
Pascale Sébillot 280
.
CC: Creative Common License, 2006
that define a word. In one of the components of this model, called the qualia structure, words are
described with semantic roles such as the purpose or function (e.g., cut for knife), or the creation mode
(build for house), etc. For a given word, each role can get numerous realizations, and the qualia structure
of each word, especially for common nouns, is mainly made up of verbal associations. Such N-V pairs, in
which V plays one of the qualia roles (function, creation mode, etc.) of N, are called qualia pairs
hereafter.
Thus, in our ILP framework, the concept to be learned is the qualia nature of a N-V pair occurring within
a sentence. Using a ML technique is especially well-suited here both because the extraction patterns for
qualia relations are not known, and statistical co-occurrence-based methods have been proved not
satisfactory for this task (see (Bouillon et al. 2002) and section 3.3). An example (resp. counter-example)
corresponds to a N-V couple manually indicated by an expert as verifying (resp. not verifying) the target
qualia relations in one sentence of a our corpus. For example, the pair (screwdriver-screw) in its context:
“The operator uses the screwdrivers to screw…”
3
can be used to code an example, whereas (tyre-
prescribe) in its context: “Inflate the tyre to the prescribed pressure... ” can form a counter-example. For
our experiment, the corpus used is a collection of helicopter maintenance handbooks, provided to us by
MATRA-CCR Aérospatiale. This French 700 KBytes technical corpus contains more than 104,000 word
occurrences, and has been Part-of-Speech and semantically tagged (see (Bouillon et al. 2002) for details
concerning this second tagging). Hierarchies of PoS and semantic tags are parts of the knowledge
inserted in the background knowledge (for generalizations).
A GL expert has thus extracted from the corpus qualia and non-qualia N-V pairs with their contexts (all
the words and their tags occurring with the pair within a sentence). A
SARES, our ILP system, has been
given this way about 3,000 examples and 3,000 counter-examples.
The ILP system automatically infers rules, i.e., extraction patterns for the target relations like:
is_qualia(N,V) :- precedes(V,N), near_verb(N,V), infinitive(V), action_verb(V). which means that a pair
composed by a noun N and a verb V will be considered as qualia if N appears in a sentence after V, N
and V are not far from each other (especially not separated by a verb), and V is an action verb in the
infinitive.
The nine rules obtained are given below:
1. is_qualia(N,V) :- precedes(V,N), near_verb(N,V), infinitive(V), action_verb(V).
2. is_qualia(N,V) :- contiguous(N,V).
3. is_qualia(N,V) :- precedes(V,N), near_word(N,V), near_verb(N,V), suc(V,X), preposition(X).
4. is_qualia(N,V) :- near_word(N,V), sentence_beginning(N).
5. is_qualia(N,V) :- precedes(N,V), singular_common_noun(N), suc(V,C), colon(C), pred(N,D),
punctuation(D).
6. is_qualia(N,V) :- near_word(N,V), suc(V,C), suc(C,D), action_verb(D).
7. is_qualia(N,V) :- precedes(N,V), near_word(N,V), pred(N,C), punctuation(C).
8. is_qualia(N,V) :- near_verb(N,V), pred(V,C), pred(C,D), pred(D,E), preposition(E),
sentence_beginning(N).
9. is_qualia(N,V) :- precedes(N,V), near_verb(N,V), pred(N,C), subordinating_conjunction(C).
Those inferred rules are then used as qualia N-V pair extraction patterns to retrieve new qualia pairs
from the corpus.
3.3 Results and discussions
As explained in (Bouillon et al. 2002), using the produced patterns to extract qualia N-V pairs from the
corpus gives good results. The evaluation has been conducted on a test-set formed by N-V pairs in which
N is one of seven domain relevant common nouns: (screw, nut, door, indicator signal, plug, cowl, cap)
4
.
3
For understanding reasons, all the examples given in this paper are translations of those obtained from our French corpus.
4
To prevent distortion of results, none of these common nouns were used as examples or counter-examples for the pattern induction
in the ILP system.