2021 Fred_Knowledge Discovery, Knowledge Engineering, and Knowledge Management

Briefing Document: Key Themes in Knowledge Discovery, Engineering, and Management

Overview: The 13th International Joint Conference on Knowledge Discovery, Knowledge Engineering, and Knowledge Management (IC3K 2021) brought together researchers, engineers, and practitioners to address pressing issues in these interconnected fields. This volume compiles revised and extended selected papers from the conference, highlighting significant trends and presenting novel approaches to complex problems, particularly in data-intensive and bias-prone domains. The conference received 103 submissions from 47 countries, with 9% included in this book.

Main Themes and Key Insights:

1. Addressing Information Overload and Data Prioritization in Clinical Settings: A significant challenge in modern data-rich environments, particularly in healthcare, is "information overload." Clinicians often face hundreds of pathology results daily, necessitating efficient prioritization.

• The Problem: "Even for experienced doctors, this procedure is complex and time consuming. One solution is to prioritise pathology results using the tools and techniques of machine learning so as to provide a “look at these first” style of help to accelerate the process of analysing pathology results." (Qi et al., p. 1-2).
• Absence of Ground Truth: A major hurdle is the lack of "ground truth" data for pathology prioritization, making traditional supervised learning difficult. Unsupervised anomaly detection, while attempted, has flaws because "just because a record was unusual, it did not necessarily mean it was a priority record (and vice-versa)" (Qi et al., p. 2).
• Proposed Solutions (Qi et al.):Proxy Ground Truth Pathology Data Prioritisation (PGT-PDP): Uses patient outcome events (Emergency Patient, In-Patient, Out-Patient) as a proxy for ground truth, correlating them with high, medium, and low priority. This approach hypothesizes that "there are patterns in patients’ historical lab test results which are markers as to where the patient “ended up”, and which can hence be used for prioritisation" (p. 2).
• Methods Explored: kNN (k=1 with Dynamic Time Warping for similarity) and Long Short-Term Memory (LSTM) recurrent neural networks.
• Findings: LSTM-based PGT-PDP consistently outperformed kNN-based PGT-PDP, with the PGT-PDP mechanism overall producing the best recall (0.79) and precision (0.58) for the high priority class, outperforming anomaly detection methods.
• Future Result Forecast Pathology Data Prioritisation (FRF-PDP): Predicts the next pathology result value in a sequence. If the predicted value falls outside the normal clinical range, the current result is prioritized.
• Methods Explored: LSTM regression and Facebook Prophet (FP). FP is notable for handling irregular time intervals in data.
• Findings: FP-based FRF-PDP showed higher overall recall than LSTM-based FRF-PDP at the local task level, suggesting a higher chance of correctly identifying priority records, though with higher MSE values. At the global level, both approaches showed higher recall than precision for high prioritization, implying a lower chance of missing high-priority results, but potentially classifying low-priority results incorrectly.

2. The Critical Role of "Understanding" in Event Tracking and Modeling: Event tracking algorithms, particularly in rapidly evolving social media contexts, require deeper "understanding" beyond mere detection.

• Historical Context: "In 1998, Topic Detection and Tracking (TDT), or event tracking, was only two years old. By then, however, event tracking’s pioneers had already realized that for algorithms to accomplish their task, to detect and track events in the news media accurately, they had to understand events" (Mamo et al., p. 21). However, early linguistic-based approaches often failed due to simplistic interpretations of understanding.
• Modern Challenges: Social media (e.g., Twitter) has amplified these challenges, requiring algorithms to "make sense of Twitter streams," "explain, not merely detect events," and "model the structure of events to allow complex reasoning" (Mamo et al., p. 22). Traditional approaches applying machine learning or text embedding have led to limited substantial improvements.
• Understanding as a Paradigm Shift: The authors argue that "understanding can be event tracking’s paradigm shift" (Mamo et al., p. 22). This involves moving beyond simple linguistic understanding (which often fails because it "understood language, not events," p. 27) to semantic understanding.
• Semantic Understanding: Focuses on identifying "which named entities mattered to an event" and "which nouns and verbs described its essence" (Mamo et al., p. 27). This type of understanding, while more deliberate to generate, offers "true knowledge because their presence in the lexicon carries a strong justification" (Mamo et al., p. 27).
• Event Models: Represent the most complete form of semantic understanding, comprehensively filling the 'five Ws and one H' structure (Who, What, Where, When, Why, How). These models can link "What happens with Who is involved" (Mamo et al., p. 28).
• Proposed Solutions (Mamo et al.):Automatic Participant Detection (APD): A six-step framework that refines Named Entity Recognition (NER) to understand "Who participates in events or Where they occur" (Mamo et al., p. 31). APD aims to extract "semantic concepts" rather than mere strings.
• Automatic Term Extraction (ATE): Adapting ATE for event domains to understand "What happens in events" (Mamo et al., p. 32). This focuses on identifying "domain-specific terms from domain-specific corpora."
• Symbiotic Relationship: Understanding allows event tracking to approach event modeling, with event tracking becoming a guide that automatically describes the 'five Ws', and event modeling providing a new purpose for event tracking.

