Semantic approaches for query expansion: taxonomy, challenges, and future research directions

Rationale

The rationale for conducting this review on semantic QE is driven by the need to address the challenges and limitations of traditional QE. While the number of search engine users increase, the challenges to meet users’ needs also increase. Semantic QE is a powerful tool for meeting this goal. It provides the search process with an effective enhancement mechanism. However, the literature does not provide a comprehensive overview for researchers in the field. The aim of this work is to guide researchers and developers in the semantic expansion field by providing them with a comprehensive review, including the pipeline and tools. In addition, it reviews and compares recent frameworks to stay up to date with the latest advancements in this field. This work covers the revolution in semantic expansion, which includes simple linguistic techniques up to contextual embedding techniques.

Organization of the study

The remainder of the article is organized as follows. “Survey methodology” presents the survey methodology, while “Related works & research motivation” presents recent works that summarize aspects of QE. “Background” discusses background and refinement techniques. “Impact and Applications of Semantic Query Expansion” presents the usage and impact of using semantic query expansion in different fields. “QE categories” presents the QE categories. “Linguistic-based approaches” discusses some recent frameworks based on the linguistic-based approach. “Ontology-based approach” presents the ontology-based approach, and “Hybrid-based approach” discusses the hybrid approach. The “Discussion” section summarizes the challenges and gives an overview of open issues. Finally, “Conclusion” concludes the article.

Preprocessing

The first step is the preprocessing of the query. This is done by removing unwanted terms, such as stop words, and keeping only the important ones. For example, if the user searches using the query “What is the capital of Saudi Arabia?”, then the words “is”, “the”, and “of” do not play an important role in the search process; instead, they can affect the efficiency of the search application by adding extra and unneeded work. Preprocessing can be accomplished using many steps based on the application’s needs.

The main preprocessing procedures are tokenization, data cleaning and stemming as shown in Fig. 2.

Tokenization

In this step, the text is transformed into tokens, which are useful and meaningful elements. For this, we split the text into words or phrases based on the application’s workflow. Any n-gram tokenization could be considered based on the purpose of the framework. The above example presents uni-gram tokenization.

Data cleaning

After breaking the text into tokens, the unnecessary ones will be eliminated. Unwanted tokens could be stop words, symbols, URLs, extra white spaces, or words in different languages if the application only deals with a specific language and the text is written in more than one language.

Stemming

In this step, all the words are changed to their root word. Using the stem form is much more useful than the entered text since most of the external knowledge sources (i.e., dictionaries, lexicons, thesauruses, and ontology) are structured using the stem form of the words.

Many tools and libraries can be used in the preprocessing phase such as Natural Language Toolkit (NLTK, https://www.nltk.org), Industrial-Strength Natural Language Processing (spaCy, https://spacy.io/) and Gensim (Gensim: Topic modelling for humans, https://radimrehurek.com/gensim/), which can be used for cleaning data and stemming.

Feature extraction

The second step in the QE process is the extraction of the expansion features. This step is the main procedure that differentiates between the QE approaches and can be done in several ways.

After gathering useful words from the query to be expanded, they are transformed into vectors. The main idea of presenting the text as vectors is to determine the weight and dimensions of each word for further comparison. This creates a representation in vector space where spatial proximity indicates similarity and eases the computational processing of natural language. As illustrated in Fig. 3, each word will be transformed into a vector with a direction and a magnitude that represents its meaning. Consequently, identifying and manipulating the linguistic similarity of the words will be possible. To construct a vector, its dimensions (i.e., characteristics or features) need to be identified. For example, if we would like to present a group of children at school, we should first define the features of each child. As is shown in Table 3, we have chosen their math grades, English grades and age as the desired features and dimensions. For instance, the word “Jack” will be represented by the values of its features, which are 5, 10 and 7 in the Table 3. Therefore, “Jack” will be located in the vector space as the point (5, 10, 7). Consequently, each record in the table represents a vector of a child. Now, by using the vectors, we can present the data in the space as a point in the vector space as illustrated in Fig. 4. From the representation, we can conclude that Sara and Suen have similar features since they are close to each other in the vector space. On the other hand, Mike and Noor have different feature since their points are far away from each other in the vector space.

For QE, the representation takes a deeper path called word embedding. Word embedding is a method that assigns numerical representations to words, facilitating their quantitative comparison and analysis. This technique converts words, originally represented as strings, into vectors within a vector space utilizing chosen features or dimensions. Usually, each word is depicted as a one-hot vector in word embedding. These vectors are subsequently positioned in a continuous space, bringing together words with similar meanings in closer proximity (Mikolov et al., 2018).

To determine the features, several useful natural language processing (NLP) algorithms can be used. In the following, some of the widely used algorithms are defined (e.g., term frequency-inverse document frequency (TF-IDF), bag of words (BOW), and word embeddings.

TF-IDF: is a statistical model that identifies the importance of each word to a specific document in a collection of documents. In doing so, it computes the weight of each word based on its appearance in the document; however, it is offset by the number of documents that contain the word. The main logic behind TF-IDF is that it gives high attention to words that rarely appear in the corpus (Sharma, Tripathi & Tripathi, 2015).

BOW: is an approach that takes advantage of the occurrences of the most frequently used words in a text, which are collected as a bag of words. It builds the vector based on the presence of the words in that bag. It assigns 1 to a word if it appears in the bag and 0 otherwise (Yan et al., 2020).

