{ "paper_id": "O14-2002", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:04:44.204821Z" }, "title": "Joint Learning of Entity Linking Constraints Using a Markov-Logic Network", "authors": [ { "first": "Hong-Jie", "middle": [], "last": "Dai", "suffix": "", "affiliation": { "laboratory": "", "institution": "Taipei Medical University", "location": {} }, "email": "hjdai@tmu.edu.tw" }, { "first": "\uf02a", "middle": [], "last": "Richard", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Tzong-Han", "middle": [], "last": "Tsai", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Central University", "location": {} }, "email": "thtsai@csie.ncu.edu.tw" }, { "first": "Wen-Lian", "middle": [], "last": "Hsu", "suffix": "", "affiliation": {}, "email": "hsu@iis.sinica.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Entity linking (EL) is the task of linking a textual named entity mention to a knowledge base entry. Traditional approaches have addressed the problem by dividing the task into separate stages: entity recognition/classification, entity filtering, and entity mapping, in which different constraints are used to improve the system's performance. Nevertheless, these constraints are executed separately and cannot be used interactively. In this paper, we propose an integrated solution to the task based on a Markov logic network (MLN). We show how the stage decision can be formulated and combined in an MLN. We conducted experiments on the biomedical EL task, gene mention linking (GML), and compared our model's performance with those of two other GML approaches. Our experimental results provide the first comprehensive GML evaluations from three different perspectives: article-wide precision/recall/F-measure (PRF), instance-based PRF, and question answering accuracy. This paper also provides formal definitions of all of the above EL tasks. Experimental results show that our method outperforms the baseline and state-of-the-art systems under all three evaluation schemes.", "pdf_parse": { "paper_id": "O14-2002", "_pdf_hash": "", "abstract": [ { "text": "Entity linking (EL) is the task of linking a textual named entity mention to a knowledge base entry. Traditional approaches have addressed the problem by dividing the task into separate stages: entity recognition/classification, entity filtering, and entity mapping, in which different constraints are used to improve the system's performance. Nevertheless, these constraints are executed separately and cannot be used interactively. In this paper, we propose an integrated solution to the task based on a Markov logic network (MLN). We show how the stage decision can be formulated and combined in an MLN. We conducted experiments on the biomedical EL task, gene mention linking (GML), and compared our model's performance with those of two other GML approaches. Our experimental results provide the first comprehensive GML evaluations from three different perspectives: article-wide precision/recall/F-measure (PRF), instance-based PRF, and question answering accuracy. This paper also provides formal definitions of all of the above EL tasks. Experimental results show that our method outperforms the baseline and state-of-the-art systems under all three evaluation schemes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Developing a system that can identify entities, such as personal names and gene or disease mentions, and that can classify the relations between them is useful for several applications in natural language processing and knowledge acquisition. There are several possible uses for such a system in different fields, e.g., improving document retrieval for specific entities, relation extraction, and attribute assignment (e.g., gene ontology annotations). In these applications, recognized entities must be linked to unique database entries. McNamee and Dang (2009b) named the task of matching a textual entity mention to a knowledge base (KB) entry Entity Linking (EL). In Figure 1 , we provide a biomedical abstract to illustrate this task. The abstract discusses the relationship of the gene \"CD59\" to other lymphocyte antigens. After EL, the gene mention \"CD59\" in the first sentence must be linked to ID 966 in the Entrez Gene database of PubMed. In the first sentence, the authors also listed other designations of the gene, including \"membrane inhibitor of reactive lysis\" and \"protectin,\" and they defined \"MIRL\" as the abbreviation for \"membrane inhibitor of reactive lysis.\" Linking these instances to the same entry is a problem related to the name variations issue. Furthermore, the gene \"CD59\" may exist in multiple species. For example, it appeared in the title of the abstract as a murine gene, but turns out to be referring to a human (patient) gene in the first sentence. Therefore, each gene must be linked to its own unique database entry. Since these instances are polysemous, they are considered entity ambiguity issues. Finally, the \"C9 complex\" in the first sentence is a protein complex, but the Entrez Gene database does not contain this type of entity. When an entity cannot be associated with any entries, it is called an absence issue (McNamee & Dang, 2009b) , and those entities are referred to as \"Nils\".", "cite_spans": [ { "start": 539, "end": 563, "text": "McNamee and Dang (2009b)", "ref_id": "BIBREF24" }, { "start": 1860, "end": 1883, "text": "(McNamee & Dang, 2009b)", "ref_id": "BIBREF24" } ], "ref_spans": [ { "start": 671, "end": 679, "text": "Figure 1", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Of all of the aforementioned issues, entity ambiguity is the most crucial problem (Dredze et al., 2010) . Take the name \"TP53\" as an example. In the Entrez Gene database, there are over 300 proteins within over 20 species possessing the same name. Several disambiguation approaches have been proposed to address the problem. For example, Dredze et al. (2010) formulated the disambiguation task as a ranking problem and developed features to link entities to Wikipedia entries. Zhang et al. (2010) used an automatically generated corpus to train a binary classifier to reduce ambiguities. Dai et al. (2010) collected external knowledge for each entity and calculated likelihoods stating the similarity of the current text with the knowledge to improve the disambiguation performance.", "cite_spans": [ { "start": 82, "end": 103, "text": "(Dredze et al., 2010)", "ref_id": "BIBREF6" }, { "start": 338, "end": 358, "text": "Dredze et al. (2010)", "ref_id": "BIBREF6" }, { "start": 477, "end": 496, "text": "Zhang et al. (2010)", "ref_id": "BIBREF37" }, { "start": 588, "end": 605, "text": "Dai et al. (2010)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Usually, a real-world EL system is constructed in a bottom-up manner, so it is necessary to make several decisions in different stages during the EL process. Figure 2 depicts the bottom-up process (Krauthammer et al., 2004) . Entity recognition marks single words (or several adjacent words) that indicate the presence of entities. As entity recognition does not determine the specific meaning of a concept, it is often combined with Entity Classification, which assigns entities to different classes, such as persons, genes, or diseases. After removing Nils (Entity Filtering), Entity Mapping maps entities to controlled database entries by calculating the similarities between the recognized entities and lexicon resources. This stage may resolve the entity ambiguity issue by a disambiguation process that uses contextual information to link entities to KB entries.", "cite_spans": [ { "start": 197, "end": 223, "text": "(Krauthammer et al., 2004)", "ref_id": "BIBREF17" } ], "ref_spans": [ { "start": 158, "end": 166, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "As shown in Figure 2 , the traditional method for dealing with Nils has been to employ an additional step to filter out entities that have no corresponding entry in a KB. For example, Bunescu et al. (2006) filtered out mentions whose confidence scores are less than a fixed threshold. J and Li et al. (2009) trained separate binary classifiers to validate linked mentions. Dredze et al. (2010) treated Nils as another KB entry candidate to train their EL ranking model. Unfortunately, the separate-stage approach ignores possible dependencies among these stages and can result in error propagation. Continuing our example in Figure 1 , in the EL stage, \"MIRL\" can be unambiguously linked to ID 996 with high confidence, because a search for the name in Entrez Gene returns only one match. Nevertheless, linking other mentions (e.g. \"CD59\" and \"protectin\") to ID 996 is not as easy, since \"CD59\" alone has 18 candidate entries. These names can be linked with more ease when considered as synonyms of MIRL. Nevertheless, a divergent filtering stage may filter out the entity mention \"MIRL\" because it is listed as an abbreviation of organization names, such as Mineral Industry Research Laboratory.", "cite_spans": [ { "start": 184, "end": 205, "text": "Bunescu et al. (2006)", "ref_id": "BIBREF0" }, { "start": 291, "end": 307, "text": "Li et al. (2009)", "ref_id": "BIBREF20" }, { "start": 373, "end": 393, "text": "Dredze et al. (2010)", "ref_id": "BIBREF6" } ], "ref_spans": [ { "start": 12, "end": 20, "text": "Figure 2", "ref_id": null }, { "start": 625, "end": 633, "text": "Figure 1", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "With a joint inference process, we can carry out both tasks simultaneously to avoid this type of error propagation (Poon et al., 2007) .", "cite_spans": [ { "start": 115, "end": 134, "text": "(Poon et al., 2007)", "ref_id": "BIBREF31" } ], "ref_spans": [], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "Joint inference has become popular recently, because it allows features and constraints to be shared among different tasks. For example, J. R. Finkel et al. (2009) integrated parsing and named entity recognition into a joint model, whereas Dai et al. (2011) created a joint model for co-reference resolution and gene normalization and Liu et al. (2012) conducted entity recognition and normalization jointly for tweets. In this paper, we use the Markov Logic Network (MLN) (Richardson et al., 2006 ), a joint model that combines first order logic and Markov networks, to capture the bottom-up decisions derived from the process illustrated in Figure 2 . This model captures the contextual information of the recognized entities for entity disambiguation, as