Human-centric and semantics-based explainable event detection: a survey

Researchers often confuse explainability with terms such as intelligibility, transparency, comprehensibility, and interpretability. Scholars often disagree with the scope and the intersection of these terminologies. An important distinction between explainability and interpretability is that explanation does not generally elucidate how a model works but provides users and practitioners with useful information in an accessible manner (Ehsan and Riedl 2020). While interpretability and explainability have human-centric properties, interpretability describes how a model works. In contrast, explainability bothers on what, why, and how the output/decision of the model is made (Hall et al. 2022).

Roscher et al. (2020) made a clear distinction between interpretability, explainability, and transparency: Transparency takes into account the AI/machine learning strategy, interpretability considers the AI alongside the data, and explainability takes into account the model, the data, and human involvement. Model transparency, design transparency, and algorithm transparency are all subcategories of transparency, which refers to the processes of building models. The transparency of the model structure, such as the number of layers, activation function, splitting criteria, and decision trees in the random forest, is referred to as model transparency. Design transparency is related to understandable, replicable, and well-motivated choices during the AI algorithm’s construction. The transparency of the algorithm is related to the uniqueness of the final solution. The user’s comprehension of the AI model is the focus of interpretability. According to Montavon et al. (2018), interpretability converts an abstract concept like a predicted class into an easily comprehendible domain. Interpretability methods reveal the crucial characteristics that are responsible for model prediction. Interpretation and additional contextual information are combined in explainability. The what, how, why, and causality questions are addressed (Miller 2019). A comprehensive comprehension of AI systems and their actions is one broad definition of explainable AI (Vaughan and Wallach 2020). According to Gunning et al. (2019), the true goal of XAI is to ensure that end users can see the results, which will help them make better decisions. With the definition of explainability, it can be inferred that the existing research has focused on interpretability and much work is still needed to achieve explainable AI models.

Users must occupy a central place and go beyond technology to comprehend contextual usage in developing AI systems (Inkpen et al. 2019). Keeping human-in-the-loop will give room to the creation of dynamic AI solutions (Syed et al. 2020). In addition, users must know the capacity of AI systems, what they can and cannot accomplish, the data on which it was trained, and for what it has been optimised (Ontika et al. 2022). Human values such as responsibility, transparency, trustworthiness, and fairness must be integrated into the design of AI solutions (Friedman and Hendry 2019). When humans are engaged in the design process of AI systems, the resulting systems will be safe, useful, ethical, reliable, fair, and adaptable (Bond et al. 2019; Liao and Varshney 2022).

XAI techniques fall into two main categories: ante-hoc and post-hoc. Ante-hoc explainability incorporates the explanation into the model itself, while post-hoc explainability attempts to generate explanations of the result of the model. Examples of ante-hoc explainability models include fuzzy inference systems, decision trees, and linear regression. Post-hoc explainability models are usually done with black box models such as neural networks (Guidotti et al. 2018). The choice between the two techniques is premised on explainability and performance. Unfortunately, the highest-performance AI models are the least explainable and vice versa (Kelly et al. 2019). According to Guidotti et al. (2018), both ante-hoc and post-hoc XAI techniques can be further subdivided into global model explanations, outcome explanations, and counterfactual inspection. Global model explanation focuses on the model’s overall logic. The distillation technique (Wood-Doughty et al. 2022) can be used to achieve a global model explanation. Outcome explanations aim at explaining a specific model output. Two main algorithms used for outcome explanations are LIME (Local Interpretable Model-agnostic Explanations) (Zafar and Khan 2021) and SHAP (Shapely Additive Explanations) (Lundberg and Lee 2017; Mangalathu et al. 2020). For deep neural network explanations, additional features to use for explanation are propagation, gradient (Linardatos et al. 2021), and occlusion (Kakogeorgiou and Karantzalos 2021). The counterfactual inspection provides an understanding of the model behaviour with alternative input. Techniques such as Partial Dependence Plot (PDP) (Szepannek and Lubke 2022) and Individual Conditional Expectation (ICE) (Rai 2020) can be used for counterfactual inspection. While any machine learning model’s explainability can be achieved with LIME and SHAP, other models like Deep Learning Important Features (DeepLIFT) (Shrikumar et al. 2017) as well as GRADient Class Activation Mapping (GRAD-CAM) (Selvaraju et al. 2017) are used for Deep Learning Models.

