University of Sussex
Browse
Ashby, Colin.pdf (4.79 MB)

Detecting deception using interview assistive technology

Download (4.79 MB)
thesis
posted on 2023-06-10, 03:40 authored by Colin AshbyColin Ashby
This thesis presents the design, implementation and evaluation of an application designed to support interviewers in detecting deception. This application is evaluated in a job interviewing study using novice interviewers, which shows it to be a highly effective method of de- ception detection, correctly identifying 68.8% of deceivers overall, an increase of 107% and 97% over two baselines without application sup- port, while reducing false positives. We follow work that suggests effective test questioning is the key to detecting deception in interviewing. The rationale behind this ap- proach is that a good breadth and depth of questioning increases cog- nitive load in deceivers, which greatly increases the chance of eliciting detectable behaviour change indicative of deception. Our application is based on Controlled Cognitive Engagement (CCE). Our motivation for supporting interviewers is the difficulty of the interviewing task. Interviewers must simultaneously manage the in- terview process, observe and control the interviewee while generat- ing probing test questions for subjects they potentially know little or nothing about. The application developed in this thesis, called Intek, for Interview Technology, is designed to assist interviewers in generating test ques- tions and providing checkable answers, while also providing a basis to keep track of interview progress. The information supplied by In- tek aims to provide unexpected tests of expected knowledge relevant to the specific personal information provided in a CV or elicited dur- ing an interview. Intek uses multiple information extraction pipelines, from multiple data sources, driven by state-of-the-art Natural Language Processing (NLP) techniques, such as BART for abstractive summarisation, spaCy for fast and accurate Named Entity Recognition (NER) and BERT fine-tuned on the CoNLL-2003 NER dataset for slower but best accur- acy NER. These pipelines integrate into a single simple user interface which may be used by an interviewer for real-time questioning. While most of the underlying NLP technology we used was "off the shelf", we discovered an opportunity to investigate a novel approach to web named entity recognition using HTML tags. Our Text+Tags ap- proach resulted in F1 improvements of between 0.9% and 13.2% over a collection of five datasets and two NER models. Our approach is suitable for extracting named entities from websites containing vary- ing amounts of HTML structure, as well as applicable to other NLP tasks.

History

File Version

  • Published version

Pages

181.0

Department affiliated with

  • Informatics Theses

Qualification level

  • doctoral

Qualification name

  • phd

Language

  • eng

Institution

University of Sussex

Full text available

  • Yes

Legacy Posted Date

2022-05-26

Usage metrics

    University of Sussex (Theses)

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC