Entity Linking is the task of identifying an entity mention in unstructured text and establishing a link to an entry in a structured Knowledge Base (KB), such as Freebase , DBpedia , etc. It is an essential component of many information extraction and natural language understanding pipelines since it resolves the lexical ambiguity of named entities.
. The state-of-the-art neural entity linking models have shown undoubted improvements over classical machine learning approaches based on feature engineering. In this survey, we systematize recently proposed neural models, distilling one generic architecture used by the majority of the neural EL models, but also discuss its prominent variations.
The important component of neural entity linking systems is entity vector representations and entity encoding methods. It has been shown that encoding the KB structure (entity relationships), entity definitions, as well as textual information in large unstructured corpora, helps to improve the generalization capabilities of EL models significantly. We summarize novel methods for entity encoding, as well as context/mention encoding techniques.
Many natural language processing systems take advantage of deep pre-trained language models like ELMo , BERT , and their modifications. EL made its path into these models as a way of introducing information stored in KBs, which helps to adopt word representations to certain text processing tasks. We briefly discuss this novel application of EL.
There are few previous surveys devoted to the EL task [61, 43, 67, 1]. In the most recent paper, Al-Moslmi et al. ieeeelsurvey review both entity recognition and general entity disambiguation/linking methods published between the years 2014-2019. Instead, we focus specifically on rapidly developing neural models presented since 2015. The previous surveys do not address the topics of entity encoding, applications of EL to deep pre-trained language models, and cross-lingual EL. We also the first to summarize the domain-independent approaches to EL, many of which are based on zero-shot techniques. More specifically, this paper makes the following contributions:
a survey of state-of-the-art neural entity linking models;
a survey of entity embedding techniques;
a discussion of recent domain-independent (zero-shot) and cross-lingual EL approaches;
a survey of EL applications to modeling word representations.
The structure of this survey is the following. We start with defining the task of EL in Section 2. In Section 3.1, the common architecture of neural entity linking systems is presented. Modifications and variations of this basic pipeline are discussed in Section 3.2. In Section 4, we summarize the evaluation results for EL and entity representation models. Section 5 is dedicated to the recently emerged application of EL for improving neural language models. Finally, Section 6 summarizes the survey and suggests a prominent direction of future work in neural EL.
2 Task Description
Consider the example presented in Figure 1 with an entity mention Scott Young. Literally, this common name can at least refer to an American football player, Welsh football player, or a writer. The EL task is to correctly reveal the mention in the text, resolve the ambiguity, and provide a link to a corresponding entity entry in a KB. Wikification  and Entity Disambiguation (ED) are considered as subtypes of EL . In this survey, we assume that entity linking encompasses both entity recognition (ER) and entity disambiguation (ED). However, only few studies suggest models that perform ER and ED jointly, while the majority of papers referring EL focus only on ED and assume that mention boundaries are given by an external entity recogniser . ER techniques that perform only recognition without disambiguation are considered in many previous surveys [52, 66, 25, 80] and are out of the scope of this work.
To learn a mapping from entity mentions in a context to entity entries in a KB, EL models use supervision signals like manually annotated mention-entity pairs. The size of KBs vary; they can contain hundreds of thousands or millions of entities. Due to their large size, training data for EL would be extremely unbalanced; training sets can lack even a single example for a particular entity or mention, e.g. AIDA training set 
. To deal with this problem, EL models should have wide generalization capabilities. Despite their large size, KBs are incomplete. Therefore, some mentions in the text cannot be correctly mapped to any KB entry. Determining such unlikable mentions is one of the EL challenges. Furthermore, it is customary to distinguish “local” and “global” EL tasks. Local EL performs disambiguation of each mention in text independently using only context near target mentions, while global EL deals with simultaneous disambiguation of the all mentions and can engage features extracted from the whole document.
3 Neural Entity Linking
We start the discussion of neural entity linking approaches from the most general structure of pipelines and continue with various specific modifications like joint entity recognition and linking, using global context, domain-independent approaches including zero-shot methods, and cross-lingual models. We also give a detailed overview of entity embedding techniques.
3.1 General Architecture
Some of the attempts to EL with neural networks treat it as a multi-class classification task, in which entities correspond to classes. However, the straightforward approach results in a large number of classes, which leads to suboptimal performance without task sharing . The streamlined approach to EL is to treat it as a ranking problem. We present the EL pipeline in Figure 2
, which is applicable to the majority of the neural approaches. Here, the entity recognition model identifies the mention boundaries in text. The next step is to produce a short list of possible entities (candidates) for a mention. The entity ranking model estimates how well a candidate entity matches the context. An optional step is to determine unlinkable mentions, for which KBs do not contain a corresponding entity.
3.1.1 Candidate Generation
The goal of this step is given an ambiguous entity mention, such as “Scott Young”, to provide a list of its possible “senses” as specified by entities in a KB. EL is analogous to the Word Sense Disambiguation (WSD) task [51, 53] in terms of addressing lexical ambiguity. One of the major differences is that, in WSD, each sense of a word can be clearly defined by WordNet , while, in EL, KBs do not provide such an exact mapping between mentions and entities. Therefore, a mention can be linked to any entity in a KB, resulting in large decision space, e.g. the notorious “Big Blue” for referring to IBM. To address this issue, preliminary filtering of an entity list, called candidate generation, is performed.
There are three prominent methods for this: a surface form matching, a dictionary lookup, and a prior probability computation. In the first approach, a candidate list is composed of entities, which simply match surface forms of mentions in the text[90, 50, 40]. For the example mention of “Big Blue”, this approach could not work well as the referent entity IBM does not contain a mention string. In the second approach, a dictionary of additional aliases is constructed using KB metadata like disambiguation/redirect pages of Wikipedia . Pershina et al. pershina provide a resource of this type used in many EL models [81, 10, 54, 60, 47]. Another well-known alternative is the YAGO ontology  – automatically constructed from Wikipedia and WordNet. Among many other relations, it provides ‘means’ relations between mentions and entities, and this mapping is utilized as a candidate generator [30, 81, 23, 37, 9, 58, 82]. In this technique, the external dictionaries would help to disambiguate “Big Blue” as IBM. In the third approach, the candidates are generated based on precalculated prior probabilities of correspondence between certain mentions and entities. These priors can be computed using the mention-entity hyperlink count statistics. Most of the studies rely on priors computed on the basis of Wikipedia anchor links [90, 73, 23, 37, 69, 9, 58, 41, 83, 82]. Another widely used option for this is CrossWikis , which is an extensive dictionary computed from the frequency of mention-entity links of web crawl data [23, 27, 37, 9, 58, 82]. The example mention string of “Big Blue” could be labeled as its referent entity IBM with precomputed priors. Recent zero-shot models [45, 24, 79] perform candidate generation without external knowledge. Section 3.2.3 describes them in detail.
