Motivations And Implications Of Veins Theory
Dan Cristea
Faculty of Computer Science of the “Alexandru Ioan Cuza” University of Iasi, Romania
Institute of Theoretical Computer Science, Iasi branch of the Romanian Academy
Abstract. The paper deals with the cohesion part
of a model of global discourse
interpretation, usually known as Veins Theory (VT). By taking from the
Rhetorical Structure Theory the notions of nuclearity and relations, but ignoring
the relations’ names, VT computes from rhetorical structures strings of
discourse units, called veins, from which domains of accessibility can be
determined for each discourse unit. VT’s constructs best fit with an incremental
view on discourse processing. Linguistics observations that lead to the
elaboration of the theory are presented. Cognitive aspects like short-term
memory and on-line summarization are explained in terms of VT’s constructs.
Complementary remarks are made over anaphora and its resolution in relation
with the interpretation of discourse.
1 Introduction
A discourse is different from a text, because a discourse is a text in the progress of
reading or hearing in a human brain. So, a discourse exists only as a process and, as
such, it has a dynamic nature. When the reading comes to an end, the discourse also
finishes and only a representation of it remains in the reader's memory.
The main concerns of the studies dedicated to discourse have been on proposing a
epresentation that best describes its structure and on understanding the relationship
existing between structure and referentiality. In Atentional State Theory (AST) [12]
the discourse is seen as having a recursive segmental structure residing in a tree-like
representation, while the dynamic interpretation uses a stack model in which the
references are allowed to occur from the top state elements towards the bottom. The
Rhetorical Structure Theory (RST) [17] gives only a static representation while
ignoring any concern on referentiality. Centering Theory (CT) [1], [11] uses the
notion of segment from AST to propose a local theory of discourse coherence.
We will review in this paper the cohesion part of a model of global discourse
terpretation, usually known as Veins Theory (VT), while also noticing some new
facts about it. By taking from RST its notions of nuclearity and relations, but ignoring
the relations’ names, VT [7] reveals a "hidden" structure in the discourse tree, called
vein, which enables to determine a domain of evocative accessibility (dea) for each
discourse unit, as that space of the discourse where all anaphors belonging to that unit
can find an antecedent. As such, the vein expression of a discourse unit gives the
minimal span necessary to understand that particular unit in the context of the whole
discourse. VT allows for an integrated explanation of the common points of AST,
Cristea D. (2005).
Motivations And Implications Of Veins Theory.
In Proceedings of the 2nd International Workshop on Natural Language Understanding and Cognitive Science, pages 32-44
DOI: 10.5220/0002565000320044
RST and CT, while also correcting some AST predictions relative to accessibility
domains (the nucleus to nucleus references and references from nuclei to left
In the following section we present linguistic observations that lead to the
formulation of VT. The basic definitions are revised in section 3. VT’s claim on
discourse cohesion is presented in section 4. The last section gives a synthesis of the
theory, exhibits a cognitive argumentation, quickly reviews applications based on the
findings of the theory and shows some possible future developments.
2 The Intuitions Underlying VT
The notion of vein was born by synthesizing observations on how references align
within the representation of a discourse as a tree. Considering the hierarchical
organization given by the tree structure and the principle of compositionality
(conforming to which, a relation that holds between two spans also holds between the
most salient units of those spans [19]), which allow long-distance sibling relations
between discourse units, these observations are collected below (to simplify the
wording, we will say that “a unit A refers to a unit B” when we mean “a referential
expression (re) belonging to the unit A refers to a discourse element (de) introduced
or referred in/from unit B”). In the examples of this section we will mark with
numbered u – the units, and with Rthe relations. An upper n or s at the shoulder of
an expression indicates that the corresponding text span is a nucleus, respectively a
satellite. The names of relations in our commentaries of the examples are taken from
a). Right satellites or nuclei can refer to their left nuclear siblings: in combinations
R u
, or u
R u
, u
can refer to u
Ex. 1:
1. John left home without an umbrella
2. although he watched the TV morning forecast announcing rain.