Word embeddings: Although previous methods were useful for presenting text, they were lacking the ability to capture the semantic properties of words. To overcome this gap, word embeddings models have emerged. These models have received increased attention from researchers recently due to their efficiency. A word embeddings model works based on the distributional hypothesis of Harris (1954), which states that words that occur in the same contexts tend to have similar meanings. This concept was adopted in leveraging large text to build useful representations of the words. This type of representation is called word embeddings and can be formed in two ways: static or contextualised. Static word embeddings is a single fixed embedding for each word in the whole vocabulary, while contextualised word embeddings generates representations of words based on the context. Word2vec is a great example of static word embeddings. It was introduced by Mikolov et al. as an open-source project in Mikolov et al. (2013a) and Mikolov et al. (2013b). Word2vec is a neural net-trained model with two layers. Without human intervention, Word2vec processes text by converting the words to vectors. It generates a full numerical representation of words by considering the text features. It is based on either predicting the words by using the context (CBOW) or predicting the context by using the words (skip-gram) (Mohamed & Shokry, 2020). As a result, Word2vec gathers vectors related to similar words in a vector space. One of the most widely used pre-trained models is Google’s Word2vec model. It has word vectors for around 3 million words that are trained using 100 billion words (Mohamed & Shokry, 2020). In fact, some frameworks, such as in Li, Wang & Yu (2018), have used the Word2vec mechanism to determine the initial similarities between the entered query and a set of documents. Moreover, the GloVe model (Pennington, Socher & Manning, 2014), which is a model for the unsupervised learning of word representations, used Word2vec for training and outperformed some models on word analogy, word similarity, and named entity recognition tasks. Recently, there has been a growing trend towards utilizing contextualized word embeddings in QE and NLP tasks in general such as in Khader & Ensan (2023) and Wang, Huang & Sheng (2023) and Bhopale & Tiwari (2024). The Bidirectional Encoder Representations from Transformers (BERT) is a popular deep neural network that has been used in the field. The intention behind its design is to pre-train deep bidirectional representations from unlabeled text by simultaneously considering both the left and right context across all layers. Consequently, a pre-trained BERT model can achieve state-of-the-art performance across a diverse array of tasks with just one additional output layer during fine-tuning (Devlin et al., 2018).

Terms selection

In this phase, the candidate terms are selected. The selection is performed based on the relationship between the original query terms and the candidate terms. The similarity between them is the key to finalizing the relevant candidate terms list. To measure the closeness of the terms with the original query terms, a similarity measurement method is used that represents the next phase in dealing with the word vectors as illustrated in Fig. 3. In the ontology-based approach, for example, the depth of the term in the ontology is used as a measurement criterion. In the linguistic structure approach, methods such as Cosine similarity, Jaccard similarity, Euclidean distance, and Latent Semantic Analysis (LSA) can be used. After defining the similarity weight of each candidate term, a threshed is determined to choose the most relevant ones. In the following, we briefly discuss the main similarity measures.

Cosine similarity: Regardless of their size, cosine similarity shows the similarity between two documents based on the cosine angle between their vectors in the vector space. The similarity range is from 1 to −1, where 1 means the documents are very similar, almost identical. The narrower the angle, the more similarity between the documents (Futia et al., 2018).

Jaccard similarity: measures the similarity between two vectors using one feature at a time by taking the intersection of both and dividing it by their union. In other words, as presented in Eq. (1), it is the size of the intersection divided by the size of the union, where A and B are the vector representing the terms.

(1)

The similarity between the features in both vectors is computed. The intersection here represents the number of times the feature equals 1 in both vectors. The union is the addition of three parts. The first part is the total number of times the first vector value of that feature equals 1 while the other one equals 0. The second part is the total number of times the first vector value of that feature equals 0 while the other one equals 1. The third part is their intersection (Verma, Agarwal & Khan, 2017; Rinartha & Suryasa, 2017).

Euclidean distance: represents the shortest path between two vectors. It is the true straight line distance between two vectors. It measures how close two word vectors are in the vector space (Zhang et al., 2019).

Latent semantic analysis: learns latent topics by performing a matrix decomposition on the document-term matrix using singular value decomposition (SVD). LSA is typically used as a dimension reduction or noise-reducing technique. LSA aims to reduce the dimensions of a document by using the topics that present the underlying information in the document (Dahir & Qadi, 2021). It decomposes the m*m documents/terms matrix, where m represents the number of matrix dimensions, into three matrices. The first one is the m*n singular matrix, which represents the terms/topics matrix that assigns the terms with their topics. The second one is the n*n diagonal matrix, where n represents the number of the matrix dimensions, which represents the topics/topics matrix that determines the importance of the topic. The third one is the n*m singular matrix, which represents the topics/documents matrix that shows the distribution of the topics across the documents. After that, it rearranges the matrices based on the descending order of the topics/topics matrix. Finally, it chooses the topics with the highest importance values (El Ghali & ElQadi, 2017).

Terms ranking

To show search results to users, they must be ordered by how close they are to the user’s original query. Since each of the selected candidate terms has a similarity value, the results will be ordered based on that value. Thus, the terms will be ranked based on how similar they are to the user query. To manage the closeness of the terms to the original query terms, a threshold could be established that is used to define the minimum value of the needed similarity score. The selection of the threshold should not be random. Since it has a direct effect on the results, it should be chosen carefully. Indeed, several experiments could be conducted to determine the best threshold value.

Query reformulation

After manipulating the query and exploring the related terms, the selected ones will be used to reformulate the query by replacing the original query terms with the new expanded ones. The replacement may concern one term at a time or multiple terms, depending on the application. Consequently, the documents will be returned based on the newly expanded queries.

Evaluation

Retrieved documents play the main role in evaluating the QE framework. Generally, the performance of QE frameworks could be measured by examining the portion of retrieved documents that are actually relevant. Recall and precision are the metrics used to test an IRS (Singh & Sharan, 2018), as they measure the desired improvement. The recall metric is used if the tests include searching in a specific database, for instance, Text Retrieval (TREC) datasets (https://trec.nist.gov/). It measures the ability of the framework to retrieve all relevant documents in the database, see Eq. (2). On the other hand, the precision metric measures the relevance among retrieved documents from a limitless database, such as the web, see Eq. (3). Moreover, average precision is used with ranked retrieval results, see Eq. (4). It is considered if more relevant documents are required to be returned earlier (Zhang & Zhang, 2009). MAP is the mean value of the average precision for multiple runs of the framework. Precision can be maximized using precise terms in the expansion process, while using general terms can maximize the recall (Selvaretnam & Belkhatir, 2011).

