Once we have a group of transformed dependency trees, we aim at finding the best node alignment for those trees to further build a graph expressing all the content from the input.
We use a simple, fast and transparent method and align any two nodes provided that the nodes they contain
• are content words;
• have the same part-of-speech;
• have identical lemmas or are synonyms.
4.3 Constructing of a Dependency Graph 43
Mathematik und
an in
studierte
subj obja pp
pp
kon cj
pn
pn Bohr
Niels
app Universitaet Kopenhagen
der det Physik
(a) Original tree
Mathematik und studierte
subj obja an
in
kon cj Bohr
Niels app
Physik
Universitaet der
det
Kopenhagen
(b) AfterPREP
studierte
subj an
in
Bohr Niels
app
Universitaet der
det
Kopenhagen Mathematik Physik
obja obja
(c) AfterCONJ
studierte
subj an
in
Bohr
Universitaet der
det
Kopenhagen Mathematik Physik
obja obja Niels
(d) AfterAPP
studierte
subj an
in
Bohr
Universitaet Kopenhagen Mathematik Physik
obja obja Niels
(e) AfterFUNC
studierte
subj an
in
Bohr
Universitaet Kopenhagen Mathematik Physik
obja obja Niels
root s
(f) AfterROOT
studierte
subj an
in
Universitaet Kopenhagen Mathematik Physik
obja obja root
s
bio
(g) AfterSEM
studieren
subj an
in
Universitaet Kopenhagen Mathematik Physik
obja obja root
s
bio
(h) AfterLEMMA
Figure 4.3: The transformations of the dependency tree of the sentence in (4.6)
We prefer this very simple method to bottom-up ones (Meyers et al., 1996;Barzilay & McK-eown, 2005;Marsi & Krahmer, 2005) mainly for two reasons. Firstly, pursuing local subtree alignments, bottom-up methods may leave identical words unaligned and thus prohibit fusion of complementary information. On the other hand, they may force alignment of two unrelated words if the subtrees they root are largely aligned. Although in some cases it helps to discover paraphrases, it also considerably increases chances of generating ungrammatical output which we want to avoid at any cost. For example, even synonymous verbs such as say and tell have different subcategorization frames, and mapping one onto another would include the possibil-ity of generating *said him or *told to him. In case of multiple possibilities, i.e., in cases when a word from one sentence appears more than once in a related sentence, the choice is made randomly. It should be noted, however, that such cases are extremely rare in our data.
By merging all aligned nodes we get a dependency graph which consists of all the de-pendencies from the input trees. If the graph contains a cycle, one of the alignments from the cycle is eliminated. Root insertion during the transformation stage guarantees that the graph obtained as a result of alignment is connected. Recall the sentences (4.3) and (4.4) from Section4.1. Figures 4.4aand4.4bshow their transformed dependency trees and Figure 4.4c presents the graph obtained as a result of their alignment (nodes shared by the input trees are in blue).
The graph we obtain covers all the dependency relations from the input sentences. More-over, these are no longer relations between words but between entities and concepts as some nodes cover several words which may differ. For example, the node bio in Figure4.4c rep-resents an entity (namely, Niels Bohr) referred to with er and Niels Bohr in (4.3-4.4). Given that some of the tree transformations reveal implicit semantic relations, the graph is not purely syntactic but is also semantically motivated. Constructing such graphs is important because it brings sentence fusion one step closer to abstractive summarization which, as the reader might remember from the introduction, proceeds by “understanding” the text – i.e., creating a semantic representation for it – and by generating a summary from this representation.
Apart from the apparent advantages, the dependency graph representation also has a num-ber of serious disadvantages. For example, time and temporal relations are not treated prop-erly. Consider, e.g., a set of two similar sentences such as (4.7-4.8)3which concern the same activity (marriage) but two different events. From the graph emerging after alignment of the respective trees (see Fig.4.5) it is no longer possible to infer whom Albert Einstein married in which year.
(4.7) [...]
[...]
am on the
2.
2nd Juni June
1919 1919
heiratete married
er he
Elsa, Elsa,
die who
ihre their
T¨ochter daughters
Ilse Ilse
und and Margot
Margot mit
VERB-PREFIX in in
die the
Ehe marriage
brachte.
brought.
3See Files 10199 and 13199 in CoCoBi.
4.3 Constructing of a Dependency Graph 45
Abitur Schule Gammelholm
Physik
Mathematik Chemie
Astronomie Philosophie 1903
Universitaet Kopenhagen bio
studieren root s
nach an
in an
subj zeit
obja obja obja
obja obja
(a) The transformed tree of the sentence in (4.3)
studieren root s
und
erlangen dort
Doktorwuerde bio
Universitaet Kopenhagen subj
1911 s
an kon
adv
zeit obja cj
(b) The transformed tree of the sentence in (4.4)
und
erlangen dort
Doktorwuerde 1911 adv
zeit obja Abitur cj
Schule Gammelholm
Physik
Mathematik Chemie
Astronomie Philosophie 1903
Universitaet Kopenhagen bio
studieren root s
nach
an
in an
subj zeit
obja obja obja
obja obja
kon s
(c) The dependency graph obtained from the trees above
Figure 4.4: Transformed dependency trees of the sentences (4.3) and (4.4) and the graph built from them
heiraten root s
2. Juni 1919 Elsa am
Mileva Maric bio obja
subj obja 1903 zeit
bringen in die obja
Tochter
Ehe subj
s
rel
Figure 4.5: Graph built from the trees of the sentences (4.7-4.8)
’[...] on the 2nd of June 1919 he married Elsa with whom he had two daughters Ilse and Margot.’
(4.8) 1903 1903
heiratete married
er he
Mileva Mileva
Maric¸, Maric¸,
eine a
Mitsch¨ulerin classmate
am at the
Polythechnicum.
Polythechnicum.
’In 1903 he married Mileva Maric¸, a classmate at Polythechnicum.’
Thus, it is important that only sentences concerning the same event are aligned and grouped together, otherwise a sentence inconsistent with the input can be generated. Fortunately, even a shallow word-based alignment algorithm turns out to be robust enough in practice, and alignment of inherently different sentences is unlikely. Still, it is important to keep in mind this limitation of our graph-based representation. Other disadvantages are due to dependency syntax in general. For example, it is impossible to express that a certain subtree modifies the proposition as a whole and not just the verb because there are no non-terminal nodes in dependency grammar.