The pronoun he in unit 2, a satellite of unit 1, refers to the entity [John],
introduced by the referential expression John in the first unit.
b). A right nucleus can refer to a left satellite: in combinations u
R u
, u
refer to u
as in:
Ex. 2:
1. Although John watched the TV morning forecast announcing rain,
2. he left home without an umbrella.
where he in 2, a nucleus, refers to [John] introduced in 1, a left satellite of it.
c). A right satellite of a nucleus u is not accessible from another, more distant, right
sibling of u, nuclear or satellite: in combinations (u
or (u
, u
can refer to u
but not to u
Ex. 3:
1. John told Mary that he loves her.
2. He was never married
3. and lived until 40 with his mother.
4. She, on the contrary, was married twice.
The sequence 2-3-4 ELABORATES on 1. The sequence 2-3 is in a relation of
CONTRAST (a paratactic relation) towards 4, while unit 3 ELABORATES on 2. The
structure is therefore: u
. For most readers, she in unit 4 must
be [Mary], and not [John’s mother], although [John’s mother] is the most recent
entity from the position of unit 4 in agreement in gender and number with the pronoun
she. The reason why the reader prefers Mary instead of the mother is because s/he
recognizes the unit 4 as being in a CONTRAST relation with unit 2 (evidenced by on
the contrary), which makes the two units to be perceived as adjacent, and having the
same status with respect to a common nucleus, unit 1. Their proximity however is not
linear but hierarchical, on the structure. This makes unit 3 to be closed for reference
from unit 4, and the pronoun she in 4 to find its antecedent in the common upper
nucleus – unit 1.
d). A nucleus blocks the reference from a right to a left satellite: in combinations
, u
can refer to u
but not to u
Ex. 4:
1. With one year before finishing his mandate as president of the company,
2. Mr. W. Ross has begun to bring about its bankruptcy.
*3. There were rumors that he has obtained it by fraud.
In this example the reader is confused on who the referent of the pronoun he in unit
3 could actually be. 1 and 3 are both satellites of unit 2: 1 is in a CIRCUMSTANCE
relation towards 2, while 3 is intended to give a BACKGROUND for 2, if it would be
perceived as referring [the mandate as president of Mr. Ross]. However this
coreferential link is found with difficulty, which lowers the understandability of the
whole discourse. It can be repaired in two ways:
Ex. 5:
1. Mr. W. Ross has begun to bring about the bankruptcy of his company.
2. with one year before finishing his mandate as president.
3. There were rumors that he has obtained it by fraud.
In Ex. 5, unit 2, expressing the positioning in time of the action expressed by the
main clause, unit 1, is a satellite of 1, and unit 3, reproducing a gossip occasioned by
an element introduced in unit 2, is a satellite of 2. The reference it=[Mr. Ross’
mandate as president] can be recuperated without difficulty. The motivation for the
failing of Ex. 4 compared to the acceptance of Ex. 5 stays not in the linearly longer
distance between the anaphor and antecedent in Ex. 4 than in Ex. 5, but in the fact that
a nuclear unit is interposed between the unit of the anaphor and the unit of the
antecedent in Ex. 4, contrary to Ex. 5 where this situation does not occur.
If the reference is eliminated, then the discourse is also repaired:
Ex. 6:
1. With one year before finishing his mandate as president of the company
2. Mr. W. Ross has begun to bring about its bankruptcy.
3. There were rumors that he has been elected by fraud.
3 VT’s Basics
The fundamental intuition underlying the unified account on discourse structure and
accessibility in VT is that the RST-specific distinction between nuclei and satellites
constrains the range of referents to which anaphors can be resolved. In other words,
the nucleus-satellite distinction, superimposed over a tree-like structure of discourse,
induces for each anaphor a dea. More precisely, for each anaphor x in a discourse unit
u, VT hypothesizes that x can be resolved by examining discourse entities from a
subset of the discourse units that precede u. If the x’s antecedent belongs to a unit that
resides beyond the dea of u, then the link anaphor-antecedent is found with difficulty
or, in order to realize it, strong referential means should be surfaced (as for instance
proper names).