(2)

Furthermore, to evaluate the QE framework, a baseline model should be chosen to use its results for comparison. The baseline model could be a state-of-art model, pseudo-relevance feedback model, Rocchio model, or a model without expansion. The latter model runs the framework with and without expansion and compares their results (Raza et al., 2019).

It is worth mentioning that user evaluation could be considered in the evaluation process to seek their satisfaction; however, it can increase the computation time since it can add turn-around time along with additional computation.

Local-source category

This approach relies entirely on user feedback. For previously discovered queries, the feedback on the relevance of retrieved documents plays a key role in choosing related terms. Pseudo-relevance feedback (PRF) is an example of this feedback method, which selects related terms from the top k relevant documents (Zingla, Chiraz & Slimani, 2016; Dahir, El Qadi & Bennis, 2021).

Query logs also fall under the local-context-based approach. This method utilizes the interactions between the user and the IRS, mapping the users’ logs to the new query (El Ghali & ElQadi, 2017). While this approach is simple and effective in some cases, it depends heavily on the user’s past behavior, making it difficult to automate in an integrated system.

Global-source category

This category includes methods that rely on external corpora, such as WordNet (Fellbaum, 1998). Terms are extracted based on the statistical probabilities of terms present in the corpus, which is done by establishing a language model. All retrieved documents are included in the extraction process, with candidate terms selected based on their relationship with the corpus terms (Zingla, Chiraz & Slimani, 2016). While both local-context-based and global-context-based methods can help expand queries, the lack of semantic linkage between terms in these approaches may cause query drifting. Using external corpora increases the generality of the terms, but since it is based on statistical calculations, there is no guarantee that the extracted terms will be meaningfully related to the original query. Additionally, because all retrieved documents are involved, the process can be computationally expensive (El Ghali & ElQadi, 2017).

While this method is effective for expanding a query, the relationship between the terms is typically statistical and context-dependent, rather than semantically meaningful. The lack of semanticity makes the distinction between global-context-based and knowledge-based methods more subtle.

Knowledge-based category (Semantic expansion)

Semantic query expansion (QE) represents a more advanced technique in the field. This approach enhances query reformulation by making the query more “intelligent.” It is a key aspect of the semantic web, where information is interlinked. By structuring the query to better capture its meaning, semantic QE enables a more precise understanding of the query. This category focuses in selecting terms that are semantically related to the original query terms, rather than based on statistical patterns. This makes knowledge-based methods more precise, ensuring that expanded terms maintain the semantic integrity of the query.

This category is typically built on a knowledge structure, which can be either a linguistic structure or an ontology. Consequently, the knowledge-based category can be further divided into three approaches: linguistic structure-based, ontology-based, and hybrid approaches.

Hybrid approach

The hybrid approach combines the strengths of both linguistic structure and ontology-based methods to enhance query expansion. In this approach, the first step involves extracting terms from the linguistic structure, followed by exploring the terms within the ontology to find related concepts (such as parent and child terms). While this approach can yield more semantically relevant terms, it is computationally expensive and may still suffer from query drifting.

Table 4 illustrates a brief comparison between these three structures based on their complexity, whether they are computationally expensive or not, and source aspects such as their availability, integration ability and whether they support domain specification.

Table 4:

Semantic expansion approaches.

Approach	Computationally complex	Source availability	Source integration	Support domain specific search
Linguistic structure approach	No	Yes	Easy	No
Ontology approach	No	Only domain specific	Difficult	Yes
Hybrid approach	Yes	Only domain specific	Difficult	Yes

DOI: 10.7717/peerj-cs.2664/table-4

Based on the above semantic expansion approaches, all the frameworks have been categorized. Figure 6 presents a proportional distribution of the frameworks based on the semantic approach they use. In the following three sections, we examine the frameworks based on the approach they use.

Linguistic models

Lu et al. (2015) proposed a method to help software maintainers perform updates on projects. Their work mainly focuses on expanding the query to improve the search for code. The main source used is WordNet (Fellbaum, 1998). The authors’ basic methodology is to apply the POS with each word in the query. Then, all the resulting tags will be matched with their related synonyms in WordNet (Fellbaum, 1998). Moreover, the methods identifiers are collected to be used in the query matching phase. Then, the results will be chosen based on their similarity with the natural language phrases that exist in these identifiers. The similarity score is calculated based on the total number of common words between the expanded query and the identifier divided by the total number of words on both. The candidate selection criteria depend on how far the tag is from the threshold set by the authors. To test their approach’s effectiveness, they performed two main evaluation stages. First, they compared the result of the expansion method vs the original one. The main comparison was based on the number of relevant and irrelevant suggested terms. The results show an improvement in precision and recall with 40% and 24%, respectively. Second, they compared their approach with the result of Conqure (Hill et al., 2014). Their approach outperforms its precision by 5% and its recall by 8%. The approach is simple, yet it showed improvement in the code query’s effectiveness due to the consideration of methods identifiers. However, the performance of the approach depends indirectly on the quality of the codes. Hence, using a specific domain ontology would be more useful and comprehensive.

