PaperlayerView exact revision on arXiv

Source-linked AI summary

Generating Natural Language Descriptions from OWL Ontologies: the NaturalOWL System

Ion Androutsopoulos, Gerasimos Lampouras, Dimitrios Galanis

arXiv:1405.6164v1cs.CLcs.AI

TL;DR

OWL statements can be difficult for non-experts to understand, while typical verbalizers produce axiom-by-axiom text rather than fluent, coherent end-user descriptions. NaturalOWL applies NLG methods to generate multi-sentence descriptions, and trials report better text with domain-dependent resources that take relatively light effort to construct.

  • Problem

    OWL syntaxes can be difficult for people unfamiliar with formal knowledge representation, and typical verbalizers produce axiom-by-axiom text rather than coherent end-user descriptions.

    Source: §1 · p. 2

  • Method

    NaturalOWL generates multi-sentence descriptions of OWL classes and individuals using NLG methods and optional domain-dependent generation resources.

    Source: §1 · pp. 2–3

  • Results

    Trials found that domain-dependent generation resources help NaturalOWL produce significantly better texts and can be constructed with relatively light effort compared with developing an OWL ontology.

    Source: §5 · p. 40

  • Takeaways & Limitations

    OWL information can be published with generated texts, making it accessible to computers and end-users speaking different languages.

    Source: §1 · p. 3

    NaturalOWL does not consider OWL property axioms, and it excludes property domain and range statements because these typically provide overly general information for end-users.

    Source: Limitations of content selection · p. 15

Abstract

from arXiv · show

We present NaturalOWL, a natural language generation system that produces texts describing individuals or classes of OWL ontologies. Unlike simpler OWL verbalizers, which typically express a single axiom at a time in controlled, often not entirely fluent natural language primarily for the benefit of domain experts, we aim to generate fluent and coherent multi-sentence texts for end-users. With a system like NaturalOWL, one can publish information in OWL on the Web, along with automatically produced corresponding texts in multiple languages, making the information accessible not only to computer programs and domain experts, but also end-users. We discuss the processing stages of NaturalOWL, the optional domain-dependent linguistic resources that the system can use at each stage, and why they are useful. We also present trials showing that when the domain-dependent llinguistic resources are available, NaturalOWL produces significantly better texts compared to a simpler verbalizer, and that the resources can be created with relatively light effort.

1. Introduction

NaturalOWL aims to turn OWL statements into fluent, coherent multi-sentence descriptions for end-users, unlike verbalizers that typically translate axioms one at a time. Its trials find that domain-dependent generation resources improve text quality significantly and can be created with relatively light effort.

1. Introduction

  • NaturalOWL aims to generate fluent, coherent multi-sentence descriptions of OWL classes or individuals for end-users, unlike verbalizers that typically translate one axiom at a time.Verbalizers often produce controlled, not entirely fluent English, primarily for domain experts.Source: §1 · p. 2
  • The open-source system supports English and Greek, with Greek generation requiring appropriate Greek linguistic resources.The same OWL statements can be used to generate texts in both languages.Source: §1 · p. 2
  • NaturalOWL uses NLG methods for information selection, ordering, aggregation, and referring expressions, and addresses OWL-specific generation issues without proposing novel theoretical NLG algorithms.Some OWL statements need simpler intermediate representations, and OWL-specific opportunities include aggregating cardinality axioms and using class hierarchies for referring expressions.Source: §1 · p. 3
  • NaturalOWL can use any OWL ontology, but high-quality output requires domain-dependent resources such as class-to-name mappings and property-to-sentence-plan mappings.A domain author familiar with OWL creates these resources using Protégé and a plug-in.Source: §1 · p. 3
  • Domain-dependent generation resources produce significantly better texts and can be constructed with relatively light effort compared with developing an ontology.The trials compare NaturalOWL using these resources with a simpler verbalizer that requires none.Source: §1 · p. 4
  • NaturalOWL builds on ideas from ILEX and M-PIRO but was developed specifically for OWL, after attempts to use M-PIRO with OWL encountered problems.ILEX did not support OWL; M-PIRO was a multilingual extension tested in several domains.Source: §1 · p. 4

2. Related Work

OWL verbalizers can map axioms to English-like sentences, but typically prioritize faithful, unambiguous expression over fluent, coherent, multilingual end-user texts. Earlier work explored controlled languages and NLG components, while NaturalOWL targets multi-sentence OWL descriptions.