3.1.2 Entity Ranking
The goal of this stage is given a list of entity candidates from a KB and a context with a mention to rank these entities assigning a score to each of them. Figure 3 depicts the typical architecture of the ranking component. To correctly disambiguate an entity mention, it is crucial to thoroughly capture the information from its context. A contextualized vector representation of a mention is generated by an encoder network. Several early techniques in neural EL utilize a convolutional encoder [72, 21], as well as attention between candidate entity embeddings and embeddings of words surrounding a mention [23, 41]. However, in recent models, two approaches prevail: recurrent networks and self-attention .
. Gupta et al. gupta concatenate outputs of two LSTM networks that independently encode left and right contexts of a mention (including the mention itself). In the same vein, Sil et al. sil encode left and right local contexts via LSTMs but also pool the results across all mentions in a coreference chain and postprocess left and right representations with a tensor network. A modification of LSTM – GRU is used by Eshel et al. eshel2017named in conjunction with an attention mechanism  to encode left and right context of a mention. Kolitsas et al. end2end represent an entity mention as a combination of LSTM hidden states included in the mention span. Le and Titov titov1 simply run a bidirectional LSTM network on words complemented with embeddings of word positions relative to a target mention. Shahbazi et al. shahbazi2019entity adopt pre-trained ELMo  for mention encoding by averaging word representations inside mentions.
for context and mention encoding. In Peters et al. knowbert, a mention representation is modeled by pooling over word pieces in a mention span. The authors also put an additional self-attention block over all mention representations that encode interactions between several entities in a sentence. Another approach to modeling mentions is to insert special tags around them and perform a reduction of the whole encoded sequence. Wu et al. wufacebook reduce a sequence by keeping the representation of the special pooling symbol ‘[CLS]’ inserted at the beginning of a sequence. Logeswaran et al. zero-shot mark positions of a mention span by summing embeddings of words within the span with a special vector and use the same reduction strategy as Wu et al. wufacebook. Yamada et al. yamada20 concatenate text with all mentions in it and jointly encode this sequence via a self-attention model based on pre-trained BERT.
The produced mention representation is compared with candidate entity representations. Entity representations can be pre-trained (see Section 3.1.3) or generated by another encoder as in some zero-shot approaches (see Section 3.2.3). The BERT-based model of Yamda et al. yamada20 simultaneously learns how to encode mentions and entity embeddings in the unified architecture. Most of the state-of-the-art studies compare mention and entity representations using a dot product [23, 27, 37, 58, 79]72, 21, 24]. The calculated similarity score is often combined with mention-entity priors obtained during the candidate generation phase [21, 23, 37] or other features including various similarities, string matching indicator, and entity types [21, 69, 65, 83]. One of the common techniques for that is to use an additional one or two-layer feedforward network [21, 23, 65]
. The final disambiguation decision is inferred via a probability distribution, which is usually approximated by a softmax function over the candidates. The local similarity score or a probability distribution can be further utilized for global scoring (see Section3.2.2).
3.1.3 Entity Representations
The linking decision requires to measure how accurately candidate entities match a corresponding mention or context based on a structured or textual information of candidate entities. Low-dimensional semantic representations of entities encode this information in such a way that spatial proximity of entities in a vector space correlates with their semantic similarity. This fact is illustrated in Figure 4 for four entities of Scott Young. All four entities could be disambiguated with their closest entities; the most similar entities of Scott_Young_(American_football) are related to American football including Alex_Henery. In contrast, for Scott_Young_(politician), the most similar entities are politicians.
The earliest methods, including Milne and Witten milne, He et al. he, and Huang et al. huang, depend on hand-engineered features such as bag-of-words or one-hot vectors to represent entities. After word2vec  gained success in word representations, its architecture was modified to be able to produce entity vectors. There are three common ways to apply this adjustment. The first one is to extend the objective function with a joint alignment function based on several features of entities, however, this requires sparse entity-entity co-occurrences statistics [17, 81, 10, 68]. To overcome this issue, the second approach provides a formulation of the objective function based on an extensive entity-word co-occurrences statistics [23, 60]. The last approach is to directly replace the raw input text with entity annotated text without any statistics [90, 50, 73].
More recently, Gupta et al. gupta aim at capturing different kinds of entity information, including entity type, description page, linked mention, and contextual information, and therefore, they generate a large encoder, which involves CNN for the entity description and alignment function for the others. Gillick et al. gillick encode entities based on their title, description page, and category information. All previously mentioned models rely on the annotated data, and a few studies are challenged with less resource dependence. In this track, Sun et al. sun and Sil et al. sil derive entity embeddings using pre-trained word2vec word vectors through description page words, surface forms words, and entity category words (Sun et al. sun continue learning representations while disambiguation and they use annotation in the disambiguation phase). Newman-Griffis et al. griffis expand the word2vec architecture with a distant supervision setup based on the terminology of Wikipedia’s page titles and redirects. Sevgili et al. sevgili build a graph from entity-entity hyperlinks and use a graph embedding method to generate entity embeddings. Logeswaran et al. zero-shot and Wu et al. wufacebook depend on the BERT architecture to create representations through the description pages.
The neural architectures for learning representations are mostly designed to keep word and entity in joint semantic space to allow a straightforward ranking of candidates with vector similarity. They are trained with three common types of score functions. The first and most commonly used one is the similarity or the prior score of entity and mention. The second direction is the alignment score function, where several features are learned independently and joined in the alignment function [81, 17, 10, 27]. The final one relies on cross- or bi-encoders of entity and mention/context, which are built over the BERT architecture [45, 79]. The representation models also vary in terms of data sources, which can be structured (e.g. hyperlinks) or textual (i.e. description of an entity, anchor texts, or annotated texts). The detailed model-wise comparison can be found in Table 3 in the appendix.