well as the constraints used when linking an entity mention to a database entry. For example, an entity mention can only be linked to a database entry when the mention has not been recognized as a Nil.", "cite_spans": [ { "start": 143, "end": 163, "text": "Finkel et al. (2009)", "ref_id": "BIBREF8" }, { "start": 240, "end": 257, "text": "Dai et al. (2011)", "ref_id": "BIBREF3" }, { "start": 335, "end": 352, "text": "Liu et al. (2012)", "ref_id": "BIBREF21" }, { "start": 473, "end": 497, "text": "(Richardson et al., 2006", "ref_id": "BIBREF32" } ], "ref_spans": [ { "start": 643, "end": 651, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "Existing EL evaluation metrics assess a system's performance in terms of the effectiveness of database curation (Morgan et al., 2008) or question answering (QA) accuracy . In addition, we evaluate our system at a fine-grained entity by entity level. Such evaluation is more relevant to information extraction tasks, such as the bio-molecular event extraction task (Kim et al., 2009) .", "cite_spans": [ { "start": 112, "end": 133, "text": "(Morgan et al., 2008)", "ref_id": null }, { "start": 364, "end": 382, "text": "(Kim et al., 2009)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "When considering EL tasks from the entity level, one challenge is the lack of contextual information for disambiguating each individual entity. The major scheme of traditional entity disambiguation approaches relies on domain knowledge derived from entities' profiles and contextual features extracted within a predefined content window. Rule-based (Dai et al., 2010; J\u00f6rg Hakenberg et al., 2008) , vector space models (Cucerzan, 2007) , and machine learning approaches (Crim et al., 2005; Mihalcea et al., 2007; Milne et al., 2008) have been proposed to disambiguate entity mentions individually. Nevertheless, the context is unclear under certain circumstances. Take the sentence \"The synthetic replicate of urocortin can bind with high affinity to Type 1 and Type 2 CRF receptors\" as an example. The sentence itself does not explicitly provide any clues to help computer programs determine the identity of the gene mention \"urocortin\", which has at least eight ambiguous Entrez Gene IDs. One approach is to expand the context window used for disambiguation to the paragraph level. Nevertheless, a paragraph described in a biomedical article usually incorporates several pieces of information in its description, which may not be related directly to a target entity instance and leads to the failure of traditional EL approaches.", "cite_spans": [ { "start": 349, "end": 367, "text": "(Dai et al., 2010;", "ref_id": "BIBREF4" }, { "start": 368, "end": 396, "text": "J\u00f6rg Hakenberg et al., 2008)", "ref_id": null }, { "start": 419, "end": 435, "text": "(Cucerzan, 2007)", "ref_id": "BIBREF2" }, { "start": 470, "end": 489, "text": "(Crim et al., 2005;", "ref_id": "BIBREF1" }, { "start": 490, "end": 512, "text": "Mihalcea et al., 2007;", "ref_id": "BIBREF27" }, { "start": 513, "end": 532, "text": "Milne et al., 2008)", "ref_id": "BIBREF28" } ], "ref_spans": [], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "Our idea of dealing with the challenge of deficient contextual information for disambiguating individual entity instances is to model dependencies among entities across sentences in the same paragraph. These dependencies have been ignored by most of the previous EL approaches. We refer to our approach as the collective EL, which is developed by considering the relational information hidden among entities. In the following sections, we first give formal definitions of the EL tasks mentioned above, followed by an introduction of MLN and a description of the main ideas of the proposed EL method with the formulation of the collective EL approach.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Figure 2. Stages in the bottom-up EL approach: Some works combine the entity recognition and the entity classification into one step.", "sec_num": null }, { "text": "This section gives formal definitions of all related EL tasks.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Entity Linking Problem Definition", "sec_num": "2." }, { "text": "Let M = (m 1 , m 2 , \u2026) denote a sequence of entities mentioned in an article A. The surface name of m i is denoted by Name(m i ). The named entity type of m i is EntityType(m i ). The surrounding context of m i can be extracted by Context(m i ). Given a KB containing a set of entries ID = {id 1 , id 2 , \u2026}, each of which organizes knowledge related to an entity, the instance-based EL problem is defined as finding a mapping function LinkTo(m i ) that maps each m i in M to a unique entry id i in ID and satisfies the constraint ( ):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 1: Instance-based Entity Linking Problem", "sec_num": null }, { "text": "i i LinkTo m m M M \uf0ce \uf03d .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 1: Instance-based Entity Linking Problem", "sec_num": null }, { "text": "In instance-based gene mention linking (GML), only entities whose EntityType(m i ) belong to \"gene\" are considered for evaluation. Both the gene normalization task in BioCreAtIvE (Morgan et al., 2008) ", "cite_spans": [ { "start": 179, "end": 200, "text": "(Morgan et al., 2008)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Definition 1: Instance-based Entity Linking Problem", "sec_num": null }, { "text": "i i LinkTo m m M M \uf0ce \uf0a3", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 1: Instance-based Entity Linking Problem", "sec_num": null }, { "text": ". We refer to this task as the article-wide EL problem.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 1: Instance-based Entity Linking Problem", "sec_num": null }, { "text": "Let M = {m 1 , m 2 , \u2026} denote a set of entities mentioned in A. Given the entries ID = {id 1 , id 2 , \u2026} in a KB and the mapping function LinkTo(m i ), the article-wide EL problem satisfies the constraint ( ):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 2: Article-wide Entity Linking Problem", "sec_num": null }, { "text": "i i LinkTo m m M M \uf0ce \uf0a3 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 2: Article-wide Entity Linking Problem", "sec_num": null }, { "text": "On the other hand, the KB population EL task only considers one certain entity m i mentioned in A. We refer to this task as the article-wide \"salient entity\" linking problem, in accordance with the Wikipedia style manual, in which only the salient entity and its related entities should be linked in wikification. Excessive links would obstruct the readers in following the article by drawing attention away from important links (Mihalcea & Csomai, 2007) .", "cite_spans": [ { "start": 429, "end": 454, "text": "(Mihalcea & Csomai, 2007)", "ref_id": "BIBREF27" } ], "ref_spans": [], "eq_spans": [], "section": "Definition 2: Article-wide Entity Linking Problem", "sec_num": null }, { "text": "Let M = m i denote the salient entity mentioned in A. Note that, in encyclopedia-style articles, 1 M \uf03d because the same surface name described in such articles should refer to the same instance. Given the entry set ID = {id 1 , id 2 , \u2026} of a KB, the purpose of the article-wide salient EL problem is to find the mapping function LinkTo(m i ) that links m i to a unique entry id i in E.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 3: Article-wide Salient Entity Linking Problem", "sec_num": null }, { "text": "Note that, in the KB population EL subtask (pertained to the article-wide salient EL problem), the salient entity is given. Nevertheless, in the instance-based GML or the BioCreAtIvE gene normalization (pertaining to the article-wide EL problem) tasks, the systems must also deal with the entity recognition/classification problem.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 3: Article-wide Salient Entity Linking Problem", "sec_num": null }, { "text": "Markov logic is a statistical relational learning language based on first-order logic (FOL) and Markov networks. In this section, we consider FOL and Markov networks in terms of the GML task.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "First-order Logic and Markov Logic Networks", "sec_num": "3." }, { "text": "In FOL, the formulae are constructed using four types of symbols: constants, variables, functions, and predicates. For GML, a constant symbol may represent a gene mention (e.g. \"CD59\") or its unique database entry (e.g. the Entrez Gene ID \"966\"). If variables and constants are type-specific, their range can only cover objects of the corresponding type. To give an example, the variable y's range covers all Entrez Gene database IDs. Predicate symbols are used to represent the relations between terms; for example, we can define the predicate, LinkTo(x, y), to indicate that a gene mention (the variable x) should be linked to an entry (the variable y). Formulae are constructed recursively from predicates applied to a tuple of terms by through use of logical connectives and quantifiers. Then, we can model the EL task by introducing a set of logical predicates. For instance, we can define the predicate Candidate(i, j) to indicate that the gene mention i can be mapped to an entry j. The predicate captures information about gene mentions and their corresponding candidate database entries. Through this predicate, we can infer whether a gene mention is unambiguous. Then, we can use the following formula", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "First-order Logic and Markov Logic Networks", "sec_num": "3." }, { "text": "Formula 1: ! . ( , ) ( , ) x id Candidate x id LinkTo x id \uf022 \uf024 \uf0de", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "First-order Logic and Markov Logic Networks", "sec_num": "3." }, { "text": "to model the concept that, when a gene mention is mapped to only one entry, it should be linked to that entry. Note that we use the symbol, ! \uf024 , to refer to a uniqueness quantification.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "First-order Logic and Markov Logic Networks", "sec_num": "3." }, { "text": "A first-order KB is a set of hard constraints on the set of ground atoms of predicates (or so-called possible worlds). If a world violates any formula, it has zero probability. In most domains, however, it is very difficult to derive non-trivial formulae that are always true. Markov logic softens these constraints to handle uncertainty by associating each formula with a weight that reflects the strength of a constraint. Ideally, if we could define a formula with a proper weight for its distribution, a world in which the formula is satisfied would have a higher probability than a world in which it is not. In Markov logic, a set of weighted formulae is called a MLN.