The discourse structure assumptions in VT are, to a great extent, the same as in
RST: a) the basic units of a discourse are non-overlapping spans of text, usually a
clause of a sentence (expressing an event, or a situation); b) discourse structures are
represented as trees. Unlike RST, in VT, without any loss of generality, the trees are
considered binary; a similar representation is used by Marcu [19]; c) terminal nodes
of the tree represent elementary discourse units (edus) and non-terminal nodes
represent discourse relations. Unlike RST, VT is not concerned with the type of
relations among textual spans, but considers only the topological structure of the
discourse; d) a polarity, established among the daughters of a relation, identifies at
least one node as nucleus, considered essential for the writer’s purpose; non-nuclear
nodes, which include spans of text that increase understanding but are not essential to
the writer’s purpose, are called satellites.
Vein expressions defined over a discourse tree are sub-sequences of the sequence
of units making up the discourse. To define vein expressions, the following notations
are used:
each terminal (leaf) node (discourse unit) has an attached label;
) is a function that takes a string of symbols
and returns each symbol in
marked in some way (e.g., within brackets);
) is the reverse function of mark(). It removes all markings attached to
symbols in the expression
. (e.g. unmark(
. mark(
) . γ) =
. γ);
simpl(x) is a function that eliminates all marked symbols from its argument, if they
exist, e.g. simpl(mark(
)) = ø, the empty string, and simpl(
· mark(
) · γ)) =
· γ;
) is a sequencing function that takes as input two non-intersecting strings of
terminal node labels,
, and returns that permutation of
concatenated with
that is given by the left-to-right reading of the sequence of labels in
the terminal frontier of the tree. The function maintains the markings, if they exist
and seq(ø,
) =
; seq(
, seq(
)) = seq(seq(
) = seq(
H(n) and V(n) are the notations for the head and vein expressions of a node n;
) retains the prefix of the expression
up to and including the symbol u.
VT computes two expressions that are attached to all nodes of a discourse
structure. The notion of head in VT is equivalent to that of Marcu’s promotion set
[19]. The intention in the head expression of a node of a discourse tree is to capture
the sequence of the most important units in the span of text covered by the node. It is
a sequence of unit labels as follows:
1. The head of a terminal node is its label.
2. The head of a non-terminal node is the concatenation of the heads of its
nuclear daughters.
Note that the recursive definition of head induces a bottom-up computation over
the tree structure.
The vein expression of a node is intended to give the sequence of edus which are
significant for summarizing, in the context of the whole text, the span of text covered
by the node. In the vein expression of any node in the discourse structure, are
included edus belonging to the span covered by the node, possibly together with edus
outside the span. By synthesis, or summary, of a text span we understand a shorter
text, which still renders the original idea of the text. Irrespective whether it is realized
by paraphrasing or by concatenating sub-sequences of the original text [16], any
summary should be comprehensible by itself (among other things, this means that it
should contain all elements that allow the resolution of anaphors). When the span to
be summarized is extracted from a larger span, in order for the summary to be
comprehensible, it should contain also elements from outside the span, which belong
therefore to the context. We have, in this case, the summary of a text span, in the
context of a larger span. Let’s note also that, in many respects, “summarizing” is
equivalent to “understanding” because what we are usually left after the reading of a
text is a synthesis of it.
In the following, the whole text is called total context. In Fig. 1, the nodes to which
the definition currently applies are depicted in grey. They are simultaneously drawn
with a rectangle and a circle in order to suggest that they can be either inner nodes or
terminal nodes.