3.1.4 Unlinkable Mention Prediction
The referent entities of some mentions can be absent in the KBs, e.g. there is no Wikipedia entry about Scott Young as a cricket player of the Stenhousemuir cricket club.111https://www.stenhousemuircricketclub.com/teams/171906/player/scott-young-1828009 Therefore, an EL system should be able to predict the absence of a reference, which is known as NIL prediction. There are four common ways to perform NIL prediction. Sometimes a candidate generator does not yield any corresponding entities for a mention; such mentions are trivially considered unlikable [73, 69]. One can set a threshold for the best linking probability (or a score), below which mention is considered unlinkable [58, 38]. Some models introduce an additional special ‘NIL’ entity in the ranking phase, so models can predict it as the best match for the mention 
. It is also possible to train an additional binary classifier that accepts mention-entity pairs after the ranking phase, as well as several additional features (best linking score, whether mentions are also detected by a dedicated NER system, etc.), and makes the final decision about whether a mention is linkable or not[50, 47].
3.2 Modifications of the General Architecture
3.2.1 Joint Entity Recognition and Disambiguation Architectures
A few systems provide a joint solution for entity recognition and linking. Undoubtedly, solving these two problems simultaneously makes the task more challenging. However, the interaction between these steps can be beneficial for improving the quality of the overall pipeline due to their natural mutual dependency. While first competitive models that provide joint solution were probabilistic graphical models [46, 55], we focus on purely neural approaches proposed recently [37, 58, 47, 8].
enumerate all spans in a sentence with a certain maximum width, filter them by several heuristics (remove mentions with stop words, punctuation, ellipses, quotes, and currencies), and try to match them to a pre-built index of entities used for the candidate generation. If a mention candidate has at least one corresponding entity candidate, it is further treated by a ranking neural network that can also discard it by considering it unlinkable to any entity in a KB (see Section3.1.2). Therefore, the decision during the entity disambiguation phase affects entity recognition.
Martins et al. martins describe the approach with tighter integration between recognition and linking phases via multi-task learning. The authors propose a stack-based bidirectional LSTM network with a shift-reduce mechanism and attention for entity recognition that propagates its internal states to the linker network for candidate entity ranking. The linker is supplemented with a NIL predictor network. The networks are trained jointly by optimizing the sum of losses from all three components.
Broscheit broscheit2020investigating goes further by suggesting a completely end-to-end method that deals with entity recognition and linking jointly without explicitly executing a candidate generation step. They formulate the task as a sequence labeling problem, where each token in a text is assigned an entity link or a NIL class. They leverage a sequence tagger based on pre-trained BERT for this purpose. This simplistic approach does not supersede  but outperforms the baseline, in which candidate generation, entity recognition, and linking are performed independently.
3.2.2 Global Context Architectures
Two kinds of context information are accepted to perform entity disambiguation: local and global. Local approaches rely on the words around the entity mention in a specified window, and each mention is disambiguated independently. This kind of method does not perform properly if the surrounding words do not carry sufficient contextual information . Semantic consistency across all entities in a context is also quite informative for the disambiguation. Global approaches use this topical coherence and disambiguate all mentions simultaneously, as illustrated in Figure 5. Although the extra information of the global context improves the disambiguation accuracy, the number of possible entity assignments is combinatorial, which results in a high time complexity . Another difficulty is to attempt to assign an entity with its coherence score, since this score is not possible to compute in advance due to the simultaneous disambiguation .
Earlier, global EL approaches typically perform PageRank or RandomWalk algorithms over a graph containing candidate entities of mentions in a context [90, 57, 26]. Another well-known solution is to maximize the Conditional Random Fields score function, which contains two terms: one evaluates an entity-context propriety, and the other one measures coherence [23, 39, 41]. However, the exact inference is NP-hard. Hofmann et al. hofmann adapt loopy belief propagation with message passing iterations using pairwise entity scores to reduce the complexity, and Le and Titov le2018 expand it with a coreference relation of mentions as latent variables (the mentions are coreferent if they refer to the same entity). Some recent studies define the global EL problem as a sequential decision task, where the disambiguation of new entities is based on the already disambiguated ones. Fang et al. reinforcement apply LSTM to be able to maintain long term memory for previous decisions. Yang et al. yang2019 execute Dynamic Context Augmentation, where basically previous decisions are collected as dynamic context to improve the following predictions, and Yamada et al. yamada20 compute confidence scores based on previous predictions. Another alternative for the recent direction is to attach an entity relatedness score to the score function of the entire model. Kolitsas et al. end2end first select a set of entities with a high local score and compute the similarity between the in-process entity embedding and an average of the selected entity embeddings. Fang et al. fang calculate the similarity between the present entity and its surrounding entity candidates in a specified window. Yamada et al. yamada and Radhakrishnan et al. elden measure the similarity first based on unambiguous mentions and then predict entities. Rather than computing entity score, Tsai and Roth crossling-wikification directly append previous entity embeddings to the model. The final common approach is to rely on global information to capture the coherence instead of directly including an entity coherence component [58, 27, 10, 50, 69]. Distinctively, Cao et al. cao2018 integrate Graph Convolutional Network into a neural model to handle the global information, which is represented as a subgraph of candidate entities.
3.2.3 Domain-Independent Architectures
Domain independence is one of the most desired properties of EL systems. Annotated resources are very limited and exist only for a few domains. Obtaining labeled data in a new domain requires much labor. Earlier, this problem is tackled by few domain-independent approaches based on unsupervised [78, 10, 54, 39] and semi-supervised models . Recent studies provide solutions based on distant learning and zero-shot methods.
The studies [40, 41] propose distant learning techniques that use only unlabeled documents. They rely on the weak supervision coming from a surface matching heuristic, and the EL task is framed as binary multi-instance learning. The algorithm learns to distinguish between positive entities set and random negatives set. The positive set is obtained by retrieving entities with high word overlap with entities and relations to other mention candidates in the sentence. While showing promising performance, which in some cases rivals fully supervised systems, these approaches require either a KB describing relations of entities  or mention-entity priors computed from entity hyperlink statistics extracted from Wikipedia .
. The zero-shot setting means that only descriptions of entities are available, while other sources of supervision such as relations between entities in a Knowledge Graph (KG) and entity types are absent. This allows applying the system to almost any source of entities, since building a database of entities and their descriptions might be far less laborious compared to the development of complex KGs with the abundance of relations.