2. Related Work

  • Sydney OWL Syntax and Attempto Controlled English provide mappings from normative OWL syntaxes to English-like controlled languages, while Rabbit and CLOnE primarily support ontology authoring.Some OWL controlled languages cannot express every kind of OWL statement.Source: §2 · p. 5
  • Verbalizers typically translate axioms individually without ensuring coherence, aggregation, referring expressions, natural fluency, domain-specific resources, or multilingual output.Their priority is often unambiguous expression of axiom meaning for domain experts.Source: §2 · p. 5
  • Some verbalizers use NLG methods such as aggregation and text planning, but those methods have generally been used only to a limited extent with OWL ontologies.Ontosum generated descriptions of individuals, but no detailed system description or trials with independently created ontologies appear to have been published.Source: §2 · p. 6
  • Prior work includes single-sentence RDF triple aggregation and OWL class content selection that can include deduced facts, whereas NaturalOWL produces multi-sentence texts.Other work explores explaining facts deduced from OWL ontologies; adding deduction and explanation to NaturalOWL is identified as an interesting possibility.Source: §2 · p. 6

3. The Processing Stages and Resources of NaturalOWL

NaturalOWL organizes generation as a pipeline and divides document planning into content selection and text planning. For individual targets, it retrieves relevant OWL statements, converts them to message triples, and selects information for readable end-user descriptions while imposing restrictions on input structures.

3. The Processing Stages and Resources of NaturalOWL

  • NaturalOWL uses a three-stage pipeline: document planning, micro-planning, and surface realization.Micro-planning includes lexicalization, sentence aggregation, and referring-expression generation.Source: §3 · p. 6

3.1 Document Planning

  • Document planning selects the information to convey and plans the structure of the resulting text.Source: §3.1 · p. 6

3.1.1 Content Selection

  • For content selection, NaturalOWL retrieves relevant OWL statements, converts them into message triples, and selects which triples to express.For an individual target, it scans for supported statement forms and retrieves statements describing that target directly.Source: §3.1.1 · pp. 6–7, OWL statements for individual targets · p. 7

OWL statements for individual targets

  • For individual targets, the input disallows nested union and intersection operators, which can make statements difficult to express and texts hard to comprehend.Nested unnamed classes can be replaced by named classes, but meaningful identifiers and natural-language names may require a domain author; the system may also omit information based on user modeling or length limits.Source: OWL statements for individual targets · pp. 7, 9

OWL statements for class targets

NaturalOWL handles named class targets and can also retrieve information about second-level targets for individuals. Its generated class descriptions may express necessary, but not sufficient, conditions for class membership.

OWL statements for class targets

  • For class targets, NaturalOWL scans for specified OWL statement forms, which require class identifiers at certain points and therefore named classes.The authors note that ontologies using unnamed classes in those positions can be modified.Source: OWL statements for class targets · p. 9
  • NaturalOWL generates the same text for EquivalentClasses and SubClassOf, so class descriptions express necessary, not sufficient, membership conditions.Some ontology information may also be omitted when user modeling indicates it is already known or the text should stay within a length limit.Source: OWL statements for class targets · p. 9

OWL statements for second-level targets

  • For an individual target, second-level targets are its named class and individuals linked directly through object properties; class targets do not use them.The system sets maximum fact distance to 2 to retrieve statements for these targets, or 1 to retrieve statements only for the current target.Source: OWL statements for second-level targets · p. 10
  • Retrieving statements about second-level targets can add descriptions of related entities, such as an aryballos’s type, period, and technique.Canned sentences can supply information that is impractical to represent fully in logical terms, but must be entered in multiple versions for different languages or user types.Source: OWL statements for second-level targets · p. 11

Converting OWL statements to message triples

NaturalOWL converts OWL statements into message triples that are designed to correspond to simple sentences. This representation can expose similarities in the sentences to generate that are less apparent in OWL or RDF triples.

