed through the semantic relations between the terms that represent the domain’s concepts, rather than by the meaning of the words alone (e.g., the semantic relations between BRCA1/protein and RNF53/gene in the biomedical field). Additionally, semantic relations are important for applications in the semantic web, mapping ontologies, text categorization, natural language understanding, etc., and a requisite for paraphrasing and machine translation, where words and expressions often must be substituted by semantic equivalents, such as synonyms between support verb constructions and single verbs (make an operation = operate; say hello to = greet ), or other type of semantic alternates.
The most studied lexico-semantic relations are: (1) synonymy, when different lexical items have the same meaning (e.g. car synonym-of automobile); (2) homonymy, when lexical items have the same orthographic form but different meanings (e.g. bank, financial institution vs. slope); (3) hyponymy, when a lexical item is a subclass or a specific kind of another (e.g. dog hyponym-of mammal ); and (4) meronymy, when a lexical item is a part, piece or member of another (e.g. wheel part-of car ).
This paper describes the first attempt
to extract cross-language semantic relations
between English and Portuguese from the
lexical resources of the OpenLogos machine
translation system described by
Dictionaries are probably the main source of lexico-semantic knowledge, as they are repositories of words, which include the description of several word senses. However, as definitions are written in natural language, dictionaries are not completely ready for being used as computational lexical resources.
Common representations of
lexicosemantic knowledge, ready for being used in
natural language processing tasks, include
thesauri, taxonomies, as well as lexical
ontologies or lexical knowledge bases. For
example, the Roget Thesaurus
As the manual creation of lexical
knowledge bases is typically an extensive and
time-consuming task, there are several works
where lexico-semantic relations are extracted
automatically from text, and then used either
to create new knowledge bases from scratch
or to enrich existing knowledge bases. Due to
their structure, dictionaries are an obvious
target for the extraction of lexico-semantic
relations (see, for example,
For Portuguese, in the latest years,
semantic relations have also been a subject
of increasing research interest.
Still, to the best of our knowledge, no research has been published on the automatic generation of cross-language semantic relations by using a linguistic method to map syntactic and semantically related words. This method can be extended to the type of relations that set equivalence between a word and a multiword unit (e.g. take a look = look ), with a relative clause (that was corrected = corrected ), with complex compounds (bottle made of plastic = plastic bottle) or even with a more complex construction, such as a possessive construction or a passive, by exploiting the morpho-syntactic and semantic relations pairs described in the dictionaries. The method has the advantage of being systematic, expandable, holding an 1See http://mwnpt.di.fc.ul.pt/ unlimited possibility to grow and improve in observance of natural language complexity and compliant to distinct languages and across languages. This is the novel aspect of the work presented in this paper in relation to the state of the art. 3
In this section, we will describe the English and Portuguese resources used to achieve cross-language semantic relations.
Eng4NooJ and Port4NooJ
The Port4NooJ resources are publicly
available2 and, at the moment, are being
used in tools such as Corp´ografo, a
corpora tool
3eSPERTo (in Portuguese, stands for Sistema de
Parafraseamento para Edica~o e Revis~ao de Texto).
It is a derivative of ReEscreve, proposed by Barreiro
(2008a), and also described in (Barreiro and Cabral,
2009). The English version of eSPERTo is called
SPIDER, standing for a System of Paraphrasing In
Document Editing and Revision (formerly ReWriter).
SPIDER uses Eng4NooJ resources and is described in
Eng4NooJ and Port4NooJ lexica were inherited from the OpenLogos system and enhanced with several new properties, which will be described in detail in Section 5.
The OpenLogos lexical entries are classified with more than 1,000 distinct categories, based on a taxonomy called SAL (Semanticosyntactic Abstraction Language)4. In the OpenLogos model, SAL is a meta-language that represents natural language, in effect, an ontology that represents things, ideas, relationships, dispositions, conditions, processes, etc., as well as the elements of grammar such as articles, prepositions, conjunctions, etc. In terms of natural language processing, the meta-language represents both syntax and semantics. SAL is an actual language, not a set of linguistic markers or primitives. This implies that natural language can be readily mapped to SAL. The granularity of the representational ontology is sufficient for translation purposes only, i.e., the ontology does not need to be especially fine-grained.