Sharma, Tripathi & Tripathi (2015) provided a semantic expansion technique to overcome the delay in searching for patents by using patent abstracts as a query and the International Patent Classification (IPC) as metadata to reduce the search time. Their method is based on three main stages. Firstly, all the relevant abstracts from the IPC are extracted. The extraction is performed by calculating the Term Frequency—Inverse Document Frequency (TF-IDF) feature. Secondly, the expansion phase occurs on two levels: extracting relevant words from WordNet and extracting reliable page titles from Wikipedia. With WordNet, a vector is established. It contains weights ranging between 0.0 and 1.0 for each word. The closer the new word is to the original word, the higher weight it holds. After that, a vector of incoming and outgoing links is calculated using the Page Rank algorithm on Wikipedia. To avoid failure, each page has from 10 up to 100 links, which means there is limited expansion. Finally, the final semantic similarity will be based on the expanded vectors. Its measurement is conducted by using cosine similarity and the Jaccard coefficient. To test their work, the authors performed experiments with and without the suggested expansion. The precision and recall for the former showed better results. The integration of using different resources is a sufficient improvement. Moreover, the usage of IPC can provide more accurate and faster results. However, in the semantic similarity calculation, the Jaccard coefficient method showed poor performance. Also, they did not clarify a threshold to accept the term based on its semantic similarity score.

El Ghali & ElQadi (2017) attempted to eliminate the ambiguity of a query. They took the environment of the query into consideration by building the context around it with the help of the language models (LM) and applying the LSA method. Their work was based on two main phases. The first phase is the query recommendation phase. This step is based on the user’s past searches. By using LM, the past queries are measured and ordered by their correlations with the new query. This method relies on the presence of a term in a query and the presence of a document among the clicked documents. The ordered list of past queries will then be handled by the next phase. The second phase is the LSA method for QE. This phase considers the conceptual meaning of the words. It constructs a word-document matrix to determine the word usage among documents by using singular value decomposition (SVD). SVD assigns a vector to each term. It focuses on the occurrence of the terms in similar documents and queries, even if they never co-occur in the same query. After that, the vectors will be ranked using the similarity score for each query term by using the Cosine similarity measure. To validate their work, the authors used the CISI text database from SRT. They used the original query results as a baseline. Their system improves the precision by 24% and F-measure by 7.76%. The method took advantage of the user’s query log to define the semanticity based on the user’s preferences. However, the log size should be considered since it can affect the time efficiency.

Goslin & Hofmann (2018) used a states-based approach to generate the candidate terms. It is based on providing a sub-state framework where the surviving related candidate terms for the current state will be moved to the next state to extract more related ones. First, it constructs the root query, which is treated as the basic reference throughout the states. After preprocessing of the query, if its length is three tokens, then a connection is made with Wikipedia pages corresponding to it. Otherwise, the N-grams feature is applied to longer queries. For each token, the term frequencies are stored. Then, each term will be passed to six data utilization modules to gather the corresponding weight of the terms. The six data utilization modules are developed linguistic modules. Consequently, a stem query is generated to be passed to the next state. The stem query is selected based on the terms with the best weights from the last step. To prevent query drift, the original query is passed to the next state along with the stem query. The final selected terms are the output of the final state. To test the approach, they conducted a comparison experiment with five existing algorithms. A total of 50 different web topics were used. Their approach had an overall mean average precision of 80%, which outperformed the other algorithms by 27%. Overall, the approach highlights the importance of using states to examine the term importance. Furthermore, query drift was prevented by appending the original query in each state. However, time-complexity-wise, the algorithm is based on revisiting Wikipedia, which could be computationally expensive.

Dao Thi Thuy et al. (2017) proposed a method for retrieving images by using an image as a query. This method relies on the user to determine the semanticity of the images (Dao Thi Thuy et al., 2017). In addition, it differentiates the features by weighting their semantic importance to gather more accurate results. The main idea of the proposed algorithm is to transfer the user image query into a multi-point query. Each query point will be considered as a cluster in the image database for further search. After the user sends their image for a search, the query image will initially be changed to four representations, which are multiple versions of the image with different color layers. Each representation is a cluster. The feature vectors will be extracted for each of them. Then, the distance between multi-point query and the images in the database will be calculated and assigned based on the minimum distance between the image, the feature importance weight and the query points. The feature importance weight is used to increase the semanticity of the image. Indeed, the feature importance is calculated as the inverse of the variance of the feature in each axis in the multi-feature space. This distance is combined with the semantic weight of each query point, which is calculated using the number of related images in each cluster. Based on the user selection, the selected images will be the clusters’ new centroids, and the above algorithm will be repeated. If the user is satisfied, then the algorithm stops; otherwise, the whole procedure will be repeated with the user intervention until they are satisfied. The method was tested with a database with 34 categories with a total of 3,400 images. It produced a higher precision compared to the other four methods. This method takes advantage of the user’s judgment to determine the semanticity. Also, it brings attention to the feature importance and how it can be used to get accurate results. However, when calculating the semantic weight of each query point, the system relies on a traditional content-based retrieval algorithm. Indeed, building the clusters around the centroids lacks any definition of real semanticity between them; this can affect the actual semanticity since the images’ relatedness is not clearly identified.

Nasir, Varlamis & Ishfaq (2019) proposed a framework to improve a method for automatically enhancing the relevant terms by using external information sources. The semantic measurement is calculated for the query and the related documents and for the query and the candidate terms. Firstly, the related documents are extracted based on the probability of the query terms being on them. Then, an automatic relevance feedback method is used to select the most semantically related documents. For the relatedness between the query and the documents, a combination between three measurements is used; these are omiotis, Wikipedia link-based and pointwise mutual information measurements. The first is based on calculating both the semantic compactness and semantic path elaboration of two words using WordNet. The second measurement is based on the relatedness between terms based on Wikipedia articles. Two terms are related when they have been referred by the same Wikipedia articles. The last measurement measures the co-occurrence of two terms in a large document collection by how often two terms occur together. After that, the top-M documents are selected, and their words and the query terms will be measured with the original query by the above measurement techniques to establish the candidate terms. To test the framework, they used three popular datasets. As a baseline, they applied unigram and Okapi methods. They measured the framework performance with precision and recall. The experiment with Omiotis measurement provided the best results for one of the datasets, while the Wikipedia measurement outperformed the other measurements in different datasets. Also, they compared the framework performance with other existing works, and it outperformed them on some datasets. This framework introduced a multilevel semantic measurement; thus, more related documents will be retrieved. In addition, it considers the relatedness between the query terms and the documents before the expansion. However, the threshold of semantic degree for selecting most related terms is not specified.

