Named entity recognition
Named entity recognition is about recognising named entities, for example proper nouns, etc. in text.
There are several translation issues that can show up when there are unknown proper nouns in the input. One is that transfer rules that work on <np>-tagged words do not apply when the word is unknown. Another is that proper nouns can be ambiguous with other, known words, and thus be translated when they should stay untranslated.
Unknown proper nouns in transfer
When working with long rules, one of the problems in having them applied can be proper nouns. For example, names, companies, places etc. that aren't in the dictionaries and thus are not analysed. So for example in a sentence like:
- Die man het John gesien.
would be analysed something like (simplifying slightly):
- Die<det> man<n> hê<vbhaver> *John gesien<vblex><past>
If we have a rule that says something like:
- <vbhaver> <noun phrase> <vblex><past> → <vbhaver> <vblex><past> <noun phrase>
This will not apply, because "John" is not detected as anything. As a result the translation will be worse because the word re-ordering has not taken place. So, instead of getting:
- The man had seen John
We would get:
- The man had John seen.
Which is less than ideal. What we need is something that can tag "John" as a proper noun (<np>), so that the rules may be applied in the appropriate fashion.
Case handling
The problem becomes more acute in other language groups where proper nouns have cases. For example in Serbo-Croatian or Polish:
- Władysława → Władysław<np><ant><m><sg><gen>
and
- Marijom → Marija<np><ant><f><sg><ins>
Pipeline
It should probably go in between tagging and transfer, and work only on unknown words.
Disambiguating unknown proper nouns from known words
The module apertium-pn-recogniser detects proper nouns in the input and marks them as unknown words so that the rest of the modules in the pipeline do not process them. This avoids the common case of wrong translations of source-language proper nouns which are also common nouns (or other parts-of-speech) according to the dictionaries. The proper noun recogniser is mainly based on the one already included in the Freeling project.
The proper noun recogniser must be invoked between the tagger and the transfer modules. Option -p is needed in the tagger, so a version from Apertium greater or equal to 3.1.0 is needed.
Check out from SVN with:
svn co https://svn.code.sf.net/p/apertium/svn/trunk/apertium-pn-recogniser
Further reading
- Babych, B. and Hartley, T. (2003) "Improving machine translation quality with automatic named entity recognition ". Procs. EACL-EAMT 2003: Improving MT through other language technology tools, Budapest, Hungary, April 2003 .