Once each node of the tree is marked for the head expression, vein expressions are
computed top-down for each node in the tree:
1. The vein expression of the root is its head expression.
The vein expression of the root node, conforming to the intention associated to the
vein expression of a node, should be made of the most significant edus that are
necessary to understand/summarize the span covered by the node (in this case – the
whole text), in the total context. But, since the covered text span in this case is the
whole text, this gives us the definition of the head expression of the root node.
2. For each nuclear node whose parent node has a vein v:
a) if the node does not have a left non-nuclear sibling, then its vein expression is
v (see Fig. 1a);
b) otherwise, if the left non-nuclear sibling has the head h, then the vein
expression of the nuclear node is seq(mark(h), v) (see Fig. 1b).
The definitions say that in order to understand/summarize, in the total context, a
nuclear span, a right satellite sibling can be ignored, while a left satellite is significant.
When positioned at the right of a nuclear unit, a satellite can be ignored, since the
same units are necessary to understand/summarize, in the total context, the nuclear
span plus the satellite span, or only the nuclear span. When positioned at the left, a
satellite helps to understand/summarize its right nucleus, but should be ignored for
any other right satellite of this nucleus (case commented in Ex. 4). The marking
function mark signals the contribution of this left satellite, in order that a subsequent
removal is operated in the vein expression of a right satellite (see 3b below). On the
contrary, twin nuclei cannot be understood/summarized one without the other,
meaning that the same units are significant to understand/summarize each one of them
as their union span.
3. For each non-nuclear node of head h whose parent node has a vein v:
a) if the node is the left daughter of its parent, then its vein expression is seq(h,v)
(see Fig. 1c);
b) otherwise, the vein expression is seq(h, simpl(v)) (see Fig. 1d).
The definitions express the fact that in the understanding/resuming, in the total
context, of a satellite span, one should add to the units that contribute to the
understanding/resuming of its parent node the most important units within the satellite
span itself (given by the sequence of units in its own head expression). Let’s note that
the vein expression of the parent node of this satellite, with one exception, inherits
only head expressions of nuclear nodes from its own ancestors, therefore the
significant units belonging to the satellite own span cannot be there and must be
included explicitly. The exception mentioned refers to exactly the case when a
satellite is placed on the left side of the nucleus towards which this node is itself a
satellite, and whose units have been recorded by markings. The simpl function will
delete this influence (see an example in Fig. 2).
V=seq(mark(h), v)
V=seq(h, v)
V=seq(h, simpl(v))
Fig. 1. Computing vein expressions. The node to which the computation applies is
depicted in dark; nuclei are underlined.
=seq(v, mark(h
, v)
Fig. 2. Simplifications in the computation of the vein expression of a right satellite:
, simpl(V
, simpl(seq(v, mark(h
)))) = seq(h
, seq(v)) = seq(h
, v).
4 The Relationship Between Discourse Structure And
If we particularize the intuition behind the vein expression to a terminal node, we
obtain: the vein expression of a terminal node u gives the sequence of edus that are
significant for understanding/summarizing u in the total context. Among other things,
which we will not discuss in this paper, this means that, within the material indicated
by the vein expression of an edu, antecedents of all anaphors belonging to that edu
must be found. More precisely, seen: – the semantic nature of the anaphoric relation
[13], – a representation of anaphoric relations in which res of a textual layer are
linked to representations of des on a semantic layer, as the one proposed by Cristea
and Dima [4] – and the common cognitive nature of anaphora and cataphora (as
discussed in section 2), which allows for a unique directionality in the search for
antecedents, always towards the beginning of the text, we are lead to the definition of
a domain of evocative referential accessibility (on short domain of evocative
dea(u) = pref(u, unmark(V(u)).
The definition of dea formalizes the first conjecture of VT (or the cohesion
conjecture), which defines for any discourse unit a specific domain of accessibility
computed in relation with the discourse structure: antecedents of the res belonging to
an edu u are mostly found among the des anchored in the edus which precede u in its
vein expression, including u itself.