Since pre-build resources for candidate generation are not available for the target domain, for candidate selection, one can rely only on textual descriptions of entities. Logeswaran et al. zero-shot use the BM25 information retrieval formula  to rank entity Wikipedia pages via mentions and keep entities with top relevant pages as candidates. Wu et al. wufacebook use the BERT bi-encoder on top of entity descriptions to select candidates. The bi-encoder consists of two networks that separately encode context and an entity description. As in supervised approaches, entity and mention representations are compared via a dot product, and the top candidates are selected for ranking. In Gillick et al. gillick, the candidate generation stage is absent at all, and ranking is performed for all entities in a KG.
For ranking entities, which are absent in a training corpus of the source domain, zero-shot methods rely on entity description encoding robust across multiple domains. One of the first studies that utilize such a technique is proposed by Gupta et al. gupta (not purely zero-shot because they used entity typings). Gillick et al. gillick propose a CNN network for encoding entity descriptions along with optional entity typing information. It is also worth noting that Gillick et al. gillick rely on annotated data while training entity representations. Some other approaches [45, 79] utilize the BERT-based cross-encoder to perform joint encoding of mentions and entities. The cross-encoder takes concatenation of context with a mention and an entity description to produce a scalar score for each candidate. In both studies, cross-encoders achieve substantially better results compared to other approaches.
3.2.4 Cross-lingual Architectures
The cross-lingual EL methods  leverage supervision signals from multiple languages for training a model in a target language. For example, the inter-lingual links in Wikipedia can be utilized for alignment of entities in multiple languages. Using such an alignment, the annotated data from high-resource languages like English can help to improve the quality of text processing for the low-resource ones.
One of the challenges in cross-lingual EL is candidate generation, since the low-resource language can lack mappings between mention strings and entities. In this case, candidate generation can be approached by: mining a translation dictionary , training a translation and alignment model , or applying a neural character level string matching model 
. The latter relies on training on a high-resource pivot language, similar to the target low-resource one. The neural string matching approach can be further improved with simpler average n-gram encoding and extending entity-entity pairs with mention-entity examples.
There are several approaches to candidate ranking that take advantage of cross-lingual data for dealing with the lack of annotated examples. Pan et al. pan2017cross uses the comparison of Abstract Meaning Representation (AMR)  statistics in English Wikipedia and mention context for ranking. To train an AMR tagger, pseudo-labeling  was used. Tsai and Roth crossling-wikification train monolingual embeddings for words and entities jointly by replacing every entity mention with corresponding entity tokens. Using the connection of entities to Wikipedia pages that exist for multiple languages, they learn the projection functions from multiple languages into the English embedding space. For ranking, context embeddings are averaged, projected into English space, and compared with entity embeddings. The authors demonstrate that this approach helps to build better entity representations and boost EL performance on low-resource languages. Sil et al. sil propose a method for zero-shot transfer from a high-resource language. The authors extend the previous approach with the least squares objective for embedding projection learning, the CNN context encoder, and a trainable re-weighting of each dimension of context and entity representations. The proposed approach demonstrates improved performance compared to previous non-zero-shot approaches. Upadhyay et al. upadhyay_joint argued that the success of zero-shot cross-lingual approaches might be highly related to mention-entity prior probabilities used as features. Their approach extends  with global context information and incorporation of typing information into context and entity representations (the system learns to predict typings during the training). The authors report a significant drop in performance for zero-shot cross-lingual EL with an excluded mention-entity prior. They also show that training on the high-resource language might be very beneficial for the low-resource settings.
We summarize critical design features for neural EL models in Table 4 in the appendix. One can note, the EL systems do not utilize the whole spectrum of available features. The mention encoders have made a shift to self-attention architectures and start using deep pre-trained models like BERT. The majority of studies still rely on external knowledge for the candidate generation step. There is a surge of models that tackle the domain adaptation problem in a zero-shot fashion. However, the task of zero-shot joint entity recognition and linking has not been addressed yet. It is shown in several works that the cross-encoder architecture is superior compared to models with separate mention and entity encoders. Many approaches rely on pre-trained entity representations, only few take advantage of a trainable entity encoder inside an EL model. The global context is widely used, but there are few recent studies that focus only on local EL.
In this section, we present and summarize the evaluation results for entity embeddings and neural entity linking systems reported by their authors.
4.1 Entity Relatedness
An entity relatedness dataset was put forward by Ceccarelli et al. ceccarelli using the dataset of Hoffart et al. hoffart. The dataset is in the form of queries, where the first entity is accepted as correctly linked and the second entity is the candidate. Here, the evaluation task is to rank entities for the target one, which is performed using cosine similarity of entity representations except for two studies: Milne and Witten introduce a Wikipedia hyperlink-based measure, known as WLM, and recently El Vaigh et al. elvaigh provide a weighted semantic relatedness measure. The evaluation of ranking quality is performed with normalized discounted cumulative gain (nDCG)  and mean average precision (MAP) . nDCG is a well-known quality metric used in information retrieval, which provides a fair evaluation by measuring the position impressiveness. Similarly, MAP measures how accurately the model performs for the target entity.
The highest score is reported by Huang et al. huang; they specifically train the embeddings based on a pairwise entity score function in a supervised way. Instead, in other models, entity embeddings are trained in a joint space with word embeddings based on the relatedness between mentions and words. Therefore, the results of Huang et al. huang are distinctively higher. Ganea and Hofmann hofmann and Cao et al. cao achieve good scores, and recently, Shi et al. shi also present an excellent performance using a large amount of data sources based on textual and KB information.
4.2 Entity Linking
We report EL performance results on widely-used datasets: AIDA , TAC KBP 2010 , MSNBC , AQUAINT , ACE2004 , CWEB [26, 22], and WW [26, 22], in Table 2. Among them, CWEB and WW are the largest datasets that are annotated automatically, while AIDA is the largest dataset, annotated manually. AIDA contains the development set AIDA-A and the test set AIDA-B. We report the results calculated for AIDA-B. Some of these results are evaluated using GERBIL  – a benchmarking platform for entity recognition and disambiguation systems. The cross-lingual EL results are reported for the TAC KBP 2015  Spanish (es) and Chinese (zh) datasets. We present accuracy and micro F1 scores. The micro F1 scores for systems that perform ER and EL jointly are different from the measures reported for disambiguation only systems due to mistakes in entity recognition.
|Accuracy||Micro F1||Micro F1||Micro F1||Micro F1||Micro F1||Accuracy||Accuracy|
Among the joint recognition and disambiguation solutions, the leadership is still owned by Kolitsas et al. end2end. This system and others that solve the ER task fall behind the disambiguation-only systems since they rely on noisy mention boundaries produced by themselves. Among published local models for disambiguation, the best result is reported by Wu et al. wufacebook. It is worth noting that this model can be used in a zero-shot setting. The global models expectedly outperform the local ones. The work of Yamada et al. yamada20 reports results that are consistently better compared to other solutions. The performance improvements are attributed to the masked entity prediction mechanism for entity embedding and the use of the pre-trained model based on BERT with the interdependent scoring function.