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "First-order Logic and Markov Logic Networks", "sec_num": "3." }, { "text": "An MLN L is a set of pairs (F i , w i ), where F i is a formula in FOL and w i is a learned weight corresponding to the F i whose value is a real number. In combination with a finite set of constants C = {c 1 , c 2 , . . . , c |C| }, it defines a Markov network M L,C as follows: M L,C contains one node for each possible grounding of each predicate appearing in L. The value of the node is 1 if the ground predicate is true, otherwise it is 0.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": "Based on the definition, we can generate a graph structure of the ground Markov network where there is an edge between two nodes of M L,C if the corresponding ground atoms appear together in at least one grounding of one formula in L. Thus, the predicates in each ground formula form a clique in M L,C . Each clique in the graph is associated with a potential function i \uf066 .The joint distribution of a set of variables X represented by M L,C then is defined by:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": "( ) { } 1 P( ) ( ) i n x i i i X x x Z \uf066 \uf03d \uf03d \uf0d5", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": "where x {i} is the state of the ith clique (i.e., the exact values of the predicates that appear in that clique F i ), n i (x) is the number of true groundings of F i in x, and", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": "{ } ( ) i w i i x e \uf066 \uf03d . Z is the partition function given by { } ( ) i i x X i Z x \uf066 \uf0ce \uf03d \uf0e5 \uf0d5", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": ". Markov networks are often represented as log-linear models in which each clique is replaced by an exponential weighted sum of the features of the state, leading to Figure 3 illustrates the input of the proposed EL system developed for GML and the FOL predicates defined for the corresponding bottom-up stages. The input is an article, such as a biomedical abstract. The given article is first processed by a gene mention recognizer to identify gene mention boundaries, and the employed gene mention mapper then maps each recognized gene to a list of candidate identifiers, based on a lexicon compiled from the Entrez Gene database. Ideally, we should be able to treat all recognized gene mentions as candidates, and proceed directly to the entity mapping task. Nevertheless, the employed recognizer may generate false positive gene mentions. Such mentions can be classified into two types: those that do not belong to any entity class and those that belong to classes that are not the curation target (Nils). In GML, Nils appear when the gene mentions are DNA polymerases, or in a specific organism that is not considered. To capture the concept, we define the predicate isSuitableForLinking(x), which indicates that the gene mention x of the article is suitable for linking to an entry. For entity mapping, we use the predicate LinkTo(x, id) to represent that the gene mention x must be linked to the database entry id. As the objective of the entity mapping task is to determine a unique KB entry for each entity, we must define a formula to ensure that the constraint is satisfied. Regarding GML, we use the following formula to prevent a gene mention from associating with more than two identifiers.", "cite_spans": [], "ref_spans": [ { "start": 166, "end": 174, "text": "Figure 3", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": "1 P( ) exp ( ) i i i X x wf x Z \uf0e6 \uf0f6 \uf03d \uf03d \uf0e7 \uf0f7 \uf0e8 \uf0f8 \uf0e5 In our implementation, f i is a binary feature, ( ) {0,1} i f x \uf0ce .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 4: Markov Logic Network", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( , )", "eq_num": "( , )" } ], "section": "Formula 2: Entity Mapping Constraint", "sec_num": null }, { "text": "i i j j LinkTo x id id id LinkTo x id \uf0d9 \uf0b9 \uf0de\uf0d8", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formula 2: Entity Mapping Constraint", "sec_num": null }, { "text": "Within the machine learning community, classification is typically done on each object independently without taking into account any underlying relation that connects the objects. In most of the individual EL formulations, an individual classifier is employed to assign a probability to the linked ID of an individual instance independently of the linked IDs of other instances. For example, the following formula expresses that, if the chromosome location information of the entity mention x, which has the KB entry id as its candidate ID, can be found in the surrounding text, x should be linked to id.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulation of the Instance-based GML", "sec_num": "4.1" }, { "text": "Formula 3: Other useful biographical information for locally disambiguating gene mentions includes tissues, gene ontology, and PPI. Some researchers (Hakenberg et al., 2007; Lai et al., 2009) have used this information for individual GML. Table 1 shows the observed predicates and formulae defined for this individual approach. We take hasPPIPartnerRank as another example. For the individual GML process, we define the following formula. implies that the gene mention x should be linked to the id that has the most PPI partners. The other predicates, including hasGOTermRank and hasTissueRank, follow a similar concept, in which the context is matched with the corresponding keywords to determine the frequency in the given abstract text. hasPPIPartnerRank (x, id, r) hasGOTermRank (x, id, r) hasTissueTermRank (x, id ,r) hasDictionaryMatchRank (x, id, r) hasPrecedingWord (x, w, l), hasFollowingWord(x, w, l) hasUnigramBetween (x, y, w) Variable Type", "cite_spans": [ { "start": 149, "end": 173, "text": "(Hakenberg et al., 2007;", "ref_id": "BIBREF13" }, { "start": 174, "end": 191, "text": "Lai et al., 2009)", "ref_id": "BIBREF19" }, { "start": 758, "end": 768, "text": "(x, id, r)", "ref_id": null }, { "start": 783, "end": 793, "text": "(x, id, r)", "ref_id": null }, { "start": 812, "end": 822, "text": "(x, id ,r)", "ref_id": null }, { "start": 846, "end": 856, "text": "(x, id, r)", "ref_id": null }, { "start": 874, "end": 910, "text": "(x, w, l), hasFollowingWord(x, w, l)", "ref_id": null }, { "start": 929, "end": 938, "text": "(x, y, w)", "ref_id": null } ], "ref_spans": [ { "start": 239, "end": 246, "text": "Table 1", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Formulation of the Instance-based GML", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( ) ( , )", "eq_num": "( ," } ], "section": "Formulation of the Instance-based GML", "sec_num": "4.1" }, { "text": "x: integer that refers to the xth gene mention in the given article (similarly y refers to the yth gene mention) id: an Entrez Gene ID, which refers to the linked KB entry. w: a word. r: integer that refers to the rank of the matching. l: integer that refers to a context window length. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulation of the Instance-based GML", "sec_num": "4.1" }, { "text": "\uf0d9 \uf0d9 \uf0d9 \uf024 \uf0d9 \uf03d \uf0de ( , ) ! . ( ) ( , ) i i i Candidate x id id MostGOTerms id id id LinkTo x id \uf0d9 \uf024 \uf0d9 \uf03d \uf0de ( , ) ! . ( ) ( , ) i i i Candidate x id id MostTissueTerms id id id LinkTo x id \uf0d9 \uf024 \uf0d9 \uf03d \uf0de ! . ( , ) ( , ) id Candidate x id LinkTo x id \uf024 \uf0de ! , . ( , ) Pr ( 1, ) \"(\" ( , 1, ) ( 1,\")\") ( 1, ) u id Candidate x id has ecedingWord x u u hasUnigramBetween x x u hasFollowingWord x LinkTo x id \uf024 \uf0d9 \uf02b \uf0d9\uf03d \uf0d9 \uf02b \uf0d9 \uf02b \uf0de \uf02b ! , . ( 1, ) Pr ( 1, ) \"(\" ( , 1, ) ( 1,\")\") ( , ) u id Candidate x id has ecedingWord x u u hasUnigramBetween x x u hasFollowingWord x LinkTo x id \uf024 \uf02b \uf0d9 \uf02b \uf0d9 \uf03d \uf0d9 \uf02b \uf0d9 \uf02b \uf0de", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulation of the Instance-based GML", "sec_num": "4.1" }, { "text": "A drawback of individual EL classifiers is that, when they decide the linking entry of an entity, they cannot utilize information about the linked entries and features of other entities in the same article. Nevertheless, those entity instances can be related, and the interrelationship can be used to improve the EL performance. Furthermore, there are strong dependencies among the unknown IDs of the instances, which could either be a true positive entity mention or a Nil. These dependencies are highly nonlocal.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulation of the Instance-based GML", "sec_num": "4.1" }, { "text": "Hong-Jie Dai et al. Collective classification refers to the task of inferring labels for a set of objects using not just their attributes, but also the relations among them (Sen et al., 2008) .", "cite_spans": [ { "start": 9, "end": 19, "text": "Dai et al.", "ref_id": null }, { "start": 173, "end": 191, "text": "(Sen et al., 2008)", "ref_id": "BIBREF34" } ], "ref_spans": [], "eq_spans": [], "section": "20", "sec_num": null }, { "text": "Given a network N, a node n in N, and the label set L, there are three distinct feature types that can be utilized to determine the label l of n, where l L \uf0ce .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 5: Collective Classification", "sec_num": null }, { "text": "1. The observed features of n.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 5: Collective Classification", "sec_num": null }, { "text": "2. The observed features (including observed labels if they are known) of nodes in the neighborhood (related nodes) of n.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 5: Collective Classification", "sec_num": null }, { "text": "3. The unobserved labels of nodes in the neighborhood (related nodes) of n.