The first conjecture hypothesizes two types of anaphoric processes: evocative (or
imediate) and post-evocative (or inferential). The evocative processes appear most
frequently, are resolved quickly and can be realized at the surface by any referential
material, including the most fragile, as empty subjects and pronouns. They give
fluency to the text and make it cohesive. The post-evocative processes are less
frequent, need a greater inferential load for their resolution and make use of strong
referential material (as proper nouns).
If we transfer this classification to the anaphoric references involved in these
processes, we will have evocative and post-evocative references (see Fig. 3). In the
evocative references, the backward-looking chain of units anchoring res that are
referentially related intersects the dea of the anaphor’s unit in at least one more unit
than the anaphor’s unit itself. In post-evocative references this double intersection is
missing. In [3] and [6] the evocative references are further detailed in direct and
indirect. In direct references the second intersecting unit (looking backward from the
anaphor’s unit) is the linearly most recent one, counting from the anaphor’s unit,
anchoring the same de as the one referred by the anaphor (in case of coreference), or a
de that is anaphorically related to the anaphor’s de (in case of functional reference). In
indirect references the two backward looking chains intersect in a unit that is not
linearly most recent from the anaphor’s unit.
Sometimes an anaphor belonging to the post-evocative class can be understood
without even having to make a connection to an antecedent. These are usually called
pragmatic references or pseudo-references. The interpretation of res in this class can
be made based on knowledge that comes from outside the test, from common
knowledge. Although the text contains at least one more re that realizes the same de
as the anaphor, the coreferential expressions may not be represented identically in
order for the text to be understood.
Fig. 3. Evocative and post-evocative references. Anaphoric chains are depicted by dotted-lines
and dea chains b
thick lines. The ana
hor’s unit is the last one to the ri
5 Discussions
The fundamental assumption underlying VT is that an inter-unit reference is possible
only if the two units are in a structural relation with one another, even if they are
distant from one another in the text stream. Furthermore, inter-unit references are
rather to nuclei than to satellites, reflecting the intuition that nuclei assert the writer’s
main ideas and provide the main “threads” of the discourse [17]. This is shown in the
computation of veins over (binary) left polarized discourse trees, where any reference
from a nuclear unit must be to entities contained in linguistic expressions from the
previous nuclei (although perhaps not any nucleus). On the other hand, satellites
depend on their nuclei for their meaning and hence may refer to entities introduced
within them.
Given the mapping of Grosz and Sidner's [12] stack-based model of discourse
structure on RST structure trees outlined by Moser and Moore [21] and Marcu [18],
the domains of referentiality defined for left-polarized trees using VT are consistent
with those defined using the stack-based model. However, in cases where the
discourse structure is not left-polarized, VT provides a more natural account of
referential accessibility than the stack-based model. In non left-polarized trees, at least
one satellite precedes its nucleus in the discourse and is therefore its left sibling in the
binary discourse tree. The vein definition formalizes the intuition that, in a sequence
of units A B C, where A and C are satellites of B, B can refer to entities in A (its left
satellite), but the subsequent right satellite, C, cannot refer to A due to the
interposition of the nuclear unit B. In stack-based approaches to referentiality, such
configurations raise problems: as B dominates A, B must appear below A on the stack,
even though it is processed after A. Even if the processing difficulties are overcome,
this situation leads to the postulation of “right” references of cataphora included in
satellites that precede their nuclei, which is counter-intuitive.
Inferential references, as defined by VT, seem to minimize the importance of the
domain of referential accessibility, because references can now “escape” from the
domain. Does the domain of accessibility have any significance anymore? Is it an
artificial invention or is it defended by a natural characteristic of the manner people
process texts? We claim that there are two significantly distinct types of anaphora
resolution processes: evocative (or associative) and post-evocative (or inferential).
The evocative resolution processes are based on associations, which are processes
of pattern-matching on feature structures decorated with morpho-semantic attributes.