Converting OWL statements to message triples

  • A message triple has a semantic subject, a property specifying the mapping, and a semantic object or filler.The subject may be the target or a second-level target; the object may be an individual, value, class, or set of these.Source: Converting OWL statements to message… · p. 11
  • Each message triple is intended to be easily expressed as a simple sentence, unlike some RDF triples, and can capture similarities between the sentences to generate.For example, a ClassAssertion and a SubClassOf statement can map to identical triples apart from the individual and class identifiers.Source: Converting OWL statements to message… · p. 12
  • ObjectIntersectionOf operators are discarded in favor of multiple message triples, while ObjectUnionOf operators become disjunctions of message triples.The resulting disjunction can be shortened during sentence aggregation.Source: Converting OWL statements to message… · p. 13
  • NaturalOWL maps EquivalentClasses and SubClassOf to identical triples using isA, producing the same text for both statement types.The paper notes that this avoids the stricter wording of the original OWL statement, which it considers inappropriate for end-users.Source: Converting OWL statements to message… · p. 13

Interest scores and repetitions

NaturalOWL uses interest scores and user models to control which message triples are conveyed and whether they are repeated. Scores may vary by user type and by the subject and property involved.

Interest scores and repetitions

  • Domain authors assign non-negative interest scores to message triples, potentially with different scores for different user types.Scores can apply by property, by subject class and property, or to a particular subject and property.Source: Interest scores and repetitions · p. 14
  • The article does not detail how interest scores are assigned, though it notes that learning methods or graph algorithms could be used.The authors refer readers elsewhere for a detailed description of the score-assignment mechanisms.Source: Interest scores and repetitions · p. 14
  • A message triple can be repeated until the user is assumed to have assimilated it; after that, it is not repeated for that user.NaturalOWL maintains a personal model recording which triples were conveyed and how many times.Source: Interest scores and repetitions · p. 14

Selecting the message triples to convey

NaturalOWL retrieves and ranks message triples, then selects a bounded number to convey. Its text planner orders the resulting sentences, with local coherence considered but global coherence omitted.

Selecting the message triples to convey

  • NaturalOWL retrieves relevant OWL statements, converts them to message triples, ranks them by interest, removes assimilated information, and selects up to maxMessagesPerPage.If an assimilated triple concerns the target, connected triples about second-level targets may also be discarded.Source: Selecting the message triples to convey · p. 15

Limitations of content selection

  • NaturalOWL excludes ObjectPropertyDomain and ObjectPropertyRange statements because they typically provide information too general and uninteresting to end-users.It also omits property axioms such as symmetry, transitivity, functionality, and subproperty statements.Source: Limitations of content selection · p. 15

3.1.2 Text Planning

  • The text planner orders message triples, each intended to be expressible as a single sentence.The ordering therefore determines the order of the corresponding sentences.Source: §3.1.2 · p. 15

Global and local coherence

  • With maximum fact distance 1, every sentence transition is a preferred continue; at distance 2, transitions are not always continue.The planner groups sentences about a second-level target immediately after the sentence introducing it, while global coherence is omitted because its structures tend to be uninteresting for these texts.Source: Global and local coherence · p. 16

Topical order

NaturalOWL groups topically related sentences into sections and lets domain authors specify section and property order. For second-level targets, it orders their triples separately and inserts them after the triple that introduces each target.

Topical order

  • Topically related sentences are grouped into sections, which can be displayed as separate paragraphs or sections in longer texts.Domain authors assign properties to sections and specify the order of sections and properties within them.Source: Topical order · p. 17

The overall text planning algorithm

Further related work on text planning

NaturalOWL's micro-planning turns selected message triples into sentence specifications using multilingual lexical entries and property-specific sentence plans. These plans define how triple content fills sentence slots.

Further related work on text planning

  • Sentence ordering draws on property and section order, related to text schemata; prior work also studies ordering and learning it from semantically tagged corpora.The passage cites McKeown (1985), Duboue and McKeown (2001), Barzilay and Lee (2004), Elsner et al. (2007), Barzilay and Lapata (2008), and Chen et al. (2009).Source: Further related work on text planning · p. 18

3.2 Micro-planning

  • Micro-planning has three sub-stages: lexicalization, sentence aggregation, and generation of referring expressions.Lexicalization turns selected message triples into abstract sentence specifications.Source: §3.2 · p. 19, §3.2.1 · p. 19

3.2.1 Lexicalization

  • For each property and supported language, domain authors can specify sentence plans; words used in those plans require lexicon entries with inflectional forms.Lexicon entries are bilingual and cover forms such as the English verb “find” and its Greek equivalent.Source: §3.2.1 · p. 19, Lexicon entries · p. 19

Lexicon entries

  • NaturalOWL stores lexicon entries for verbs, nouns, and adjectives, and a few dozen entries sufficed in the ontologies considered.English inflectional forms could often be generated automatically; no entries are needed for closed-class words.Source: Lexicon entries · pp. 19–20

Sentence plans

  • A sentence plan is a sequence of slots that can specify referring expressions, inflected words, fixed text, and expressions for triple values.For the usedDuringPeriod property, a plan specifies the owner's referring expression, a past passive verb, a preposition, and the object expression; it can also permit sentence aggregation.Source: Sentence plans · pp. 20–21

Default sentence plan

When a property-specific sentence plan is unavailable, NaturalOWL uses a three-slot default plan, optionally using an rdfs:label for the property. Built-in plans cover domain-independent properties, while modified-property plans are generated from unmodified ones.

Default sentence plan

  • Without a property-specific plan, the default plan combines a referring expression for the owner, a tokenized property identifier, and an expression for the filler.If available, an rdfs:label replaces the tokenized identifier; tokenization handles CamelCase and underscore-style identifiers.Source: Default sentence plan · pp. 21–22

Using rdfs:label strings

  • An rdfs:label can make the default plan's property phrase more natural, but the plan may still produce disfluencies and lacks grammatical-category information for many aggregation rules.The default plan also does not allow additional phrases before the owner slot or after the filler slot.Source: Using rdfs:label strings · p. 22

Sentence plans for domain-independent and modified properties

  • Domain-independent properties use built-in sentence plans, and plans for modified properties are generated from those for unmodified properties.The built-in plans have fixed, domain-independent semantics; they can be modified, for example, to use past rather than present tense.Source: Sentence plans for domain-independent… · pp. 22–23, Table 3 · p. 23

Specifying the appropriateness of sentence plans

NaturalOWL lets authors score alternative sentence plans for different user types. Its plans simplify authoring relative to some generic surface realizers, while leaving some linguistic decisions in the plans themselves.

Specifying the appropriateness of sentence plans

  • Alternative plans for the same property and language can receive user-type-specific appropriateness scores, such as preferring a more common verb for children.Automatically constructed plans inherit the scores of the plans from which they are derived.Source: Specifying the appropriateness of… · p. 23

Related work on sentence plans

  • NaturalOWL plans require only elementary linguistic concepts and no upper-model expertise, but can encode information available from large grammars or corpora.Unlike some template systems, these plans do not allow conditionals or recursive invocation of other templates; statistical methods cited in the passage were limited to single sentences from flat records.Source: Related work on sentence plans · p. 24

Specifying natural language names

NaturalOWL uses author-specified natural-language names, then applies aggregation and referring-expression rules to produce concise descriptions. These mechanisms allow terminology to vary by audience, while referring expressions remain limited in range.