Riyahi & Sohrabi (2020) proposed a hybrid recommender system to generate recommendations in discussion groups by utilizing information about both the content and the user. The method has three main stages: content-based filtering, collaborative filtering and hybrid filtering. Based on the content of the posts and tags, similar posts will be collected. The tags will be extracted, and a semantic hierarchical structure of them will be built by using WordNet as an expansion source. In the collaborative-based filtering phase, the implicit ratings of all users are obtained to find the users most similar to the active users. The users’ ratings will be based on their behavior such as their comments or their favorite posts. Consequently, the similarity between the active user and the users will be determined by using Cosine similarity measure or Pearson correlation coefficient. As a result, users similar to the active user will be collected. In the content-based filtering phase, the system will generate a recommendation list of the user’s questions from the hierarchical structure of posted question tags. If the question tags do not exist in the hierarchical structure, a search for their synonyms is conducted using WordNet. As a result, posts similar to the active user’s question will be collected. Finally, in the hybrid stage, the most similar posts with the most interactions from similar users will be returned. The method was compared with other recommendation systems, and it achieved an improvement in precision by 35%. This approach is useful for enhancing the quality of discussion group websites. Also, by suggesting similar existing posts, it saves on system resources by minimizing the need for saving posts that already exist. However, using similarity measurement between the new tag and its WordNet synonyms can detect more related tags and thus more related posts.

In the work of Dhokar, Hlaoua & Romdhane (2021), the authors presented a framework for tweet contextualization. It aims to determine the context of a given tweet by retrieving related documents. To do that, they used semantic QE techniques, where they treat the tweet as a single query. The main aim of employing semanticity is to retrieve as many relevant documents as possible to avoid any mis-contextualization. The main purpose of this work is to find an answer to the question “what is this tweet about?” by viewing a summary of information. This summary is constructed from related documents that represent the tweet concept. To collect such documents, a semantic QE approach (SemQEx) is represented. It relies on WordNet as an expansion source. The query (tweet) is annotated via TreeTagger, which is an NLP tool that applies POS on the query terms. Then, candidate terms are selected from WordNet based on their similarities with the query terms. This is accomplished by considering the synsets and their definitions in WordNet. Indeed, the similarity score is calculated based on the occurrence frequency of the query term in the synsets and its definition. Thus, it is considered as a candidate term if the original term appears as a synset or in its definition. To test their work, they applied a dissimilarity score to test the improvement of their work and succeeded in decreasing the score.

Linguistic-embeddings models

General domain

Beirade, Azzoune & Zegour (2019) proposed a Quran-based search engine and built a Quranic ontology. The Holy Quran has special characteristics, and words used inside the Quran may have different meanings compared with when they are used outside of it. The authors constructed and built the Quranic ontology with respect to its linguistic and semantic concepts. Relationships have been established between the key concepts in the Quran (e.g., relation of synonym, relation of antonym, relation of hyponymy and relation of similarity). Their search engine works through several steps. After applying lemmatization for each word, all candidate terms are collected from the Quranic ontology (i.e., all child nodes of the term in the ontology will be collected). Then, the system will collect the terms that are semantically related to the original query terms. Two Quranic search engines were used to evaluate their approach: the Alfanous Quran search engine and Corpus Quran. The approach system gives 70% average precision, which is the highest precision among other search engines. Indeed, using recall as an evaluation metric is required since the system uses a limited dataset, the Quran, to search. The approach’s strength is in the terms expansion phase. However, it does not provide a ranking mechanism since it expands the query with various terms, which can result in a massive number of verses.

Furthermore, Al-Smadi et al. (2019) attempted to build a bridge between Arabic questions and DBpedia. They proposed a framework to create SPARQL queries from the entered question. For preprocessing, they performed tokenization, stemming and named entity recognition. The latter is the most important since it could identify the names in the question. To perform the preprocessing steps, they used a Java library called FARASA. After that, they collected the resource identification and labels, DBpedia in their case. Then, they acquired the properties from the question, which connect the subject with the objects. Since DBpedia does not have Arabic labels for most of its represented properties, they performed the properties collection with different approaches: the baseline approach, Wikidata-based approach, and dependency parsing-based approach. The dependency parsing-based approach generated the best results for their framework. After that, they classified the triples to subject-based triple and object-based triple and formulated the SPARQL query based on that. After evaluating their work, they achieved 84% precision.

The work of Jain, Seeja & Jindal (2021) considered working with fuzzy ontology to show its impact on the QE process. To perform this, their framework has three main phases: fuzzy ontology construction, QE using the fuzzy ontology and retrieving the needed information based on the expanded query. First, to build the fuzzy ontology, a dictionary is created by applying text mining on a certain corpus with the help of external ontologies. The main purpose of the text mining is to extract the main concepts on the selected corpus. Then, the ConceptNet ontology is used to define the relationships of the concepts, whether they are synonymous, hierarchical or functional relationships. Then, based on the type of relationship, a semantic weight is assigned for each pair of the concepts to complete the creation of the fuzzy ontology. After that, the fuzzy ontology is used to expand the user’s query. This is performed by selecting the top three related concepts based on their weights. Then, the information retrieval process is performed using the expanded query. To test their work, they constructed a fuzzy ontology using the Solar domain using a database from the UCI Repository. Then, the expanded query is tested throughout the web search engines. Their framework reached 80% precision among Yahoo, Google and Bing.

