Semantic Web technologies and bias in artificial intelligence: A systematic literature review

4.1.1.Bias affecting specific groups of people

KGs can help assess recommendation disparities in user groups that are less active [29] (e.g. economically disadvantaged users). This problem constitutes a population bias because it affects groups that are underrepresented in the data with respect to the most active users. Therefore, their historic user-interaction data is less visible in the recommendation system. The harmful effects of this bias impact the recommendation quality and diversity of the results, posing a fairness problem. A fairness-aware algorithm that leverages entities, relationships and paths in KGs is proposed to explicitly model the recommendations in reasoning paths and apply constraints that impose fairness across users. In this case, the Amazon item e-commerce KG of entities and relations is used to quantify the richness and evenness of recommendations, revealing disparities of groups with historically less user-item interaction data. This semantic knowledge is crucial since it is a setting where users do not disclose the personal information required to deal with possible discriminatory treatments (i.e. sensitive features such as gender, age, or religion). Richness and evenness disparities are measured with each user’s number of graph patterns and the relative importance of each pattern across users, respectively. The paper shows how these measures can also be used as fairness constraints to improve the quality and diversity of recommendations for these vulnerable user groups.

The under-representation of certain demographic groups was assessed in drug exposure studies. Due to the heterogeneity and lack of metadata in the gene expression databases resulting from these medical studies, it is challenging to examine large-scale differences in sex representation of the data. This assessment is essential, as women are 50% more likely to suffer from adverse drug effects. In [27], they were able to assess sex bias in public repositories of biological data by mapping existing and inferred metadata (using ML models) to existing medical ontologies. Specifically using Cellosaurus and DrugBank databases to identify cell lines and drugs, respectively. They could label drug studies from all publicly available samples using named entity recognition to identify drug mentions in the metadata and normalisation to map every instance to all its possible names. This analysis generated a new resource with unduplicated and normalised data that allows examination across study platforms. As a result, they identified that sex labels are inconsistently reported, with most samples lacking this information. More importantly, they report the existence of sex biases in drug data (e.g., female under-representation in studies of nervous system drugs). This study draws attention to the lack of study and the importance of including sex as a study variable in future analyses.

Another use of KGs relevant for this task is to assess disparities in the presentation of news reported by different sources, i.e., with different political leaning [5]. This media bias can affect groups or individuals who are part of the story or use these web search systems to build an opinion, since the news is written with the reporter’s or media outlet’s perception, which can be done, in some cases, partially or unfairly. As a result, bias compromises the reliability of the news source and may raise concerns closely related to the growth of misinformation, polarisation, or online hate. The structure of a KG could help to uncover such disparities in reporting between different media outlets, e.g., of specific stakeholders (politicians, political parties) advocating or opposing the same issue depending on which source the news appeared. The YAGO KG proved to help extract holders, opinions, and topics and store them, allowing to compare topics and visualise biased news. Identifying potentially contradictory information is vital to raise awareness and encourage critical thinking because we, as web users, are exposed to a massive amount of information.

4.1.2.Bias affecting individuals

The assessment of polarised web search queries is also essential because users are prone to look for information that reinforces their existing beliefs [18]. Especially when it comes to controversial topics, the confirmation bias of web users often conveys views that can lead to a strong division of opinion, affecting individuals and society at large. The impact of bias in this type of user-generated content is closely related to the above mentioned media bias concerns. To prevent these issues, an approach to identifying the sentiment of queries could improve web search systems. This natural language processing (NLP) task (i.e. sentiment analysis) incorporates the support of the SentiWordNet lexical resource with a two-fold goal. First, it aims to improve the quality of results by including recommendations from less popular queries but with similar sentiments. More importantly, providing results from queries of opposite sentiment improves the diversity of opinions and shows the viability of query sentiment analysis to deal with the problem of bias and polarisation in web search.