Specifying natural language names

  • Natural-language names are author-specified slot sequences with a required lexicon-backed head, whose number and case—and adjective gender—can inflect by context.Other slots can contain articles, additional nouns or adjectives, prepositions, or fixed strings.Source: Specifying natural language names · p. 25
  • Authors can provide audience-specific alternative names; absent names fall back to an rdfs:label or tokenized identifier.Names can also specify default article and number choices.Source: Specifying natural language names · pp. 24–25

3.2.2 Sentence Aggregation

  • Aggregation greedily combines sentence plans under a user-type message cap, while cardinality rules may combine nonadjacent sentences and ignore that cap.Rules operate almost exclusively on adjacent sentences in the same topical section.Source: §3.2.2 · pp. 26–27
  • When coordinated objects share adjective sequences and a head noun, aggregation can omit repeated head nouns.For example, “the Classical, the Hellenistic, and the Roman period” avoids repeating “period.”Source: §3.2.2 · p. 26
  • Rules fold class statements with passive or prepositional properties and adjectives, and combine clauses sharing or differing in verb form.Related rules also handle conjunctions and disjunctions.Source: §3.2.2 · pp. 27–28

3.2.3 Generating Referring Expressions

  • Referring expressions use names, identifiers or labels, pronouns, and demonstrative-plus-class phrases, but do not describe referents through their properties.Pronouns are used for a repeated sentence subject, a constraint that helps readers resolve them as intended.Source: §3.2.3 · p. 28
  • For an anonymous primary target, NaturalOWL can use a demonstrative phrase based on its most specific named, nonanonymous class; this setting affects only that target’s references.Source: §3.2.3 · p. 29

3.3 Surface Realization

NaturalOWL’s sentence plans already specify each sentence’s ordered, final surface form, so surface realization mainly converts those specifications into text. The stage also adds punctuation and capitalization, and can be modified to add application-specific markup or images.

3.3 Surface Realization

  • Unlike systems that leave constituent order or word forms underspecified, NaturalOWL’s sentence plans already specify each sentence’s ordered final form.Source: §3.3 · p. 29
  • Surface realization mainly converts fully specified sentence plans into final text, adding punctuation and capitalization.Source: §3.3 · p. 29
  • Modifying NaturalOWL’s surface-realization code can add application-specific markup, such as HTML tags or hyperlinks, or images.Source: §3.3 · p. 29

4. Trials

The trials assess whether NaturalOWL’s domain-dependent resources improve end-user texts and how much effort they take to build across the Wine and Consumer Electronics domains. They evaluate English and Greek descriptions of development and test products and examine the contribution of sentence plans and section organization and ordering.

4. Trials

  • The trials used independently developed ontologies and compared NaturalOWL with the SWAT verbalizer and with NaturalOWL lacking domain-dependent resources.Earlier cultural-heritage trials produced texts indistinguishable from human-authored texts, but the authors noted that their involvement in ontology development may have favored generation quality.Source: §4 · p. 30

4.1 Trials with the Wine Ontology

  • For the Wine Ontology, a 10-student blind evaluation rated 95 texts per system, and creating NaturalOWL’s resources took the second author three days.The evaluation used English texts; the 43 class and 52 individual descriptions were generated with maximum fact distance 1, and the resources used five sentence plans and seven properties.Source: §4.1 · pp. 31–32
  • NaturalOWL(+) produced more fluent sentences and better referring expressions; it scored highest on interest, significantly above both alternatives, and highest on clarity, but not significantly above SWAT.All three systems reported the same information except that NaturalOWL(+) blocked obvious sentences using zero interest scores, which the paper attributes to its higher interest scores.Source: §4.1 · pp. 33–34
  • SWAT scored higher than NaturalOWL(−) except on text structure, but only their referring-expression score difference was statistically significant.Both systems received particularly low referring-expression scores; the paper suggests SWAT’s better tokenization of OWL identifiers explains its slightly higher score.Source: §4.1 · p. 34