Domain-specific

In Yan et al. (2017), the authors focused on searching in the ontology itself using RDF queries. They tried to find solutions to queries with too few results or empty results. They calculate the similarity between the query and its candidate by considering the similarity of two triples in the ontology and the IDF score of the candidate triple. They developed a relaxation algorithm to produce relaxed queries and ranked their importance based on their semantic similarity with the original query. Their methodology was based on ontology relaxation of RDF triples. This means dealing with the entailment in RDF. They studied the semantic similarity with two factors: the semantic overlap-ratio and semantic depth. The former represents the number of mutual elements in two RDF triples. The second depends on the ontologies’ hierarchical depths of the concepts. It is measured by the distance between the hierarchical depths. Also, every RDF is weighted by using TF-IDF. The final similarity degree will be computed by using the semantic overlap-ratio, semantic depth and the RDF triple weight. In the relaxation method, the top-k relaxed queries are then chosen to be executed based on the total similarity degree and a defined threshold. They tested the method by using a dataset with 100 K distinct triples. Queries with empty answers were used. They compared their work with other published methods and got the minimal running times. Their algorithm outperforms other algorithms with higher recall; however, based on their precision results, their ranking methodology needs further development.

Deepak & Priyadarshini (2018) proposed a feedback-based approach using appropriate ontologies for homonyms to retrieve images semantically. The first phase of the system focuses on building the ontology. The ontology is constructed for the homonyms of the search keywords. Multiple ontologies can be added as required. The system saves the homonymous list to guide the homonym LookUp Directory, which is a HashMap that has a key for multiple values. These keys are considered as indices to the ontology tree. When the user enters the query, after it gets cleaned, the remaining words will be passed to the semantic algorithm. Based on homonym LookUp content, this algorithm is used to choose the related ontologies. After that, they will be classified using the SVM classifier. After the classification, the semantic similarity is measured between the terms and the class labels of the ontologies. An OntoEntity is then formulated by adding the matching class label, the matching homonym LookUp content and the similarity score. Further, the similarity is measured between the OntoEntity and the metadata (i.e., the description of the image). When the user sees the results and clicks on one of them, this decision will be considered later in the algorithm and used as content information. Thus, the algorithm will work based on the lookup directory along with the content information and construct the semantic similarity by using semantic equivalence matching. The content-based analysis is done by keyword matching between the query original terms and the selected image URL words. Based on that, the classes of ontologies will be prioritized and the most semantically related ones will be displayed. The semantic measurement is done using pointwise mutual information measurement. The proposed framework was tested and compared with four other similar works. It achieved 95.33%, 96.41% and 95.87% on precision, accuracy and F-measure, respectively. The framework constructed a separate phase for building an ontology repository, which makes updating it much easier. It incorporates user intervention in building the ontology, which may be misleading if the user enters inconsistent data into the ontology. It also incorporates user intervention in the semantic measurement but without relying on it to determine the semantic score.

Al-Nazer, Albukhitan & Helmy (2016) worked on utilizing existing semantic web repositories as well as building a bridge between non-English speakers and English resources. This was done by implementing a natural language interface to the English ontologies. This interface will match the entered query with the most relevant semantic web query written in SPARQL, which can be used to get RDF results. The SPARQL query is created based on the expanded queries of the entered query. The authors’ architecture is built in four main stages. First, the entered query will be preprocessed to eliminate any useful words and then get translated to English. After that, in the query mapping phase, the query is classified to create the correct SPARQL query. After that, if any measurement is listed in the query, it will be determined whether it should use the SPARQL query. Next, semantic mapping is performed between the query terms and ontology name entities to identify the ones most relevant to the ontology topics. Finally, based on the relationship between the terms in the ontology, the relationship between the terms will be identified. As a final step, the SPARQL query will be generated based on all the expanded queries terms and be able to be executed to gain the related RDF results. To test their work, they built an integrated ontology from food and health ontologies. They compare the performance of the approach with a manual approach. A total of 389 questions were used. First, their concepts were manually identified. After that, the same experiment was conducted using the proposed approach. Their approach’s precision, recall and F-Measure are 89%, 79% and 84%, respectively. The proposed framework could capture more semanticy if the original terms were expanded before matching them with the entity names’ components. In other words, if the query terms do not match any topic from the entity names, the SPARQL query could be empty.

In Dahir & Qadi (2021), the authors attempted to utilize DBpedia attributes values by choosing the most appropriate ones. To achieve that, they used an integration of topic modeling to find out how topics and attributes are semantically related. Their methodology has two main phases. In the first phase, the relationship between the query and the document is determined by a language model. Then, Bo1 is used to assign the probability distribution to select the top Bo1 expansion terms from the returned documents. In the second phase, for each Bo1 expansion term, DBpedia attributes are selected, whether they are single or multi-valued. After that, latent Dirichlet allocation (LDA) is applied to Bo1 expansion terms from DBpedia by using two attributes to collect the final candidate expansion terms. LDA is used to find the distribution of the topic in documents and the distribution of words of the document to topics. Their experiments showed that LDA can solve the problem of multi-valued attributes of DBpedia, which means increasing the efficiency of the selected terms.

Table 6 presents a summary of the frameworks based on their techniques, sources, datasets, language used and evaluation metrics.

General domain

He et al. (2016) worked on an image retrieval system based mainly on the CYC knowledge base. Their method is constructed through three phases. First, an interface is provided to the user to pass the initial query. Second, in the QE candidates generator, using the CYC knowledge base, all related terms are extracted. This is done by digging further into CYC ontologies to find related concepts and passing them to the QE candidates ranking system as the expansion candidates array. Third, in the QE candidates ranking system, all the array elements’ semantic similarities will be measured. The semantic similarity is calculated based on the WordNet (Fellbaum, 1998) similarity measurements. After that, the most relevant candidates will be returned to the user to expand their options. Then, based on their deeper selection, the final results will emerge. To prove their approach’s efficiency, the authors used two comparison strategies. First, they tested the system with and without their solution. The results showed that the expansion approach had 27% higher precision than the other one. For the second comparison, they compared the results of their approach with the results of Bing, a well-known search engine. The proposed approach produced more effective candidates. The effectiveness of their approach is 6.4% higher than Bing. The proposed model demonstrates a good interaction between the user and the system. It depends on their dynamic feedback rather than their search history. Also, it uses multiple expansion resources to assess the semantic relationship between the query and the candidates. However, it does not consider the image features in the semantic measurement. It focuses only on analyzing the query text and its related concepts.

