Organisational Ontology Framework for Semantic Business Process Management
Unsupervised and Open Ontology-Based Semantic Analysis
Transcript of Unsupervised and Open Ontology-Based Semantic Analysis
Unsupervised and Open Ontology-based Semantic
Analysis
Amal Zouaq1, Michel Gagnon2, Benoit Ozell2
1
Simon Fraser University, School of Interactive Arts and Technology, 13450 102 Ave.
Surrey, BC V3T 5X3, Canada 2
Ecole Polytechnique de Montréal, C.P. 6079, succ. Centre-ville, Montréal (QC), H3C 3A7
[email protected], {michel.gagnon, benoit.ozell}@polymtl.ca
Abstract. This paper presents an unsupervised and domain independent
semantic analysis that outputs two types of formal representations: discourse
representations structures and flat scope-free logical forms. This semantic
analysis is built on top of dependency relations produced by a statistical
syntactic parser, and is generated by a grammar of patterns named α-grammar.
The interest of this grammar lies in building a clear linguistically-grounded
syntax-semantic interface using a representation (dependencies) commonly
used in the natural language processing community. The paper also explains
how semantic representations can be annotated using an upper-level ontology,
thus enabling further inference capabilities. The evaluation of the α-Grammar
on a hand-made gold standard and on texts from the STEP 2008 Shared Task
competition shows the interest of the approach.
Keywords: semantic analysis, α-Grammar, patterns, upper-level ontology
1 Introduction
Computational semantics aims at assigning formal meaning representations to natural
language expressions (words, phrases, sentences, and texts), and uses these meaning
representations to draw inferences. Given the progress made in computational syntax,
with the availability of robust statistical parsers, it is now possible to envisage the use
of syntactic parsers for a semantic analysis.
This paper introduces a semantic analysis pipeline based on dependency grammars
that generates two types of semantic representations: flat and scope-free logical forms
and discourse representation structures (DRS) [9]. The pipeline itself includes a
syntactic analysis, a semantic analysis and a semantic annotation which extract
respectively dependency relations, meaning representations and ontology-based
annotations of these meaning representations. The pipeline is modular by nature and
enables an easy change and update of the components involved at each step. The
semantic analysis itself is performed through a grammar of patterns called α-
Grammar. One main interest of such grammar is its ability to provide a syntax-
semantic interface between dependency grammars (which gain more and more
importance in the current NLP research [2]) and semantic formalisms, thus enabling
future reuse from a practical and theoretical point of view.
These formalisms are then annotated using an ontology which defines formally a
reusable set of roles and promotes the interoperability of the extracted representations
between semantic analyzers. In particular, we focus here on upper-level ontologies,
which are independent from any domain and define concepts at a high-level.
After explaining the motivation and theory behind our research (section 2), the
paper presents the α-Grammar which outputs our semantic representations, details
some of its patterns called α-structures and gives examples of the obtained
representations (section 3). Section 4 details the annotation of the semantic
representations and briefly explains the word-sense disambiguation algorithms
involved at this step. Finally section 5 evaluates the logical forms and the discourse
representation structures extracted from two corpora. Section 6 analyzes the obtained
results, draws some conclusions and introduces further work.
2 Motivation, Theory and Practice
The goal of this research is to create an open and unsupervised semantic analysis.
Open analysis means that it can be applied on many types of texts and many domains.
Unsupervised analysis means that we do not provide any training example to the
system.
Open information extraction is one recent challenge of the text mining community
[15] and it is also one objective of the computational semantics community. In fact,
Bos [3, 4] underlines that the availability of robust statistical syntactic analyzers
makes it possible to envisage a deep and robust semantic analysis. One way to
perform this analysis is to build a syntax-semantic interface that is to create semantic
representations from syntactic representations, which are generated by a statistical
syntactic parser. Here we focus on dependency grammars. In fact, dependencies are
recognized as the optimal base for establishing relations between text and semantics
as they abstract away from the surface realization of text and they can reveal non-
local dependencies within sentences [12]. Moreover, there are many semantic theories
based on dependency grammars such as DMRS [5] and the meaning-text theory [10].
