Knowledge Graphs (KGs) play a pivotal role in the field of artificial intelligence, yet the construction of such graphs often requires significant human resources. Automated KG construction technologies are key to achieving large-scale KGs construction. To address this, we have developed an automated Knowledge Graph Construction Pipeline (KGCP) and applied it to the enhancement of the Australian National University (ANU) Scholarly Knowledge Graph (ASKG), which comprehensively represents not only the metadata but also the scholarly knowledge encapsulated in the academic papers. This study introduces an innovative, automatic approach to KGs construction using an array of Natural Language Processing (NLP) techniques. Leveraging Named Entity Recognition (NER) models, key academic entities related to computer science are eficiently identified, such as Research Problems, Named Entity Linking (NEL) with Wikidata, keyword extraction, automatic summarisation, and the integration of entities from the Metadata Extractor & Loader and The NLP-NER Toolkit (MEL & TNNT).
CEUR ceur-ws.org
Academic KGs have been a focus in the field of cognitive intelligence. However, these KGs often concentrate on high-level metadata of papers, such as the author, date, venue, etc., while the in-depth exploration of paper content is often overlooked. This limitation hinders the full interpretation and utilisation of detailed knowledge within academic papers.
Addressing this issue is crucial as it can guide deeper analysis, identify emerging academic
trends, reduce Large Language Models (LLMs) hallucination problem as well as enhance the
training outcome of LLMs [
Numerous projects have been developed in the domain of academic KGs, such as AMiner [
ORKG [
This paper presents an innovative approach to constructing KGs, emphasising the extraction of fine-grained knowledge from scholarly papers to enrich ASKG. Unlike the above-mentioned systems, our methodology entails section-wise parsing of academic papers adhering to the IMRaD (Introduction, Method, Results, and Discussion) structure. Many academic papers essentially adhere to the IMRaD structure. For those that do not follow the IMRaD format, we are in the process of implementing new tools and ontologies that can be customized according to the specific structure of each paper. We employ NLP techniques such as NER, NEL, automatic summarisation, and keyword extraction, individually applied to each IMRaD segment. This specific strategy distinctly positions ASKG from other KGs, ofering a significant edge in gathering and processing academic data. Detailed comparisons with these platforms and tools can be found in our GitHub repository. Initial results suggest our method significantly enriches academic knowledge graphs, ofering a more comprehensive and diverse data set, thus exemplifying the eficiency of our decomposition and refinement approach in knowledge graph construction.
Our ultimate goal is to expand the academic KGs by automatically extracting fine-grained knowledge through the structural decomposition of the documents (research papers). To achieve this, we are implementing and extending our KGCP1 pipeline. As a key component of KGCP, the PARSE component is specifically focused on enriching ASKG by extracting meaningful knowledge from academic papers related to computer science.
As shown in Figure 1, firstly, we utilise web crawling to access ANU’s target sources (academic
web pages), MAKG, ScholarlyData, etc., automatically extracting information on researchers
and their papers to build an academic paper dataset. We subsequently generate JSON files
depicting paper metadata. PARSE operates in two primary phases. The first entails importing
papers into the MEL & TNNT systems [
In the second phase, PARSE extracts scientific field-specific knowledge from academic papers. Paper metadata described in JSON files is fed into a statistical analyser to obtain document set metadata and identify computer science papers. Using the targeted paper list, we send HTTP requests to the TNNT RESTful API, fetching the papers’ original text. Then PARSE processes
academic papers, segmenting them based on the IMRaD structure. We design and employ transformer-based NER models with RoBERTa, SciBERT, LinkBERT, etc. The text is sent to the NER module to identify computer science-related academic entities which are categorised as Research Problems, Methods, Solution, Tool, Resource, Dataset, and Language. Academic entities from the NER module are linked with Wikidata entities to enhance our knowledge graph. Meanwhile, we send diferent parts of the paper to the automatic summarisation model, BRIO, to generate summaries, and to the keyword model, KeyBERT, for keyword identification. All the outputs are processed to enrich the academic knowledge graph.
The comparison between the original and enriched ASKG, as shown in Table 1, reveals significant growth in aspects such as the number of relation types, entity types, entities, and triples, indicating enhanced structural diversity and information capacity. However, the information density has decreased, suggesting the enriched ASKG has become more sparse, posing a new research direction.