Recently, a research project presented a similar approach to assess confirmation bias and other similar group phenomena that occur when analysing, visualising and disseminating information on the Internet (group polarisation [8] and the belief echo chamber [60]). The ontology is the primary data structure for modelling groups and individuals in a network structure [67] and is used as a tool to find similarities and anomalies in the profiles resulting from the data collected by the system. They present a web information system to evaluate such effects, integrating NLP methods into this data structure to identify themes and sentiments towards them. This architecture allows tracking individual and group responses to different events (e.g., COVID restrictions, vaccination) by simply adding new terms to the ontology (e.g., from specific vaccines such as Pfizer, Moderna or Astra Zeneca), inferring the sentiment towards them. These results are promising for integrating AI methods for natural language understanding (NLU) into distributed open ecosystems, as this is fundamental to understanding these phenomena in the broader societal domain.

4.1.3.Bias affecting AI systems

Bias can cause many problems affecting AI systems.

Semantics are used to assess inconsistencies in the predictions of AI systems due to the use of a small, domain-specific corpus for training. Model overfitting can compromise the quality of AI systems and the extent to which they can fulfil their purpose. It is a significant challenge in the AI community due to the potential harms that may arise from using models that are black-boxes to us, especially when we do not understand why they make specific predictions. This survey found two use cases that focus on this problem. In the first example, sentences that entail the same action but are different may drift the model towards undesired behaviours, i.e., paying attention to the wrong words in the sentence [57]. The use of external knowledge from the FrameNet lexical database helped to reveal these biased predictions from the mismatch between the words in a sentence with the highest value in the attention layer of the model, and the action they should trigger, as captured in this semantic resource. This analysis revealed patterns with no theoretical basis but which the model systematically followed, i.e., recurrently giving more attention to words that were not relevant to trigger the action implied by the sentence. This paper also showed how this knowledge could be included as additional examples in the training data to make the model more consistent with the linguistic theory and help it generalise beyond the annotated examples of the training corpus.

Similar artefacts in datasets are evaluated in pre-trained masked linguistic models, which are increasingly used in factual knowledge bases to extract information from a query string [14]. The task consists of using queries such as “Steve Jobs was born in [MASK]”, where Steve Jobs is the subject of the fact and was born in a prompt string for the relation “place of birth”, to predict the object placed as [MASK]. However, their study demonstrates that many of the successes of prompt-based approaches are due to spurious correlations between similar prompts (Fig. 2). As a result, predictions on completely different datasets are similar, and this is because the dataset has been over-fitted to specific prompts. The authors reveal that current case-based approaches that aim to improve performance by providing illustrative cases mainly succeed in providing a “type guidance”. They reveal that performance is enhanced primarily by recognising the object type in the illustrative cases. Therefore, models can effectively make analogies between entities of the same type but not predict facts based on their internal knowledge and the illustrative cases. This analysis is possible due to the use of the Wikidata taxonomy to infer the object type of each relation. It allows us to probe into the behaviours of these black-box models and better understand the critical factors underlying their task performance, which is crucial for building trust in the predictions of these systems in benchmarks and closed-world studies.

Fig. 2.

Example of two dataset artifacts (i.e. prompt bias, type guidance) that can overfit pre-trained masked language models in factual knowledge extraction tasks [14].

Recently, a new framework for automatic decision making in the medical domain was proposed to meet the requirements of explainability, robustness, and reduced bias in machine learning models [38]. Through a series of experiments to address three fundamental challenges in medical research, the authors reason that a multimodal, decentralised and explainable infrastructure is needed, where KG can play a crucial role. In this second use case, a series of arguments are presented as to why KGs can benefit future human-IA interfaces to be effective in this field (e.g., integrating the characteristics of different medical data modalities). It concludes with the basics of counterfactual graphs, which store the path from the feature to the changing class to enable the exploration of different counterfactual decision paths (bringing the “human-in-the-loop”) and serves as a communication channel with black-box models. This work has a significant impact, as it provides knowledge-based constraints to regularise the training process of deep learning models and the possibility to contest them. This mechanism for opening the black box is critical, as future human-AI interfaces must enable medical experts to understand the causal pathways of automated decision-making systems.

Training corpora also have limitations due to errors in manual annotations that compromise the reliability of the corresponding AI systems. Subjectivity and errors in human annotations constitute a significant problem in developing benchmarks for AI systems, so the assessment of bias in annotation tasks is vital. In [17], background knowledge from the WordNet lexical database is used to support the annotation task where the context is missing, i.e., the set of neighbouring words that provide domain information. Notably, a comparison of the precision of two lexicographers in a context-agnostic scenario for a word sense disambiguation annotation task and using WordNet parameters to provide context revealed that annotations consistently shift towards the most frequent sense of a word in the absence of context. Even though this analysis used few semantic parameters (conceptual and semantic distance and belonging to the dominant concept), its binding machine versus human annotation study could help demonstrate the importance of context in human annotation tasks.

We found three other examples of assessment of interpretation bias in training data. One case study focused on assessing subjectivity in the interpretation of the analytical variables used to explain weather conditions in forecast texts, as these may vary due to humour, fatigue, or mood [31]. Using this as training data may compromise the truthfulness of the resulting weather prediction systems. They base their approach on the identification of numerical values and properties of different atmospheric variables in texts in order to be able to compare them with observational data. An ontology supports an information extraction model, as it can represent this domain knowledge. Specifically, they developed a proprietary ontology (AEMIX) using the Web Ontology Language (OWL) to extract the linguistic information of the critical events detected in the text and could reveal the inconsistencies of these texts with the objective information of the interpreted mathematical models.

Similarly, there can be bias when using user product reviews, blog posts and comments to support search engines, recommender systems, and market research applications, due to the subjectivity and ambiguity of this content [46]. As with the previous use case, this may compromise the quality and functionality of NLP systems, especially the degree to which they can detect mixed opinions. They propose a lexical induction approach because mapping subjectivity scores to opinion words in the text can detect review sentiment independently of individual language use. They use SentiWordNet to obtain these sentiment scores, which are used as additional input features for sentiment analysis. Despite the proposed method can only exploit several senses of each word (the overall score or the value of the first sense) without incorporating semantic relations, it shows the value of ontology-based approaches to avoid human biases arising from the use of machine-learned annotations.

Finally, we have found an excellent example of how inconsistencies in human-generated data used as ground truth to train automated decision-making systems can compromise the capability and effectiveness of these systems. The use of data-driven neural networks for automated radiology report generation is becoming a critical task in clinical practice [51]. In particular, image captioning approaches trained on medical images and their corresponding reports can significantly improve diagnostic radiology. However, the large volume of images that is a heavy workload for radiologists and, in some cases, lack of experience hinders the generation of these reports. The problem with using previous reports to train automation models is the variability and redundancy between the sentences used to describe the image, especially in describing the normal regions. For example, as shown in Fig. 3, the Blue text corresponds to the description of all normal image elements, while only the Red text indicates the abnormality. Since normal images are already overrepresented in the dataset, these deviations and repetitions aggravate the data imbalance and make the generation of sentences to describe normal regions more predominant. As a result, this bias in the data leads to errors where rare but significant abnormalities are not described (Underlined text) and repeated sentences describing the same normal region (Italic text). The use of prior medical domain knowledge captured in a KG showed an improvement in the reports generated, as seen in their quantitative and qualitative results (the text under “Our” in Fig. 3). In particular, the medical KG covering the most common abnormalities and findings can be used as an attention mechanism when exploring new input images. Its framework significantly improves abnormality detection, especially when the occurrence of normal reports dominates the entire dataset. We believe these enlightening examples can motivate future research, as similar data bias issues affect many AI applications.

Fig. 3.

Example of bias impact in image captioning for automatic radiology report generation when using human-generated data as ground truth for training automated decision-making systems [51].

4.2.1.Bias affecting specific groups of people

KGs could help to represent systematic preferences that are consistently applied across the examples used as training data [61]. This is the second example of population bias, as these preferences that influence the model, in the same way, are not individual features but domain categories representing specific groups. Mapping the influential input features to a KG (Wikidata) allowed them to be categorised and described with facts so that groups corresponding to individual entities were captured as counter-intuitive rules. For example, that an Italian origin reduces the value of painters’ works. Thus, this additional semantic reasoning capability revealed predictions based on undesirable input data features, such as race or gender, which can be critical for identifying the modification requirements that a model may need to mitigate biases.

There are three other examples of population bias where minority groups are disadvantaged by their lack of representation in the data. The first example focuses on the under-representation of minority cultures in music platforms, which leads to a dominance of commercial music and a lack of diversity [47]. Linked web data can be used to represent contextual features of music data (author biographical information or social connections between artists with similar singing patterns) and create a more relevant navigation space for the cultural background of music. Their prototype is based on a multimodal knowledge base, in which an Open Information Extraction system is used to extract contextual features of music data from the Linked Open Data (LOD). The combination of contextual features together with content features, extracted from audio recordings, can better contextualise the data and reveal non-trivial and deeper relationships between musical entities to lead to more meaningful music discovery and recommendations.

The use of ontologies can be advantageous in improving the representation of minority groups, as shown in the following two examples. In particular, being able to represent differences in the way of expressing and perceiving the affection conveyed by an image across languages can avoid the discrimination of generalist models that only fit a majority language [43]. Language is one of the main characteristics of a culture, so not paying attention to the context of each language can end up damaging entire ethnic groups. In this example, an ontology (Multilingual Visual Sentiment Ontology, MVSO) is constructed to represent a training dataset for visual sentiment analysis with a broader scope (including 12 different languages). Using NLP techniques, social media data in these languages and semantic resources (in particular, SentiWordNet and the SentiStrength ontology), they show that including this knowledge in the training datasets improves the degree to which image classification systems can predict sentiment for visual concepts in different languages. This work empirically demonstrates differences in model performance depending on the language used to express the sentiment used in training, as predictions from models trained in a specific language cannot generalise to image data collected in another language. Ensuring diversity in the training data is critical to avoid biased downstream applications to data from the predominant group.

On the other hand, the interpretation of which verbs trigger a given action varies from language to language [34]. The development of an ontology with videos to represent different actions enables identifying groups of verbs inclusive of different languages since videos have annotations in ten languages. The IMAGACT Ontology of Action is a video-based disambiguation framework that can help clustering algorithms not to be specific to a predominant language since they can thus rely on the multilingual lexical features of each action.

4.2.2.Bias affecting AI systems

We present a use case that uses a KG to improve the representation of users’ interests beyond the items captured in the training data [20]. This example falls back to the model overfitting problem, which in this case can compromise the quality of recommender systems. Their framework incorporates a KG (DBpedia) to expand the user vector representation of a relational graph convolutional network used in the content-based RS. This structure encodes structural and relational information about the neighbouring nodes of the items already part of the training data to provide recommendations consistent with the users’ needs. The propagation of relevant knowledge could enhance the performance of the recommender and dialogue systems.

