Using DepPattern to Correct the PoS Tagged Input Text

DepPattern is provided with tools suited to correct errors of the input PoS tagged text. DepPattern allows a linguist to elabore syntactic rules in order to correct systematic mistakes made by the PoS tagger. For this purpose, we are provided with 3 new elements:

• A new type of dependency, “Head”, which represents a unary relation (arity 1). In the default configuration file, dependencies.conf, we declared one type unary relation, called “Single”.
• A new operation, “Corr”, whose aim is to correct all information associated to a lexical unit: type of PoS tag and morpho-syntactic features. It is similar to the operation “Add”. The main difference is that “Corr” allows to change the PoS tag itself.
• A new output format obtained using flag -c. Instead of generating as output the dependency triplets identified by the grammar (flag -a), we can use flag -c to rewrite the same input, but containing all corrections made by operations such as “Corr”, or “Inherit”, or “Add”.

Let's see an example. Suppose that the PoS tagger systematically tag as a subordinate conjunction the word that following a noun, even if in this context that is, in general, a relative pronoun. To solve the problem, we can write a rule as follows:

Single : [NOUN] CONJ<lemma:that&type:S>

Corr: tag:PRO, type:R

%

This way, the information introduced by the operator “Corr” is used to change the head expression of the unary relation “Single”. It substitutes tag PRO and type R for the information contained in the head (tag CONJ and type S). More precisely, this rule identifies as head a subordinate conjunction with lemma that following a noun (its context), and transform this head entry into a relative pronoun. Notice that there there is no dependent expression involved in the rule, since the relation type of “Single” is Head.

“Corr” also allows correcting attibutes by using the values of other attributes:

Corr: lemma:=token

It means that the value of the lemma is the value of the token. In other words, the lemma attribute inherits the value of the token attribute.