Explainability is now more than just trying to understand models; it’s now a crucial requirement for people to trust and use AI solutions in various fields (Liao and Varshney 2022). Value Sensitive Design (VSD) is an approach to XAI that assumes that designers should design all technologies principally and comprehensively, accounting for human values throughout the design process (Friedman and Hendry 2019). It has been argued that VSD will influence how solutions are designed in the future (Friedman et al. 2017; Umbrello and de Bellis 2018). A thorough understanding of XAI and hands-on expertise in XAI techniques are needed to make informed decisions (Gill et al. 2022). Figure 2 provides an overview of XAI approaches, and the description of the XAI approaches is presented in Table 2.

A human-centric explanation is key when measuring or comparing XAI tools and techniques. Future AI must exhibit properties such as transparency, accountability, fairness, performance, trustworthiness, and causality (Shin 2021). Such properties will promote the comprehensibility of humans’ operability of specific applications. Supporting a wider acceptance requires explainability, encompassing more than interpretability (Hall et al. 2022).

Much research has been done to provide explainability, especially for black box models. Some of the existing explainable models include Local Interpretable Model Agnostic Explanations (LIME), GraphLIME, Deep Taylor Decomposition (DTD), Anchors, SHapley Additive exPlanations (SHAP), Prediction Difference Analysis (PDA), Layer-wise Relevance Propagation (LRP), Asymmetric Shapley Values (ASV), Explainable Graph Neural Network (XGNN), Break-Down, Testing with Concept Activation Vectors (TCAV), Shapely Flow, X-NeSyL, Integrated Gradients, Meaningful Perturbations, Causal Models, Textual Explanations of Visual Models, and more ((Holzinger et al. 2022a). This section compares XAI techniques in terms of the extent to which they have satisfied Human-centric explanations. Figure 3 presents popular XAI toolboxes. Table 3 presents the analysis of different XAI techniques and to what extent they cover human-centricity properties.

From Table 3, no existing XAI models have captured the entire dimensions of human-centric event detection. In addition, most of the explanations provided by these XAI methods are logical axioms understandable by experts and not common users. More research efforts are still needed to provide explanations that are comprehensible to users.

Moreover, it can be observed that more research efforts are still needed to invent explainable AI methods that address causal dependencies, as few XAI methods currently capture causal relationships (Holzinger et al. 2020). Research efforts should also gear at XAI models that encourage contextual understanding and answer questions and counterfactuals such as “what-if”. This allows human-in-the-loop where conceptual knowledge and human experience are utilised in the AI process (Holzinger et al. 2021). The existing XAI methods barely scratch the ‘black box’ surface (by stressing on features or localities, for instance, within an image) and do not provide explanations understandable to humans. This is quite different from how humans reason, evaluate similarities, make decisions, draw an analogy or make associations (Angelov et al. 2021). The best AI algorithms still lack conceptual understanding, so there is room for the XAI research community to contribute to this open problem.