SAL elements are divided in a hierarchical scheme of supersets, sets and subsets, distributed by all parts-of-speech. SAL comprises 12 supersets for nouns: Concrete (CO), Mass (MA), Animate (AN), Place (PL), Information (IN), Abstract (AB), Process intransitive (PI), Process transitive (PT), Measure (ME), Time (TI), Aspective (AS), and Unknown (UN). For example, the concrete nouns superset consists of countable physical things, either man-made or natural, including parts of the human body. Concrete (count5) contain both sets and subsets. The principal sets of concrete nouns are functional things and agentive things. Other sets are: natural things (COnat); impulses/lights (COlight); marks/blemishes 4The full description of the multiple SAL categories can be found at the Logos System Archives (http://logossystemarchives.homestead.com/) and all the resources (and descriptions) are downloadable from OpenLogos website at DFKI (http://logosos.dfki.de/).
5Concrete nouns are always count nouns and, unless in the plural, generally cannot occur without a preceding article or quanti er. For example: Computers are effective. *Computer is effective. (COblem); edibles non-mass (COednm); edibles/color (COedcol); classifiers (COclass); amorphous (COamorph); and atomistic (COatom). For example, the set of natural things (COnat) includes subsets such as: minute flora (COflora) (e.g. algae, spore); plants (COplant) (e.g. rose, weed ); trees (COtree) (e.g. apple, willow ); trees/wood (COtrwd) (e.g. oak, maple); and miscellaneous natural things (COmnat) (e.g. pebble, iceberg ).
The SAL meta-language is semanticosyntactic in nature, representing natural language at a second-order abstractions (common nouns are first-order abstractions). Syntax and semantics are seen as a continuum. This semantico-syntactic continuum is always taken into account when classifying each lexical entry within SAL. The classification was done through the years by trial and error. For example, when classifying elements into the functional (COfunc) or agentive (COagen) of the concrete noun superset, the following reasoning is taken into consideration: functional things tend to be passive, i.e. typically do not act of their own accord and generally require an agent to use them. Hence, they are more instrumental in nature. Agents typically do work in and of themselves. This distinction may sometimes seem arbitrary. For example, hinge is a fastener under functional things and clearly does work of itself, but is not coded as an agent. Airplane, on the other hand, obviously does require an agent and yet is coded under agentives as a vehicle. As a rule, agentives have a source of power or energy in themselves, while functionals do not. Parts of the human/animal body are also classified as concrete. Words like heart, brain, digestive tract, stomach, and organs in general are machines/systems under agentives. Words like teeth, ngernail, toes, lips, tendons, ligaments, bones, etc. belong to various subsets under functionals.
SAL categories contain domainindependent ontological (lexical-contextual) and semantico-syntactic relations (the same word form can be mapped to different concepts) are assigned to general language words or domain-specific terms. The general language dictionary contains many lexical entries which are broadly classified, which could be considered to pertain to a more specific domain. For example, the lexical entries dog IS HYPONYM OF animal ca˜o E HIPONIMO DE animal dog IS HYPONYM OF mammal ca˜o E HIPONIMO DE mam´ıfero dog IS HYPONYM OF non-human being ca˜o E HIPONIMO DE ser n˜ao humano dog IS HYPONYM OF invertebrate ca˜o E HIPONIMO DE ser vertebrado dog IS HYPONYM OF animate being ca˜o E HIPONIMO DE ser vivo/animado for HIV (immunology), manic-depressive disorder, bipolar disorder (mental health) and asthma (pulmonology) are all classified under the superset Abstract and subset State (also for conditions and relationships). This subset corresponds to abstract nouns that describe something about a thing or person that is not inherent to its nature (e.g. cancer, coma, circumstance, condition, disease, fatherhood, inequality, insolvency, loneliness, parity, poverty, status). Being more extrinsic, these states, conditions or relationships could conceivably change without altering the nature of the thing or person. This is not a strict rule but is indicative of the difference between this subset and the properties/qualities/nature subset.
