File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/04/w04-0904_metho.xml
Size: 26,570 bytes
Last Modified: 2025-10-06 14:09:09
<?xml version="1.0" standalone="yes"?> <Paper uid="W04-0904"> <Title>OntoSem and SIMPLE: Two Multi-Lingual World Views</Title> <Section position="3" start_page="0" end_page="1" type="metho"> <SectionTitle> 2 Overview of SIMPLE </SectionTitle> <Paragraph position="0"> The SIMPLE project is developing 10K-sense &quot;harmonised&quot; semantic lexicons for 12 European Union languages (Catalan, Danish, Dutch, English, Finnish, French, German, Greek, Italian, Portuguese, Spanish, Swedish), continuing the earlier PAROLE project, which developed 20Ksense morphological and syntactic lexicons for these languages. The lexicons are monolingual and are developed independently, with the word stock based on corpus evidence for each language. To ensure some overlap of lexical senses, certain Base Concepts of EuroWordNet must be covered in each language (462 nominal, 187 verbal and 185 adjectival Base Concepts that were culled and cleaned from EuroWordNet). This overlap will permit direct interlinking among languages; interlinking of the rest of the lexical stock is slated as future work. Pustejovsky's four Qualia (which are, essentially, properties expressing formal, agentive, constitutive and telic meanings; see Pustejovsky 1995) are used to specify certain aspects of word meaning, and a common library of 140 template types is used to guide acquisition in all languages (Lenci et al 2000a, 2000b).</Paragraph> <Paragraph position="1"> Lenci et al. 2000b (p. 5) summarize the information that can be represented in a SIMPLE lexicon entry: &quot;i) semantic type, corresponding to the template the SemU (semantic unit) instantiates; ii) domain information; iii) lexicographic gloss; iv) argument structure for predicative SemUs; v) selectional restrictions on the arguments; vi) event type, to characterise the aspectual properties of verbal predicates; vii) link of the arguments to the syntactic subcategorization frames, as represented in the PAROLE lexicons; viii) Qualia Structure; ix) information about regular polysemous alternation in which a word sense may enter; x) cross-part-of-speech relations (e.g. intelligent intelligence; writer - to write); xi) synonymy.&quot; Below is the SemU for a sense of lancet, instantiating the template Instrument (from Palmer et al. 2000).</Paragraph> <Paragraph position="2"> While the SIMPLE project is certainly producing useful resources, we would suggest that the lexical information and structure are being overly constrained by the frameworks selected, which we will comment on briefly in preparation for an extended comparison between OntoSem and SIMPLE in section 4.</Paragraph> <Paragraph position="3"> EuroWordNet is being used as the anchor for semantic description in SIMPLE. However, like the original English WordNet, it is not a propertyrich ontology but, rather, a hierarchical net of lexical items whose use in NLP has the same pitfalls as any non-ontological word net (e.g., lack of disambiguating power and lack of sufficient relations between entities; see Nirenburg 2004c for a discussion of the insufficiency of WordNet for NLP). In order to make up for the sparsity of information in the semantic substrate, the SIMPLE lexicons contain what would, we believe, be more efficiently recorded in a single, sufficient ontology. For example, when the lexicon acquirers for each language use the Instrument template to describe lancet, they must rerecord in the lexicon of each L all of the language-independent property values for this lexical item, like the values for the four Qualia (formal, agentive, constitutive, telic), the domain, the unification path, etc. This is significant redundancy and, moreover, there is no guarantee that acquirers will arrive at the same decisions, either through error, oversight or competing analyses of the phenomena in question.</Paragraph> <Paragraph position="4"> Another, in our view, insufficiently explained aspect of SIMPLE is the priority given to Qualia as descriptors of lexical items. The original inventory of Qualia (from Pustejovsky 1995) consists of only four properties of the hundreds that can usefully be used to link concepts for purposes of NLP. Lenci et al. (2000b) address this issue as follows: &quot;Although they [the four Qualia] clearly do not exhaust the semantic content of lexical items, Pustejovsky (1995) has convincingly shown that these four Qualia dimensions play a particularly prominent role in determining the linguistic behavior of word senses, as well as in the explanation of the generative mechanisms at the basis of lexical creativity. Qualia-based information can be specified for all the parts of speech, although prima facie it seems to be more directly suitable for the characterization of certain types of nominals&quot;. However, there is large gap between theoretical interest and practical application: in fact, because of this, the SIMPLE project has moved toward an Extended Qualia Structure with more fine-grained subtypes of given Qualia.