They are performed between a feature structure projected by the anaphor re and a de
that already exists in the dea of the unit the anaphor belongs to [4]. These are fast
processes, direct ones being faster and more frequent than indirect ones. When
hierarchical adjacency is considered, an anaphor may be resolved to a referent that is
not the closest in a linear interpretation of a text. Because co-referential expressions
are organized in equivalence classes, it is sufficient if an anaphor is resolved to some
member of the set. This is consistent with the distinction between direct and indirect
On the other hand, the post-evocative processes are inferential processes that are
developed in memory, based on the knowledge accumulated by the preceding
discourse, or based on the cultural knowledge the subject owns. We believe these
inferences swing the semantic space in an order that is also dictated by the discourse
structure. Eventually, the target entity can be found based on a pattern-matching
process between the projected structure of the anaphor and the semantic
representation of the antecedent. They are slow – computationally and cognitively
(compel to more inference load), require more powerful referencing means (like
proper nouns), and are less frequent.
An aspect not described in this paper is VT’s account on discourse coherence [7].
Starting from deas, the notion of segment in a hierarchical sense is introduced, which
generalizes the classical notion of segment as employed in AST [12] and CT [11]. By
this, VT generalizes CT from a local theory of coherence to a global one.
Empirical evidences on the VT’s claims on cohesion and coherence have been
reported in [6], [7] and [14] with experiments developed on corpora annotated to
discourse structure and coreferentiality in English, French and Romanian. In
particular, these studies reveal the following: in most cases the references are direct;
in less cases the references are indirect; in very few cases the references are
pragmatic; inferential references which are not pragmatic signal a hard-to-make
inference or a failed discourse. Moreover, it can be proved that VT’s assumptions
regarding the cohesion are stable to the change of granularity (the limit below which
material edus are considered) from lower to upper.
A side effect of corpus research motivated by the evaluation of VT claims was the
notice that there is a strong relationship between the different kinds of referential
expressions and their distribution with respect to the three kinds of references put in
evidence by VT. It was revealed an alignment between the evoking power and the
percentage of different types of referential expressions that did not corresponded to a
vein reference (inferential). Four types of inferential references have been discovered:
pragmatic, proper nouns, common nouns and pronouns, which revealed to have
descending frequencies, in this order. Pragmatic and proper nouns references are
easily resolved, which makes their use much less restricted by the placement of an
antecedent on a current dea. At the other pole, pronouns are very fragile evoking
means, and, as such, a message emitter employs them when s/he is certain that the
current structure of the discourse allows for easy recuperation of the antecedent on the
dea of the anaphor. The alignment of the evoking power of referential expressions
with the percentage of exceptions of references outside the deas shows that the
predictions made by VT in the cohesion conjecture are correct. Practically, except for
the cases when the pronoun can be understood without an antecedent, it becomes
impossible to use a pronoun as an anaphor to refer an antecedent that is outside the
Scholars dealing with the interpretation of discourse and reading in connection
with the cognitive science [2], [15], [23], [26] generally, agree on three types of
memory: immediate memory (IM), short term memory (STM) and long term memory
(LTM). IM is a sensorial storage of information, which allows the retaining of traces
from the last half second. STM keeps information for few seconds. According to
Miller [20], the length of this memory seems to be of 7±2 signs (words, figures,
letters – depending on the context), while others estimate this ”buffer” to an average
of 13-15 words [22]. In [9] and [10] an incremental discourse parsing model is
described in which the developing structure is updated with a new auxiliary tree after
the reading of each sentence. The discourse tree becomes bigger and bigger as the text
In the human memory, as well as in automatic discourse parsing systems,
summarization processes must evolve in parallel with the building of the discourse
structure. We believe that the STM should be linked to the dea of the last edu
processed: either the last 7±2 edus in this sequence, or the same number of event
structures – as representations of edus, or only words picked up from this buffer.
