Semantic annotation of natural language text requires a certain degree of understanding of the document in question. Especially the resolution of unclear reference is a major challenge when detecting relevant information units in a document. The ongoing KunDoc project examines how domain specific ontologies can support the task of Coreference chaining in order to enhance applications such as automatic annotation, information extraction or automatic summarization. In this paper, we present a robust methodology for acquisition of semantic contexts that does not depend on a thorough syntactic parsing as necessary tools often are unavailable for “smaller” languages. Based on a shallow corpus-analysis, verb-subject relations constitute the framework for the extraction of semantic contexts. Our approach either adds the semantic contexts to concepts and instances in an existing ontology or builds up the domain knowledge necessary for coreference chaining from scratch.
Automatic semantic annotation of natural language text such as web documents
requires a certain degree of text understanding. An important task in order to
constitute a coherent semantic representation of a document is the resolution of
anaphoric expressions and coreference chains. In the current NLP landscape there are
numerous approaches for anaphora resolution based on either heuristics, such as
Most of the heuristics-based methods for Coreference chaining rely mainly on morpho-syntactic features such as number and gender agreement, syntactic function, topicalisation etc. Only few efforts have been made – besides the knowledge-based methods in the early days of AI, such as frames or scripts – in order to utilize semantic cues in order to resolve anaphoric expressions. In the KunDoc project, these semantic cues are retrieved from domain-specific ontologies. The semantic features used in KunDoc consist of: • Class/subclass relations according to the taxonomy used (e.g. person, organization, etc.) • Verbs or adjectives with which the concept in question frequently co-occurs
As a methodological framework for extraction of possible referents and their
antecedents we used the CORPORUM system
During the text analysis, possible antecedents are collected and stored in a candidate list. These candidates are enriched with their semantic features that are retrieved from the ontology using the JENA2 interface. As soon potential anaphoric expressions are encountered (definite noun phrases or pronouns), their features are matched with the candidates in order to find the correct antecedent, as shown in Figure 1. Whereas class/subclass relations are inherent in existing taxonomies, the semantic contexts for coreference chaining may have to be added. The following section presents a methodology for generating these contexts.
Approach 3
The idea of deriving semantic classes from noun phrase/verb co-occurrences is not
new in itself. Most of the work in this area is based on the distributional hypothesis,
i.e. that nouns are similar to the extent that they share context. We assume that certain
actions – denoted by verbs – are typically performed by a semantically restricted set
of entities. Notable work in this area has been done by both Hindle
The extraction of explicit Predicate-Argument Structures requires somewhat accurate parses of the sentences in the corpus, which can be problematic, due to the availability of the necessary tools for various languages. Especially for the “smaller” languages, such as Norwegian, this is a well known obstacle for the development of tools for information extraction or annotation. As an alternative – and more robust – approach, a shallow parsing of the text was chosen by using the Oslo-Bergen Tagger (OBT), a PoS-Tagger developed within a cooperation between the Universities of Oslo and Bergen, Norway. The OBT consists of a morphologic tagger and a CGbased module for disambiguation of tags. The CG component gives all found options that cannot be excluded. This gives a fair recall, but low precision. As the analysis of the data set will show, some of these mistakes will be filtered out as noise, whereas others may obscure the results.
The data set was extracted from a corpus of newspaper articles about a murder case in the village of Førde, Norway. All 94 texts were published in the Norwegian online newspaper VG Nett (www.vg.no), yielding a total of 1619 subject-verb structures. In order to provide a basic benchmark for semantic classification all subjects were grouped manually into 6 conceptual classes: • politi (police) • offer (victim) • etterforskning (investigation) • spor (trace) • pårørende (relatives) • gjerningsmann (perpetrator)
We assume that subjects in most cases denote the agent of the action described by
the respective verbs. Therefore, in our first experiments, a co-occurrence score is
calculated for subject-verb pairs only. An alternative approach to Hindle’s similarity
measure – also based on the distributional hypothesis – is presented in
SIM =
(∑v∈A(n1 )∩ A(n2 )Csubj (v n1 )* Csubj (v n2 )) ∑v∈A(n1 )Csubj (v n1 )2 * ∑v∈A(n1 )Csubj (v n2 ) 2 where for each subject n, A is the set of verbs v that share a subject-verb structures with n. The results for the concept “police” are depicted in Table 1.
politi politi politi politi politi politi politi politi politi politi politi politi politi (police) etterforsker (detective) lensmann (seargeant) Politiet (police, definite form) Broberg (person name) tekniker (technician) VG (name of newspaper) mannskap (squad) Fonn (person name) vitne (witness) Borlaug (person name) Naustdal (person name)
This approach seems promising as there is only little noise in the ten most similar subjects, such as VG. In addition to the semantic classification of the concepts in the corpus, based on the co-occurrence measure between verbs and subjects, we are able to establish probable semantic contexts that can be added to concepts and instances in the ontology as depicted in Table 2 for the three top terms in the “police” cluster: etterforsker lensmann Broberg fatte overse etterforske utdype avtale erfare antyde fastholde oppfordre bekrefte Table 2. Semantic contexts for concepts in the “police” cluster.
Approach 5
As
(3)
Hun skal ha hørt rop.
(She is supposed to have heard cries.)
This analysis was extended by Eiken by a clustering of concepts, which implies that concepts no longer need to be matched perfectly, but the coreferent must be part of a concept group. In (4), for example the pronoun hun (she) is first linked to the concept kvinne (woman), which is not among the candidates. However, the correct concept Slåtten, which is among the candidates is clustered together with kvinne and can therefore be selected.
(4)
Hun ble funnet omkommet.
(She was found dead.)
In this way, a certain fuzziness of the matching is achieved, which enhances the possibility of finding coreferents in a set of candidates. In the KunDoc project this analysis is again extended by using the relations in the extracted ontologies.
We have presented a method for the enhancement of information extraction and semantic annotation through Coreference chaining. We have shown how a shallow and robust analysis of a domain specific text corpus yields Verb-Subject-Structures that can be exploited in order to extend domain-specific ontologies by adding semantic contexts to the concepts and instances in the taxonomy. The robustness of the proposed methodology is constituted by the fact that it does not require tools for extensive parsing of natural language, but only a part-of-speech tagger and rather simple statistical models. This will also ensure an easy transfer to other languages.
The next steps in the KunDoc project will be a thorough evaluation of precision and recall of the resolution of anaphoric expressions and an evaluation on how this improves the quality of information extraction. Future work in the KunDoc project will look into the feasibility of the methodology for other knowledge domains and text genres.
Stuckardt, R. (2000). Robust Anaphor Resolution: Design and Evaluation of the ROSANA System. 1st workshop on RObust Methods in Analysis of Natural language Data. Lausanne.