Thus, developing a formal method to transform dependency formalism into a
semantic representation is desirable from a practical and theoretical point of view.
Now the question is what kind of semantic representation should be adopted. Here we
focus on two types of representations: predicative logical forms and discourse
representation structures. Depending on the depth of analysis required for a particular
application, one can choose flat and scope-free logical forms or discourse
representation structures, which are powerful representations that cover a wide range
of linguistic phenomena in a unified framework [4].
In order to implement these ideas, we used the Stanford dependency parser [7] to
obtain the syntactic representations. The Stanford dependencies have been used
successfully in several areas [6, 15] and they can be distinguished as stated by [6] by
their rich grammatical hierarchy (with the possibility of under-specifying the relation
with the label “dep”) and the fine-grained description of NP-internal dependency
relations. This last characteristic enables a better handling of the meaning of NP
phrases. Next, we decided to use Prolog to implement a grammar for semantic
analysis, also called the α-Grammar.
3 The α-Grammar, a Pattern-based Grammar
Following the approach of [15], we propose an α-Grammar which transforms
dependency representations into logical forms and discourse representation structures
in a compositional way. This grammar is based on sound and clear linguistic
principles and identifies patterns in dependency representations (named α-structures).
An α-structure is a representation whose nodes represent variables coupled with
syntactic parts-of-speech and whose edges represent syntactic dependency relations.
These dependency relations and parts-of-speech constitute constraints on the patterns.
Discovering a pattern in a dependency graph means instantiating an α-structure with
the current sentence lexical information. Each pattern is linked to a rewriting rule that
creates its semantic representation.
A rewriting rule implements a series of transformations on the pattern including
node fusion, node destruction and predicate creation. The application of a rule creates
a semantic representation which uses a knowledge model. The knowledge model
defines general and universal categories such as Entity, Named Entity, Supertype,
Event, Statement, Circumstance, Time, Number, Measure and Attribute. The category
is determined by the part-of-speech (POS) and the grammatical relationships detected
in the syntactic structure.
A set of examples representing α-structures and their transformation through
rewriting rules is shown below:
• An example of an α-structure involving a predicative event:
• An example of an α-structure involving potential anaphora resolution through the
predicate resolve:
An α-grammar can be divided into two modules: the first module involves the
transformation of dependency relations into tree-like representations, and the second
module implements a semantic compositional analysis.
entity(id_X, X), event(id_e, Y, id_X)
nsubj Y
/ v X
/ n
det entity(id_Y, Y), resolve(id_Y)
Y
/ n
X
/ d
3.1. A Tree-like representation conversion
In order to ease the subsequent process of compositional analysis, the grammar runs a
tree converter to create a tree-like representation from the dependency relations and
from the parts of speech. This tree converter is currently composed of 14 rules that
modify the structure of dependencies. Each rule specifies a tree segment that should
be modified and implements modification operations, namely removal, add, copy and
aggregate, which can be performed on nodes and/or relations. For example, one α-
structure recognizes clausal complements without a subject such as in the sentence
Paul likes to eat fish. The transformation associated with this rule consists in copying
the subject in the clausal complement as shown below using the add operation on a
node (Paul) and a link (eat, Paul):
�
The interest of this transformation is to facilitate the compositional analysis, since
the clausal complement augmented by its subject can then be interpreted in an
independent manner. In addition to this type of transformations, the tree converter
tackles the processing of various linguistic structures including compound noun
aggregation, verb modifiers aggregation, conjunctions, and negation.
Compound Noun aggregation: they are identified through the “nn” relationship
(noun compound modifier) and are aggregated to form a single entity. For e.g.
Guatemala army in the sentence “the Guatemala army announced …” is considered as
an entity after the aggregation. Noun can also be modified by adverbial modifiers
such as in the example “Genetically modified food”. In this case, there is also an
aggregation operation which is performed between the head and its children.