Listing 1 shows a portion of the output from PARSE. Unlike most traditional academic knowledge graphs and previous ASKG, the enriched ASKG not only includes high-level abstract metadata such as authors and publication dates, but also contains more detailed academic information. This information includes, but is not limited to, keywords and a summary in each academic paper section, as well as more specific academic concepts like academic entities in each sentence and their locations in the academic paper. @prefix askg − d a t a : < h t t p s : / /www. anu . edu . au / d a t a / s c h o l a r l y / > . @prefix askg − o n t o : < h t t p s : / /www. anu . edu . au / onto / s c h o l a r l y #> . @prefix domo: < h t t p s : / /www. anu . edu . au / onto / domo#> . . . . . . . askg − d a t a : P a p e r −5003681 f a 6 a 9 1 4 a askg − o n t o : P a p e r ; r d f s : l a b e l ‘ ‘ [ SPICE: Semantic P r o p o s i t i o n a l Image Caption E v a l u a t i o n ] −[ P e t e r Anderson ] − [ 2 0 1 6 ] ’ ’@en ; askg − o n t o : h a s S e c t i o n askg − d a t a : A b s t r a c t −1 f 3 5 f 0 4 2 4 3 f 7 3 0 , askg − d a t a : D i s c u s s i o n − fc3bb8b300771b , askg − d a t a : E x p e r i m e n t − dc48c6d08186a7 , . . . . . .
askg − o n t o : p a p e r L i n k ‘ ‘ h t t p : / / a r x i v . org / abs / 1 6 0 7 . 0 8 8 2 2 v1 ’ ’ ^^ x s d : s t r i n g . askg − d a t a : A b s t r a c t −1 f 3 5 f 0 4 2 4 3 f 7 3 0 a askg − o n t o : A b s t r a c t ; r d f s : l a b e l ‘ ‘ Paper −[ SPICE: Semantic P r o p o s i t i o n a l Image Caption E v a l u a t i o n ] −[ P e t e r Anderson ] − [ 2 0 1 6 ] | S e c t i o n −[
A b s t r a c t ] ’ ’@en ; domo:keyword askg − d a t a : K e y w o r d O f S e c t i o n −0619 f 5 f d 0 a b 6 a 4 , askg − o n t o : c o n t a i n s askg − d a t a : E x c e r p t − e d1 f c3 d d5 c 08 ab , askg −onto:summary ‘ ‘ SPICE: Semantic P r o p o s i t i o n a l Image Caption i s a new automated c a p t i o n e v a l u a t i o n m e t r i c . . . . . . ’ ’ ^^ x s d : s t r i n g . askg − d a t a : E x c e r p t − e d 1 f c 3 d d 5 c 0 8 a b r d f s : l a b e l ‘ ‘ Paper −[ ’ SPICE: ␣ Semantic ␣ P r o p o s i t i o n a l ␣ Image ␣ Caption ␣ E v a l u a t i o n ’ ] |
S e c t i o n −[ ’ A b s t r a c t ’ ] | Excerpt − [ 2 0 7 ] − [ 2 0 8 ] ’ ’@en ;
askg − o n t o : i n S e n t e n c e ‘ ‘ t h e r e i s c o n s i d e r a b l e i n t e r e s t i n t h e t a s k o f g e n e r a t i n g a u t o m a t i c a l l y image c a p t i o n s
image c a p t i o n s [
With the enhanced ASKG, we propose a range of innovative use cases. One such use case is knowledge graph-based research trend analysis, illustrated in Table 2. While our study primarily focuses on capturing the dynamic evolution of academic research trends at the ANU, the methodology is designed to be adaptable and can be applied to other institutions as well. By executing SPARQL queries, we extract relevant data from the KGs and carry out a quantitative analysis, identifying the most mentioned academic entities and research problems, which can be interpreted as current research trends of the university’s academic sources.
up to Jun. 2022
up to Dec. 2022
1 2 3 4 5
Optical Flow Modal Logic Image Captioning Blur Kernel Action Recognition 230 231 144 101 168
Research Trend Analysis can be used for academic performance management, resource allocation, etc. It’s worth noting that performing this level of refined analysis is challenging within traditional academic KGs that only include paper metadata. This is mainly because the metadata typically does not encompass in-depth descriptions of specific research problems or other academic knowledge, limiting our ability for a deep understanding of the dynamics within the research field. In contrast, our enriched ASKG can capture more information, thereby facilitating more detailed trend analysis.
Moreover, the enriched ASKG has a wider range of application scenarios, such as research relationship mining. By integrating diverse data including authors, research interests, academic entities, and summaries, it enables the discovery of overlooked patterns and potential crossdisciplinary collaborations between researchers through graph mining.
Currently, we continue applying the PARSE to other disciplines, such as astronomy and physics. Simultaneously, we are developing an innovative semantic query processing system (as an additional component of the KGCP) that combines LLMs with the enriched ASKG, aiming to improve the eficiency of academic information queries and the accuracy of context-based information retrieval from LLMs. In this system, user queries are translated into triple formats and then processed using SPARQL for graph matching, thereby supplying LLMs with more accurate and complete academic information.
We continue investigating and optimising the application of the LLMs and KGs in semantic searches and KG construction-related tasks, further advancing the fields of information retrieval and knowledge representation.