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Definition 5: Collective Classification", "sec_num": null }, { "text": "In our formulation for EL, for a given article, the candidate database entries of all recognized entities form the network N. A mention's candidate entry and order form the node n = (id, order) in N, and an edge exists between two nodes if they have dependencies. In this work, the dependencies are constructed based on two main ideas: the discourse salience property in centering theory (Grosz et al., 1995) and the protein-protein interaction (PPI) association.", "cite_spans": [ { "start": 388, "end": 408, "text": "(Grosz et al., 1995)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Definition 5: Collective Classification", "sec_num": null }, { "text": "Discourse salience is a phenomenon where, in a given discourse, there is precisely one entity that is the center of attention. This entity is mentioned over and over again, which makes it more salient than others. We utilize this phenomenon to improve the instance-based EL confidence. Suppose that id is a candidate database entry for several entities in a discourse, we then can assume that id is more salient than other database entries. If the EL system can link one of these mentions to id with high confidence, then the system is more likely to be able to link all of the other mentions to id as well.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discourse Salience", "sec_num": "4.1.1" }, { "text": "Similarly, the idea of employing the PPI association allows us to express the concept that a gene mention y should be linked to id j if another gene mention x has been linked to id i and id i forms an interaction with id j .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Protein-protein Interaction", "sec_num": "4.1.2" }, { "text": "In order to capture the concepts above, the order of all individual instances described in an abstract are leveraged to build dependencies in our formulation. The lack of local contextual information then can be resolved by the constructed dependencies. In our work, the salience collective is written as follows in Markov logic. If the database entry id is linked to an entity x that precedes the current mention y and id is a candidate entry of y, then the current entity y should also be linked to id. This formula is similar to the transition feature of the linear-chain conditional random fields (Lafferty et al., 2001) , which can be implemented in Markov logic as follows.", "cite_spans": [ { "start": 601, "end": 624, "text": "(Lafferty et al., 2001)", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Protein-protein Interaction", "sec_num": "4.1.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( , )", "eq_num": "( , ) ( , ) ( , )" } ], "section": "Formula 5: Salience collective", "sec_num": null }, { "text": "Transition feature:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formula 5: Salience collective", "sec_num": null }, { "text": "( , ) ( , ) ( , ) Pr e cede x y Label x L Label y L \uf0d9 \uf02b \uf0de \uf02b", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formula 5: Salience collective", "sec_num": null }, { "text": "Note that the symbol \"+\" in the above formula directs the MLN learning algorithm to associate the formula with a different weight depending on variables containing the \"+\" notation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formula 5: Salience collective", "sec_num": null }, { "text": "We define the predicate PPIPartner(id i , id j ), whose value is true if id i and id j form a PPI pair. We then use the following formula to capture the PPI association concept. Based on these two collective formulae, Figure 4 compares the ground Markov network (b) of our collective formulation with the traditional individual approach (a). In Figure 4 (a) , the individual approach considers the likelihoods stating the similarity of the current context with the domain knowledge of the recognized entity, including chromosome location (ChromosomeInfo) and gene ontology (MostGOTerm). Comparing Figure 4 (b) with (a), our collective formulation captures the dependencies among entities, allowing the information to be employed in the GML decision. ", "cite_spans": [], "ref_spans": [ { "start": 218, "end": 226, "text": "Figure 4", "ref_id": "FIGREF5" }, { "start": 345, "end": 357, "text": "Figure 4 (a)", "ref_id": "FIGREF5" }, { "start": 597, "end": 605, "text": "Figure 4", "ref_id": "FIGREF5" } ], "eq_spans": [], "section": "Formula 5: Salience collective", "sec_num": null }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( , ) ( , )", "eq_num": "( , ) ( ," } ], "section": "Formula 6: PPI", "sec_num": null }, { "text": "The hidden predicate isSuitableForLinking captures the decisions made after the entity recognition/classification stage. When the gene mention x is linked to an identifier id, we employ the following formula to ensure that it is an entity suitable for linking.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "Formula 7:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "( , ) ( ) LinkTo x id isSuitableForLinking x \uf0de", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "Note that the formula models the bottom-up decision, as shown in Figure 2 . The identifier id does not have to be linked to the gene mention x proposed by the entity recognition/classification stage. Nevertheless, the id cannot be assigned to the gene mention x that has not been proposed as a potential entity.", "cite_spans": [], "ref_spans": [ { "start": 65, "end": 73, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "Our first formula for isSuitableForLinking treats all gene mentions as potential entities:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "( , ) ( ) hasName x n isSuitableForLinking x \uf0de .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "The other formulae are constructed using the observed predicates defined in Tables 1 and 2 to check the contextual information. For example,", "cite_spans": [], "ref_spans": [ { "start": 76, "end": 91, "text": "Tables 1 and 2", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "( , ) ( ) ( ) hasFirstWord x w isSpeciesTerm w isSuitableForLinking x \uf02b \uf0d9 \uf02b \uf0de", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "implies that the suitability of a certain gene mention for linking depends on whether or not the first word is a keyword for a certain species. hasName(x, n) hasFirstWord (x, w), hasLastWord(x, w) hasPrefix (x, ch, d, l) : the xth gene mention has a prefix ch of length l, and the prefix's case is the same as its following character (d = 0) or different (d = 1).", "cite_spans": [ { "start": 171, "end": 196, "text": "(x, w), hasLastWord(x, w)", "ref_id": null }, { "start": 207, "end": 220, "text": "(x, ch, d, l)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "isSpeciesTerm(w), isAllUpperCase(i), hasPartOfSpeech (x, k, p) isContainedMoreSpecificMentions (x) Variable Type n: a word or a sequence of words that refer to the surface name of a gene mention. ch: a character. d: an integer. k: the kth index of the gene mention. p: a part-of-speech", "cite_spans": [ { "start": 53, "end": 62, "text": "(x, k, p)", "ref_id": null }, { "start": 95, "end": 98, "text": "(x)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Formulae for Entity Recognition/Classification and Filtering", "sec_num": "4.2" }, { "text": "We use three metrics to evaluate our approach and compare it with other GML methods. The first and second metrics used the standard precision, recall, and F-measure metrics (PRF) at two resolutions (article and instance).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "5.1.1" }, { "text": "Article-wide evaluation used the standard used in the BioCreAtIvE challenge (Hirschman et al., 2005) , which is designed to determine an GML system's performance as an aid for the curation of biological databases. The GML system outputs a list of IDs for a given article, and this list is compared to the gold standard ID list. The PRF scores are calculated based on the sums of true/false positives/negatives (TP, TN, FP, FN) .", "cite_spans": [ { "start": 76, "end": 100, "text": "(Hirschman et al., 2005)", "ref_id": "BIBREF14" } ], "ref_spans": [ { "start": 410, "end": 426, "text": "(TP, TN, FP, FN)", "ref_id": null } ], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "5.1.1" }, { "text": "Instance-based evaluation measures the GML performance at a fine-grained resolution (Dai et al., 2011) . In contrast to the first metric, the PRF scores are calculated based on the sums of TP, TN, FP, and FN for all instances in the test dataset. Therefore, under this criterion, an FP can link a true gene mention to the wrong KB entry or link a false gene mention to any entry, while an FN can link a true gene mention to the wrong entry or fail to recognize a true gene mention. For TP/FP/FN, we need to determine when the predicted boundary matches that of the gold standard. Most pure entity recognition tasks use \"exact-matching\" as the primary criterion. Under this criterion, a candidate gene mention can only be counted as a TP if both its left and right boundaries fully coincide with the gold answer. In a real case, however, a gene mention can be tagged in several ways (e.g., \"serum LH \" and \" serum LH \" are both correct), which are intrinsic to the annotation of any gene mention corpus, whether developed by humans or machines (Tsai et al., 2006) , and may depend on the annotator's perspective (Franz\u00e9n et al., 2002) . Furthermore, for the GML task, the correctness of the linked entry is more important than its boundary. Therefore, we used the approximate-match to determine the boundary criterion. For example, a TP is counted when a machine-linked gene mention is a substring of the gold standard-linked gene mention or vice versa and the linked entry is equal to the gold entry.", "cite_spans": [ { "start": 84, "end": 102, "text": "(Dai et al., 2011)", "ref_id": "BIBREF3" }, { "start": 1081, "end": 1100, "text": "(Tsai et al., 2006)", "ref_id": null }, { "start": 1149, "end": 1171, "text": "(Franz\u00e9n et al., 2002)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "5.1.1" }, { "text": "The third metric, the mean accuracy across all queries, considers the QA perspective. EL is important in QA systems because the systems rely on data from multiple sources, so name ambiguity will lead to wrong answers and poor results. We adopt the evaluation metrics used in text analysis conference KB population (KBP) track 2009 (McNamee & Dang, 2009a) to report Accuracy micro and analyze the results from the QA perspective 1 .", "cite_spans": [ { "start": 331, "end": 354, "text": "(McNamee & Dang, 