Bilingual Semantic Role Labeling Inference via Dual Decomposition 15:5
Fig. 3. An argument output of monolingual SRL systems: location loc, label l, and classification probability p.
2.3. Output of Monolingual SRL System
One reason why we choose a maximum entropy classifier as our classifier is that it can
easily incorporate arbitrary features. The other reason is that the maximum entropy
classifier can directly output classification probabilities without any postprocessing
that will be used in our joint inference stage. For clear illustration, we wrap the
classification probabilities into the output of monolingual SRL systems. As shown in
Figure 3, an argument output of monolingual SRL systems has three attributes: its
location in the sentence loc, represented by the position of its first word and last word;
its semantic role label l; and its classification probability p. Therefore, we annotate an
argument output of a monolingual SRL system as a triple (loc, l, p). For example, the
A0 argument in Figure 3 is ((6, 6), A0, 0.85).
2.4. Complementarity of Chinese and English SRL Systems
The semantic equivalence of the parallel bi-texts means that they should have the
same predicate semantic structure, which implies that the aligned argument pairs in
bi-texts (e.g., the pair of ‘’ and ‘Recently’ in Figure 1) should have consistent labels.
Nevertheless, the monolingual SRL systems of Chinese and English work indepen-
dently. Thus, some inconsistent cases often occur on both sides of monolingual SRL
systems. However, there are many complementarity clues that are beneficial to correct
these mistakes. For example, in Figure 1, if ‘’ is missed or annotated with a wrong
label by a Chinese SRL system, but ‘Recently’ is labeled correctly by an English SRL
system, intuitively, with the help of the English side, we can recognize the argument
of ‘’ on the Chinese side and give it a correct label.
From these observations, we can see that there are mainly two types of complemen-
tarity that can be utilized in the bilingual SRL system.
Location Complementarity means that, if one of an argument pair is found by the
monolingual SRL system but the other is missed, we can track down the missed ar-
gument in the other language according to word alignment information. It is worth
noting that there are also some exceptional cases, for example, the argument ellipsis
on one side. These cases have little impact on the location complementarity, however,
because, in these cases, the aligned points are null in the other language.
Label Complementarity means that an aligned argument pair in a bi-text should
have consistent labels. For instance, in Figure 1, the labels for ‘’ and ‘Recently’ are
AM-TMP, which are consistent. However, if the label on one side is wrong but the label
on the other side is right, we obtain inconsistent results for both sides. In these cases,
we should abandon this result and search for a pair of consistent results. It is worth
noting that the consistent labels for the aligned argument pair do not mean that they
have to have the same label. For a brief illustration, an example is given in here. In
ACM Trans. Asian Low-Resour. Lang. Inf. Process., Vol. 15, No. 3, Article 15, Publication date: December 2015.