5 Applications of Entity Linking for Training Word Representation Models
. In addition to such end tasks, a new trend is the use EL information for representation learning. Namely, several studies have shown that contextual word representations could benefit from information stored in KBs by incorporating EL into deep models for transfer learning.
KnowBERT  injects between top layers of the BERT architecture one or several entity linkers and optimizes the whole network for multiple tasks: the masked language model (MLM) task, next sentence prediction, and EL. The authors adopt the general end-to-end EL pipeline of  but use only the local context for disambiguation and use an encoder based on self-attention over the representations generated by underlying BERT layers. If the EL subsystem detects an entity mention in a given sentence, corresponding pre-built entity representations of candidates are utilized for calculating the updated contextual word representations generated on the current BERT layer. These representations are used as input in a subsequent layer and can also be modified by a subsequent EL subsystem. Experiments with two EL subsystems based on Wikidata and WordNet show that presented modifications in KnowBERT help it to slightly surpass other deep pre-trained language models in tasks of relationship extraction, WSD, and entity typing.
ERNIE  expands the BERT  architecture with a knowledgeable encoder (K-Encoder), which fuses contextualized word representations obtained from the underlying self-attention network with entity representations from a pre-trained TransE model . EL in this study is performed by an external tool TAGME . For model pre-training, in addition to the MLM task, the authors introduce the task of restoring randomly masked entities in a given sequence keeping the rest of the entities and tokens. Using English Wikipedia and Wikidata as training data, the authors show that introduced modifications provide performance gains in entity typing, relation classification, and several GLUE tasks.
Wang et al. wang2019kepler propose to train a disambiguation network using the composition of two losses: regular MLM and a Knowledge Embedding (KE) loss based on TransE  objective for encoding graph structures. In KE loss, representations of entities are obtained from their textual descriptions encoded with a self-attention network , and representations of relations are trainable vectors. Although the system exhibits a significant drop in performance on general NLP benchmarks such as GLUE , it shows increased performance on a wide range of KB-related tasks such as TACRED , FewRel , and OpenEntity .
6 Conclusion and Future Directions
In this survey, we have analyzed the recently proposed neural entity linking models. The majority of studies still rely on external knowledge for the candidate generation step. The mention encoders have made a shift to self-attention architectures and start using deep pre-trained models like BERT. There is a surge of models that tackle the domain adaptation problem in a zero-shot fashion. It is shown in several works that the cross-encoder architecture is superior compared to models with separate mention and entity encoders. Many approaches rely on pre-trained entity representations, only few take advantage of a trainable entity encoder inside an EL model. The global context is widely used, but there are few recent studies that focus only on local EL. Among the joint recognition and disambiguation solutions, the leadership is still owned by Kolitsas et al. end2end. Among published local models for disambiguation, the best result is reported by Wu et al. wufacebook. It is worth noting that this model can be used in a zero-shot setting. The global models expectedly outperform the local ones. The work of Yamada et al. yamada20 reports results that are consistently better compared to other solutions. The performance improvements are attributed to the masked entity prediction mechanism for entity embedding and to the usage of the pre-trained model based on BERT with the interdependent scoring function. It is also worth noting that several studies have demonstrated some benefits for deep transfer learning models of using information stored in KBs by incorporating EL into these models.
In the future work, we expect that zero-shot EL will rapidly evolve engaging other features like global coherence across all entities in a document, NIL prediction, joining ER and EL steps together, or providing completely end-to-end solutions. The latter would be an especially challenging task but also a fascinating research direction.
-  (2020) Named entity extraction for knowledge graphs: a literature overview. IEEE Access 8, pp. 32862–32881. External Links: Cited by: §1.
-  (2007) DBpedia: a nucleus for a web of open data. In Proceedings of the 6th International The Semantic Web and 2nd Asian Conference on Asian Semantic Web Conference, ISWC’07/ASWC’07, Berlin, Heidelberg, pp. 722–735. External Links: Cited by: §1.
-  (2015) Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San-Diego, California, USA. Cited by: §3.1.2.
-  (2013) Abstract meaning representation for sembanking. In Proceedings of the 7th linguistic annotation workshop and interoperability with discourse, Sofia, Bulgaria, pp. 178–186. Cited by: §3.2.4.
-  (2014) Semantic parsing via paraphrasing. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Baltimore, Maryland, pp. 1415–1425. External Links: Cited by: §5.
-  (2008) Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data, SIGMOD ’08, New York, NY, USA, pp. 1247–1250. External Links: Cited by: §1.
-  (2013) Translating embeddings for modeling multi-relational data. In Advances in neural information processing systems, Stateline, Nevada, USA, pp. 2787–2795. Cited by: §5, §5.
-  (2019) Investigating entity knowledge in BERT with simple neural end-to-end entity linking. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), Hong Kong, China, pp. 677–685. External Links: Cited by: §3.2.1.
-  (2018) Neural collective entity linking. In Proceedings of the 27th International Conference on Computational Linguistics, Santa Fe, New Mexico, USA, pp. 675–686. External Links: Cited by: §3.1.1.
-  (2017) Bridge text and knowledge by learning multi-prototype entity mention embedding. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vancouver, Canada, pp. 1623–1633. External Links: Cited by: §3.1.1, §3.1.3, §3.1.3, §3.2.2, §3.2.3.
-  (2013) Relational inference for Wikification. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, Seattle, Washington, USA, pp. 1787–1796. External Links: Cited by: §2.
-  (2018) Ultra-fine entity typing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia, pp. 87–96. External Links: Cited by: §5.
Empirical evaluation of gated recurrent neural networks on sequence modeling. In NIPS 2014 Workshop on Deep Learning, Montréal, Canada. Cited by: §3.1.2.
-  (2011) Natural language processing (almost) from scratch. Journal of machine learning research 12 (Aug), pp. 2493–2537. Cited by: §1.
-  (2007) Large-scale named entity disambiguation based on Wikipedia data. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), Prague, Czech Republic, pp. 708–716. Cited by: §4.2.
-  (2019) BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Minneapolis, Minnesota, pp. 4171–4186. External Links: Cited by: §1, §3.1.2, §5.