Verbs and Verb Modifiers: Verbs can be modified by particles “prt” such as
“made up” or “climb up” and by auxiliaries. In general, they are considered as events
predicates such as “banners flap” or as statement predicates such as “the gates of the
city seem getting closer”.
Conjunctions should be identified to build various sub-trees from the input
dependencies. These sub-trees constitute representations of the phrases linked by the
conjunctions. For example, in the sentence “There are hens and pigs in the farm”, the
tree converter extracts two sub-trees « there are hens in the farm » AND « there are
pigs in the farm ». This distributive interpretation of the conjunction can be erroneous
in some sentences depending on the intended meaning, and further versions of the tree
converter will consider the various possible interpretations of a given conjunction.
Negation: In order to handle negation, the tree converter places the negation node
not as a parent of the verb and removes the subject from the scope of the negation.
This way of handling negation is required by the embedding of structures in the
resulting DRS, as shown in the following DRS, which represents the semantics of
« the cat did not eat the mouse »:
--------------------------------- [id1] --------------------------------- resolve(id1) entity(id1,cat) NOT: ------------------------------- [id2,e1] ------------------------------- resolve(id2) entity(id2,mouse) event(e1,eat,id1,id2) ------------------------------- ---------------------------------
The resulting tree-like representation can then be processed by the compositional
semantic analysis. An example of such representation (in Prolog) is: root/tree(token(flap,2)/v,
[nsubj/tree(token(banners,1)/n,[]),
prep/tree(token(in,3)/prep,
[pobj/tree(token(wind,5)/n,
[det/tree(token(the,4)/d,[])])]),
prep/tree(token(outside,6)/prep,
[pobj/tree(token(walls,8)/n,
[det/tree(token(the,7)/d,[]),
prep/tree(token(of,9)/prep,
[pobj/tree(token(city,11)/n,
[det/tree(token(the,10)/d,[])])])])])]).
Banners flaps in the wind outside the walls of the city
3.2. A compositional analysis
An α-grammar uses compositional analysis to output semantic representations. Fig. 1
shows how a compositional analysis coupled with logical forms is performed on the
sentence “Banners flap in the wind”. The grammar starts by examining the children of
the head word “flap”. The pattern nsubj(Verb, Noun) is detected and involves the
creation of an event predicate event(Id,Node,IdAgent) where Node is the verb, Id an
identifier for the event, and IdAgent the identifier of the agent. The agent itself, the
sub-tree nsubj/tree(token(banners,1)/n,[]), should then be explored to identify its
label and eventually its modifiers and determiners. Here, banners is a leaf node which
corresponds to a noun, thus leading to a predicate entity(id1, banners). This predicate
has never been encountered before in the sentence, thus leading to a predicate new
(id1).
As can be noticed, the compositional nature of an α-grammar makes it possible to
use the result of a sub-analysis (namely the created predicates and the variables) and
to refer to these variables in higher level analysis. This is the case for prepositional
relations on events such as in(e1, id2) in fig.1.
Fig. 1. Compositional Semantic Analysis [14]
There are two kinds of α-structures in an α-grammar: core α-structures (table 1)
and modifiers α-structures (table 2). Core α-structures are primary linguistic
constructions that are organized into a hierarchy of rules where more specific rules are
fired first. For instance, the pattern “nsubj-dobj-iobj” is higher in the hierarchy than
“nsubj-dobj”. This avoids the possibility of misinterpreting a particular syntactic
construction by neglecting one essential grammatical relationship. Interpreting “Mary
gave Bill a book” has indeed not the same logical interpretation as “Mary gave Bill”,
which means nothing. Modifiers α-structures (table 2) are auxiliary patterns that
complement the meaning of core α-structures such as temporal modifiers or adverbial
clause modifiers.