Zingla, Chiraz & Slimani (2016) attempted to overcome short queries issues on microblog platforms. Their approach is based on extracting terms using text mining. Their methodology has two main phases. In the first stage, all the candidate terms are generated. In this stage, three steps are used. First, unstructured texts of articles from Wikipedia that are related to the query will be selected. The selection is done by using TF-IDF, and then association rules are applied between query terms and expansion terms to choose the candidate terms. The second step focuses only on the definition of the Wikipedia-related articles. The terms are extracted from the first sentence and the first paragraph of the article. In the third step, SPARQL is used to match related terms of the query with the DBpedia using the description of the concept. In the second stage, the candidate terms will be filtered based on a semantic relatedness measure that includes semantic analysis of Wikipedia and the resulting confidence of the association rules. All the terms that have semantic relatedness measures that are greater than a defined threshold will be selected. To evaluate their work, they performed several experiments on Twitter. The experiments showed good performance when the extraction techniques were combined and both filtering techniques were used.

Futia et al. (2018) illustrate a framework to semi-automate semantic graph generation. They structured the system by using a neural network and used SPARQL queries as a training set. SPARQL is extracted by retrieving all the properties that exist in the queries and stored as URIs. Also, the data of these properties will be retrieved. After that, triples from the queries data are extracted. Then, for each concept among the queries, the label of the highest level classes of the concept is extracted. Before the clustering phase, the neural language engine takes the SPARQL triples and transforms them into sentences for training. After that, vector representation is generated for SPARQL variables, which will be clustered based on the close proximity using cosine similarity. At this point, by using the GUI, the user can edit the results of the clustering. Finally, in the semantic graph mapping phase, the median value of Levenshtein distances is used to assign an attribute in the data source with its close cluster. To test the system, the authors compared the precision of their semantic mapping with the semantic mapping generated by domain experts. The ontology used was DBpedia. The result showed a complete overlap between them. The approach introduced the use of AI in the semantic expansion field. However, to make the framework more intelligent and effective, there should be less user intervention involved and linguistic sources such as WordNet could be useful. In addition, Levenshtein distances focus on the spelling of variables; therefore, it does not consider the semantic relationships. Semantic analysis could be performed by stemming the variables in the cluster and then using semantic measurement to accept the new ones.

Later, Song, Huang & Sun (2019) presented a framework that makes a similar contribution as Al-Nazer, Albukhitan & Helmy (2016). They proposed a semantic query graph to construct the information semantically with the help of a QE approach. The main idea behind this is to translate the NL question query to SPARQL. The framework works in four phases. First, in the query dependency parsing, the question structure will be analyzed by using Stanford Parser, an NLP parsing tool. This analysis includes determining the dependency relations of the question words. Then, the dependency parsing tree is generated based on the relationship between the words. After that, the tree structure will go through an optimization process that includes merging words, removing less semantic dependencies and joining dependencies; then, a dependency parsing graph is constructed. Secondly, based on the graph from the previous step, the semantic query graph is constructed with respect to the main entity on the question, the question type and the verb type if the question contains one. After that, using string similarity score, the semantic similarity is calculated by using WordNet to match the semantic query graph with the knowledge base: BDpedia Spotlight. Then, the relation edges in the graph are mapped to predicates and the entity nodes to the attribute value. Finally, the corresponding SPARQL is generated. This is done by traversing the semantic query graph where the edge represents the SPARQL conditional statement. The SPARQL variable will be the variable node, and the resource will be the entity node. Four different datasets were used to test the approach. The results were obtained by using their approach and three other ones. When their approach was used with two datasets, it showed a good performance; however, the approach did not introduce an optimal solution, as it failed in capturing and translating some questions. Using a different knowledge base may produce better performance of the approach since BDpedia stores the resources in a format that may differ from the ones in the question.

Furthermore, Nehar et al. (2022) established a hybrid semantic statistical QE for an Arabic IRS. They relied on two modules: semantic and statistical. The semantic module is based on using the AraBERT model (Antoun, Baly & Hajj, 2021) by preparing both the queries and the documents to follow the AraBERT representation. Afterward, they trained the MLP model as a relevance classifier. To do so, they used the EveTAR dataset to train and test the model. For the training phase, each query had 10 positive samples (i.e., relevant documents) and 10 negative samples (i.e., irrelevant documents). For the testing phase, they used the documents retrieved by the BM25 retrieval model. On the other hand, the statistical module relied on using the PRF approach by selecting the 1,000 top documents from the initial set resulting from BM25 model. Then, they used RM3 to select the 10 top terms. Finally, to perform the expansion, both terms collected from both modules were used. Their framework reached 72% MAP.

Kumar & Sharma (2022) presented a hybrid approach for QE and applied optimization algorithms to obtain the optimal terms for expansion. Their framework employed a topic mapping dataset to use it as an expansion source. They establish a topic map in order to use it later for terms matching. To do so, they used Wikipedia articles and newsgroup datasets. Once the main topics are extracted, Wordnet is used to add related terms to each topic. Therefore, when the user enters a query, the framework tries to find the related terms from the topic map by applying Jaccard similarity. Then, the resulted terms must be passed to the optimization module. To accomplish that, they used two optimization algorithms: spatial bound whale and binary moth flame. The algorithms work in parallel and, finally, produce the optimal set of related terms for the expansion. They achieved 40% MAP@top-3.

