Incorrectly assigns references to citations

[themeo]

Cermine seems to incorrectly assign references to citations, even for perfectly recognized PDFs, and for correctly recognized bibliography. A citation co-authored by X is being assigned to other publications also including author X that occur in the bibliography. This is a very frequent bug and it often happens for obvious cases where the name of the first author is different for the citation and for the actually assigned item in the bibliography.

I can provide examples with specific PDFs if required, although this behaviour should be very easy to elicit using any PDF that includes citations co-authored by the same person.

[dtkaczyk]

@themeo Thanks for reporting. Please provide some example files or links to them, so that I can reproduce this issue. I will take a look at it within a couple of days.

[themeo]

Under this link is one PDF and a corresponding .cermxml file: https://app.sugarsync.com/iris/wf/D9346138_08836977_601970

I traced just one citation: ref17 , corresponding to (Gathercole, Frankish, Pickering, & Peaker, 1999). It is possible that many other citations are mis-assigned. This is a representative situation for many of my PDFs.

Below is a list of all mentions of ref17 (with the corresponding line number in .cerxml):

155: correctly detected ref17, incorrectly ref28 199: incorrect detection of ref17 and one other (3 papers cited, 5 references assigned) 258: incorrect detection of ref17 and two other (2 papers cited, 5 references assigned) 308: incorrect detection of ref17 and one other (2 papers cited, 4 references assigned) 563: correctly detected ref17, incorrectly ref28 720: incorrect detection of ref17 and one other (2 papers cited, 4 references assigned)

hope this helps!

[dtkaczyk]

Thanks. In general this style of citing papers (by author names and years) is the most problematic one, due to matching issues and possible ambiguities.

According to manual inspection I did, originally in this paper we had precision of 72%. This refers to the fraction of reported citation-reference pairs that are indeed correct and seems to be the most informative measure in this case. I was able to increase the precision to nearly 90% (commit https://github.com/CeON/CERMINE/commit/2df02a858ac54bea5eb2f91e015a93a8e2efe45e). Further increase is more complicated and might result in correct extractions missing from the output, so I would leave it as is for now.

Please test it against your cases and let me know what you think.

[themeo]

Thanks, this indeed improved the matching accuracy.

With respect to this citation style and what can be improved - this is APA style - one of the most popular styles in scientific journals. It is supposed to be unambiguous. Going back to the examples I provided, CERMINE still isn't sure if:

(Gathercole, Frankish, Pickering, & Peaker, 1999) (ref17)

stands for this reference or for: (Masoura & Gathercole, 1999) (ref 28)

So in this case it ignores the order of the authors that is disambiguating the two citations.

It also fails to detect any reference in: (e.g., Masoura & Gathercole, 1999, 2005)

In this latter case I suspect the problem might be related to the fact that the references are not separated with a semicolon but with a colon - which admittedly is an exception to the rule.

I suspect these issues are solvable but probably going beyond regex and thus associated with extra time cost. But given the popularity of this style I think this effort might be worthwhile.

In any way, thanks for this amazing library!

[themeo]

Just to add to my previous comment. At least in my usage of CERMINE, both false positives and false alarms are bad (I extract phrases used to reference specific papers), so measuring accuracy with something akin to F1 would be a more appropriate index of usefulness in my application of CERMINE.

[dtkaczyk]

About the format: it is commonly used in social sciences and less popular in other disciplines. It general, I feel it is more human- than machine-readable format and is definitely more problematic for automated parsing than other styles. One of the problems is matching authors names, where you'll have issues with encoding, transliterations, accents, etc. Also there are a lot of variations of a citation (letters attached to the years for disambiguation, author names full or partial lists, using "et al", the way a few citations are combined together). It is not a trivial task to take all possible variants into account, and even then I am sure some things will be missed.

BTW, much easier style is for example IEEE, which uses numbers as identifiers of references within papers. It is unambiguous by design and in practice, plus it is so much easier to parse numbers that author names.

About the evaluation: I chose to calculate precision mainly because originally you reported precision-related issues (incorrectly assigned references). Also, after inspecting the example file it seemed to me precision is the biggest problem and the biggest area for improvement. I did not calculate recall, it seemed high to me, much higher than precision.

Improving the accuracy further at this point would require a different, more sophisticated approach. It would also require gathering a larger, representative set of papers, annotating citations within them, so that the solution and further variants can be evaluated against it. This way we could be sure that along with improving precision we do not decrease recall significantly. If you are interested in this task, maybe you could prepare such a set?

[themeo]

I'm afraid I'm too busy to do that. But I can jot down instances of incorrect assignment as I use CERMINE-provided data, which could be than used as test cases. Please let me know if that would be helpful for you.

[dtkaczyk]

I'm afraid it wouldn't be enough, and would still require a lot of manual work on my side. And I also have too much on my plate right now. I will close this issue, if there is time in the future, I'll reopen.