When we replace the current unit u
with the next unit u
, actually we replace the
STM dea(u
) with the STM dea(u
). Sometimes this means a simple prolongation of
the preceding dea, other times it means the deletion of the most distant in time unit
and the inclusion of a new unit – the current edu. STM is therefore made of a chain of
edus (or of microstructures corresponding to edus), which is projected from the
dynamic evolving discourse structure. The alterations affecting the STM string reflect
the updates of the sub-discourse in focus, while reading. When the interest has moved
along another direction, the content of the current vein and, consequently, of the
current dea, is updated too. The inclusion and deletion from STM of certain mini-
structures, therefore these “recall” and “oblivion” processes, resemble the calling in
attention of Walker’s [26] cash memory model. The recall processes are possible from
the discourse structure that is kept in a summarized form in the LTM. Evocative
anaphoric processes are thus developing in the STM, while post-evocative processes
are of an inferential type, and necessitate greater inference load to recover des from
memory or evoke entities kept in the generic cultural sphere of the individual. We
belief these processes evolve also on the developing discourse structure, but leaving
the dea when the resolution failed there.
There are as many ways to read a text as there are edus in it. These different
readings are given by the edus’ vein expressions. Each vein represents a summary of
the text focused on the respective unit. When the reader is focused on a certain
episode or entity mentioned by the text s/he can skip entire fragments and look for the
manner in which the element of interest integrates in the whole discourse. Summaries
focused on different events or entities can contain elements in common, while each of
them has also specific elements, although strongly correlated to the main line of the
discourse. All these sub-discourses are coherent and, generally, there are no anaphoric
references whose interpretation to necessitate elements outside the summary itself.
We believe that the processes of anaphora resolution and discourse structure
building are interdependent to such a degree that discourse analysis should make use
of them in tandem, and combine their partial results to acquire the best discourse tree.
In the same way that anaphora resolution can benefit from the discourse structure,
already solved anaphora can be used in determining the best structure, which in turn
contributes to the resolution of further anaphora. The constraints evidenced act as
forces that, in a well-understood discourse, give rise to a sort of state of equilibrium,
resembling the minimum potential energy of a physical system. Humans have an
innate cognitive mechanism that allows them to obtain naturally the most plausible
interpretation of a text. When arrived there, they are invigorated by the reach of a
“comfortable” mental state, which should be based on the maximal satisfaction of a
constraints system. In [9], a model and an implementation that mimics this behavior
are described. Scores contributed by the cohesion conjecture are combined with
scores contributed by the coherence conjecture of VT (hierarchical generalization of
CT) in order to obtain the most “fluid” possible discourse structure (maximum of
cohesion and of coherence).
VT’s account on the relationship between discourse structure and referentiality can
be exploited in three ways:
to constrain a simultaneous parsing and anaphora resolution process towards that
interpretation that requires minimum inferential load in building the structure and
in identifying the antecedents of referential expressions [3], [5], [8], [9];
to correct discourse structure when referential links are known [25];
to guide a process aimed at producing focused summaries [9], [10].
The notice that slightly modified texts can display the same vein structure
(although not the same tree structure) can lead to the idea that veins could be seen as a
kind of sub-specification representation [24], a direction which has not been
investigated yet. Also, as trees annotated at discourse structure and veins can lead to
almost instantaneous computation of focused summaries on any discourse entity or
event mentioned in the text, it would be worth investigating an RDF representation of
vein structures obtained by processes of automatic parsing, with interesting
applications in Semantic Web.
1. Brennan,S.E.; Walker Friedman,M. and Pollard, C.J.: A centering approach to pronouns.
Proc. of the 25
Annual Meeting of ACL, Stanford (1987) 155-162
2. Cornea, P.: Introduction in the Theory of Reading (in Romanian), Polirom Publishing
House, Iaşi (1988)
3. Cristea, D.: An Incremental Discourse Parser Architecture. D. Christodoulakis (Ed.)
Proceedings of the Second International Conference - Natural Language Processing -
NLP 2000, Patras, Greece Lecture Notes in Artificial Intelligence 1835, Springer (2000)
4. Cristea, D. and Dima, G.E.: An Integrating Framework for Anaphora Resolution.
Information Science and Technology, Romanian Academy Publishing House, Bucharest,
vol. 4, no. 3 (2001)
5. Cristea, D., Dima, D.E., Postolache, O.D., Mitkov, R.: Handling complex anaphora
resolution cases. Proceedings of the Discourse Anaphora and Anaphor Resolution
Colloquium, Lisbon, Portugal (2002)
6. Cristea, D., Ide, N., Marcu, D., and Tablan, M.V.: Discourse Structure and Co-Reference:
An Empirical Study. Proceedings of the 18th International Conference on Computational
Linguistics COLING'2000, Saarbrueken (2000)
7. Cristea, D., Ide, N., and Romary, L.: Veins Theory: A Model of Global Discourse
Cohesion and Coherence. Proceedings of the 17th Coling and the 36th Annual Meeting of
the ACL (COLING-ACL'98). Montreal, Canada, (1998) 28185
8. Cristea, D. Postolache, O.D., Dima, D.E., Barbu C.: AR-Engine – a framework for
unrestricted co-reference resolution. Proceedings of the LREC’2002, Las Palmas, Spain
9. Cristea, D., Postolache, O., Pistol, I.: Summarisation through Discourse Structure. In
Proceedings of CiCling 2005, Springer LNSC, vol. 3406 (2005)
10. Cristea, D., Postolache, O., Puşcaşu, G., Ghetu, L.: Local and global information exploited
in producing summaries. Proceedings of the International Symposium on Reference
Resolution and Its Aplications to Question Answering and Summarization, Venice, Italy,
June (2003)
11. Grosz, B.J.; Joshi, A.K. and Weinstein, S.: Centering: A framework for modeling the local
coherence of discourse. Computational Linguistics, 12(2), (1995) 203-225
12. Grosz, B.J., and Sidner, C.: Attention, intentions, and the structure of discourse.
Computational Linguistics, 12(3), (1986) 175204.
13. Halliday, M.A.K. and Hassan, R.: Cohesion in English, Longman, London and New York
14. Ide, N., and Cristea, D.: A Hierarchical Account of Referential Accessibility. Proceedings
of the 38th Annual Meeting of the Association for Computational Linguistics, ACL'2000,
Hong Kong (2000)
15. Kintsch, W. and Van Dijk, T.A.: Comment on se rappelled et on résume des histories,
Langages, 40 (1975)
16. Mani, I.: Automatic Summarization, Amsterdam, John Benjamins (2001)
Mann, W.C., and Thompson, S.A.: Rhetorical Structure Theory: Toward a Functional
Theory of Text Organization. Text 8(3), (1988) 243281
18. Marcu, D.: A formal and computational synthesis of Grosz and Sidner's and Mann and
Thompson's theories. Proceedings of the Workshop on Levels of Representation in
Discourse, Edinburgh (1999)
19. Marcu, D.: The theory and practice of discourse parsing and summarization, The MIT
Press, Cambridge, Massachusetts (2000)
20. Miller, G.: The magical number Seven, Plus or Minus Two: Some Limits on Our Capacity
for Processing Information, The Psychological Review, vol
. 63, (1956) 81-97
21. Moser, M., and Moore, J.D.: Toward a synthesis of two accounts of discourse structure.
Computational Linguistics, 22(3), (1996) 409419
22. Richadeau, F.: La lisibilité. Langage-Typographie-Signes-Lecture, Paris (1969)
23. Schank, R. and Abelson, R.: Scripts, plans, goals and understanding, Hillsdale, N.J (1977)
24. Schilder, F. Robust Discourse Parsing Via Discourse Markers, Topicality and Position.
Natural Language Engineering 1, (1), (2001) 1-22
25. Sereţan, V. and Cristea, D.: The Use of Referential Constraints in Structuring Discourse.
Proceedings of the LREC’2002, Las Palmas, Spain (2002)
26. Walker, M.A.: Limited attention and discourse structure. Computational Linguistics,
(1996) 22-2