Table 1. Core verbal α-structures
Core
α-structures
Examples
Verb-iobj-dobj Mary gave {Bill}iobj a {raise}dobj
Verb-dobj-xcomp The peasant carries {the rabbit}dobj, {holding it by its
ears}xcomp
Verb-ccomp John saw {Mary swim}ccomp
Verb-expletive {There}expl is a small bush.
Verb-acomp Amal looks {tired}acomp
Verb-prep-pcomp
They heard {about {Mia missing classes }pcomp} prep
Verb-dobj {The cat}nsubj eats {a mouse}dobj
Table 2. Modifiers α-structures
Modifiers α-structures Examples
Verb-prep-pobj Banners flap {in {the wind}pobj}prep
Verb-tmod Vincent arrived {last night}tmod
Verb-advcl The accident happened {as the night was falling}advcl.
nsubj
pobj
det
Banners/n
The/d
in/prep
Wind/n
prep
entity(id1, banners), new(id1)
in(e1, id2)
entity(id2, wind), resolve(id2)
event(e1, flap, id1)
Flap/v
Verb-purpcl Benoît talked to Michel {in order to secure the
account}purpcl.
Verb-infmod The following points are {to establish}infmod.
Verb-advmod The grass bends {abruptly} advmod.
For the moment, the grammar is composed of 36 core α-structures and 17
modifiers α-structures for a total of 53 α-structures. There are also some specific
grammatical relations that we would like to further explain:
Determiners: they help to identify the state of the referred object in the discourse.
Some determiners, such as “the” implies that the object has already been encountered
in the discourse, hence leading to a predicate “resolve (id)”. Other determiners such as
“a” describes a new entity and are referred by a predicate “new (id)”. These two
predicates help resolve anaphora and enable to consider the sentence elements at a
discourse level rather than at a sentence level.
Proper Nouns: Proper nouns, which are identified by the nnp relationship, are
transformed into named entities.
Prepositions: Prepositions are generally considered as modifiers patterns when
they modify a verb such as in the sentence “Banners flap in the wind” or a noun such
as “the gates of the city”. In both cases, a predicate representing the preposition is
created such as “of (id1, id2)” where id1 is the identifier of city and id2 refer to gates.
There could also be a predicate between an event id, and an object id (e.g. … flap in
…). Some particular patterns such as the preposition “of” as in the example above
lead to the creation of an attribute relationship. E.g. attribute (city, gate).
Possessive Pronouns: Possessive pronouns enable the creation of implicit
possessive relationships. For example “Benoit washes his car” implies that “Benoit
has a car”, which is represented in the logical form. The same kind of deduction is
used for constructions such as “The peasant’s eyes…” This enables building world
knowledge.
3.3. Examples
The α-grammar outputs either resolved or underspecified representations in the form
of flat scope-free logical expressions and in the form of discourse representation
structures. Underspecified representations means that certain ambiguities are left
unresolved in the semantic output such as for example the predicate “resolve (id)” for
anaphora resolution. An independent component can then be used to deal with these
ambiguities. The following table illustrates the DRS-output of our α-grammar on two
sentences from the competition STEP 2008 Shared Task. This competition is meant to
compare the results of semantic analyzes on a shared corpus of small texts.
As can be seen, the α-grammar correctly identifies the first part of the sentence in
A “An object is thrown with a horizontal speed of 20 meters per second from a cliff”
and discovers correctly the entities, events (with a correct handling of the passive
voice), prepositional relations, attributes and numerical relations. However, the
fragment “that is 125 m high” is ignored by the grammar. We also would like to
emphasize that the relation from(e1,id6) correctly applies the preposition “from” on
the event “thrown”, despite the long distance dependencies. In B, the entire sentence
is correctly analyzed.
4 Ontology-based Semantic Analysis
Once the semantic representations are obtained either through logical forms or
through discourse representation structures, there is a need to annotate these
representations using a formal and interoperable structure. In fact, one of the
drawbacks of current semantic analysis is the multiplicity of the adopted formal
representations, which hinder their comprehension, exchange and evaluation.
Standardizing these representations through ontological indexing may help these
issues. Moreover, one of the goals of computational semantics is the ability to
perform inferences on the obtained representations. Using ontologies to describe the
various predicates, discourse referents and conditions enables further reasoning, and
builds a bridge with the semantic web community and with other semantic analysis
initiatives such as semantic role labeling and textual entailment.
Upper-level ontologies can provide this formal definition of a set of roles and
enable the indexing of semantic representations in an interoperable way. One of these
upper-level ontologies is the Suggested Upper Merged Ontology (SUMO) [11], which
is widely used in the NLP community and which has gone through various
development stages and experimentations, making it stable and mature enough to be
A) An object is thrown with a horizontal
speed of 20 meters per second from a cliff
that is 125 m high.
B) The object falls for the height of
the cliff.
---------------------------------
[id2,e1,id1,id3,id4,id5,id6]
---------------------------------
entity(id2,object)
entity(id1,undefined)
event(e1,thrown,id1,id2)
entity(id3,speed)
entity(id4,meters)
entity(id5,second)
per(id4,id5)
num(id4,20)
of(id3,id4)
attribute(id3,horizontal)
with(e1,id3)
entity(id6,cliff)
from(e1,id6)
---------------------------------
------------------------------
[id1,e1,id2,id3]
------------------------------
---
resolve(id1)
entity(id1,object)
event(e1,falls,id1)
resolve(id2)
entity(id2,height)
resolve(id3)
entity(id3,cliff)
of(id2,id3)
for(e1,id2)
------------------------------
taken as a “standard” ontology. Moreover, SUMO has been extended with a Mid-
Level Ontology (MILO), and a number of domain ontologies, which allow coverage
of various application domains while preserving the link to more abstract elements in
the upper level. One interesting feature of SUMO is that its various sub-ontologies are
independent and can be used alone or in combination. In our current semantic
analysis, we only exploit the upper level, meaning that we take into account only the
SUMO ontology itself. Another interesting point of SUMO is its mapping of concepts
and relations to the WordNet lexicon [8], a standard resource in the NLP community.
The SUMO-WordNet mapping associates each synset in WordNet to its SUMO sense
through three types of relationships: equivalent links, instance links and subsumption
links. One drawback of these mappings is that they are not always consistent:
sometimes verbs are mapped to SUMO relationships and sometimes to concepts.
Although these mappings cannot be considered perfect in their original form, they
constitute an excellent demonstration of how a lexicon can be related to an ontology
and exploited in a semantic analysis pipeline.
To annotate the semantic representations and obtain SUMO-based DRS and/or
SUMO-based logical forms, we tested multiple word sense disambiguation algorithms
mainly inspired from Lesk algorithm and its derivatives such as [1]. We also used the
most frequent sense baseline as this is commonly done in WSD competitions. These
algorithms had to be applied based on a given context. In this respect, we tested
various contexts such as word windows, sentence windows, and graph-based contexts
extracted from the semantic logical representations [13] obtained in the semantic
analysis.
An example of SUMO-based logical form annotation is: outside(e1, id3), of(id3,
id4), entity(id4, SUMO:City), resolve_e(id4), entity(id3, SUMO: StationaryArtifact),
resolve_e(id3), in(e1, id2), entity(id2, SUMO: Wind), resolve_e(id2), event(e1,
SUMO: Motion, id1), entity(id1, SUMO: Fabric), new_e(id1).
5 Evaluation
The evaluation of this research is not a simple task as it involves various modules and
outputs and requires a multi-dimensional evaluation: logical forms evaluation, DRS
evaluation and Word sense disambiguation evaluation. One other issue that faces the
evaluation of semantic analysis is the lack of gold standard on which to compare our
representations. In order to tackle this issue, this evaluation relies on two corpora: 1) a
first corpus of 185 sentences that we have manually analyzed and annotated to build a
complete gold standard. This corpus is extracted from children stories such as Alice in
Wonderland; 2) a corpus of seven texts that have been used in the STEP 2008 shared
task [3] and whose objective is the evaluation of a semantic analysis. However, there
is no defined gold standard on this corpus, and the task consists mainly in defining
criteria to judge the effectiveness of the extracted representations based on the advice
of a human expert. We are aware of the limited size of this evaluation but as can be
noticed in the STEP 2008 shared task, this is a limit common to all semantic analysis
systems.
