Most of scientific texts give arguments about a subject, that is, they give logic sentences ' well-formed sentences'.
I want to find out if the arguments of a text are really arguments about the subject of the text, or if that writing is only text that it has not relation with the subject, and that writing makes difficult to understand the reading in some cases
I like this question because it contains the word "automatically". I make a step furher and I ask if there are algorithms (or even implemented programs) able to analize a text and to find out which parts of the text do not link with the subject, which of them are contradicting each other, which of them are fallacious, etc. As far as Iknow there are no such algorithms, and the question posed like this goes into Artificial Inteligence, with the mention that AI is not so far and is still not able to deal with texts written by humans. There are some progresses made not for texts but for programs - the so called "verification tools". Some of them verify if a program and the same program after some modification still compute the same function for all arguments. However, it is known that the equivalence of programs (of Turing Machines) is an undecidable problem - that means, that there is no algorithm able to verify this in general, in a finite number of steps. I am pretty sure that your problem, posed algorithmically, is also undecidable.
There are however some nice empiric rules to filter out sentences which do not match with the subject. For example:
- Many autors start to introduce autobiographic elements. They tell the story of their evolution in contact with the subject, justify themselves why do they write so late or so early about something, etc. Most of such paragraphs are not helpful at all.
- Other autors make the same think discussing about biographic elements of some other guy, some celebrity or classic of the subject. Same effect.
- The story of the own illumination is in general not so interesting. In more than 90 % of cases that means only that the author still have doubts about the things she writes about.
- In a talk is very interesting when someone starts to tell jokes about some historical context. In an article this generates, in most cases, unimportant text.
- and so on...
The field that studies how to find automatically the arguments in a text is called 'Argument mining'. It's an exciting subfield of NLP, but very much in its infancy.
I suggest you to check the work of these people:
Chris Reed, Simone Teufel, Manfred Stede, Francine Moens, Patrick St. Dizier
A relevant journal is 'Argument and Computation'.
One important issue is that argument mining presupposes that you have a theory of what an argument is. This brings you to the field of Argumentation Theory. Your theory also needs to be quite sophisticated if you want to see whether arguments are really relevant to the issue being discussed as you seem to hint in your question. This means that you need a theory of argumentative relevance and of argument quality.
On Argumentation Theory I suggest reading the works of
Frans H. van Eemeren, Douglas Walton, James Freeman, Eddo Rigotti, Mark Vorobej.
Hi George,
Argument extraction and argumentation mining are a relatively recent, but also quite active and very complex area of research in NLP. In addition to the work of the people mentioned by Andrea (which are all very valid and helpful links), I would suggest to also look into the following:
Stab and Gurevych, 2014: Identifying argumentative discourse structures in persuasive essays.
Rahwan et al., 2011: Representing and classifying arguments on the Semantic Web.
Feng, 2010: Classifying Arguments by Scheme. (or Feng and Hirst, 2011)
Habernal et al., 2014: Argumentation Mining on the Web from Information Seeking Perspective
Ravenscroft, 2014: Digging by Debating - Final Report
Moens, 2014: Tutorial Argumentation Mining (Slides)
Stede and Sauermann, 2008: Linearization of arguments in commentary texts
Peldszus, 2014: Towards segment-based recognition of argumentation structure in short texts
Niekler, 2014: Creating dictionaries for argument identification by reference data
Just last year there have been two workshops on this topic, the proceedings of which you might want to review:
- ArgNLP 2014 Frontiers and Connections between Argumentation Theory and Natural Language Processing
- ACL 2014: Proceedings of the First Workshop on Argumentation Mining
And finally, here are links to some freely available corpora with training data:
(unfortunately, not all of the from the works above is readily available)
- AIFdb: http://www.arg.dundee.ac.uk/aif-corpora/
- NoDE: http://www-sop.inria.fr/NoDE/index.html
- UKP corpora: https://www.ukp.tu-darmstadt.de/ukp-home/research-areas/argumentation-mining/ (under "Resources")
HTH.
Cheers,
Martin