Explainable AI is still in the infancy stage. As such, there is no standard agreement on how human-centric explanations should be evaluated (Li et al. 2022) due to the subjectivity of the explainability concept, perception, and interest of users (Carvalho et al. 2019). Because the model’s inner workings are unknown, there is no ground truth for evaluating post-hoc explanations (Samek et al. 2019). Clearly defining evaluation goals and metrics is necessary to advance research on explainability. Additional efforts in this area are still required (Ribera and Lapedriza 2019). Most existing systems skip or provide an informal evaluation (Danilevsky et al. 2020). According to Mohseni et al. (2021), three human-centred evaluation methods for XAI are user satisfaction and trust, usefulness, and mental models. User interview allows for the measurement of user trust and satisfaction. The user’s performance can be used to determine the usefulness, for example, event detection with the aid of XAI systems. Mental models show how the user comprehends the system, and it can be measured by asking the user to predict the output of the model. Future research on human-centric XAI evaluations should center on investigating innovative methods and be effective for collecting subjective measures for explanation evaluation of user experiment designs (Zhou et al. 2021).

The only available tool for the quantitative evaluation of explanation methods is Quantus (Hedstrom et al. 2022). Despite the numerous explanation strategies that have been developed, it is necessary to quantify their quality and determine whether or not they achieve the established goals of explainability. More research efforts are still needed to fill this gap. The common evaluation metrics for explainable AI are Accuracy, Fidelity, Sparsity, Contrastivity, and Robustness (Li et al. 2022).

Accuracy refers to the proportion of correct explanations. Accuracy can be measured in two ways. First, you can use the percentage of the important number of features identified by the explainable method to the truly important number of features (Luo et al. 2020). However, due to the absence of dataset ground truth explanations, this is typically not possible in real-world situations. The accuracy measure is depicted as follows:

where \(\left|{s}_{i}\right|\) represents the important features identified by the explainable method, \(|{S}_{i}{|}_{gt}\) is the truly important number of features, and \(N\) is the total number of samples. Second, you can derive accuracy measures through the perspective of the model’s prediction (Yu et al. 2022).

Fidelity measures how faithful the explanations provided are important to the prediction of the model (Yuan et al. 2021). The main idea behind fidelity is that removing salient features should degrade the performance of AI models, such as getting higher prediction error rates or lower classification accuracy. The formal definition of fidelity is given as follows:

Sparsity measures the proportion of important features by the explanation methods (Pope et al., 2019) defined as follows:

where \(\left|{s}_{i}\right|\) represents the important features identified by the explainable method, and it is a subset of \(\left|{S}_{i}\right|\), where \(|{S}_{i}{|}_{total}\) is the total number of features in the model, and \(N\) is the total number of samples.

According to Pope et al. (2019), contrastivity is the ratio of the Hamming distance between binarised heat maps for negative and positive classes. Contrastivity is based on the assumption that an explanation method’s highlighted features should vary across classes.

Robustness looks at the consistency of explanations despite input perturbation/corruption, model manipulation, and adversarial attack (Zhang et al. 2021).

An event is considered complete when all the 5W1H components can be deduced and answers the question of Who did what, when, where, why and how. We subsequently define these components according to Muhammed et al. (2021).

where L_e is the spatial information about the location (where) of the event (for example, longitude-latitude pairs); T_e is the temporal information about when the event occurs, such as the time it takes to create content; S_e represents the textual semantic description of what occurred, such as a name, title, tag, or content description; P_e stands for participants (such as a person or organisation) who describe their participation in the event; C_e is a causal description is one that explains why the event occurred (or which event is causing it) and thus demonstrates the relationship between two events, with event1 being identified as the cause and event2 as the effect, and M_e is the textual details about how an event was carried out.

L =< ∅, λ, h >.

where \({C}_{e}\) represents causal dimension; \({E}_{i}\) represents the causal event, \({E}_{j}\) represents the effect of the causal event (\({E}_{i}\)), and \({R}_{n}\) represents the relationship among events.

No doubt, the existing machine learning tools/algorithms for event detection have recorded success. However, most machine learning approaches use latent features to detect events effectively. Still, they cannot explain why a piece of social media text is classified as an event (Shu et al. 2019). Capturing an explainable event’s six dimensions (5W1H) is important because new knowledge and insights are originally hidden from the users, and practitioners can be derived from such explanations. In addition, extracting explainable features can further improve the event detection performance and gain users’ trust.