</Paragraph> <Paragraph position="5"> In conclusion, we believe that SIMPLE is pursuing useful goals that could be pursued in even more useful ways by shifting the focus from lexicon-only work to integrated work within an environment in which ontological and lexical resources are developed together and where extant types of processors can be used to test the value of resources as they are developed.</Paragraph> </Section> <Section position="4" start_page="1" end_page="3" type="metho"> <SectionTitle> 3 Overview of OntoSem </SectionTitle> <Paragraph position="0"> The OntoSem approach to lexicon and ontology acquisition differs from that used in SIMPLE As an aside, we see a parallel between focusing on Qualia in lexical description and, for example, focusing on classes of verbs with respect to their alternations, as is done, for example, in Levin 1995. While the descriptions that derive from theoretically-driven research such as this can certainly be useful, when it comes to writing &quot;well-rounded&quot; semantic descriptions of words for large-scale systems, there is no distinction between a Quale and other properties. Similarly, the fact that a verb belongs to some group with respect to alternations is no more or less important than its other potential group membership along other parameters. See Nirenburg and Raskin 2004 for further discussion of these and related issues.</Paragraph> <Paragraph position="1"> because OntoSem is an integrated text processing environment, meaning that knowledge resources are crafted hand-in-hand with each other and with processors such that responsibility for various analysis tasks can be distributed in an ideal (to the degree of our understanding) way.</Paragraph> <Paragraph position="2"> OntoSem takes as input unrestricted raw text and carries out preprocessing, morphological analysis, syntactic analysis and semantic analysis, with the results of semantic analysis represented as formal text-meaning representations (TMRs) that can then be used as the basis for a wide variety of NLP applications. Text analysis relies on: * The OntoSem language-independent ontology, which is written using a metalanguage of description and currently contains around 5,500 concepts, each of which is described by an average of 16 properties. In all, the ontology contains hundreds of properties (which cover the same territory as the Qualia plus much more).</Paragraph> <Paragraph position="3"> Fillers for properties can be other ontological concepts or literals.</Paragraph> <Paragraph position="4"> * An OntoSem lexicon for each language processed, which contains syntactic and semantic zones (linked using variables) as well as calls to &quot;meaning procedures&quot; (i.e., programs that carry out procedural semantics, see McShane et al.</Paragraph> <Paragraph position="5"> 2004a) when applicable. The semantic zone most frequently refers to ontological concepts, either directly or with property-based modifications, but can also describe word meaning extraontologically, for example, in terms of modality, aspect, time, etc. The current English lexicon contains approximately 12K senses, including all closed-class items and the most frequent verbs, as indicated by corpus analysis. This English lexicon took less than 1 person year to build and can (as described below) be ported to other languages.</Paragraph> <Paragraph position="6"> preprocessing, syntactic analysis, semantic analysis, and creation of TMRs. They are largely parameterizable and thus can be ported to other languages.</Paragraph> <Paragraph position="7"> * The TMR language, which is the metalanguage for representing text meaning. A very simple example of a TMR (simple because most of the sentences we process are much longer), which reflects the meaning of the sentence He asked the UN to authorize the war, is as follows: Details of this approach to text processing can be found, e.g., in Nirenburg et al. 2004a,b. The ontology itself, a brief ontology tutorial, and an extensive lexicon tutorial can be viewed at http://ilit.umbc.edu.</Paragraph> <Paragraph position="8"> OntoSem has been used with languages including English, Spanish, Chinese, Arabic and Persian, to varying degrees of lexical coverage (e.g., earlier, less fine-grained English and Spanish lexicons contained 40K entries and were used for MT in the Mikrokosmos project). What makes OntoSem amenable to efficient cross-linguistic usage is that many of the resources are either fully language independent (the ontology, the fact repository, the TMR metalanguage) or parameterizable in well understood ways. Here we focus on exploiting cross-linguistic similarity for lexical acquisition, but a similar analysis could be applied to the OntoSem analyzers.</Paragraph> <Paragraph position="9"> 3. 1 OntoSem Lexicons A basic verbal lexicon entry in OntoSem looks as follows (in presentation format): The syntactic structure (syn-struc) says that this is a transitive sense of watch and the semantic structure (sem-struc) says that a VOLUNTARY-VISUAL-EVENT - which is a concept in our ontology - must be instantiated in the TMR. The variables are used for linking, so, for example, the syntactic subject is linked to the meaning of the AGENT of the VOLUNTARY-VISUAL-EVENT (^ is read 'the meaning of').</Paragraph> <Paragraph position="10"> Apart from mapping directly to an ontological concept, there are many other - and more complex - ways to express meaning in OntoSem. For example, one can map to an ontological concept with modified property values: e.g., * Zionist is described as a POLITICAL-ROLE that is the AGENT-OF a SUPPORT event whose THEME is Israel.