You are on page 1of 12

Structured Semantic Modeling of

Scientific Citation Intents


Authors: Roger Ferrod, Luigi Di Caro and Claudio Schifanella

Presentor: Thuy Pham Trong


Course: Knowledge Technology

03/03/2022 - Knowledge Technology - 1


1 Motivation

2 Background

CONTENT 3 What’s new?

4 Evaluation

5 Conclusion

03/03/2022 - Knowledge Technology - 2


Motivation Background What’s new? Evaluation Conclusion

Citations represent a fundamental mechanism Remaining problems


for both making new research and keeping track
of what is going on within a scientific field. The search for relevant information withing
large scholarly databases is becoming an
unaffordable task

Current proposed works remain at a


theoretical level only or simply reduces
problems to short-text classification

Some solutions are difficult to be employed


in concrete applications

Recent works with ML have limitations

03/03/2022 - Knowledge Technology - 3


Motivation Background What’s new? Evaluation Conclusion

Inheriting
Scaffold model for
CiTO* identifying citation intents

- An ontology written in OWL 2 DL


- Characterized the nature or types of
citations and permit these descriptions
to be published on the Web
- Contains the object property cito:cites
and sub-properties, inverse property
cito:isCitedBy
*: http://purl.org/spar/cito

03/03/2022 - Knowledge Technology - 4


Motivation Background What’s new? Evaluation Conclusion

Subject-oriented role in citations


Active role A relation of a certain intent between a source paper A and a
(A-subject) referenced paper B, focus on research objects in B

A relation between a referenced paper B (mentioned by the


Passive role
citing paper A) and some research objects presented by
(B-subject)
unknown third-party paper

03/03/2022 - Knowledge Technology - 5


Motivation Background What’s new? Evaluation Conclusion

Object and Context fields


Object as the minimal span of text to which the
quotation refers

Context is represented by a usually larger text


surrounding the object helping disambiguate or
specify it

To deeply evaluate the To leave room for future


model through inter- automated labeling
annotation agreements in technologies (Ontology
absence of physical/ Learning, Transformer-
narrow lexical constraints based ML…)

03/03/2022 - Knowledge Technology - 6


Motivation Background What’s new? Evaluation Conclusion

Semantic Structuring of Citation Intents

Proposes

“For faces, we used FaceScrub dataset from [CIT]” ---> A-subject


Uses context object

Intents Extends “This MMSE representation was used in [CIT], to prove the EPI” ---> B-subject
object context
Compares

Analyzes

03/03/2022 - Knowledge Technology - 7


Motivation Background What’s new? Evaluation Conclusion

Data selection and Annotation


276 instances per each class Annotation
Dataset 10000 papers from Semantic Scholar corpus - 3 annotators validated the class label
randomly and annotated the A, B – subject role,
object and context fields for each
Removed the sentences longer than 40 words
instance
- In order to calculate inter-annotation
agreement by using Bleu score

Notice: ordered list of the classes


Extend > analyze > compare > use > propose

To guide the disambiguation of ambiguous cases

03/03/2022 - Knowledge Technology - 8


Motivation Background What’s new? Evaluation Conclusion

Inter-annotation Agreement
Calculated by averaged the
scores obtained from three
pairs of annotations

Context has a less constrained


definition with object

Among citation intents,


propose resulted to be easy to
classify by the annotators

Averaged inter-rater agreement among


the three annotators for each citation intent

03/03/2022 - Knowledge Technology - 9


Motivation Background What’s new? Evaluation Conclusion

Compare with other dataset when classifying on the annotated data

Can be applied a neural network for


performing role classification. For
example, using biGRU architecture,
obtaining the better results as F1-score
77.6%, Precision and Recall of 80.6%
and 74.8% respectively.

Intent classification results on CiTelling and


ACL-ARC data

03/03/2022 - Knowledge Technology - 10


Motivation Background What’s new? Evaluation Conclusion

Summary:
- Proposed a new semantic representation for modeling citations within a corpus of scholarly articles
- Obtain state-of-the-art classification results with straightforward neural network models and
outperform with other dataset
- Enable the construction of a directed citation semantic graph (used in advanced analyses rather than
semantic web search applications)

Future works:
- Making the integration of further articles metadata
- Manage the sentences with multiple types of citations

03/03/2022 - Knowledge Technology - 11


Thank you for listening!

03/03/2022 - Knowledge Technology - 12

You might also like