4.2.3.Bias affecting individuals

As final examples, we present three case studies that use semantics to represent consistent and predicable errors that can compromise how data is used and analysed. This psychological bias can affect groups developing AI systems to support the search, interpretation, selection and visualisation of information needed to draw conclusions from large masses of data (Intelligence Activity, IA). The first example deals with its impact in the evaluation of evidence, in the search for hypotheses, and argumentation of scientific methods [80]. A domain ontology (TIACRITIS) is developed in a collaborative effort to represent all the reasoning steps, probabilistic assessments and assumptions of analysts in data-driven evidence analyses. Similarly, bias can affect planning, collection, processing and exploitation, analysis and production, dissemination and integration activities [53]. The CBOntology is an application ontology that captures the cognitive patterns known to affect these tasks in order to render them explicit and support experts who may experience them. It covers more than 400 classes of such patterns extracted using string, semantic, logical, and topological matching similarities of existing ontologies. These two assistance tools aim to recognise known biases, advise the user to counter them and argue for the need to make biases explicit in AI systems and the experts who use them. The most recent surveyed paper related to this type of bias aims to support and reduce the psychological bias that occurs in decision making under risk and uncertainty [22]. Building on the Core Ontology on Decision Making (CODM), the authors extend this knowledge with the decision preferences that are bound to specific circumstances. For this, they use descriptive decision-making theory to extend the ontology with the concepts of intuitive decision-making so that choices made with deliberation or intuition can be explicitly represented to improve understanding of risk preferences and the situation in which they occur. All these works ultimately aim to develop decision support systems that help humans understand their own preferences to make better decisions.

4.3.1.Bias affecting specific groups of people

Semantic knowledge can be used to mitigate stereotypical perspectives of marginalised groups that are shown to be learned by automated decision-making systems [6]. This is another example of population bias because it reflects over-generalised beliefs about specific groups of people that can cause the model to shift towards incorrect predictions. In this example, hate speech detection systems are prone to be overly sensitive to the presence of specific demographic identity terms (gay, female, black) due to the large amount of hate content that exists against these communities. Their technique is based on replacing these bias-sensitive words with more abstract concepts, e.g., gay is a person, to prevent them from being incorrectly learned by the model as indicators of hate. They use WordNet’s lexical relations to find suitable substitution candidates and demonstrate empirically that systematic deviations towards the hate class of these terms can be reduced without losing effectiveness in detecting hate speech. This initial data pre-processing bias mitigation technique reduces bias towards a closed list of words representing vulnerable groups.

4.3.2.Bias affecting individuals

KGs have shown advantages in mitigating biases due to humans’ heuristic way in web searches. We present two use cases in which the structure of KGs can help counteract the confirmation biases that affect web users. On the one hand, one approach focuses on investigating the search environment to improve users’ knowledge and attitudes on controversial topics (in particular, vaccination) [54]. The authors investigate including factual information extracted from Wikidata in a knowledge box in the search environment interface. It showed that users exposed to this information were significantly more informed, less sceptical about vaccination and more critical in discerning quality information after a simulated web search.

Similarly, the advantage of using a KG in the search interface is addressed in another study that aims to increase the efficiency, quality and user satisfaction with the information obtained after a web search [74]. To this end, they developed a KG-based interface prototype using the Open Information Extraction system to generate the entity-relationship-entity triplets of the text. Their qualitative study, based on a post-experiment evaluation, revealed that the KG interface helps to reduce the number of times required to view the source content during exploratory searches with respect to general hierarchical tree interfaces. These user-based studies serve to uncover important notions that shape the use of AI systems.

4.3.3.Bias affecting AI systems

Bias can cause many problems affecting AI systems.

Similar errors can affect the manual annotations often needed to train AI systems. The scarcity of such data compromises the development of systems to automate specific tasks. This use case assesses human annotators’ errors due to a possible lack of knowledge to provide reliable annotations in extracting information from texts [3]. In particular, in coreference resolution tasks, annotators have to identify different mentions corresponding to the same entity. A reinforcement learning approach is proposed to address the lack of examples to train specific neural systems leveraging information from a knowledge base. Wikidata instances check the consistency of facts extracted from the text. Using this information to tune the model produces better results than other state-of-the-art methods and paves the way for assisting in the difficult task of obtaining human annotations needed in many AI systems.

