Identifying multidisciplinary problems from scientific publications based on a text generation method

Figure 1.

Figure 2.

Figure 3.

Comparison of stacking method and other methods in disciplinary classification_

Algorithm	Macro-Precision	Macro-Recall	Macro-F1
SVM	0.81	0.69	0.74
NB	0.64	0.77	0.68
LSTM	0.67	0.65	0.66
Stacking	0.81	0.79	0.80

Discipline distribution of the number of papers in the CPCN dataset_

Main category	Data volume of main category	First-level category	Data volume of first-level category
07 Science	1,917	0703 Chemistry	1,334
		0706 Atmospheric Sciences	583
		0805 Materials Science and Engineering	736
		0807 Power Engineering and Engineering Thermophysics	1,008
		0813 Architecture	638
08 Engineering	15,322	0817 Chemical Engineering and Technology	5,309
		0819 Mining Engineering	767
		0820 Oil and Gas Engineering	1,008
		0823 Transportation Engineering	750
		0828 Agricultural engineering	2,055
		0830 Environmental Science and Engineering	3,051

Examples of multidisciplinary research problems_

Multidisciplinary research problems	The first-level disciplines involved
Catalytic, Cracking, Hydrogenation	0703 Chemistry, 0817 Chemical Engineering and Technology, 0820 Oil and Gas Engineering
Oxidation, Desulfurization, Catalytic	0817 Chemical Engineering and Technology, 0820 Oil and Gas Engineering, 0830 Environmental Science and Engineering
Rare earths, Catalysts, Environmentally friendly	0805 Materials Science and Engineering, 0820 Oil and Gas Engineering
Coal Combustion, Flue Gas, Distribution	0817 Chemical Engineering and Technology, 0823 Transportation Engineering
Communities, Microorganisms, Carbon Sources	0828 Agricultural Engineering, 0830 Environmental Science and Engineering

Text pattern of abstracts and titles of scientific papers_

Research objective	Abstract features	Abstractive title
US	Study/investigate/test + individual object + structure/state/performance	Research/analysis of the performance/characteristics of problem
SO	To address/tackle + problem + based on/utilizing + method + construct/propose/build	Study of problem based on method
EXP-S	Summarize/review/introduce + individual object + current status/progress	The current status/overview of research on problem
EXP-RG	Investigate/explore/analyze/discuss + the relationship/interaction mechanism/influence + multiple objects	The impact /mechanism of the problem

Manual Evaluation Results_

Research problem	Quantities
Multidisciplinary research problems	34
Single-discipline research problems	16

Comparison of different methods for research objective classification_

Algorithm	Macro-Precision	Macro-Recall	Macro-F1
SVM	0.85	0.84	0.84
NB	0.81	0.81	0.81
Random forest	0.77	0.75	0.75
LSTM	0.69	0.62	0.65
FastText	0.71	0.67	0.68

Comparison of abstractive title generation between BART and ChatGLM_

Research Objective	Model	1-Gram	2-Gram	3-Gram	BLEU	Exact Match	Unigram
US	ChatGLM	0.560	0.462	0.371	0.402	0.182	0.417
	BART	0.582	0.474	0.376	0.411	0.145	0.369
SO	ChatGLM	0.612	0.494	0.387	0.440	0.299	0.441
	BART	0.631	0.498	0.374	0.437	0.356	0.438
EXP-S	ChatGLM	0.501	0.436	0.359	0.351	0.186	0.346
	BART	0.597	0.502	0.413	0.436	0.233	0.422
EXP-RG	ChatGLM	0.588	0.487	0.401	0.422	0.197	0.441
	BART	0.610	0.509	0.422	0.428	0.201	0.434
ALL	BART	0.577	0.463	0.372	0.408	0.203	0.367