3. Advancements in Semantic Similarity for Legal Text Processing: The application of Natural Language Processing (NLP) in legal tech is gaining momentum, with a focus on semantic similarity for identifying relationships within legal documents.

• Context: Investments in legal technologies are rapidly growing, but legal NLP faces challenges due to diverse legal systems and a lack of publicly available data, especially in Europe (Darji et al., p. 37-38).
• Problem: Manual annotation of legal citations is not always feasible for building citation networks or conducting further analysis like differentiating text types or clustering reasons. Textual similarity techniques are needed.
• Proposed Approach (Darji et al.):Utilizing sentence transformers (modifications of pre-trained BERT networks) to generate semantically meaningful sentence embeddings.
• Calculating cosine similarity between these embeddings to determine semantic similarity.
• Specific Transformers Used: Cross English & German RoBERTa and all-MiniLM-L6-v2.
• Experiments:Similarity between Law Text and Legal Sentences: Aimed to find semantic similarities between legal sentences referring to a law and the actual law text.
• Semantic Similarity Between Legal Sentences: Examined similarity between legal sentences that refer to the same base law.
• Findings: "The results from our experiment confirm that semantic similarity can be exploited to reference similar law texts" (Darji et al., p. 38). Cross English & German RoBERTa achieved a median similarity of approximately 0.45, while all-MiniLM-L6-v2 yielded about 0.4. Furthermore, t-SNE visualizations showed that "sentences that refer to the same law are more semantically similar and grouped together" (Darji et al., p. 49). This indicates that "sentence embeddings retrieved from Cross English & German RoBERTa for Sentence Embeddings are better than all-MiniLM-L6-v2 as embeddings from prior have shown better grouping and thus higher similarity" (Darji et al., p. 49).
• Future Implications: These findings suggest improvements for "predicting law references, identifying similar court cases and court decisions, and legal text entailment" (Darji et al., p. 49).

4. Mitigating Bias through Multidimensional Fairness in Recommendation Systems: Acknowledging persistent bias in academic peer review, the research explores fair algorithms for paper recommendation that promote author diversity.

• The Problem: Despite measures like double-blind review, "studies show that bias still exists" in paper review and selection processes (Alsaffar et al., p. 51). This bias can persist in recommendation algorithms. Statistics reveal underrepresentation of women and minorities in computer science and academia generally.
• Solution Approach (Alsaffar et al.):Developing three fair recommendation algorithms that explicitly consider author diversity across multiple protected variables concurrently.
• Multidimensional Author Profiles: Incorporate five demographic characteristics: "gender, ethnicity, career stage, university rank, and geolocation" (Alsaffar et al., p. 51).
• Methods Proposed:Overall Diversity: Ranks papers by an aggregate diversity score, prioritizing diverse authors until demographic parity goals are met.
• Multifaceted Diversity: Creates separate ranked queues for each demographic feature, prioritizing the rarest features in the pool first. "By selecting papers from a pool of SIGCHI 2017, DIS 2017, and IUI 2017 papers, we recommend papers for SIGCHI 2017" (Alsaffar et al., p. 51).
• Round Robin Diversity: Selects papers by iterating through individual demographic feature queues in a round-robin fashion.
• Feature Weighting: Both Boolean (1/0 for protected/non-protected) and Continuous-valued attributes (reflecting complement of proportion in CS professionals) were used for demographic profiles.
• Key Findings: "All three techniques boost diversity while experiencing just a slight decrease in paper quality or no decrease at all" (Alsaffar et al., p. 51).
• Best Performer (Diversity Gain): Round Robin with Boolean profiles "increased diversity by 72.6 % with only a 4.8 % reduction in utility" (Alsaffar et al., p. 70).
• Best Performer (Demographic Parity): Multifaceted Diversity with continuous features "managed to achieve 95.12 % similarity to the demographics of the pool of authors of all submitted papers" (Alsaffar et al., p. 70). Critically, this approach also resulted in a "2.45% boost in paper quality as measured by the h-index" compared to actual selections (Alsaffar et al., p. 51).
• Broader Applicability: This approach could be used to reduce bias in "grant proposals, conference papers, journal articles, and other academic duties" (Alsaffar et al., p. 51).