4.2 Trials with the Consumer Electronics Ontology

  • The Consumer Electronics trial split 60 products into development and test sets; resources built from 30 development products were used to describe the 30 unseen test products.Resource construction took approximately six days for English and Greek; sentence plans covered 42 development-set properties, while two additional test-set properties used NaturalOWL’s defaults.Source: §4.2 · p. 35
  • On English development texts, domain-dependent resources improved fluency, referring expressions, and sentence ordering; NaturalOWL(+) received perfect clarity and interest scores.The other systems’ texts jumped between topics, and their clarity and interest scores were lower; their scores were generally similar to the previous trial.Source: §4.2 · pp. 36–37, Table 7 · p. 37
  • English test-set results closely matched development results, with no significant differences between corresponding cells for any system.The findings support using resources built from sampled individuals to describe new individuals of the same types when the rest of the ontology is unchanged; English product labels may have contributed to NaturalOWL(+)’s test results.Source: Table 8 · p. 37, §4.2 · p. 37
  • Greek NaturalOWL(+) results did not differ significantly from English results or between development and test sets.English product names were acceptable in Greek in this domain, but other domains may require Greek names for new individuals.Source: Table 9 · p. 37, §4.2 · p. 38

4.3 Ablation Trials with the Consumer Electronics Ontology

  • With all components and resources enabled, NaturalOWL(+) achieved the best scores and was judged near-perfect across criteria.Removing interest-score assignments caused the biggest decrease in interest; other scores were also affected, while referring-expression and text-structure scores changed very little.Source: Table 10 · p. 38, §4.3 · pp. 38–40
  • Disabling pronouns and demonstrative noun phrases substantially reduced referring-expression scores; removing natural-language names caused a further, smaller decrease.The changes also affected fluency and clarity, while the first change affected interest; natural-language names were needed to generate pronouns and demonstratives.Source: §4.3 · p. 39
  • Turning off aggregation further reduced referring-expression and fluency scores, while its overall contribution to perceived text quality appeared small.Text structure and interest decreased slightly, and clarity did not change.Source: §4.3 · p. 39
  • Removing sentence plans sharply reduced fluency and clarity, while removing sections and ordering information sharply reduced text-structure scores.Both changes also affected perceived interest; the sentence-plan change slightly affected text structure, while removing sections and ordering also affected fluency and clarity.Source: §4.3 · pp. 39–40

5. Conclusions and Future Work

NaturalOWL aims to generate fluent, coherent multi-sentence descriptions of OWL individuals and classes for end-users, unlike simpler axiom-by-axiom verbalizers. Trials found that domain-dependent generation resources significantly improve its texts with relatively light effort; future work includes further evaluation and architectural changes.

5. Conclusions and Future Work

  • NaturalOWL is an open-source system for English and Greek descriptions of OWL individuals or classes, intended for end-users rather than domain experts.Unlike simpler verbalizers, it aims to produce fluent, coherent multi-sentence texts.Source: §5 · p. 40
  • Trials showed that domain-dependent generation resources help NaturalOWL produce significantly better texts and can be constructed with relatively light effort compared with developing an OWL ontology.The trials compared resource-construction effort with the effort typically needed to develop an ontology.Source: §5 · p. 40
  • Future work could compare resource-construction effort with manually editing lower-quality texts, which can be tedious for many individuals, though post-editing may suit some applications.The authors describe the comparison as future work and note that repeated fixes are burdensome when many individuals of a few types are described.Source: §5 · p. 40
  • The authors aim to replace NaturalOWL’s pipeline with global optimization and test it on biomedical ontologies such as Gene Ontology and SNOMED.The proposed architecture would consider NLG stages in parallel to avoid greedy stage-specific decisions.Source: §5 · p. 40
Loading 1405.6164v1…