-  (2016) Entity disambiguation by knowledge and text jointly embedding. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, Berlin, Germany, pp. 260–269. External Links: Cited by: §3.1.3, §3.1.3.
Joint entity linking with deep reinforcement learning. In The World Wide Web Conference, WWW ’19, New York, NY, USA, pp. 438–447. External Links: Cited by: §3.1.1, §3.1.2, §3.2.2.
-  C. Fellbaum (Ed.) (1998) WordNet: an electronic lexical database. MIT Press. Cited by: §3.1.1.
-  (2010) TAGME: On-the-Fly annotation of short text fragments (by wikipedia entities). In Proceedings of the 19th ACM International Conference on Information and Knowledge Management, CIKM ’10, New York, NY, USA, pp. 1625–1628. External Links: Cited by: §5.
Capturing semantic similarity for entity linking with convolutional neural networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego, California, USA, pp. 1256–1261. Cited by: §3.1.2, §3.1.2.
-  (2013) FACC1: freebase annotation of ClueWeb corpora, version 1 (release date 2013-06-26, format version 1, correction level 0). Note: Note: http://lemurproject.org/clueweb09/ Cited by: §4.2.
-  (2017) Deep joint entity disambiguation with local neural attention. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark, pp. 2619–2629. External Links: Cited by: §3.1.1, §3.1.2, §3.1.2, §3.1.2, §3.1.3, §3.2.2.
-  (2019) Learning dense representations for entity retrieval. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), Hong Kong, China, pp. 528–537. External Links: Cited by: §3.1.1, §3.1.2, §3.2.3.
Recent named entity recognition and classification techniques: a systematic review. Computer Science Review 29, pp. 21–43. Cited by: §2.
-  (2018) Robust named entity disambiguation with random walks. Semantic Web 9 (4), pp. 459 – 479. External Links: Cited by: §3.2.2, §4.2.
-  (2017) Entity linking via joint encoding of types, descriptions, and context. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark, pp. 2681–2690. External Links: Cited by: §3.1.1, §3.1.2, §3.1.2, §3.1.3, §3.2.2, §3.2.3.
-  (2018) FewRel: a large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, pp. 4803–4809. External Links: Cited by: §5.
-  (1997) Long short-term memory. Neural computation 9 (8), pp. 1735–1780. Cited by: §3.1.2.
-  (2011) Robust disambiguation of named entities in text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pp. 782–792. External Links: Cited by: §2, §3.1.1, §4.2.
-  (2011) Knowledge-based weak supervision for information extraction of overlapping relations. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, Portland, Oregon, USA, pp. 541–550. Cited by: §5.
-  (2002-10) Cumulated gain-based evaluation of ir techniques. ACM Trans. Inf. Syst. 20 (4), pp. 422–446. External Links: Cited by: §4.1.
-  (2010) Overview of the TAC 2010 knowledge base population track. In Third Text Analysis Conference (TAC), Gaithersburg, Maryland, USA. Cited by: §4.2.
-  (2015) Overview of TAC-KBP2015 tri-lingual entity discovery and linking. In Proceedings of the 2015 Text Analysis Conference, TAC 2015, Gaithersburg, Maryland, USA, pp. 16–17. External Links: Cited by: §3.2.4, §4.2.
-  (2000) A probabilistic model of information retrieval: development and comparative experiments part 2. Information Processing & Management 36 (6), pp. 809–840. External Links: Cited by: §3.2.3.
Task-specific representation learning for web-scale entity disambiguation.
The Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, Louisiana, USA, pp. 5812–5819. Cited by: §3.1.
-  (2018) End-to-end neural entity linking. In Proceedings of the 22nd Conference on Computational Natural Language Learning, Brussels, Belgium, pp. 519–529. External Links: Cited by: §3.1.1, §3.1.2, §3.1.2, §3.1.2, §3.1.4, §3.2.1, §3.2.1, §3.2.1, §5.
-  (2015) Plato: a selective context model for entity resolution. Transactions of the Association for Computational Linguistics 3, pp. 503–515. External Links: Cited by: §3.1.4, §3.2.3.
-  (2018) Improving entity linking by modeling latent relations between mentions. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia, pp. 1595–1604. External Links: Cited by: §3.2.2, §3.2.3.
-  (2019-07) Distant learning for entity linking with automatic noise detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 4081–4090. External Links: Cited by: §3.1.1, §3.1.2, §3.2.3.
-  (2019) Boosting entity linking performance by leveraging unlabeled documents. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 1935–1945. External Links: Cited by: §3.1.1, §3.1.2, §3.2.2, §3.2.3.
Pseudo-label: the simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, Vol. 3, Atlanta, USA, pp. 2. Cited by: §3.2.4.
-  (2015) Design challenges for entity linking. Transactions of the Association for Computational Linguistics 3, pp. 315–328. External Links: Cited by: §1.
-  (2019) RoBERTa: a robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Cited by: §5.
-  (2019) Zero-shot entity linking by reading entity descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 3449–3460. External Links: Cited by: §3.1.1, §3.1.2, §3.1.3, §3.2.3, §3.2.3.
-  (2015) Joint entity recognition and disambiguation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, Lisbon, Portugal, pp. 879–888. Cited by: §3.2.1.
-  (2019) Joint learning of named entity recognition and entity linking. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, Florence, Italy, pp. 190–196. External Links: Cited by: §3.1.1, §3.1.2, §3.1.4, §3.2.1.
-  (2013) Distributed representations of words and phrases and their compositionality. In Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2, NIPS’13, Red Hook, NY, USA, pp. 3111–3119. Cited by: §3.1.3.
-  (2008) Learning to link with Wikipedia. In Proceedings of the 17th ACM Conference on Information and Knowledge Management, CIKM ’08, New York, NY, USA, pp. 509–518. External Links: Cited by: §4.2.
-  (2017) Combining word and entity embeddings for entity linking. In Extended Semantic Web Conference (1), Lecture Notes in Computer Science, Vol. 10249, pp. 337–352. Cited by: §3.1.1, §3.1.3, §3.1.4, §3.2.2.
-  (2014) Entity linking meets word sense disambiguation: a unified approach. Transactions of the Association for Computational Linguistics 2, pp. 231–244. External Links: Cited by: §3.1.1.
-  (2007) A survey of named entity recognition and classification. Lingvisticae Investigationes 30 (1), pp. 3–26. Cited by: §2.
-  (2009) Word sense disambiguation: a survey. ACM Comput. Surv. 41 (2), pp. 10:1–10:69. External Links: Cited by: §2, §3.1.1.
-  (2018) Jointly embedding entities and text with distant supervision. In Proceedings of The Third Workshop on Representation Learning for NLP, Melbourne, Australia, pp. 195–206. External Links: Cited by: §A, §3.1.1, §3.2.3.
-  (2016) J-NERD: joint named entity recognition and disambiguation with rich linguistic features. Transactions of the Association for Computational Linguistics 4, pp. 215–229. Cited by: §3.2.1.
-  (2017) Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vancouver, Canada, pp. 1946–1958. Cited by: §3.2.4.
-  (2015) Personalized page rank for named entity disambiguation. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Denver, Colorado, USA, pp. 238–243. External Links: Cited by: §3.2.2.
-  (2019) Knowledge enhanced contextual word representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 43–54. External Links: Cited by: §3.1.1, §3.1.2, §3.1.2, §3.1.4, §3.2.1, §3.2.1, §3.2.2, §5.
-  (2018) Deep contextualized word representations. In The Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, Louisiana, USA, pp. 2227–2237. Cited by: §1, §3.1.2.
-  (2018) ELDEN: improved entity linking using densified knowledge graphs. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), New Orleans, Louisiana, pp. 1844–1853. External Links: Cited by: §3.1.1, §3.1.3.
-  (2013) Entity linking: finding extracted entities in a knowledge base. In Multi-source, Multilingual Information Extraction and Summarization, pp. 93–115. External Links: Cited by: §1.
-  (2011) Local and global algorithms for disambiguation to Wikipedia. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies - Volume 1, HLT ’11, Portland, Oregon, USA, pp. 1375–1384. External Links: Cited by: §4.2.
-  (2019-01) Zero-shot neural transfer for cross-lingual entity linking. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, Honolulu, Hawaii, USA, pp. 6924–6931. Cited by: §3.2.4.
-  (2014-05) Benchmarking the extraction and disambiguation of named entities on the semantic web. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), Reykjavik, Iceland, pp. 4593–4600. External Links: Cited by: §2.
-  (2019) Entity-aware elmo: learning contextual entity representation for entity disambiguation. arXiv preprint arXiv:1908.05762. Cited by: §3.1.2.
-  (2014) Named entity recognition: a literature survey. Center For Indian Language Technology. Cited by: §2.
-  (2015) Entity linking with a knowledge base: issues, techniques, and solutions. Transactions on Knowledge & Data Engineering 27 (2), pp. 443–460. External Links: Cited by: §1.
-  (2020) Joint embedding in named entity linking on sentence level. ArXiv abs/2002.04936. External Links: Cited by: §3.1.3.
-  (2018) Neural cross-lingual entity linking. In The Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, Louisiana, USA.. External Links: Cited by: §3.1.1, §3.1.2, §3.1.2, §3.1.4, §3.2.2, §3.2.4.
-  (2012) A cross-lingual dictionary for English Wikipedia concepts. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), Istanbul, Turkey, pp. 3168–3175. External Links: Cited by: §3.1.1.
-  (2007) YAGO: a core of semantic knowledge. In Proceedings of the 16th International Conference on World Wide Web, WWW ’07, New York, NY, USA, pp. 697–706. External Links: Cited by: §3.1.1.
-  (2015) Modeling mention, context and entity with neural networks for entity disambiguation. In Proceedings of the 24th International Conference on Artificial Intelligence, IJCAI’15, pp. 1333–1339. External Links: Cited by: §3.1.2, §3.1.2.
-  (2016) Cross-lingual Wikification using multilingual embeddings. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego, California, USA, pp. 589–598. External Links: Cited by: §3.1.1, §3.1.3, §3.1.4.
-  (2018) Learning better name translation for cross-lingual Wikification. In Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, Louisiana, USA. Cited by: §3.2.4.
-  (2015) GERBIL: general entity annotator benchmarking framework. In Proceedings of the 24th International Conference on World Wide Web, WWW ’15, Florence, Italy, pp. 1133–1143. External Links: Cited by: §4.2.
-  (2017) Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, Red Hook, NY, USA, pp. 6000–6010. External Links: Cited by: §3.1.2.
-  (2018) GLUE: a multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, Brussels, Belgium, pp. 353–355. External Links: Cited by: §5.
-  (2015) Language and domain independent entity linking with quantified collective validation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, Lisbon, Portugal, pp. 695–704. External Links: Cited by: §3.2.3.
-  (2020) Zero-shot entity linking with dense entity retrieval. ArXiv abs/1911.03814. External Links: Cited by: item 4, §3.1.1, §3.1.2, §3.1.2, §3.1.3, §3.2.3, §3.2.3.
-  (2018-08) A survey on recent advances in named entity recognition from deep learning models. In Proceedings of the 27th International Conference on Computational Linguistics, Santa Fe, New Mexico, NM, USA, pp. 2145–2158. External Links: Cited by: §2.
-  (2016) Joint learning of the embedding of words and entities for named entity disambiguation. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, Berlin, Germany, pp. 250–259. External Links: Cited by: §3.1.1, §3.1.3, §3.1.3, §3.2.2.
-  (2020) Global entity disambiguation with pretrained contextualized embeddings of words and entities. arXiv preprint arXiv:1909.00426v2. External Links: Cited by: §3.1.1, §3.1.2.
-  (2019) Learning dynamic context augmentation for global entity linking. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 271–281. External Links: Cited by: §3.1.1, §3.1.2, §3.2.2.
-  (2015) Semantic parsing via staged query graph generation: question answering with knowledge base. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Beijing, China, pp. 1321–1331. Cited by: §5.
-  (2018) Recent trends in deep learning based natural language processing. IEEE Computational Intelligence Magazine 13 (3), pp. 55–75. Cited by: §1.
-  (2007) A support vector method for optimizing average precision. In Proceedings of the 30th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’07, New York, NY, USA, pp. 271–278. External Links: Cited by: §4.1.
-  (2017) Position-aware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP 2017), Copenhagen, Denmark, pp. 35–45. External Links: Cited by: §5.
-  (2019) ERNIE: enhanced language representation with informative entities. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, pp. 1441–1451. External Links: Cited by: §5.
-  (2020) Improving candidate generation for low-resource cross-lingual entity linking. Transactions of the Association for Computational Linguistics 8, pp. 109–124. Cited by: §3.2.4.