The information noun superset is comprised of nouns that denote data, information, or knowledge, which might be considered more specific to certain domains. But, this category also includes the medium on which the information is recorded, represented or communicated; i.e., spoken, written, dramatized, sung, etc. Table 1 presents a list of terms classified as Instructional/legal (INinst) under the information noun superset (IN). 4
Both in Eng4NooJ and Port4NooJ, each lexical entry is described with semanticosyntactic properties, which represent relations between words or expressions. These relations can be synonymy, hyponymy, actionof, result-of, process-of, made-of, propertyof, member-of, among others. Table 2 illustrates several semantic relations for the concrete English and Portuguese nouns dog and ca~o, respectively. These relations were inferred from the SAL hierarchical categories. abolishment IS ACTION OF abolish aboli¸c˜ao E´ AC¸ A˜ O DE abolir abuse IS ACTION OF abuse abuso E´ AC¸ A˜O DE abusar happening IS ACTION OF happen acontecimento E´ AC¸ A˜ O DE acontecer agreement IS ACTION OF agree acordo E´ AC¸ A˜ O DE acordar lit IS RESULT OF light aceso E´ RESULTADO DE acender stuffed IS RESULT OF stuff embalsamado E´ RESULTADO DE embalsamar rotten IS RESULT OF rotten podre E´ RESULTADO DE apodrecer interdicted IS RESULT OF interdict interditado E´ RESULTADO DE interditar
In addition to the taxonomical classification inherited from OpenLogos, which allowed the establishment of hyponymy relations, both Eng4NooJ and Port4NooJ resources include regular derivational, morphosyntactic and semantic relations, such as synonymy, action-of, and result-of. The morphosyntactic and semantic relations are established between words of a different part-ofspeech, as for example, between an adjective and its derived adverb (e.g. quick > quickly - rapido > rapidamente), between a noun and an adjective (e.g. enthusiasm > enthusiastic - entusiasmo > entusiasmado), or between a noun and an adverb (e.g. imagination > imaginatively = with imagination - imaginaca~o > imaginativamente = com imaginac~ao).
Table 3 illustrates action-of and result-of semantic relations. Action-of relations are established between a noun and a verb, where the noun is a morphological derivation of the verb. Result-of relations are established between an adjective and a verb, where the adjective is morphologically derived from the verb. 5
In order to obtain hyponymy relations from the OpenLogos properties in Port4NooJ and Eng4NooJ dictionaries, we created a local grammar that matches on the SAL code and presents, as an output, one or more words from the description of that specific SAL code. For the examples in Table 1, the NooJ local grammar recognizes the property [SAL=ANmamm], standing for Aniintimac~ao,N+FLX=CANC A~O+INinst+EN=summons arrendamento,N+FLX=ANO+INinst+EN=lease autorizac~ao,N+FLX=CANC A~O+INinst+EN=fiat autorizac~ao,N+FLX=CANC A~O+INinst+EN=license autorizac~ao,N+FLX=CANC A~O+INinst+EN=permit autorizac~ao,N+FLX=CANC A~O+INinst+EN=warrant c^anone,N+FLX=ANO+INinst+EN=canon clausula,N+FLX=CASA+INinst+EN=clause condic~ao,N+FLX=CANC A~O+INinst+EN=proviso contrato,N+FLX=ANO+INinst+EN=contract credo,N+FLX=ANO+INinst+EN=credo declarac~ao,N+FLX=CANC A~O+INinst+EN=affidavit decreto,N+FLX=ANO+INinst+EN=decree diretiva,N+FLX=CASA+INinst+EN=guideline estatuto,N+FLX=ANO+INinst+EN=bylaw estatuto,N+FLX=ANO+INinst+EN=statute garantia,N+FLX=CASA+INinst+EN=guarantee garantia,N+FLX=CASA+INinst+EN=warranty lei,N+FLX=CASA+INinst+EN=law licenca,N+FLX=CASA+INinst+EN=license mandato,N+FLX=ANO+INinst+EN=mandate moratoria,N+FLX=CASA+INinst+EN=moratorium norma,N+FLX=CASA+INinst+EN=norm norma,N+FLX=CASA+INinst+EN=standard ordem,N+FLX=MARGEM+INinst+EN=order ordem,N+FLX=MARGEM+INinst+EN=ordinance pacto,N+FLX=ANO+INinst+EN=pact patente,N+FLX=CASA+INinst+EN=patent renuncia,N+FLX=CASA+INinst+EN=waiver testamento,N+FLX=ANO+INinst+EN=will tratado,N+FLX=ANO+INinst+EN=treaty veredicto,N+FLX=ANO+INinst+EN=veredict mate, Mammal and retrieves, as its output, words that will be used as hypernyms of the words dog or c~ao, in English or Portuguese, respectively. These words are: animal, mammal, non-human being, invertebrate, animate being. If the description of the SAL category included more hypernyms, these could, of course, be easily added to the list of pairs of the semantic relation IS HYPONYM OF for dog /ca~o.
Table 4 shows distinct types of dictionary entries with implicit semantic information, namely the support verb construction that can be synonymous to a verb entry (impressionar = causar impress~ao – impress = make an impression; ficar azedo = azedar – turn sour = sour ), the semantic relation between an adjective and a semantically related adverb (aesthetic – aesthetically ), and the semantic relation between a noun and a semantically related adverb (skepticism – skeptically ). These relations are established by means of grammar rules. We have focused on the most regular rules, which are the ones that allow transformation of part-of-speech through the process of derivation.