2009a)", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "Evaluation Metrics", "sec_num": "5.1.1" }, { "text": "In the experiments, we used the training and test sets (281 and 262 abstracts, respectively) released by the BioCreAtIvE gene normalization task (Morgan et al., 2008) for article-wide evaluation. The corpus contains annotations for human genes that are linked to IDs in the Entrez Gene database. Although the gold answers contain each ID's surface name, they do not give the exact location of the corresponding gene mention in the abstract. To obtain fine-grained evaluation results, our in-lab biologists compiled an instance-based GML corpus by annotating the exact location and the boundary of the IDs' gene mentions (Dai et al., 2012) . After compiling the corpus, we performed three-fold cross validation (CV) on the training dataset to optimize the weights and formulae and to evaluate its performance on the test set.", "cite_spans": [ { "start": 145, "end": 166, "text": "(Morgan et al., 2008)", "ref_id": null }, { "start": 620, "end": 638, "text": "(Dai et al., 2012)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Datasets", "sec_num": "5.1.2" }, { "text": "In QA evaluation, as defined in KBP, the associated document is used to provide contextual information that might be useful for linking. We paired the surface names with their corresponding documents as the input query. For each query, the corresponding gold answer could be 1) an Entrez Gene ID or 2) a Nil in cases where our biologists annotated the entity as a mention without associating it with any IDs. Table 3 shows the generated query/answer pairs based on the BioCreAtIvE Corpus. ", "cite_spans": [], "ref_spans": [ { "start": 409, "end": 416, "text": "Table 3", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Datasets", "sec_num": "5.1.2" }, { "text": "To assess the performance of our models and determine the possible gains that can be achieved by considering a collective model and a joint model of the bottom-up stages, we designed two configurations. The first configuration was the collective model, which used all of the disambiguation formulae defined in Section 4.1 (denoted as CM). The constructed Markov network resembles Figure 4 (b) with additional grounding for the predicates and individual formulae defined in Table 1 . The second configuration further included the formulae defined in Section 4.2 on CM to build a joint model (denoted as JCM). This work used the 1-best Margin Infused Relax online learning Algorithm (MIRA) (McDonald et al., 2005) for learning weights and employed cutting plane inference (Riedel, 2008) with integer linear programming as its base solver for inference at test time as well as during the MIRA online learning process.", "cite_spans": [ { "start": 688, "end": 711, "text": "(McDonald et al., 2005)", "ref_id": "BIBREF22" }, { "start": 770, "end": 784, "text": "(Riedel, 2008)", "ref_id": "BIBREF33" } ], "ref_spans": [ { "start": 380, "end": 388, "text": "Figure 4", "ref_id": "FIGREF5" }, { "start": 473, "end": 480, "text": "Table 1", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Model Configurations", "sec_num": "5.1.3" }, { "text": "In addition, we compared the first configuration (CM) with two GML approaches: Lai et al. (2009) 's rule-based approach and Crim et al. (2005) 's maximum entropy (ME) approach, which handled the GML task as an individual classification problem. The ME approach was adopted with the individual features described in Section 4.1.", "cite_spans": [ { "start": 79, "end": 96, "text": "Lai et al. (2009)", "ref_id": "BIBREF19" }, { "start": 124, "end": 142, "text": "Crim et al. (2005)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Model Configurations", "sec_num": "5.1.3" }, { "text": "Finally, to compare the stage-based approach with the second configuration (JCM), which performs joint filtering and linking, we trained a separate ME model via the features designed for the isSuitableForLinking stage in Figure 3 . This model then was combined with the first configuration, the Rule-based approach and ME approach, which we denote as CM stage , Rule-based stage , and ME stage , respectively.", "cite_spans": [], "ref_spans": [ { "start": 221, "end": 229, "text": "Figure 3", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "Model Configurations", "sec_num": "5.1.3" }, { "text": "For the above configurations, we employed Lai et al.'s system 2 to recognize all gene mentions and generate mapped candidate IDs for each mention. All configurations were based on the same candidate ID sets.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Model Configurations", "sec_num": "5.1.3" }, { "text": "In the next sub-section, we first discuss the fine-grained resolution results. Then, we derive BioCreAtIvE's evaluation results by simply merging the linked identifiers in all indices and removing duplicated identifiers. Finally, the results are displayed from the QA perspective. Table 4 shows the fine-grained results derived on the training set. The employed system's linking performance without applying any disambiguation approaches is shown in the first row (No Disambiguation.) Our model can simulate the same PRF scores when only Formula 1 is applied. The last column shows the improvement of F-score over the baseline after implementing different GML disambiguation methods.", "cite_spans": [], "ref_spans": [ { "start": 281, "end": 288, "text": "Table 4", "ref_id": "TABREF8" } ], "eq_spans": [], "section": "Model Configurations", "sec_num": "5.1.3" }, { "text": "It can be observed that, by adding the salience collective (Formula 5) without any disambiguation formulae and domain knowledge, the recall rate is improved by 1.6%, which results in an improved F-score. This demonstrates that a scientific article often contains repetitive information, such as key genes, which can be captured by the formula. Furthermore, the PPI collective combining the domain knowledge achieves a higher PRF-score, even outperforming the Rule-based and ME. The results derived on the test set by fine-grained resolution and aided for curation and QA are shown in Tables 5 and 6, respectively. In summary, we observe that the collective GML method consistently outperforms the compared methods under the three criteria. Moreover, the comparison of our joint model (JCM) and the stage models shows that the joint model performs better under all evaluation metrics. We also observe that adding the salience collective reduces the recall rate in the aid for curation evaluation. According to our analysis, the salience collective improves the recall in the fine-grained evaluation. In contrast, for database curation, the collective tends to improve the overall precision. Nevertheless, it reduces the recall. By removing the salience collective from CM, the P improved by 0.8% but the R reduced by 0.7% in the test set. This phenomenon is reasonable because adding the dependency causes the model to link mentions with previous linked IDs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiment Results", "sec_num": "5.2" }, { "text": "Finally, the results also reveal the performance gap (approximately 15.8%) when we want to employ the GML system, which is evaluated in terms of the database curation criterion with 80+% F-score on advanced IE tasks, such as relation or event extraction.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiment Results", "sec_num": "5.2" }, { "text": "Evaluating EL from the fine-grained perspective allows us to analyze the task in detail. In this section, we describe our findings and propose potential research directions.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "One advantage of employing MLN in our EL modeling is that it is easy to model arbitrary longer range dependencies, as expressed by Formula 5 and Formula 6. It is difficult to model such dependencies using ME. As shown in Tables 4 and 5, adding the collectives improves the fine-grained EL performance.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "Another advantage of our GML approach is that it is flexible and can be applied quickly in real-world applications. The EL task usually is defined as linking a mention to a unique entry. Nonetheless, in the biomedical field, there are some mention descriptions that cannot be linked to unique IDs. The following are some examples extracted from our corpus:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "1. ABCB9 protein appears to be most highly expressed in the Sertoli cells of the seminiferous tubules in mouse and rat testes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "2. cDNA cloning and chromosomal localization of the human and mouse isoforms of Ksp-cadherin.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "3. p63 was detected in a variety of human and mouse tissues.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "The GML system cannot link each of the gene mentions in the above sentences to just one ID. Our model can deal with these cases by modifying the constraint of Formula 2 with a larger cardinality or introducing additional formulae to determine the cardinal constraint dynamically.