I would like to extract attributes of a table which are mentioned in Plain text. What is a best approach to be followed : Is it supervised, semi-supervised or unsupervised ?
I have some sample case studies but I have not had big training set.
below is example of a case study:
"Consider the following relational database for Fester Zoo. Fester Zoo wants to maintain information about its animals, the enclosures in which they live, and its zookeepers and the services they perform for the animals. In addition, Fester Zoo has a program by which people can be sponsors of animals. Fester Zoo wants to track its sponsors, their dependents, and associated data. Each animal has a unique animal number and each enclosure has a unique enclosure number. An animal can live in only one enclosure. An enclosure can have several animals in it or it can be currently empty. A zookeeper has a unique employee number. Every animal has been cared for by at least one and generally many zookeepers; each zookeeper has cared for at least one and generally many animals. Each time a zookeeper performs a specific, significant service for an animal the service type, date, and time are recorded. A zookeeper may perform a particular service on a particular animal more than once on a given day.
A sponsor, who has a unique sponsor number and a unique National Insurance number, sponsors at least one and possibly several animals. An animal may have several sponsors or none. For each animal that a particular sponsor sponsors, the zoo wants to track the annual sponsorship contribution and renewal date. In addition, Fester Zoo wants to keep track of each sponsor’s dependents. A sponsor may have several dependents or none. A dependent is associated with exactly one sponsor."
Any books or online resources are appreciated.
What you want to do sounds like what has been attempted in software/requirements engineering for the past 20 years or so where concepts and their attributes are extracted from natural language text. There have been a number of prototype tools, such as NL-OOPS, CM-Builder, CIRCE, NL2SBVR (and its related tools); unfortunately they are usually very old, have never left prototype stage, or are not generally available. The provided link is a paper that surveys a number of these tools.
The most common approach to what you want to do that I have seen requires no training as they are attempts to automate the object-oriented analysis and design techniques described in 'Object-Oriented Analysis and Design with Application', 2nd ed, by Booch (1994). (If I recall correctly there is a 3rd edition that is much more recent, but I haven't actually read it myself.) These approaches typically use the following pipeline:
- tokenisation and sentence splitting
- morpho-syntactic analysis (either separately or together depending on the parser being used), often using the Stanford (Dependency) Parser
- (optional) semantic-role labelling
After these generic steps, the approaches tend to use manually defined rules that map Nouns/Noun Phrases to concepts/entities, Verbs/Verb Phrases to relationships/attributes, etc. with varying complexity and sophistication. A couple of examples are [DOI: 10.1007/978-3-642-29958-2_15] and [DOI: 10.1145/1953355.1953378].
The thing to watch out for is whether or not constraints are extracted or not. But if all you want is to identify attributes, you could potentially do that with a reasonably simple set of rules. Also, from what I have seen, approaches like these only obtain a recall of about 85% at the most.
I hope that helps.
http://dx.doi.org/10.1007/s00766-010-0111-y
I am not sure if I understand the problem correctly as there are some elements missing. I would "disagree" a bit with the previous answer, but it really depends on what exactly you are looking for.
What I understand from the problem as you very briefly present it is that you are need to look into the domain of "Information Extraction" and in particular Named Entity Recognition and Relation Extraction. Information Extraction is often defined as an effort to transform information that is found in (unstructured) natural language text into structured representations, such as tables, databases, graphs, etc.
Since you are talking about tables, I would assume that you know what types of "things"/objects (named entities) you are looking for, as well as what type of relations you are looking for. If this is the case, I have a number of references I can give you. This would be a case where supervised or semi-supervised approaches would be most relevant. At the same time, looking at the example of textual data you gave (the case study), I am not sure I can identify the target information (Named Entities or Relations), so if it is the case that you do not really know what type of relations and entities you are going to find in your text, I would suggest unsupervised learning. Do a search for information extraction from wikipedia, you will find a lot of information.
If you need more information, please give some more details on the problem.
Dear Dialekti Valsamou I think you are closer about what I am going to do. You have mentioned that you have a number of references which you can give them to me. Please send me any referent references.
If you have access, start with some very fundamental ones like this one by R. Grishman (Information Extraction Techniques and Challenges), or R. Grishman's notes for the Tarragona winter school in 2012 which present the domain very well (link below). The latter provides a good bibliography section, additionally. I believe this is a good starting point, instead of a long list of specific papers.
If you need more information on a specific topic covered (or not) by these introductions, don't hesitate to follow up with more questions.
http://cs.nyu.edu/grishman/tarragona.pdf
Article Information Extraction: Techniques and Challenges
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining