Open Information Extraction (OIE) aims to identify all the possible assertions within a sentence. Recent and thus the most e cient OIE-tools use the grammatical dependencies or the syntactic tree of the sentence to perform extraction. When they provide a wrong extraction it is mainly due to parsing errors. In this paper, we propose to handle these parsing errors before doing OIE itself. To achieve our goal we focus on multi-word expressions (MWE). They represent more than 45% of wrong extractions. We show how the MWE-problem can be handle in a given domain and how MWE-unbreakable property is a good lter for OIE.
In recent years, researchers have tackled the problem of Open Information
Extraction in di erent manner: from machine learning [
As mentioned in the introduction, we want to perform \Open" Information Extraction, but from sentences which describe a precise eld. This constraint makes us have pieces of information about terminology as described in next sections. Nevertheless, our ideas can be used in an \open" manner (MWE will be mainly named entities, formulae, etc.) and thus be compared to \traditional" OIE-systems.
During the recent years, many systems were developed to perform OIE. It is
the case of ReVerb [
Researchers commonly agree that parsing errors lead to major incorrect extractions in Open Information Extraction. In a sample set of sentences select from various regulatory texts in the eld of building engineering (see Sect. 4.1 for more details about the corpus), the percentage of errors due to MWE is 46:15% using CSD-IE. To handle problems caused by MWE is a thus relevant way to improve result of IE. Our solution to improve the quality of OIE-tools when they face the MWE-problem is a three-step operation: (i) Detect MWE, (ii) compress each MWE and (iii) expand each MWE at the end of the extraction step. This process is illustrated by Fig. 1.
For us, a MWE is every phrase which the meaning will be modi ed (even become meaningless) by the addition or the deletion of any of its word. Consequently, a domain term, an idiomatic expression, a phrasal verb, a named entity, a formula, a quotation etc. is a MWE. These examples of MWE make us foresee that MWE are more easily and reliably identi able in a given domain. One can have also domain-independent terms that are not related to the eld of study but are frequently found in the corpus. It is the case of operators (example: less than, less than or equal to, as much as), idiomatic expressions (example: \Loose your head", \Jump in feet rst"), units of measurements, etc. So, a set of MWE in a precise domain can be made up of the terminology of the eld and frequent terms. This last category of terms can be obtain by means of existing statistical methods and the help of human experts. At this stage we identify the MWE present in the original sentence. We thus have a list of possible MWE in our corpus (see Sect. 4.1 for more details).
The reason why precision of OIE-tools is a ected by MWE is that the latter is
considered by the former to be non atomic. Hence, to limit potential hazardous
fragmentation of expressions, we propose to extract information from a new
version of sentences where each MWE will have been replaced by a shortened
version. So, now the question is: how do we get this short version of MWE?
When trying to answer this question, we must have in mind that the shortened
sentences must always be semantically and syntactically correct to be
appropriately handled by OIE-tools. We propose the following steps for shortening a
MWE (using its syntactic parse tree):
1. if the MWE is a clause (list of labels for clauses is available in [
After this stage, we now perform OIE itself, which is the third step. This OIE is done by using existing OIE-systems. Consequently, the following steps come after OIE and take as input results of OIE, i.e triples.
Earlier in this work, we have pointed out a set of things which degrades precision of OIE-tools. We have focused on the problematic role caused by multi-word expressions. Now, we use the only characteristic of MWE to nalise our OIEprocess. Indeed, a MWE is unbreakable. Consequently, when a triple contains only a fragment of a MWE, it is considered as incorrect. This ltering is done before the expansion stage, so the MWE are in their \shortened" form.
After the OIE has been done from the shortened version of the sentence, we now have to reconcile extracted facts with the original (long) sentence. We then look into the list of the extracted facts to replace shorten version of MWE by their initial long form. This is the aim of this step. 4 4.1
After making some statistics on the factors which lead to incorrect Information
Extraction, we have decided to tackle the multi-word expressions-problem. The
rst step of the approach we propose is to identify them in the input sentence. Be
able to perform such identi cation implies to have a list of possible MWE. That
is why we hypothesize that the sentence describes the realities of a speci c eld
of interest. Actually, to know that we are working in a speci c domain implies
to have a good idea of the terminology of this domain. For our evaluation, we
have taken the list of terms (labels of the concepts in the eld) as the set of our
MWE. These terms have been obtained through a key terms extraction process.
1 An exhaustive list of labels for phrases is available in the Penn Treebank [
To perform OIE itself after preprocessing tasks, we have used ClausIE of
Del Corro and Gemulla [
CSD-IE performance in this domain-speci c corpus (58.26%) is less good than
in \open datasets" like the Wikipedia (70.0%) and New York Times dataset
(71.5%) [
We have seen that handle MWE, in a given domain, helps to improve OIE on sentences of that domain. However we see that our method to handle MWE has to be improved. Indeed 30% of remaining errors after the shortening of MWE are due to that operation. Indeed: { when we choose the rst noun of a noun phrase-MWE to replace this MWE it is not always its suitable representative. Indeed some nouns can sometimes be tagged as verb and thus a potential predicate (e.g: \ re" in the expression re extinguisher, \means" in the term means of access, etc.) and it can cause wrong extractions. Consequently, when we have more than a noun in a noun phrase, we must have more criteria to choose the representative. 2 http://www.alchemyapi.com/ { In some sentences, parsers correctly identify the prepositional modi ers of all verbs, nouns, adverbs, etc. Consequently the presence of MWE is a priori not a problem for OIE-systems. Unfortunately, the deletion of prepositions (found for example in a noun phrase-MWE) during the shortening may lead to parsing errors. Indeed, parsers will try to identify new relations which may be wrong leading to incorrect extractions as illustrated below: 1. Original sentence : \A stair is a xed means of access." 2. Shortened version : \A stair is a xed means." 3. OIE : CSD-IE!(A stair, means, is) & ClausIE!(A stair, a xed means) One of the possible solutions to avoid the shortening of MWE to escape from their multi-word problem is to replace a MWE by a synonym. Ideally, this synonym should have less words (why not a single word?) than the original MWE. Such synonyms could be found in Linked Open Data or in lexical databases like Wordnet, etc. 5
The goal of this work is to see how we can leverage domain knowledge (mainly terminology) to improve Open Information Extraction. We have focused on multi-word expressions which cause more than 45% of errors in existing OIEtools. We have thus decided to reduce their length by proposing a shortening algorithm for multi-word terms. First results of our approach are very promising. In our goal to take advantage as much as possible of domain knowledge we can go further in facts ltering. For instance, simple domain and range constraints could help in detecting wrong facts. When we take the following fact (building, has width, door), we can state that it is incorrect by exploiting the fact that the range of the \predicate" has width is xsd: oat.