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "Our experiment results also raise an interesting question: What causes the huge performance gap between the fine-grained and database curation evaluations? A closer look at the bottom-up approach is useful in answering this question. Several works have studied the boundary issue in entity recognition (J. Finkel et al., 2005; Tsai et al., 2006) , and this issue was found to have a significant effect on the performance of GML. For example, consider the following sentence: \" Hepatocyte growth factor (HGF) activator is a serine protease responsible for proteolytic activation of HGF in response to tissue injury\" All of the employed gene mention recognition systems and the three open available gene mention recognition systems 3,4,5 separate the first gene mention (ID:3083) into at least one mention (\"hepatocyte growth factor\" or \"HGF\"). The incorrect boundary leads to errors in the entity mapping stage, and it could result in the extraction of an incorrect self-activation event: HGF activates HGF . An experiment conducted on the test set showed that our MLN model could achieve an F-score of 79.4% from the fine-grained IE perspective if we replaced the predicted mentions' boundaries with their corresponding overlapped gold standard boundaries. These results show that a hybrid approach combined with entity-centric boundary expansion is required before entity mapping. For instance, if we input the example sentence to a syntactic parser like Enju 6 and find that the adjacent words \"Hepatocyte growth factor (HGF) activator\" belong to the same noun phrase and the word \"activator\" is a legal suffix for a gene mention, it implies that we can expand the boundary.", "cite_spans": [ { "start": 306, "end": 326, "text": "Finkel et al., 2005;", "ref_id": "BIBREF7" }, { "start": 327, "end": 345, "text": "Tsai et al., 2006)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "The result also motivates us to reconsider the bottom-up EL approach. Are the results of entity recognition/classification a prerequisite for GML? We raise this question because, under the bottom-up approach, the entity mapping process still needs to deal with the boundary issue in order to generate more candidate identifiers, as shown by the previous example sentence. Moreover, the disambiguation process needs to look for knowledge, such as species information, surrounding the gene mention's boundary, which is usually located in the same noun phrase. It has been shown that joint learning of multiple types of linguistic structures in models can produce more consistent outputs. A feasible approach would be to treat noun phrases as potential candidate gene mentions and employ a mapping algorithm to generate identifiers for each noun phrase. Using the proposed approach to model biographical information and the dependencies between noun phrases, we can perform joint learning and inference for gene mention recognition and linking. This issue will be the major direction of our future research. For chunking parsing, there are several openly available tools, such as GENIA tagger 7 , OpenNLP 8 , and Lingpipe 9 package. Kang et al. (2011) have reported that the OpenNLP package performs noun-phrase chunking, best among the six state-of-the-art chunkers specifically for the biomedical domain. Therefore, we will use the OpenNLP package as the first step to generate candidate noun phrases, and we may consider combining more results from different chunkers through a voting strategy to further improve the chunking performance. Finally, we report our observations on the hardest queries under QA evaluation. Table 7 lists the queries that could not be answered correctly by any of the employed methods. After checking each error event carefully, we found that our model outputs Nil for Query #1 due to the absence of gene biographical information in the context. The gene mention of Query #2 is an abbreviation, and the query is affected by a problem similar to the ambiguous acronym problem discussed in the KBP 2009 track (McNamee, Dang, et al., 2009 . Our model fails to output correct IDs for #3 and #4 because no distinction is made between matches of official symbols and synonyms when searching for candidate IDs. In our current work, the matches of official symbols and those of synonyms share the same predicate. We believe that appending more predicates and formulae corresponding to these two types of matches will improve our system's accuracy.", "cite_spans": [ { "start": 1230, "end": 1248, "text": "Kang et al. (2011)", "ref_id": "BIBREF15" }, { "start": 2120, "end": 2128, "text": "KBP 2009", "ref_id": null }, { "start": 2129, "end": 2163, "text": "track (McNamee, Dang, et al., 2009", "ref_id": null } ], "ref_spans": [ { "start": 1719, "end": 1726, "text": "Table 7", "ref_id": "TABREF11" } ], "eq_spans": [], "section": "Discussion", "sec_num": "5.3" }, { "text": "In this paper, we give formal definitions for EL tasks, including instance-based EL, article-wide EL, and article-wide salient EL. We then present a novel approach that employs MLN to jointly model bottom-up decisions in a specific EL task-GML. A collective formulation for instance-based GML is introduced with several useful formulae, including the dependencies among IDs, which can be used for GML disambiguation. Moreover, the benefit of predicting suitable mentions and their IDs jointly in contrast to the stage-based approach is illustrated, which selects mentions before linking IDs. Our experiments provide the first comprehensive gene mention evaluation results from three different perspectives and highlight problems that need to be addressed in the future, including the assignment of non-unique identifiers, the boundary issue, and the direction for joint entity recognition and linking.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "For details please refer to http://apl.jhu.edu/~paulmac/kbp.html.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The employed system can be downloaded from https://sites.google.com/site/potinglai/downloads.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://pages.cs.wisc.edu/~bsettles/abner/ 4 http://www-tsujii.is.s.u-tokyo.ac.jp/GENIA/tagger/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This research was supported in part by the National Core Facility Program for Biotechnology (Taiwan Bioinformatics Consortium of Taiwan) grant NSC-100-2319-B-010-002, the National Science Council of Taiwan grant NSC-102-2218-E-038-001, as well as the research grant TMU101-AE1-B55 of Taipei Medical University.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgement", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Using encyclopedic knowledge for named entity disambiguation", "authors": [ { "first": "R", "middle": [], "last": "Bunescu", "suffix": "" }, { "first": "M", "middle": [], "last": "Pasca", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bunescu, R., & Pasca, M. (2006). Using encyclopedic knowledge for named entity disambiguation. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, Trento, Italy.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Automatically Annotating Documents with Normalized Gene Lists", "authors": [ { "first": "J", "middle": [], "last": "Crim", "suffix": "" }, { "first": "R", "middle": [], "last": "Mcdonald", "suffix": "" }, { "first": "F", "middle": [], "last": "Pereira", "suffix": "" } ], "year": 2005, "venue": "BMC Bioinformatics", "volume": "", "issue": "6", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Crim, J., McDonald, R., & Pereira, F. (2005). Automatically Annotating Documents with Normalized Gene Lists. BMC Bioinformatics, 6(Suppl 1), S13.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Large-scale named entity disambiguation based on Wikipedia data", "authors": [ { "first": "S", "middle": [], "last": "Cucerzan", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cucerzan, S. (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, Prague, Czech Republic.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Integration of gene normalization stages and co-reference resolution using a Markov logic network", "authors": [ { "first": "H.-J", "middle": [], "last": "Dai", "suffix": "" }, { "first": "Y.-C", "middle": [], "last": "Chang", "suffix": "" }, { "first": "R", "middle": [ "T" ], "last": "Tsai", "suffix": "" }, { "first": ".-H", "middle": [], "last": "Hsu", "suffix": "" }, { "first": "W.-L", "middle": [], "last": "", "suffix": "" } ], "year": 2011, "venue": "Bioinformatics", "volume": "27", "issue": "18", "pages": "2586--2594", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dai, H.-J., Chang, Y.-C., Tsai, R. T.-H., & Hsu, W.-L. (2011). Integration of gene normalization stages and co-reference resolution using a Markov logic network. Bioinformatics, 27(18), 2586-2594.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Multistage Gene Normalization and SVM-Based Ranking for Protein Interactor Extraction in Full-Text Articles", "authors": [ { "first": "H.-J", "middle": [], "last": "Dai", "suffix": "" }, { "first": "P.-T", "middle": [], "last": "Lai", "suffix": "" }, { "first": "R", "middle": [ "T" ], "last": "Tsai", "suffix": "" }, { "first": ".-H", "middle": [], "last": "", "suffix": "" } ], "year": 2010, "venue": "IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS", "volume": "7", "issue": "3", "pages": "412--420", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dai, H.-J., Lai, P.-T., & Tsai, R. T.-H. (2010). Multistage Gene Normalization and SVM-Based Ranking for Protein Interactor Extraction in Full-Text Articles. IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, 7(3), 412-420.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "From Entity Recognition to Entity Linking: A Survey of Advanced Entity Linking Techniques", "authors": [ { "first": "H.-J", "middle": [], "last": "Dai", "suffix": "" }, { "first": "C.-Y", "middle": [], "last": "Wu", "suffix": "" }, { "first": "R", "middle": [ "T" ], "last": "Tsai", "suffix": "" }, { "first": ".-H", "middle": [], "last": "Hsu", "suffix": "" }, { "first": "W.-L", "middle": [], "last": "", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 26th Annual Conference of the", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dai, H.-J., Wu, C.-Y., Tsai, R. T.-H., & Hsu, W.-L. (2012). From Entity Recognition to Entity Linking: A Survey of Advanced Entity Linking Techniques. In Proceedings of the 26th Annual Conference of the Japanese Society for Artificial Intelligence, Yamaguchi, Japan.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Entity Disambiguation for Knowledge Base Population", "authors": [ { "first": "M", "middle": [], "last": "Dredze", "suffix": "" }, { "first": "P", "middle": [], "last": "Mcnamee", "suffix": "" }, { "first": "D", "middle": [], "last": "Rao", "suffix": "" }, { "first": "A", "middle": [], "last": "Gerber", "suffix": "" }, { "first": "T", "middle": [], "last": "Finin", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dredze, M., McNamee, P., Rao, D., Gerber, A., & Finin, T. (2010). Entity Disambiguation for Knowledge Base Population. In Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010), Beijing.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Exploring the boundaries: gene and protein identification in biomedical text", "authors": [ { "first": "J", "middle": [], "last": "Finkel", "suffix": "" }, { "first": "S", "middle": [], "last": "Dingare", "suffix": "" }, { "first": "C", "middle": [], "last": "Manning", "suffix": "" }, { "first": "M", "middle": [], "last": "Nissim", "suffix": "" }, { "first": "B", "middle": [], "last": "Alex", "suffix": "" }, { "first": "C", "middle": [], "last": "Grover", "suffix": "" } ], "year": 2005, "venue": "BMC Bioinformatics", "volume": "", "issue": "6", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Finkel, J., Dingare, S., Manning, C., Nissim, M., Alex, B., & Grover, C. (2005). Exploring the boundaries: gene and protein identification in biomedical text. BMC Bioinformatics, 6(Suppl 1), S5.