In the examples illustrated in Table 4, the properties in bold correspond to the derivational rule and inflectional paradigm. Accordingly, DRV=NDRV01:CAN C¸A˜ O is a dictionary property that calls the rule to derive (through the process of nominalization) the predicate noun impress~ao (impression) from the verb impressionar (impress) and assigns it the inflectional paradigm CANC¸ A˜O (the noun impress~ao inflects in the same way as the noun canc~ao; i.e., following the same process and using the same morphemes to form the plural, etc.); DRV=ADRV00:ALTO is a dictionary property that calls the rule to derive the predicate adjective azedo (sour ) from the verb azedar (sour ) and assigns it the inflectional paradigm ALTO (the adjective azedo inflects like the adjective alto).
DRV=AVDRV03 is a dictionary property that calls the rule to derive the adverb aesthetically from the adjective aesthetic; and, finally, DRV=NAVDRV02 is a dictionary property that calls the rule to derive the adverb skeptically from the noun skepticism.
The lexical entries for the verbs impressionar (impress), adaptar (adapt ), azedar (sour ), have the property VSUP, that is, the description of the support verb that occurs with the predicate nouns impressa~o (impression), adaptac~ao (adapt ) and with the predicate adjective azedo (sour ), which derive from the corresponding cited verbs. The combination of the description in the properties VSUP and DRV allows the semantic association between these verbs and their equivalent support verb constructions, namely fazer /causar impress~ao (make/cause impression), fazer adaptac~ao (make adaptation), and car azedo (turn sour ).
Table 5 shows the transformational rules to associate morpho-syntactic and semantically related words of different partsof-speech, extracted individually from the Eng4NooJ and Port4NooJ rule databases.
Rules are indexed according to different types of transformation. NDRV transforms verbs into nouns, ADRV transforms verbs into adjectives, and AVDRV transforms adjectives into nouns. The rules of each type are numbered. For example, NDRV04 is the rule number 04 that transforms a verb into a noun. The slash (/) after each ending inimpressionar,V+FLX=FALAR+SAL=PVPCpleasetype+EN=impress+VSUP=fazer+VSUP=causar+DRV=NDRV01:CANCA~O adaptar,V+FLX=FALAR+Aux=1+INOP57+Subset=132+EN=adapt+VSUP=fazer+DRV=NDRV00:CANCA~O azedar,V+FLX=LIMPAR+Aux=1+OBJTRundif98+Subset=740+EN=sour+VSUP=ficar+DRV=ADRV00:ALTO aesthetic,AFLX=NATURAL+SAL=AVstate+PT=estetico+DRV=AVDRV03 skepticism,N+FLX=BOOK+SAL=ABcause+PT=cepticismo+DRV=NAVDRV02 troduces the part-of-speech of the derived word. The plus sign (+) introduces information about a specific noun or adjective. For example, Npred and Apred stand for predicate noun and predicate adjective, respectively. The capital letters between the lessthan and the greater-than signs (<, >) correspond to commands. The command <B> means “backspace one character and add the string that follows the command, assigning it a new part-of-speech”. The command <B2> means “delete the last two characters of the word from which the new word derives and add the string that follows the command”, and so on and so forth. The strings that follow a command are the endings of the new generated words (e.g. -ion for the noun acceleration, -tically for the adverb realistically, etc.). The command <E> means that no character needs to be deleted. The command <A> means “delete the acute accent in the word from which the new word derives”.
Eng4NooJ and Port4NooJ grammars are the devices used to recognize words or expressions and generate new ones, paraphrase or translate them. For example, the grammar in Figure 1, is used to recognize adverbial compounds in Portuguese and transform them into equivalent single adverbs. This grammar transforms multiword adverbs such as de (um) modo rapido (in a fast/quick way ) into single adverbs such as rapidamente (quickly ).
This type of transformation is allowed by operations like the one represented in the first path of the graph. The box calls a new graph to recognize the strings de (um) modo, de (uma) forma/maneira (in a (ADJ) way ), which make up the multiword adverbial. The output $A ADV retrieves the adverb that is linked to the adjective $A. The adjective is transformed in the equivalent adverb by means of the derivational rules. The same grammar also recognizes multiword adverbs whose head is a noun, such as por acidente (by accident ) or com entusiasmo (with enthusiasm), following the second and third paths.