The rest of the use cases in this section deal with other limitations regarding how data is used to train AI systems. One of the problems of recommender applications is data sparsity, as less popular items are more challenging to deal with and may cause users to interact only with some of the most popular items [65]. This can pose a fairness problem because less popular items are under-represented, and so are the users that prefer to interact with less popular items. In this example, a framework is proposed to improve knowledge-based RS by including the specific semantic properties of the KG. Extracting each DBpedia property corresponding to user-item interactions allows computing similarity metrics between entities that consider each property’s meaning. These property-specific interactions are included in the vectors that model the past interactions of each user to allow making more specific recommendations, e.g., movies that are related by the actors acting even if they do not deal with the same topic. This additional semantic knowledge of user-item interactions improved recommendations, especially on less popular data. Specifically, increasing the accuracy of recommended items after discarding the most obvious ones (serendipity) and the accuracy of unknown items that are part of the long tail of the catalogue (novelty).

Having many features in the recommendation motivates the proposal of an entropy-based method for obtaining only the meaningful historical data of each user. The sparse factorisation approach proposed in [2] facilitates the training process by exploiting a higher level of expressiveness in the feature embeddings of the items provided by a KG. Facts and knowledge extracted from DBpedia provide customised recommendation lists, filtering out items with low information gain. Their method allows incorporating the implicit information provided by the KG into the latent space of features, showing an improvement in the quality of results on three benchmark data compared to other state-of-the-art methods. More importantly, their experimental evaluation shows that this method improves item diversity, which is critical for measuring popularity bias mitigation.

On the other hand, another limitation imposed by the platforms to collect information for recommendations is the small number of negative samples in the data, as most interactions are positive comments (clicks, purchases) [85]. This constitutes non-symmetric missing data, thus compromising population representation and potentially leading to biased analysis. A KG is proposed to provide informative negative signals to a collaborative filtering algorithm based on matrix factorisation. A negative sampler is constructed using reinforcement learning over Freebase to infer these signals from items related to positive interactions, assuming that these are more likely to be known to the user but were not chosen and, therefore, have a higher probability of being true negatives. Figure 4 is shown as an example. Given that a user (u1) has watched two movies (i1, i2) with the same director (p1) and genre (p2), it is more likely that the user knows other movies (e.g., i4) of the same director but different genre, for which the user has less interest. The reinforcement learning agent over a KG improved the top-K recommendation and preference ranking metrics of seven benchmark methods, which also used KGs, but only to leverage positive signals.

Fig. 4.

Example of bias due non-symmetric missing data in a Recommender System (RS) that only collects positive samples [85].

Data imbalance can also affect the system’s ability to infer new data from existing information. The following example aims to mitigate the bias caused by value propagation methods for sentiment analysis due to the imbalance between positive and negative seeds in the training data [88]. This imbalance causes new inferred values to drift towards the average value gradually. An additional step in the method is proposed to mitigate the bias on the basis that the propagation of values differs depending on the relationship between concepts, e.g., the relationship isA has a higher probability of concepts having the same sentiment value than other relations such as one concept Desires another. Their method uses a sequential forward search over ConceptNet to select neighbouring concepts with the most relevant type of relationships to propagate new sentiment values. Next, the sentiment value concepts from a manually annotated sentiment dictionary (Affective Norms of English Words, ANEW) are used to align all inferred values with the mean and variance of the concepts that are in the dictionary, assuming that the difference between their inferred and original sentiment values is a shift that occurs due to the imbalance between the initial seeds.

Similarly, bias towards majority samples is a major challenge in current natural language generation (NLG) architectures. As a result, current neural approaches have difficulty generating coherent, grammatically correct text from structured data. The “divide-and-conquer” approach proposed in [24] is based on inducing a hierarchy from a corpus of unlabelled examples using a KG of entity and relation embeddings. The notion of similarity is used to show only the most relevant examples during training to avoid bias due to imbalances in the training data. Specifically, they apply this idea to two datasets containing linked data and textual descriptions (biography paragraphs with semantic mappings to Wikipedia infoboxes). Applying similarity of embedded inputs generates effective input-output pairs that consistently outperform competitive baseline approaches. Of particular interest in this study is partitioning the dataset according to the semantic and lexical similarity of the entries for training specialised models for each particular similarity group. This general idea can be transferred to other domains to address problems in data sparsity, such as image captioning or question answering applications.