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Joint parsing and named entity recognition", "authors": [ { "first": "J", "middle": [ "R" ], "last": "Finkel", "suffix": "" }, { "first": "C", "middle": [ "D" ], "last": "Manning", "suffix": "" } ], "year": 2009, "venue": "Proceedings of NAACL", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Finkel, J. R., & Manning, C. D. (2009, June). Joint parsing and named entity recognition. In Proceedings of NAACL 2009.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Protein names and how to find them", "authors": [ { "first": "K", "middle": [], "last": "Franz\u00e9n", "suffix": "" }, { "first": "G", "middle": [], "last": "Eriksson", "suffix": "" }, { "first": "F", "middle": [], "last": "Olsson", "suffix": "" }, { "first": "L", "middle": [], "last": "Asker", "suffix": "" }, { "first": "P", "middle": [], "last": "Lid\u00e9n", "suffix": "" }, { "first": "J", "middle": [], "last": "C\u00f6ster", "suffix": "" } ], "year": 2002, "venue": "International Journal of Medical Informatics", "volume": "67", "issue": "1-3", "pages": "49--61", "other_ids": { "DOI": [ ":10.1016/s1386-5056(02)00052-7" ] }, "num": null, "urls": [], "raw_text": "Franz\u00e9n, K., Eriksson, G., Olsson, F., Asker, L., Lid\u00e9n, P., & C\u00f6ster, J. (2002). Protein names and how to find them. International Journal of Medical Informatics, 67(1-3), 49-61. doi: Doi: 10.1016/s1386-5056(02)00052-7", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Centering: A framework for modeling the local coherence of discourse", "authors": [ { "first": "B", "middle": [ "J" ], "last": "Grosz", "suffix": "" }, { "first": "S", "middle": [], "last": "Weinstein", "suffix": "" }, { "first": "A", "middle": [ "K" ], "last": "Joshi", "suffix": "" } ], "year": 1995, "venue": "Computational Linguistics", "volume": "21", "issue": "2", "pages": "203--225", "other_ids": {}, "num": null, "urls": [], "raw_text": "Grosz, B. J., Weinstein, S., & Joshi, A. K. (1995). Centering: A framework for modeling the local coherence of discourse. Computational Linguistics, 21(2), 203-225.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Inter-species normalization of gene mentions with GNAT", "authors": [ { "first": "J", "middle": [], "last": "Hakenberg", "suffix": "" }, { "first": "C", "middle": [], "last": "Plake", "suffix": "" }, { "first": "R", "middle": [], "last": "Leaman", "suffix": "" }, { "first": "M", "middle": [], "last": "Schroeder", "suffix": "" }, { "first": "G", "middle": [], "last": "Gonzalez", "suffix": "" } ], "year": 2008, "venue": "Bioinformatics", "volume": "24", "issue": "16", "pages": "126--132", "other_ids": { "DOI": [ "10.1093/bioinformatics/btn299" ] }, "num": null, "urls": [], "raw_text": "Hakenberg, J., Plake, C., Leaman, R., Schroeder, M., & Gonzalez, G. (2008). Inter-species normalization of gene mentions with GNAT. Bioinformatics, 24(16), 126-132. doi: 10.1093/bioinformatics/btn299", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Inter-species normalization of gene mentions with GNAT", "authors": [ { "first": "J", "middle": [], "last": "Hakenberg", "suffix": "" }, { "first": "C", "middle": [], "last": "Plake", "suffix": "" }, { "first": "R", "middle": [], "last": "Leaman", "suffix": "" }, { "first": "M", "middle": [], "last": "Schroeder", "suffix": "" }, { "first": "G", "middle": [], "last": "Gonzalez", "suffix": "" } ], "year": 2008, "venue": "Bioinformatics", "volume": "24", "issue": "", "pages": "126--132", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hakenberg, J., Plake, C., Leaman, R., Schroeder, M., & Gonzalez, G. (2008). Inter-species normalization of gene mentions with GNAT. Bioinformatics, 24, i126 -132.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Me and my friends: gene mention normalization with background knowledge", "authors": [ { "first": "J", "middle": [ "O" ], "last": "Hakenberg", "suffix": "" }, { "first": "L", "middle": [], "last": "Royer", "suffix": "" }, { "first": "C", "middle": [], "last": "Plake", "suffix": "" }, { "first": "H", "middle": [], "last": "Strobelt", "suffix": "" }, { "first": "M", "middle": [], "last": "Schroeder", "suffix": "" } ], "year": 2007, "venue": "Proceedings of Second BioCreAtIvE Challenge Evaluation Workshop", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hakenberg, J. o., Royer, L., Plake, C., Strobelt, H., & Schroeder, M. (2007). Me and my friends: gene mention normalization with background knowledge. In Proceedings of Second BioCreAtIvE Challenge Evaluation Workshop.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Overview of BioCreAtIvE: critical assessment of information extraction for biology", "authors": [ { "first": "L", "middle": [], "last": "Hirschman", "suffix": "" }, { "first": "A", "middle": [], "last": "Yeh", "suffix": "" }, { "first": "C", "middle": [], "last": "Blaschke", "suffix": "" }, { "first": "A", "middle": [], "last": "Valencia", "suffix": "" } ], "year": 2005, "venue": "BMC Bioinformatics", "volume": "6", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hirschman, L., Yeh, A., Blaschke, C., & Valencia, A. (2005). Overview of BioCreAtIvE: critical assessment of information extraction for biology. BMC Bioinformatics, 6(1).", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Comparing and combining chunkers of biomedical text", "authors": [ { "first": "N", "middle": [], "last": "Kang", "suffix": "" }, { "first": "E", "middle": [ "M" ], "last": "Van Mulligen", "suffix": "" }, { "first": "J", "middle": [ "A" ], "last": "Kors", "suffix": "" } ], "year": 2011, "venue": "J Biomed Inform", "volume": "44", "issue": "2", "pages": "354--360", "other_ids": { "DOI": [ "10.1016/j.jbi.2010.10.005" ] }, "num": null, "urls": [], "raw_text": "Kang, N., van Mulligen, E. M., & Kors, J. A. (2011). Comparing and combining chunkers of biomedical text. J Biomed Inform, 44(2), 354-360. doi: 10.1016/j.jbi.2010.10.005", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Overview of BioNLP'09 Shared Task on Event Extraction", "authors": [ { "first": "J.-D", "middle": [], "last": "Kim", "suffix": "" }, { "first": "T", "middle": [], "last": "Ohta", "suffix": "" }, { "first": "S", "middle": [], "last": "Pyysalo", "suffix": "" }, { "first": "Y", "middle": [], "last": "Kano", "suffix": "" }, { "first": "J", "middle": [ "I" ], "last": "Tsujii", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Workshop on Current Trends in Biomedical Natural Language Processing: Shared Task", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kim, J.-D., Ohta, T., Pyysalo, S., Kano, Y., & Tsujii, J. i. (2009). Overview of BioNLP'09 Shared Task on Event Extraction. In Proceedings of the Workshop on Current Trends in Biomedical Natural Language Processing: Shared Task, Boulder, Colorado.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Term identification in the biomedical literature", "authors": [ { "first": "M", "middle": [], "last": "Krauthammer", "suffix": "" }, { "first": "G", "middle": [], "last": "Nenadic", "suffix": "" } ], "year": 2004, "venue": "Journal of Biomedical Informatics", "volume": "37", "issue": "6", "pages": "512--526", "other_ids": {}, "num": null, "urls": [], "raw_text": "Krauthammer, M., & Nenadic, G. (2004). Term identification in the biomedical literature. Journal of Biomedical Informatics, 37(6), 512-526.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data", "authors": [ { "first": "J", "middle": [], "last": "Lafferty", "suffix": "" }, { "first": "A", "middle": [], "last": "Mccallum", "suffix": "" }, { "first": "F", "middle": [], "last": "Pereira", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the 18th International Conference on Machine Learning (ICML)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lafferty, J., McCallum, A., & Pereira, F. (2001). Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the 18th International Conference on Machine Learning (ICML).", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Using Contextual Information to Clarify Gene Normalization Ambiguity", "authors": [ { "first": "P.-T", "middle": [], "last": "Lai", "suffix": "" }, { "first": "Y.-Y", "middle": [], "last": "Bow", "suffix": "" }, { "first": "C.-H", "middle": [], "last": "Huang", "suffix": "" }, { "first": "H.-J", "middle": [], "last": "Dai", "suffix": "" }, { "first": "R", "middle": [ "T" ], "last": "Tsai", "suffix": "" }, { "first": ".-H", "middle": [], "last": "Hsu", "suffix": "" }, { "first": "W.-L", "middle": [], "last": "", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the IEEE International Conference on Information Reuse and Integration", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lai, P.-T., Bow, Y.-Y., Huang, C.-H., Dai, H.-J., Tsai, R. T.-H., & Hsu, W.-L. (2009). Using Contextual Information to Clarify Gene Normalization Ambiguity. In Proceedings of the IEEE International Conference on Information Reuse and Integration (IEEE IRI 2009), Las Vegas, USA.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Incorporating rich background knowledge for gene named entity classification and recognition", "authors": [ { "first": "Y", "middle": [], "last": "Li", "suffix": "" }, { "first": "H", "middle": [], "last": "Lin", "suffix": "" }, { "first": "Z", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2009, "venue": "BMC Bioinformatics", "volume": "10", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Li, Y., Lin, H., & Yang, Z. (2009). Incorporating rich background knowledge for gene named entity classification and recognition. BMC Bioinformatics, 10(1), 223.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Joint Inference of Named Entity Recognition and Normalization for Tweets", "authors": [ { "first": "X", "middle": [], "last": "Liu", "suffix": "" }, { "first": "M", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "F", "middle": [], "last": "Wei", "suffix": "" }, { "first": "Z", "middle": [], "last": "Fu", "suffix": "" }, { "first": "X", "middle": [], "last": "Zhou", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Liu, X., Zhou, M., Wei, F., Fu, Z., & Zhou, X. (2012). Joint Inference of Named Entity Recognition and Normalization for Tweets. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, Jeju, Republic of Korea.