</Paragraph> <Paragraph position="11"> * asphalt (v.) is described as a COVER event whose INSTRUMENT is ASPHALT.</Paragraph> <Paragraph position="12"> * recall (v. as in they recalled the high chairs) is described as a RETURN-OBJECT event that is CAUSED-BY a FOR-PROFIT-CORPORATION and whose THEME is ARTIFACT, INGESTIBLE or MATERIAL.</Paragraph> <Paragraph position="13"> There are also a number of fully or partially non-ontological ways of describing meaning, like the use of parametric values of mood or aspect. For example, the auxiliary might as in He might come over is described using the modality 'epistemic', which deals with the truth value of a statement: This meaning procedure reassigns a case-role if the listed AGENT case-role is inappropriate considering the meaning of $var1 and/or $var2: e.g., in the truck might come, truck is a THEME of a MOTION-EVENT, not an Another set of extra-ontological semantic descriptors is used for time expressions, as shown by the example of yesterday below.</Paragraph> <Paragraph position="14"> As already shown in the examples of might and yesterday, calls to procedural semantic routines (which may or may not be listed in the meaningprocedure zone of the lexicon entry) are used widely in OntoSem lexical description. This reflects the fact that many aspects of meaning cannot be statically described but, rather, must be computed. An advantage of developing lexical resources within a processing environment is being able to assign responsibility for portions of semantic composition to resources best suited for them.</Paragraph> <Paragraph position="15"> In addition to the means of lexical expression described above, OntoSem lexicon entries can include entities of any degree of complexity, including phrasals of any profile, as reported in McShane et al. 2004b.</Paragraph> <Section position="1" start_page="2" end_page="2" type="sub_section"> <SectionTitle> 3.2 Porting OntoSem Lexicon Entries Across Languages </SectionTitle> <Paragraph position="0"> As is clear from the examples above, OntoSem provides significant expressive power semantically (not to mention syntactically, which we do not pursue here). Expressive means include mapping to the ontology (which itself is rich in property-value descriptors), mapping to the ontology with lexical supplementation of properties, or referring to extra-ontological microtheories like those that treat time, reference resolution, comparison, ellipsis resolution, modality, aspect, etc. What must be emphasized, however, is how language neutral - and therefore portable across languages the semantic descriptions are. Whereas it is typical to assume that lexicons are language-specific whereas ontologies are language-independent, most aspects of OntoSem sem-strucs are languageindependent, apart from the linking of specific variables to their counterparts in the syn-struc.</Paragraph> <Paragraph position="1"> AGENT, and in I might get sick, I am an EXPERIENCER of a DISEASE event, not an AGENT of it.</Paragraph> <Paragraph position="2"> Stated differently, if we consider sem-strucs - no matter what lexicon they originate from - to be building blocks of the representation of word meaning (as opposed to concept meaning, as is done in the ontology), then the job of writing a lexicon for L2 based on the lexicon for L1 is in large part limited to a) providing an L2 translation for the head word(s), b) making any necessary syn-struc adjustments and c) checking/modifying the linking among variables in the syn- and sem-strucs.</Paragraph> <Paragraph position="3"> This conception of cross-linguistic lexicon development derives in large part from the Principle of Practical Effability (Nirenburg and Raskin 2004), which states that what can be expressed in one language can somehow be expressed in all other languages, be it by a word, a phrase, etc.</Paragraph> <Paragraph position="4"> Apart from this theoretical justification for conceptualizing the sem-strucs as building blocks for lexical representation, there are two practical rationales: supporting consistency of meaning representation across languages and using acquirer time most efficiently in large-scale lexical acquisition.</Paragraph> <Paragraph position="5"> As regards consistency, the potential for paraphrase must be considered when building multi-lingual resources. For instance, 'weapons of mass destruction' can be described as the union of CHEMICAL-WEAPON and BIOLOGICAL-WEAPON, or it can be described as WEAPON with the ability to KILL > 10,000 HUMANs (the actual number recorded will be treated by the analyzer in a fuzzy fashion; however, it would be less than ideal for a lexicon for L2 to record 10,000 while a lexicon for L3 recorded 25,000). While both representations are valid, it is desirable to use the same one in all languages covered. In addition, the decision of how to describe a notion - whether by ontologizing it, describing it using extra-ontological means, describing it using an existing concept with additional properties and values defined - is often a judgment call. It would not be desirable for the acquirer of German to map the word Schimmel 'white horse' to the concept HORSE with the lexical restriction COLOR: WHITE, while the acquirer of some other language that also has a word for 'white horse' introduced an ontological concept specifically for this entity. Again, while both representations are valid and, in this case, semantically equivalent, the general tendency should be to strive toward uniformity where possible.