The following three examples address the problem of model overfitting by relying on the use of probabilistic models to generalise better cases that are not included in the training data. In image retrieval [41], relations to concepts in a KG can improve the reasoning power of the model in cases where examples of images with a given caption are not part of the dataset. In particular, the use of the ConceptNet commonsense base can be used to extend the search to images with related captions that are relevant, e.g., the concepts kitchen and restaurant can be informative of Chef. This approach can incorporate this rule-based knowledge source and enrich a language model widely used in multimedia-related tasks for NLP. Related concepts, i.e., relations that are relevant in visual space, are included in the object detection function of the model and show improvements in qualitative and quantitative results that are promising for the study of knowledge representation and computer vision.

Second, a similar approach has been applied in an image captioning framework to allow implicit image relationships to be captured in the caption that may be relevant to describe the image. For example, if the image shows a “woman standing with her luggage” next to a sign, then it makes sense to speculate that she is waiting for the bus [39]. The ConceptNet commonsense base helps discover these relationships, so a similar strategy is incorporated into the caption generator output to increase the likelihood of latent concepts related to specific objects in the image. This example leverages semantic knowledge to allow the system to generalise beyond the training examples.

Third, the incorporation of external knowledge can help mitigate the errors of systems that respond to questions related to an input image when dealing with answers that did not appear during their training phase or that are not contained within the image scene [21]. In real-world contexts, most techniques fail to address answers if they are not within the image content and require external knowledge. For example, Question 1 (Q1) in Fig. 5 requires such external knowledge, as responses such as “dog” (an entity that desires the frisbee) cannot be inferred from the image. A KG allows an understanding of the open-world scene beyond what is captured in the image. The framework incorporates prior knowledge to guide the alignment process between the feature embeddings of the image-question pair and the corresponding target response. Using a pre-existing subset of DBpedia, ConceptNet and WebChild, this knowledge component represents the possible set of answers and concepts to enrich the relations between them. Their quantitative results and their comparison with other current methods support the exploration of knowledge-based systems to overcome overfitting errors.

Fig. 5.

Example of answer bias in a Visual Question Answering (VQA) system when dealing with concepts not seen during training or in the image scene [21].

Finally, two case studies have used semantics to mitigate problems in query formulations due to expressiveness limited to a small corpus leading to irrelevant and incorrect results. In image retrieval based on the semantic representation of scene graphs [19], WordNet and ConceptNet can be used to increase the precision of searches that include more complex concepts. For example, in cases where the system cannot infer that the entities “dog” and “cat” are relevant in the query “animals running on grass”. Their approach introduces a set of rules to find images with fuzzy descriptions and infer the name of the concepts they express to help with more complex searches and enhance semantic and knowledge-based methods for image retrieval processing.

In order to reduce the number of irrelevant results of a web crawler to retrieve social media information, the development of a domain ontology is proposed [76]. Specifically, the Travel&Tour ontology is developed using the Protege-OWL editor to model the specific domain of travel and tourism. The aim is to enrich the content of social media data with the properties and relations of the ontology to improve context-specific searches and take advantage of the domain knowledge provided by this type of data. For example, a search for an expedition-type tourist destination in South America (expedition+South+America) may not return results because the extracted data does not explicitly mention those terms. However, there may be examples with mentions of related terms relevant to the search, e.g., Amazon river tours offered in the city of Brazil, even though there is no mention of the Amazon being in South America. Although these last two use cases presented only validate their results on a limited set of queries, they are initial works that favour data enrichment to alleviate the lack of expressiveness of the query methods.