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Online Large-Margin Training of Dependency Parsers", "authors": [ { "first": "R", "middle": [], "last": "Mcdonald", "suffix": "" }, { "first": "K", "middle": [], "last": "Crammer", "suffix": "" }, { "first": "F", "middle": [], "last": "Pereira", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "McDonald, R., Crammer, K., & Pereira, F. (2005). Online Large-Margin Training of Dependency Parsers. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, Ann Arbor, Michigan.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Overview of the TAC 2009 Knowledge Base Population Track", "authors": [ { "first": "P", "middle": [], "last": "Mcnamee", "suffix": "" }, { "first": "H", "middle": [ "T" ], "last": "Dang", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Second Text Analysis Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "McNamee, P., & Dang, H. T. (2009a). Overview of the TAC 2009 Knowledge Base Population Track. In Proceedings of the Second Text Analysis Conference (TAC 2009), Gaithersburg, Maryland USA.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Overview of the TAC 2009 Knowledge Base Population Track", "authors": [ { "first": "P", "middle": [], "last": "Mcnamee", "suffix": "" }, { "first": "H", "middle": [ "T" ], "last": "Dang", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Second Text Analysis Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "McNamee, P., & Dang, H. T. (2009b). Overview of the TAC 2009 Knowledge Base Population Track. In Proceedings of the Second Text Analysis Conference (TAC 2009), Gaithersburg, Maryland USA.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "An Evaluation of Technologies for Knowledge Base Population", "authors": [ { "first": "P", "middle": [], "last": "Mcnamee", "suffix": "" }, { "first": "H", "middle": [ "T" ], "last": "Dang", "suffix": "" }, { "first": "H", "middle": [], "last": "Simpson", "suffix": "" }, { "first": "P", "middle": [], "last": "Schone", "suffix": "" }, { "first": "S", "middle": [ "M" ], "last": "Strassel", "suffix": "" } ], "year": 2009, "venue": "Proceedings of Test Analysis Conference", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "McNamee, P., Dang, H. T., Simpson, H., Schone, P., & Strassel, S. M. (2009). An Evaluation of Technologies for Knowledge Base Population. In Proceedings of Test Analysis Conference 2009 (TAC 09), Gaithersburg, Maryland USA.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Wikify!: linking documents to encyclopedic knowledge", "authors": [ { "first": "R", "middle": [], "last": "Mihalcea", "suffix": "" }, { "first": "A", "middle": [], "last": "Csomai", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the sixteenth ACM conference on Conference on information and knowledge management", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mihalcea, R., & Csomai, A. (2007). Wikify!: linking documents to encyclopedic knowledge. In Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, Lisbon, Portugal.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Learning to link with wikipedia", "authors": [ { "first": "D", "middle": [], "last": "Milne", "suffix": "" }, { "first": "I", "middle": [ "H" ], "last": "Witten", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 17th ACM conference on Information and knowledge management", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Milne, D., & Witten, I. H. (2008). Learning to link with wikipedia. In Proceedings of the 17th ACM conference on Information and knowledge management, Napa Valley, California, USA.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "Overview of BioCreative II gene normalization", "authors": [ { "first": "L", "middle": [], "last": "Hirschman", "suffix": "" } ], "year": 2008, "venue": "Genome Biology", "volume": "9", "issue": "S3", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hirschman, L. (2008). Overview of BioCreative II gene normalization. Genome Biology, 9(Suppl 2), S3.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "Joint inference in information extraction", "authors": [ { "first": "H", "middle": [], "last": "Poon", "suffix": "" }, { "first": "P", "middle": [], "last": "Domingos", "suffix": "" } ], "year": 2007, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Poon, H., & Domingos, P. (2007). Joint inference in information extraction.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Markov logic networks", "authors": [ { "first": "M", "middle": [], "last": "Richardson", "suffix": "" }, { "first": "P", "middle": [], "last": "Domingos", "suffix": "" } ], "year": 2006, "venue": "Special Issue: Multi-Relational Data Mining and Statistical Relational Learning", "volume": "62", "issue": "", "pages": "107--136", "other_ids": {}, "num": null, "urls": [], "raw_text": "Richardson, M., & Domingos, P. (2006). Markov logic networks. Machine Learning, 62(Special Issue: Multi-Relational Data Mining and Statistical Relational Learning), 107-136.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "Improving the Accuracy and Efficiency of MAP Inference for Markov Logic", "authors": [ { "first": "S", "middle": [], "last": "Riedel", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 24th Conference on Uncertainty in Artificial Intelligence (UAI 2008)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Riedel, S. (2008). Improving the Accuracy and Efficiency of MAP Inference for Markov Logic. In Proceedings of the 24th Conference on Uncertainty in Artificial Intelligence (UAI 2008), Helsinki, Finland.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Collective classification in network data", "authors": [ { "first": "P", "middle": [], "last": "Sen", "suffix": "" }, { "first": "G", "middle": [], "last": "Namata", "suffix": "" }, { "first": "M", "middle": [], "last": "Bilgic", "suffix": "" }, { "first": "L", "middle": [], "last": "Getoor", "suffix": "" }, { "first": "B", "middle": [], "last": "Galligher", "suffix": "" }, { "first": "T", "middle": [], "last": "Eliassi-Rad", "suffix": "" } ], "year": 2008, "venue": "AI Magazine", "volume": "29", "issue": "3", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., & Eliassi-Rad, T. (2008). Collective classification in network data. AI Magazine, 29(3), 93.", "links": null }, "BIBREF36": { "ref_id": "b36", "title": "Various criteria in the evaluation of biomedical named entity recognition", "authors": [ { "first": "W.-L", "middle": [], "last": "Hsu", "suffix": "" } ], "year": 2006, "venue": "BMC Bioinformatics", "volume": "7", "issue": "92", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hsu, W.-L. (2006). Various criteria in the evaluation of biomedical named entity recognition. BMC Bioinformatics, 7(92), 14.", "links": null }, "BIBREF37": { "ref_id": "b37", "title": "Entity Linking Leveraging Automatically Generated Annotation", "authors": [ { "first": "W", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "J", "middle": [], "last": "Su", "suffix": "" }, { "first": "C", "middle": [ "L" ], "last": "Tan", "suffix": "" }, { "first": "W", "middle": [ "T" ], "last": "Wang", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhang, W., Su, J., Tan, C. L., & Wang, W. T. (2010). Entity Linking Leveraging Automatically Generated Annotation. In Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010), Beijing, China.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "text": "Structure of the CD59-encoding gene: further evidence of a relationship to murine lymphocyte antigen Ly-6 protein ABSTRACT: The gene for CD59 [membrane inhibitor of reactive lysis (MIRL), protectin], a phosphatidylinositol-linked surface glycoprotein that regulates the formation of the polymeric C9 complex of complement and that is deficient on the abnormal hematopoietic cells of patients with paroxysmal nocturnal hemoglobinuria, consists of four exons spanning 20 kilobases. \u2026 PMID [1381503]", "uris": null, "num": null }, "FIGREF1": { "type_str": "figure", "text": "An example of entity linking.", "uris": null, "num": null }, "FIGREF2": { "type_str": "figure", "text": "MLN hidden predicates defined for each stage.", "uris": null, "num": null }, "FIGREF3": { "type_str": "figure", "text": "", "uris": null, "num": null }, "FIGREF4": { "type_str": "figure", "text": "(a) Individual EL formulation. (b) Collective EL formulation.", "uris": null, "num": null }, "FIGREF5": { "type_str": "figure", "text": "Ground Markov network obtained by applying example formulae tothe constants x, y = {1, 2}, c = {0, 1}, and id = {966}.", "uris": null, "num": null }, "TABREF2": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
! . id hasPPIPartnerRank x id ( , ,1) \uf024\uf0d9 \uf024. w hasWord w isPPIKeyword w ( ) ( ) \uf0d9\uf0de( , ) LinkTo x id
" }, "TABREF3": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
Candidate(x, id)
hasChromosomeInfo(id)
hasWord(w): the abstract contain a word w.
isPPIKeyword(w), isPPIPartner(id 1 , id 2 )
" }, "TABREF6": { "text": "", "type_str": "table", "html": null, "num": null, "content": "" }, "TABREF7": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
Dataset# of queries# of Nil# of Entities
Training1073871132
Test1266661154
" }, "TABREF8": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
Config.P (%)R (%)F (%) Diff (F)
No Disambiguation81.048.060.3-
Salience Collective79.949.661.3+1.0
PPI Collective79.351.262.2+1.9
Rule-based71.754.061.6+1.3
ME79.848.960.5+0.5
CM73.555.963.5+3.2
Rule-based stage71.754.061.6+1.3
ME stage86.446.960.8+0.8
CM stage73.555.963.5+3.2
JCM79.954.965.1+4.8
" }, "TABREF9": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
MetricsFine-grained Resolution (%)Aids for Curation (%)
Config.PRFDiffPRFDiff
No Disambiguation 80.7 56.3 66.3077.3 71.5 74.30
Salience Collective 79.5 59.0 67.7 +1.477.2 71.3 74.1-0.2
Rule-based72.9 63.9 68.1 +1.882.6 83.4 83.0+8.7
ME79.2 58.2 67.1 +0.888.8 79.0 83.6+9.3
CM73.8 64.3 68.7 +2.486.1 83.0 84.5 +10.2
Rule-based stage73.7 64.2 68.7 +2.484.1 83.7 83.9+9.6
ME stage80.2 58.4 67.6 +1.390.2 79.0 84.3 +10
CM stage74.3 64.3 69.0 +2.787.9 83.2 85.5 +11.2
JCM77.5 63.7 69.9 +3.687.7 83.8 85.7 +11.4
" }, "TABREF10": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
Config.Accuracy
No Disambiguation65.7
Salience Collective67.2
Rule-based72.4
ME66.8
CM73.1
Rule-based stage73.0
ME stage67.0
CM stage73.3
JCM73.5
" }, "TABREF11": { "text": "", "type_str": "table", "html": null, "num": null, "content": "
<query id, name, docid>
<#1, AIP1, 9647693>
<#2, TR1, 10455115>
<#3, PAGE1, 9651357>
<#4, UGT2B11, 8333863>
" } } } }