5. Foundational Ontologies and Architectural Approaches for Knowledge Management Systems: The development of coherent and interoperable knowledge-based systems requires robust ontological foundations and clear architectural frameworks.

• The Need for Ontologies: Ontologies are crucial for capturing intended interpretations of terms and relations, facilitating "information/data exchange, retrieval, discovery, integration, and analysis" (Olsina, p. 76). However, many existing domain-specific ontologies are not based on higher-level (core/foundational) ontologies, leading to interoperability challenges.
• Foundational Ontologies (FO): These are domain-neutral conceptualizations that serve as a "useful reference conceptualization for any science" (Olsina, p. 74). They should be formally simple, transparent, concise, and balanced in taxonomic and non-taxonomic relationships.
• Proposed Framework (Olsina):Foundational, Core, Domain, and instance Ontological Architecture (FCD-OntoArch): A five-tier architecture (Foundational, Core, Top-Domain, Low-Domain, Instance) emphasizing modularity, extensibility, and reuse of ontology elements (p. 77).
• ThingFO: A proposed foundational ontology within FCD-OntoArch.
• Key Concepts:Thing: Represents particular, perceivable/conceivable objects (e.g., a human agent). Includes Property (intrinsic constitution) and Power (behavior/capabilities). "A Thing is not a particular entity without its Properties and its Powers, so according to Fleetwood [8] “things, properties and powers all emerge simultaneously to form a unity”" (Olsina, p. 85).
• Thing Category: Represents universal, abstract categories for classification (e.g., Product Category, Process Category).
• Assertion: Represents "a positive and explicit statement or expression that somebody makes about something concerning Things, or their categories" (Olsina, p. 87). Subtypes include Assertion on Particulars and Assertion on Universals, and more specific types like Situation-related Assertion or Constraint-related Assertion.
• Applicability of ThingFO: ThingFO semantically enriches lower-level ontologies, such as:
• SituationCO: Enriches concepts like "Target Entity" and "Context Entity" from "Thing," and "Situation" from "Situation-related Assertion."
• ProcessCO: Enriches terms like "Work Entity" and "Work Product" from "Thing," and "Process Perspective" from "Assertion on Particulars."
• Benefits: This structured approach ensures semantic consistency and facilitates the development and harmonization of diverse domain-specific ontologies. It emphasizes a clear delegation of concerns (e.g., process, event, situation) to core ontologies rather than overloading the foundational level.

6. Systematic Approaches to Evidence in Policymaking: Policymakers face "wicked problems" characterized by complexity and uncertainty, necessitating systematic approaches to incorporating evidence.

• The Challenge: Policymakers need to design "best policy alternative[s]" and "best political strategy[ies]" for implementation, but complex social problems lack clarity and optimal solutions (Vargas et al., p. 100).
• Sources of Policy Failure: Identified sources include overreaching government, failure to anticipate consequences, implementation failures (lack of oversight), and failure to learn from past experiences (Vargas et al., p. 102).
• Role of Evidence-Based Policymaking: Proponents suggest that improved information management can address these failures by providing "greater amounts of evidentiary policy-relevant data collected by means of rigorous research and evaluation of both one’s own and similar programs" (Vargas et al., p. 102-103).
• Proposed Methodology (Vargas et al.): A Decision Support System (DSS) methodology that combines expert knowledge and policy-relevant evidence.
• Twofold Coding Process: Facilitates pattern identification in empirical evidence and serves as a rubric for data collection.
• Parametrization of Solution Space: Defines relevant factors:
• Contextual Parameters: Define the environment of the problem (e.g., GDP, demographics, transport information).
• Control Parameters: Shape the policy strategically (e.g., system design/technology, governance, user needs, mobility services).
• Time: Helps define the dynamics of the policy process and describes sequential events.
• Definition of Policy Life Cycle: A "customized" policy cycle (Ideation, Design, Implementation, Operation) with identified "policy gates" (Decision-making, Adoption, Commissioning).
• Application (Urban Vehicle Access Regulations - UVARs in Europe): A small-scale experiment with 11 city case implementations of UVARs.
• Findings: The methodology "could facilitate case data analysis, the identification of alternative policy configurations and key implementation aspects of the policy strategy, and the drafting of alternative scenarios based on possible configurations and strategies" (Vargas et al., p. 100). For example, the short span of the implementation phase across all UVAR cases suggests a strategic choice by decision-makers. Monitoring activities are common in the ideation phase, and complementary measures (e.g., public transport improvements) are crucial for restrictive policies.

7. Design Principles for Business-Oriented Digitalized Innovation Environments (DIEs): As digital transformation progresses, companies need to enhance innovation capacity and employee digital competencies, making Digitalized Innovation Environments (DIEs) a promising concept.