</Paragraph> <Paragraph position="6"> As concerns acquirer time, composing sem-strucs is, by far, the most time- and effort-intensive aspect of writing OntoSem lexicon entries. This derives from the wealth of expressive means; the fact that microtheories of time, reference, etc., are naturally built during lexicon development (recall that our environment is fully integrated with processors); and the fact that ontology development occurs hand-in-hand with lexicon development. Therefore, work on the first lexicon entry that describes a word sense - regardless of the language of origin - takes much more time than editing a word sense for a new language.</Paragraph> <Paragraph position="7"> Moreover, although in the worst case some editing of entries is necessary for L2, L3, etc., in most cases no such editing is needed. Although one might hypothesize this state of affairs based on cross-linguistic principles, we have tested it in the lexicon-porting experiment described below.</Paragraph> </Section> <Section position="2" start_page="2" end_page="3" type="sub_section"> <SectionTitle> 3.3 An English to Polish Lexicon Porting Experiment </SectionTitle> <Paragraph position="0"> For the experiment, a bilingual English/Polish computational linguist took the English OntoSem lexicon as a seed and experimented with various porting methods into Polish.</Paragraph> <Paragraph position="1"> The primary insight was that while manually porting individual lexical senses is quite straightforward and will save time over acquisition from scratch, porting lexicons wholesale is rather more complex. That is, manually providing translations for the senses in L1 is a conceptually relatively simple task, complicated only by the need for the occasional remapping of variables, editing of syntactic structures, omission of given senses due to language lacunae (e.g., a phrasal encoded in L1 might not occur in L2 in a fixed form), etc. However, if one attempts either to (semi-)automate the acquisition process and/or use L1 as a seed lexicon for more &quot;creative&quot; acquisition of L2, the space of options becomes quite broad and must be constrained programmatically in order to actually benefit from the reuse of semantic descriptions.</Paragraph> <Paragraph position="2"> For example, if a well-trained acquirer of L2 is using L1 as a seed, questions that arise include: Should the base lexicon be left as is (considering that it is known to have incomplete coverage) or should one attempt to improve its quality and coverage while building L2? Should L2 acquisition be driven by correspondences in head words or simply by the content of sem-struc zones (e.g., all English senses of table will be in one head entry, and typically will be acquired at once; should all senses of all L2 translations of table be handled at once during L2 acquisition or should the L2 acquirer wait until he comes upon sem-strucs that represent the given other meanings of the L2 words)? To what extent should the regular acquisition process - including ontology supplementation - be carried out on L2? The answers to all of these, and more such, questions depend entirely upon available resources and should be informed by (a) experiments to determine what works best for a given acquirer, and (b) the goals of a given project.</Paragraph> <Paragraph position="3"> As regards automation, the experiment found that automatically mapping L2 words to L1 OntoSem entries works very well (at well over 90%) when the machine-tractable L1-L2 resource used to support this process has one sense of the given word in the given part of speech and the OntoSem lexicon also has one sense for the given part of speech. The extraction and matching of such senses represents a well-defined, extremely time-efficient task, especially for specialized terminology that tends to have only one sense in any language. When the mapping between senses in the L1-L2 lexicon and the OntoSem lexicon is more than one to one, manual linking of senses (which do not always correspond among the languages) has proved necessary, with the potential benefits of a time-saving interface becoming immediately clear.</Paragraph> </Section> </Section> <Section position="5" start_page="3" end_page="3" type="metho"> <SectionTitle> 4 Pudding in SIMPLE and OntoSem </SectionTitle> <Paragraph position="0"> Now we return to the comparison between SIMPLE and OntoSem. We use the example of pudding, which is cited in numerous documents related to SIMPLE. The Qualia (in italics) and their values (in boldface) for this word are: formal - substance; constitutive - ingredients; telic eat; agentive - make. The stated rationale for encoding these qualia values in SIMPLE lexicon entries is that they are needed to understand the semantics of the sentences like the following (from Lenci et al. 2000b): a) John refused the pudding (= refused to eat: telic); b) That's an easy pudding (= easy to make: agentive); c) There is pudding on the floor (= substance: formal); d) The pudding came out well (= has been made well: agentive); e) That was a nice bread pudding (= made of/ingredient: constitutive) We would suggest, as before, that the lexicon is not the best place for this information and, further, that this information is incomplete. For comparison, we present our approach to describing and processing pudding in the OntoSem environment. Since OntoSem uses a full ontology (not a word net), the ontological specification of the concept PUDDING contains much of the needed Some automation of the mapping between L1-L2 multi-sense words is possible as demonstrated by Pianta, et al. 2002, but the results still require intensive manual work by an acquirer.