Explainable event detection can generate human-understandable explanations without knowing the detailed knowledge of the underlying models for the event detection (Ribeiro et al. 2016). Explainable event detection can increase users’ comprehensibility and trust (Adadi and Berrada 2018).

The need for explainable systems, according to Samek et al. (2017), is to (1) verify the system or comprehend the rules governing the decision-making process to identify biases; (2) improve the system, avoid failures by doing a comparison of different models with the intended model and dataset; (3) extract the distilled knowledge from the system by learning from it; (4) comply with applicable legislation by responding to legal inquiries and informing those who the model decision will impact. By revealing the logic behind a model decision, an explanation can be used to prevent errors and ascertain the appropriate use of certain criteria. An explanation will force trade secrets to be revealed (Novelli et al. 2023).

Machine learning systems that can provide human-centered explanations for decisions or predictions have recently gained much attention. No matter how good and efficient a model is, it is difficult for the users or practitioners to trust the model if such users or practitioners cannot understand the model or its behaviours (Mishima and Yamana 2022). In event detection, explainability is crucial for practitioners and users to ensure that models are widely accepted and trusted. Human-centric explainable event detection will achieve trustworthiness, explainability, and reliability, which are currently lacking. Achieving such a human-centric event detection system will necessitate designing and developing more explainable models. Optimising models or regularisers would only be worthwhile if they can solve the human-centric task of providing explanations (Narayanan et al. 2018). Incorporating explainability that is human-centric in event detection systems is significant for building a decision-making process that is more trustworthy and sustainable (Vemula 2022). Unfortunately, even the developer who wrote the code for the model does not understand why a decision was made and, therefore, cannot assure the user to trust such a model. This huge gap necessitates the development of explainable event detection models that are human-centric to promote trust and wider adoption.

Many early AI systems used the expert system approach (i.e., a rule-based) approach to provide explanations that address what, why and how the decision of AI systems. However, such systems did not address user context when they generated user explanations. Today there are XAI (with deep learning approaches) that focus on explaining the underlying mechanisms of these black boxes. While this is appreciable, it is not enough to provide tailored, personalised, and trustworthy explanations to the AI systems’ users or consumers (Yang et al. 2022. Moreover, machine learning models use scores for their prediction. While using a score for prediction may be useful to gain some confidence level, it lacks context and therefore is inadequate for explanation without additional information. Semantic web and reasoning technology is well suited to fill this gap (Chari et al. 2020). AI systems need to include provenance to improve the confidence and trust of AI systems. The strength of the semantic web combined with AI will contribute significantly to XAI systems.

A user-centric semantic-based explanation should be (1) understandable: explanation should include capabilities that will define terms for unfamiliar terminologies; (2) appeal to the user: a resourceful explanation should appeal to the user’s current need and mental cognition; (3) adapt to users’ context: explanations should be tailored to the current user’s context and scenario by leveraging on the user information; and (4) include provenance: a property that is either absent or has not yet received the proper emphasis. Provenance aims to include domain knowledge utilised along with the method used in obtaining the knowledge (Chari et al. 2020).

Doran et al. (2017) presented a better perspective of explainability. They opine that rather than focusing solely on mathematical mappings, XAIs should provide justification or reason for their outputs. In addition, they argue that to produce explanations that humans can understand, truly XAI systems must use reasoning engines that can run on knowledge bases that contain explicit semantics. This notion is also supported by Chari et al. (2020). Using logical reasoning on the model’s output is inadequate as it performs an explanation guided only by the knowledge base axioms. No explicit link is used from the knowledge base concepts and model learned features in that case. Hence, logical reasoning must be tied to semantics to have a human-centric explanation as it links the model’s output and human concepts. Semantic-based XAI can present a sufficient explanation comprehensible by a human. We believe that human-comprehensible explanations cannot be achieved without domain knowledge and that data analysis alone is insufficient for full-fledge explainability. There is a need to integrate semantic web technologies with AI systems to provide explanations in natural language (Ai et al. 2018; Lecue 2020).