-  (2016) Robust and collective entity disambiguation through semantic embeddings. In Proceedings of the 39th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’16, New York, NY, USA, pp. 425–434. External Links: Cited by: §3.1.1, §3.1.3, §3.2.2.
A Entity Embeddings
|Huang et al. huang||✓||✓||✓||✓||
|Sun et al. sun||✓||✓||✓||word2vec||✓|
|Fang et al. fang||✓||✓||✓||✓||-||✓|
|Yamada et al. yamada||✓||✓||-||✓|
|Zwicklbauer et al. zwicklbauer||✓||✓||-||only entity|
|Tsai and Roth crossling-wikification||✓||✓||-||✓|
|Ganea and Hofmann hofmann||✓||word2vec||✓|
|Cao et al. cao||✓||✓||✓||-||✓|
|Moreno et al. moreno||✓||-||✓|
|Gupta et al. gupta||✓||✓||✓||GloVe||✓|
|Sil et al. sil||✓||word2vec||✓|
|Upadhyay et al. upadhyay_joint||✓||✓||✓||-||✓|
|Newman-Griffis et al. griffis||✓||✓||-||✓|
|Radhakrishnan et al. elden||✓||✓|
|Rijhwani et al. rijhwani2019zero||✓||✓||✓||-||✓|
|Logeswaran et al. zero-shot||✓||BERT||✓|
|Gillick et al. gillick||✓||✓||✓||✓||GLove||✓|
|Sevgili et al. sevgili||✓||✓||-||only entity|
|Shi et al. shi||✓||✓||✓||-||✓|
|Zhou et al. zhou2020improving||✓||✓||✓||✓||-||✓|
|Wu et al. wufacebook||✓||✓||BERT||✓|
In Table 3, we present the EL models that generate entity representations for a part of their solutions. Annotated text data contain a sequence of terms, where entities are labeled. It is a powerful resource to catch textual information of an entity and a mention, however, many specialized domains are lacking such annotations . Alternatively, it could be used as a sequence of entities by removing all words in the context, as Zwicklbauer et al. zwicklbauer did. Entity description pages are another text resource provided by KBs. They contain a textual description for each entity, however, it is unable to catch any relation, like mention-entity or entity-entity. Entity-entity links and entity-mention links are quite informative about the relational knowledge. Entity titles, types, and redirect pages are dictionary-like information, where redirect pages are the most useful feature in terms of entity-entity relations.
The following notations were used to present the external features of the model:
Mention and context are represented in a common representation.
Only entities are remained in the anchor text.
Mention and entity are paired together and represented in a common representation, known as cross-encoder.
They use bi-encoders, which use two independent encoders  for entity and context, and generate entity embedding using these encoders.
They use bi-encoders and cross-encoders for processing mention, entity, and context.
B Features of Neural Entity Linking Models
|Sun et al. sun||
|Francis-Landau et al. francis2016capturing||✓||
|Fang et al. fang||✓||
|Yamada et al. yamada||✓||
|Zwicklbauer et al. zwicklbauer||✓||✓||
|Tsai and Roth crossling-wikification||✓||✓||
|Pan et al. pan2017cross||✓||✓||
|Cao et al. cao||✓||
|Eshel et al.eshel2017named||
|Ganea and Hofmann hofmann||✓||
|Moreno et al. moreno||✓||✓||
|Gupta et al. gupta||✓||
|Le and Titov le2018||✓||pre-trained||prior||
|Newman-Griffis et al. griffis||
|Radhakrishnan et al. elden||✓||
|Kolitsas et al. end2end||✓||✓||pre-trained||prior||✓|
|Sil et al. sil||✓||✓||
|Cao et al. cao2018||✓||pre-trained||prior||✓|
|Raiman et al. deeptype||✓||n/a||
|Shahbazi et al. shahbazi2019entity||E-ELMo||prior||✓|
|Logeswaran et al. zero-shot||✓||
|Gillick et al. gillick||
|Peters et al. knowbert||✓||✓||✓||pre-trained||prior||
|Le and Titov titov1||
|Le and Titov titov2||✓||pre-trained||prior||
|Fang et al. reinforcement||✓||pre-trained||dictionary||✓|
|Martins et al. martins||✓||✓||pre-trained||dictionary||✓|
|Yang et al. yang2019||✓||pre-trained||prior||✓|
|Wu et al. wufacebook||
|Yamada et al. yamada20||✓||
In Table 4, we systematize the EL models in terms of their modifications and data requirements. Each column corresponds to a model feature.
The global column shows whether a system uses a global solution (see Section 3.2.2).
The recognition column refers to joint entity recognition and disambiguation models, where recognition and disambiguation of entities are performed collectively (Section 3.2.1).
The NIL prediction column points out models that also label unlinkable mentions (Section 3.1.4).
The entity embedding column presents how the entity disambiguation model uses entity representations, where joint architecture means the entity representations and the parameters of the disambiguation model are learned in the unified architecture, separate architecture indicates that representations are trained in one model and disambiguation parameters are learned in another, pre-trained denotes that the model uses pre-trained entity representations (Section 3.1.3).
In the candidate generation column, the candidate generation methods are noted (Section 3.1.1). It contains the following options:
surface-form – simple surface-form matching heuristic;
dictionary – a dictionary with supplementary aliases for entities;
prior – filtering with precalculated mention-entity prior probabilities;
type classifier – deeptype filter candidates using a classifier for an automatically learned type system;
tf-idf – zero-shot extract tf-idf scores based on entity description pages;
neighbours – the highest similar entities for the given mention are selected as candidates based on representations.
The zero-shot column displays if an EL system provides a zero-shot approach (see Section 3.2.3).
The annotated text data column shows whether a model uses an annotation or not. ‘In entity embedding’ denotes the models that do not use the annotated data for training, but the annotated data is used for training entity representation (see Section 3.2.3).
The cross-lingual column refers to models, which provide cross-lingual EL solutions (Section 3.2.4).
Besides, the following superscript notations were used to denote specific features of methods:
In classification, the prior is checked by a threshold. This can be considered as a candidate selection step.
While training, they detect mentions, while testing, they assume mentions are detected.
They provide EL as a subsystem of language modeling.
They use document-level mention contexts while encoding.
Zero-shot in the sense of model adaptation to a new language using English annotated data, while the other zero-shot works solve the problem of model adaptation to a new domain without switching the language.