Domain-specific

In the work of Malik et al. (2022), a hybrid QE framework was proposed for biomedical field queries. It aimed to overcome the vocabulary mismatch problem, which can occur due to the diverse lexical variants in the biomedical field. They used three main sources to enrich the query: clinical diagnosis information (CDI), Wikipedia and Mayo Clinic. Their framework consists of five stages. In the first stage, the main goal is to extract biomedical concepts. This is performed by using MetaMap annotation, which extracts the desired concepts from the unstructured clinical notes, and each one is assigned by an ID called CUI (Concept Unique Identifier). In the second stage, queries are created from the extracted concepts and sent to the Google Custom Search Engine (GCSE). Then, the GCSE fetches both Wikipedia and Mayo Clinic results and adds them to the collected related results. After that, the original query is expanded further by using the pre-trained word embedding model by collecting candidate terms that are semantically related to the original query via cosine similarity measurement. They use three types: domain-specific, domain-agnostic and hybrid. Furthermore, the resulted candidate terms and CDI are merged and sent to the PubMed corpus, a pre-processed indexed biomedical corpus, to obtain the best combination for the query. Finally, based on the final query, the related biomedical literature is retrieved. Their experiments showed an increase in precision when they used the hybrid word embedding model.

The work of Lechtenberg et al. (2022) represented a query-by-document framework that aims to take the seed document as a seed and construct a query string. This query will then be sent to an API to fetch for related scientific documents. Their methodology presents a QE mechanism with a different order and purpose than the previous works reviewed. First, they used the seed corpus to formulate the targeted search domain. To do so, a list of keywords are extracted from that corpus to determine the characteristic of the domain knowledge. These keywords are chosen based on their TF-IDF weights. Then, to fetch for the related documents, a query is constructed from the list of the extracted keywords. To accomplish this, they used the Monte Carlo (MC) sampling principle, where the query is generated by choosing candidate keywords from a list that have a probability distribution based on their TF-IDF values. After that, the document frequency is calculated based on its appearance in every MC iteration. Indeed, the document frequency will be used to rank the relevance of the resulted documents. In addition, the semantic rank of the results is calculated based on cosine similarity measurement. To test their work, they used two case studies. In the first one, they targeted two different search fields and attached a high recall value, 83.9%. In the second case study, they compared the methodology with different sampling techniques and concluded that MC reached the highest linguistic relevance results.

Table 7 presents a summary of the frameworks based on their techniques, sources, datasets, language used and evaluation metrics.

Challenges

Based on an overview of recent works on semantic QE, summarized in Tables 8, 10, 11, various challenges and open issues can be identified and explored. These challenges and issues must be addressed as they may significantly impact the system’s performance. It is imperative to carefully analyze and address these challenges to ensure the optimal performance of the system.

Open issues and future research directions

The revolution of semantic QE has enriched the field of semantic web and IRSs in various ways; however, there are still open issues that should be the focus of future research directions and improvements.

• Constructing an ontology can be a time-consuming and demanding task. Therefore, it is important to develop an algorithm that can automate this process while integrating the expertise of ontology developers and experts. For instance, in Deepak & Priyadarshini (2018), an approach is proposed that accomplishes this process in a more controlled and careful manner.

• The availability of ontologies and lexicons is limited in languages other than English, such as Arabic, which can hinder the expansion of candidate queries. Building a multilingual ontology or lexicon can improve the QE to overcome this limitation. In Al-Smadi et al. (2019), an automated question/answer system is introduced that bridges the gap between Arabic questions and linked data by translating natural language questions to SPARQL queries to retrieve answers from linked data, such as Dbpedia. To achieve this, the named entity of the original query is discovered with the help of Wikipedia labels available in Arabic. In contrast, AlAgha (2015) utilizes the statistical parser of the Arabic Toolkit Service to identify the subject, object, and predicate of the Arabic text and construct the corresponding RDF triple. To improve the reliability of linguistic resources, slang phrases must also be considered. N-gram and word/sentence embedding can play a significant role in query processing, aiding in catching such phrases with the help of reliable sources.

• Applying AI to the expansion field is a challenge, but it could add intelligence to the approaches and generate accurate and useful expanded queries. Deep learning, for instance, could help in recognizing hidden patterns among texts or images, which can provide more semanticity. Furthermore, it can provide contextual analysis of the text and match it with an accurate topic; thus, the framework can proceed to fetch the closest topic. For instance, Mohamed & Shokry (2020) represents a framework that applies a machine learning approach that aims to train a Word2vec model using CBOW on an Arabic corpus and use it for searching the Quran. It scored a high precision by increasing the intelligence of the vector structure. In addition, Fang, Zhang & Yin (2018) used a trained Word2vec model to capture both sequential and semantic information of biomedical texts. Furthermore, in applications that focus on searching in domain-specific areas, machine learning will be useful since the training dataset is precise and determined. For instance, in Zhong et al. (2020) the authors used a deep learning methodology to construct an answer/question system for building regulations to help engineers in retrieving needed information. In addition, Cakir & Gurkan (2023) and Khader & Ensan (2023) showed a great potential in using generative models that can expand the query in an intelligent way.

• In the realm of semantic QE, incorporating real-time data is an area that is yet to be fully explored. However, a real-time semantic search has the potential to provide valuable insights on trending topics, especially during emergencies and major events, when people tend to scour social media platforms for information. By leveraging the user’s semantic attributes, a framework described in Zhu et al. (2017) was able to effectively search Twitter stream data, yielding more accurate and interesting results.

• The semantic search for multimedia content, such as images and videos, is still in its nascent stages. Existing frameworks rely on user intervention to determine semanticity, as seen in Dao Thi Thuy et al. (2017), which can be misleading and fails to provide an actual semantic expansion approach. Incorporating deep semanticity with minimal user judgment can lead to better semantic frameworks. One way to achieve this is by constructing a vector space of the images based on multiple features to create a digital description of the multimedia content. This approach was used in Tautkute & Trzcinski (2021), where the authors aim to extract semantic information from different data forms for image searching. They use a GAN framework to generate a synthetic-related image based on a multimodal query containing both an input image and textual description, which is then used to search for more related images.