</Paragraph> <Paragraph position="1"> information for processing all the above sentences containing pudding. Moreover, since the OntoSem ontology, lexicons and processors are developed together, their known mutual contributions drive resource acquisition. Obviously, one cannot expect the same approaches to be used in a lexicon-only project like SIMPLE. However, a non-trivial question, considering the expense of manual resource acquisition, is to what extent should we be developing resources separately from processors that can use them, especially when the nature of processors crucially affects what is needed of knowledge resources? Below is a subset (for reasons of space) of the properties and values for the concept PUDDING in the OntoSem ontology; the first 4 are locally specified while the others are inherited.</Paragraph> <Paragraph position="2"> Since all of the necessary information about PUDDING is encoded in the ontology, the OntoSem lexicon entry for pudding need only contain a direct link to the concept.</Paragraph> <Paragraph position="3"> The analysis of sentences (a)-(e) in OntoSem is carried out as follows. For (a), there is a lexical sense of refuse that expects an OBJECT (not an EVENT, as in the main sense) as its direct object. This sense expects the semantic ellipsis of a verb and, as such, is supplemented with a meaning procedure called 'seek-specification', which searches for the elided event. There are two sources it searches: previous TMRs, for a recent semantically viable event, and the ontology itself, for an EVENT (or EVENTs) whose default AGENT is HUMAN and default THEME is PUDDING. This search procedure in some cases returns more than one candidate event to reconstruct the semantic ellipsis. While this is not always ideal, it does reflect precisely the type of lexical ambiguity that can be resolved only by contextual clues. For example, the sentence John refused the pudding could be used in a supermarket context to describe a situation where John refused to take/accept a free box of pudding that was being pushed upon him by a promoter. The desire to be able to treat this second reading of the sentence is the reason for treating constraints in OntoSem abductively. As far as one can tell, the constraints in SIMPLE are rigid: &quot;telic = eat&quot; for pudding is a hard constraint. In fact, the example John refused the pudding is representative of a much broader class of phenomena known as semantic ellipsis, the treatment of which must be carried out by procedural semantic routines (see McShane et al.</Paragraph> <Paragraph position="4"> 2004a for details).</Paragraph> <Paragraph position="5"> Example (b) is another case that OntoSem handles through lexical and procedural semantics working in tandem. The NP easy pudding is actually a construction {a value on the scale DIFFICULTY + ARTIFACT} that is known to involve semantic ellipsis. Thus, we prepare for it in the OntoSem lexicon by associating this construction with the seek-specification meaning procedure, described above, which handles with equal efficacy easy pudding (PREPARE-FOOD), easy song (PERFORM-MUSIC), etc.</Paragraph> <Paragraph position="6"> Example (c) is handled trivially based on the fact that PUDDING is a PHYSICAL-OBJECT and, like all PHYSICAL-OBJECTs, is ontologically defined for LOCATION.</Paragraph> <Paragraph position="7"> Example (d) is analyzed using the information that PUDDING is a PREPARED-FOOD and, as such, is the THEME-OF PREPARE-FOOD, which in turn is a child of CREATE-ARTIFACT. The lexicalized phrasal {ARTIFACT + come out + a value of evaluative modality} is mapped to CREATE-ARTIFACT, with the THEME being the given ARTIFACT and the evaluative modality being concretized based on the evaluative value of the lexical item (e.g., 'well', as in 'the pudding came out well' is mapped to 'evaluative .7'). This phrasal, of course, works for any ARTIFACT and any value of evaluative modality, so lexicalizing it once is a real savings in time and effort.</Paragraph> <Paragraph position="8"> Example (e) has two possible treatments in OntoSem: on the one hand, the lexical item 'bread pudding' could (and, ultimately, should - though it is not in the OntoSem lexicon at the moment) be listed as a phrasal in the lexicon, described as PUDDING: HAS-OBJECT-AS-PART BREAD. However, if it is not listed, it is treated by our productive rules for treating noun-noun compounds. One of the N-N compound rules is that the pattern MATERIAL + N is analyzed as N:HAS-OBJECT-ASPART:MATERIAL. null</Paragraph> </Section> class="xml-element"></Paper>