5.1. Logical Form Evaluation
Our logical form evaluation was carried on the first corpus. This corpus helped us in
performing the logical form evaluation as well as the semantic annotation evaluation.
Two metrics from information retrieval were used: precision and recall.
Precision = items the system got correct / total number of items the system generated
Recall = items the system got correct / total number of relevant items (which the
system should have produced)
Here the items designate entities and events (Table 3).
Table 3. The logical form analysis results in terms of entities and events
Precision % Recall %
Entities 94.98 80.45
Events 94.87 85.5
From these experiments, it is clear that our semantic analysis is promising. Most of
the time, the incorrect entities and events are due to a wrong syntactic parsing from
the Stanford Parser. There are also some patterns that are not yet identified which
make the recall lower. These results should be later completed with an evaluation of
the whole logical representation and not be limited to entities and events.
5.2. DRS Evaluation
The DRS evaluation was carried on the STEP 2008 shared task corpus. This corpus
was enriched by two texts taken randomly from Simplepedia leading to a total of 58
sentences. Overall, 51 sentences were semantically analyzed by our grammar and 7
were ignored due to an erroneous syntactic analysis and to the lack of the appropriate
patterns in the grammar. These 51 sentences were analyzed using 38 of our 53 α-
structures, resulting in a rate of 72% for α-structures effective usage.
In order to evaluate the obtained DRS, we first calculated a metric of precision for
the conditions of each DRS in the following way:
Precision = number of correct conditions / overall number of generated conditions in
a DRS
Second, the expert assessed the recall of each DRS by extracting the conditions of
a DRS that should have been generated but that were missing due to a wrong analysis.
Recall = number of correct conditions / overall number of conditions in a DRS
Table 4 summarizes the average score per sentence of overall conditions, correct
conditions and missing conditions and presents the obtained precision and recall
values.
Table 4. Mean values per sentence
#conditions # correct
conditions
# missing
conditions
Precision
(%)
Recall
(%)
6,7 5,5 3,6 81 67
We can notice a very reasonable value of precision on these real-world examples
considering that our grammar is in its first development phase. Our results also show
that half of the sentences obtain more than 90% precision. However, the recall is still
to be improved.
Table 5 shows a more fine-grained analysis of the obtained DRS by calculating the
precision obtained on the various condition categories especially entities, events and
attributes. All the other categories are classified under the label “other”. These results
are those obtained after the analysis of the 51 sentences.
Table 5. Mean values by DRS condition categories
# conditions # correct
conditions
Precision %
Entities 152 139 91
Events 56 43 77
Attributes 44 37 84
Others 81 54 64
Total 333 273 82
We can notice that entities are generally well-identified, followed by attributes and
then by events. The errors made by our grammar in event recognition are mostly due
to missing α-structures (7 cases over 13), to errors in the syntactic analysis (3 cases)
or to an erroneous conversion into a tree structure (2 cases). Regarding attributes, all
the errors are made on the same text and are related to particular modifiers wrongly
interpreted (e.g. “The other gas giants are Saturn and Uranus”, “only…”, etc.). Finally
the results of the “other” label indicate that further development should be made to
enhance our grammar.