Linking explanations to ontologies (i.e., structured knowledge) has multiple advantages, such as enriched explanations with semantic information, facilitation of effective knowledge transmission to users, and provision of potential for customisation of the level of specificity and generality of explanations to specific user profiles (Hind 2019). Integrating a domain background knowledge, such as a knowledge graph with AI models, can provide more insightful, meaningful, and trustworthy explanations (Tiddi and Schlobach 2022). Semantic technologies provide easy access to web knowledge sources. In contrast, symbolic representations such as knowledge bases, ontologies, and graph databases formalise and capture data and knowledge for specific or general domain knowledge.

where \({\left\{O\right\}}_{1}^{n}\) is the set of output symbols of the AI model and \({\left\{A\right\}}_{1}^{m}\) is the corresponding semantic attributes or features. Let us see how this applies to an AI system’s main tasks, such as regression, multiclass, and multi-label classification.

Integrating semantics into XAI aims to facilitate situational understanding by human analysts ((Holzinger et al. 2022b). Exploiting semantic relationships between model outputs and human-in-the-loop processes can improve generated explanations (Harbone et al. 2018). It has already been established that for AI solutions to have their full potential in terms of usability, such solutions must be explainable, requiring semantic context (Pesquita 2021). Semantic technologies and artefacts such as knowledge graphs and ontologies can provide a human-centric explanation.

While progress is being made in the AI community to address explainability issues, AI systems are still far from self-explainability. Self-explainability can adapt automatically to any machine learning algorithm, data, model, application, user, and context (Lecue 2020). Semantic representations and connections in the form of knowledge graphs such as DBpedia, OpenCyc, Wikidata, Freebase, YAGO, NELL, ConceptNet, WordNet, and Google Knowledge Graphs can be armed to move XAI closer to human comprehension (Tiddi and Schlobach 2022). Knowledge graphs natively and readily expose connections and relations, encode contexts, and support inference and causation (see Fig. 4). Integrating knowledge graphs with XAI systems can be of great potential (d’Amato 2020). Knowledge-driven structures can adapt to constraints, variables, and search space. Knowledge graphs can also capture knowledge from heterogeneous domains, which makes them a great candidate for explanation (Pakti et al. 2019). In addition, semantic descriptions inspired by knowledge graphs can bridge the missing gap from brute-force machine learning approaches on text analysis to improve explainability. This has yielded positive results on natural language processing tasks such as event extraction, relation extraction, or text classification (Ribeiro et al. 2018).

With knowledge graphs resources, event causality inference can be achieved. The causal relations among events can be inferred in many ways. The simplest approach is determining the likelihood that event y occurs after x has occurred (Zhao 2021). The causal-effect relationship can also be formulated as a classification task, where the cause-and-effect candidate events represent the inputs by incorporating contextual information from knowledge sources (Kruengkrai et al. 2017). Some other methods employ NLP techniques to identify causal relations, such as causal prepositions, connectives, and verbs (Cekinel and Karagoz 2022). Often, cause-and-effect identification using the above methods may result in low generalisation. One way to improve this is to adopt ontology or external knowledge bases to establish the underlying relationships among event candidates. The computation of similarities between two cause-and-effect pairs \(\left({c}_{i},{e}_{i}\right),\left({c}_{j},{e}_{j}\right)\)is given as:

One method of incorporating explainability into XAI involves the following steps: ontology selection, semantic annotation, semantic integration, and semantic explanation. Ontology selection determines the optimal set of ontologies that adequately describe the data. Semantic annotation links the data to ontologies. Semantic integration establishes links between ontologies, while semantic explanation explores background knowledge afforded by the knowledge graphs (Pesquita 2021).