The grammar in Figure 1 is monolingual, because there is no specification of the output for a different language. However, both Eng4NooJ and Port4NooJ resources contain Portuguese and English transfers for each lexical entry, i.e., they are in fact bilingual dictionaries. As a result, any grammar used to obtain monolingual transformations can be reused to generate bilingual (or multilingual) transformations. That is, the same grammar can be used to retrieve the output in English or in any other language (separately or together) as long as the words of that language are in the bilingual or multilingual dictionary and there are rules associated to the relevant dictionary properties. This means that, the grammar can generate translations from one to many languages, i.e., it can be used to create cross-language semantic relations.
For monolingual transformations, no output language is specified. For bilingual or crosslanguage transformations, the parameter for the specification of the output language needs to be added. The parameter $EN for English, $IT for Italian, $SP for Spanish, etc. specifies the retrieval of the output in one of these languages or in all of them simultaneously. Similarly, the grammar presented in Figure 2, can be used for cross-language semantic relations. This grammar matches on a support verb construction of the type [Predicate Noun Construction] (dar um abraco (a) – give a hug (to)) (in the figure represented in a box that calls a sub-graph) and paraphrases it into a single verb (abracar – hug ). Eng4NooJ NDRV04 = <B>ion/Npred e.g. accelerate > acceleration ADRV02 = <B>icable/ADJ e.g. apply > applicable AVDRV01 = <E>ly/ADV e.g. frequent > frequently AVDRV04 = <B>tically/ADV e.g. realism > realistically
Port4NooJ NDRV02 = <B>nca/N+Npred e.g. mudar > mudanc¸a ADRV02 = <B2>o/A+Apred e.g. azedar > azedo AVDRV00 = <B>zmente/ADV e.g. veloz > velozmente AVDRV05 = <A> <B>amente/ADV e.g. r´apido > rapidamente
Figure 3 illustrates the output of a grammar that generates cross-language semantic relations between Portuguese support verb constructions and English single verbs. At present, the semantic relations included in Eng4NooJ and Port4NooJ are mostly used to generate paraphrases and integrated in the paraphrasing tools SPIDER and eSPERTo.
However, cross-language relations such as those illustrated in Figure 3 can be used directly in machine translation and are fuelling the ParaMT bilingual paraphrasing tool. At the current stage of development, ParaMT translates mostly multiword units, performing well in the translation of Portuguese support verb constructions into English verbs, and vice-versa, the linguistic phenomena most researched when applying the current methodology.
Relation Hyponymy Synonymy between nouns between verbs between adjectives between adverbs Action-of Result-of In theory, the exploitation of the lexicon in combination with SAL allows the establishment of numerous relations between words and expressions. For the current paper, we focused only on a few of those relations which cover a larger number of items and could be extracted in a simple and easy way.
The result of extraction for Portuguese (not yet reviewed) is publicly available6. Currently, Port4NooJ contains more than 30,000 morpho-syntactic relations between semantically related elements. Table 6 presents some preliminary results, which do not refer to paraphrasing capabilities, but simply to relations between lexical items. The total results for paraphrasing are significantly higher. Local grammars, applied to information (properties) described in the dictionary, enable the recognition and analysis of expressions such as de (um) modo rapido, de (uma) forma/maneira rapida (in a fast/quick way ) (which could be considered as relations between an adjective and an adverb, but which were not counted), and also inflected forms such as dar uns passeios (go for some walks ), etc.
Port4NooJ contains approximately 600 derivational rules, most of them transforming verbs into predicate nouns (587). 119 of
6See http://www.linguateca.pt/Repositorio/ Port4NooJ/relacoes semanticas explicitas/ these rules are productive, covering nominalizations. 486 rules correspond to verb relations between verbs and autonomous predicate nouns. At this point in the research, rules were only superficially evaluated. 7
This paper presented semantic relations, namely domain-independent semanticosyntactic and ontological relations, suitable for paraphrasing and cross-language tasks, including machine translation. We have demonstrated that given the appropriate linguistic resources, the generation of semantic relations can become very systematic. Any grammar to generate monolingual semantic relations can be reused to generate crosslanguage relations, rules can be standardized and often re-used across close languages, etc. Even though the methodology adopted was applied to the OpenLogos resources, it is compliant with the exploitation of other lexical resources with semantic relations, for any language besides English and Portuguese, studied in this research.
Future work would gather and combine open source available semantic resources, enhance properties on the existing resources, and enlarge the linguistic phenomena coverage. Intelligence Research Society Conference (FLAIRS), pages 360–364.