6 Conclusion and further work
This paper presented the α-Grammar, a semantic analysis pattern-based grammar that
produces discourse representation structures and logical forms from free texts. With
the increase in the use of dependency grammars as a syntactic formalism, building a
conversion process from dependency relations to semantic representations is justified
from a practical point of view. Moreover, our approach proposes a semantic analysis
pipeline where the various modules (the syntactic analysis, the semantic analysis and
the ontology-based annotation) are independent, meaning that they can be easily
evolved or replaced from a software engineering perspective. The other interest of our
work is the ability to standardize the generated semantic representations through the
use of an upper-level ontology. This also enhances the inference capabilities over the
extracted representations. Finally, our computational semantics approach is domain-
independent and unsupervised, which enables better reuse in multiple domains and
applications.
In future work, we plan to enhance the grammar by discovering new patterns using
manual analysis of texts but also automatic pattern learning approaches. This will help
us improve the precision and recall of the semantic analysis. We also plan to handle
more complex discourse structures and anaphora resolution. Finally, we would like to
extend the scale of the corpora used for the evaluation and to compare our DRS with
the DRS extracted by Boxer [4] on the same corpora.
Acknowledgements. The authors would like to thank Prompt Quebec, UnimaSoft
Inc. and the FQRNT for their financial support. Amal Zouaq is funded by a
postdoctoral fellowship from the FQRNT.
References
1. Banerjee, S. and Pedersen, T. (2003). Extended gloss overlaps as a measure of semantic
relatedness. In Proc. of the 18th Int. Joint Conf. on AI, Mexico, pp.805-810.
2. BONFANTE G., GUILLAUME B., MOREY M. & PERRIER G. (2010). Réécriture de
graphes de dépendances pour l’interface syntaxe-sémantique. In Proc. of TALN 2010,
Montreal.
3. Bos, J. (2008a). Introduction to the Shared Task on Comparing Semantic Representations.
In STEP 2008 Conference Proceedings, pp.257–261. College Publications.
4. Bos, J. (2008b). Wide-Coverage Semantic Analysis with Boxer. In STEP 2008 Conference
Proceedings, pp 277–286, Research in Computational Semantics, College Publications.
5. Copestake A. (2009). Slacker semantics : Why superficiality, dependency and avoidance
of commitment can be the right way to go. In Proceedings of the 12th Conference of the
European Chapter of the ACL (EACL 2009), p. 1–9, Athens, Greece : Association for
Computational Linguistics.
6. De Marneffe, M-C and Manning, C.D. 2008. "The Stanford typed dependencies
representation." In COLING Workshop on Cross-framework and Cross-domain Parser
Evaluation.
7. De Marneffe, M-C, MacCartney, B. and Manning. C.D. (2006). Generating Typed
Dependency Parses from Phrase Structure Parses. In Proc.of LREC, pp.449-454.
8. Fellbaum, C. (1998). WordNet: An Electronic Lexical Database, MIT Press.
9. Kamp, H. and Reyle, U. (1993). From Discourse to Logic. Introduction to Model-theoretic
Semantics of Natural Language, Formal Logic and Discourse Representation Theory,
Studies in Linguistics and Philosophy.
10. Melcuk I. (1988). Dependency Syntax : Theory and Practice. Albany : State Univ. of New
York Press.
11. Pease, A., Niles, I., and Li, J. (2002). The Suggested Upper Merged Ontology: A Large
Ontology for the Semantic Web and its Applications. In Proc. of the AAAI Workshop on
Ontologies and the SW, Canada.
12. Stevenson, M. and Greenwood, M.A. 2009. Dependency Pattern Models for Information
Extraction, Research on Language & Computation, pp.13-39, Springer.
13. Zouaq, A., Gagnon, M. & Ozell, B. (2010): Can Syntactic and Logical Graphs help Word
Sense Disambiguation? In Proc. of LREC 2010.
14. Zouaq, A., Gagnon, M. & Ozell, B. (2010): Semantic Analysis using Dependency-based
Grammars and Upper-Level Ontologies, International Journal of Computational
Linguistics and Applications, 1(1-2): 85-101, Bahri Publications.
15. Zouaq, A. (2008). An Ontological Engineering Approach for the Acquisition and Exploi-
tation of Knowledge in Texts, PhD Thesis, University of Montreal (in French).