Using transcripts from Kansas State University Libraries’ (KSUL) virtual reference services (VRS), an experimental design was created to investigate the value and impact of different natural language processing techniques in the context of creating predictive models. Models were created to use only the first few word tokens supplied by VRS patrons and predict if the overall VRS interaction would be labelled as “easy” or “hard” by VRS operators. The experimental design incorporated machine learning methods (LDA and Doc2Vec), rules-based text processing (TF-IDF), and ontology structures as parameters in the modelling processes. With a specific focus on ontology structures, experimental results indicate that incorporating domain knowledge into predictive modelling processes contributes in significant and positive ways to overall model performance. Results also demonstrated that machine learning processes like Doc2Vec are capable of capturing meaningful representations of domain knowledge in abstract quantified vectors.

Last modified

• 12/17/2019

Creator

• Jeremy Walker

DOI

• https://doi.org/10.21985/N20V2W

Subject

• Data Science

Keyword

• academic libraries
• virtual reference
• predictive modelling
• natural language processing

Date created

• 2019

Resource type

• Masters Thesis

Rights statement

• In Copyright