• The Phenomenon: DIEs (e.g., FabLabs, Makerspaces) are "physical spaces that provide both traditional and digital tools and state-of-the-art technologies to support collaborative and interdisciplinary innovation and knowledge transfer" (Hellwig et al., p. 153). They facilitate faster iteration cycles and involve diverse users, but their potential is underexploited in business.
• Research Question: "What factors influence the design of a business-centered Digitalized Innovation Environment?" (Hellwig et al., p. 152).
• Multi-Perspective Approach: Factors influencing DIE design are identified from four dimensions: Business, Employee, Process, and Technology.
• Business Perspective: DIEs offer opportunities for new innovation approaches (digital, open/user innovation) and can serve as platforms for communication, digital fabrication, competency incubation, and market research. Key factors include "Stakeholder Mindset" and "Willingness to Change" (Hellwig et al., p. 158).
• Employee Perspective: User motivation is crucial. Factors influencing usage include the DIE's image within the company, its relevance to daily business, and the presence of a central, competent contact person. "Intrinsic motivation" and "openness" are prerequisites (Hellwig et al., p. 159).
• Process Aspects: DIE development processes differ from conventional R&D. Key factors include "Integratability" into existing processes, minimizing "Bureaucracy," enabling "Democracy" in decision-making, ensuring effective "Communication," fostering "Agility" (non-linear, iterative processes), and promoting an "Error Culture" (Hellwig et al., p. 160).
• Technology Aspects: Emphasis on "Usability" (user-friendliness of hardware and software), "Functionality," and "Availability" of tools (e.g., 3D printers, laser cutters). "Business Relevance" ensures technologies align with commercial interests (Hellwig et al., p. 161).
• Key Findings: A total of 19 influencing factors were identified, reflecting the complexity of DIEs as socio-technological constructs.
• Internal vs. External DIEs: Internal DIEs offer optimal adaptation but may struggle with hierarchy and bureaucracy; external DIEs offer lower financial barriers and broader expertise but may have less direct business relevance and availability issues.
• Central Role of Contact Person: A competent contact person is critical for technical support, administrative tasks, and promoting visibility, though their role needs careful design to avoid hierarchical restrictions.
• Mindset: The "Mindset" of involved actors (decision-makers, employees, contact person) is crucial for successful integration.
• Physical Space & Event Formats: An inviting and inspiring physical space and offering targeted "event formats" (workshops, innovation formats) significantly influence motivation and competency building.
• Conclusion: No single universal concept applies to all companies; individualized adaptation is necessary. However, the identified factors and their interdependencies provide a "sound basis for future concept developments" and contribute to understanding DIEs as complex ecosystems.

8. Ensemble Learning and Data Augmentation for Arabic Sentiment Analysis: Addressing challenges in Arabic Sentiment Analysis (ASA), particularly for low-resource languages, through advanced machine learning techniques.

• The Challenge: Sentiment Analysis (SA) in Arabic is complex due to its "ambiguity and rich morphological system," diverse dialects, and a "lack of resources," which "represent a challenge to the progress of Arabic sentiment analysis research" (Chouikhi & Jarray, p. 118-119).
• Proposed Approach (Chouikhi & Jarray):Data Augmentation: Applied back-translation (Arabic to English and back to Arabic) to "artificially increas[e] the size of the training data set" (p. 121), aiming to improve robustness and diversity.
• Refined BERT Model (ASA-Medium BERT): A fine-tuned Arabic BERT model, optimized with a grid search strategy for hyperparameters, achieved strong performance independently. The Arabic-BERT tokenizer was found to be more appropriate for Arabic morphology compared to mBERT.
• Ensemble Learning (Stacking BERT Models): A hierarchical model ensemble framework where predictions from multiple base models feed into a meta-model for optimal combination.
• Base Models: Medium Arabic BERT, AraBERT, and multilingual BERT (mBERT).
• Meta-Model: A fully connected layer.
• Stacking Scenarios: Automatic stacking (model with itself), pairwise stacking, and full stacking of three BERT models.
• Key Findings:The ASA-medium BERT model, through refine tuning, "outperform[s] the state-of-the-art for ASA" (Chouikhi & Jarray, p. 126).
• "The experiment results showed that the stacking strategy improves the accuracy" (Chouikhi & Jarray, p. 126).
• "The BERT-based ensemble method achieves an accuracy score of 96%" (Chouikhi & Jarray, p. 119).
• Data augmentation significantly improved accuracy for ensemble models, as "the ensemble learning model perform better as the number of training samples will increase" (Chouikhi & Jarray, p. 125).