{
"paper_id": "W11-0106",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T05:45:10.054075Z"
},
"title": "A Model for Composing Semantic Relations",
"authors": [
{
"first": "Eduardo",
"middle": [],
"last": "Blanco",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The University of Texas at Dallas",
"location": {}
},
"email": "eduardo@hlt.utdallas.edu"
},
{
"first": "Dan",
"middle": [],
"last": "Moldovan",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "The University of Texas at Dallas",
"location": {}
},
"email": "moldovan@hlt.utdallas.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper presents a model to compose semantic relations. The model is independent of any particular set of relations and uses an extended definition for semantic relations. This extended definition includes restrictions on the domain and range of relations and utilizes semantic primitives to characterize them. Primitives capture elementary properties between the arguments of a relation. An algebra for composing semantic primitives is used to automatically identify the resulting relation of composing a pair of compatible relations. Inference axioms are obtained. Axioms take as input a pair of semantic relations and output a new, previously ignored relation. The usefulness of this proposed model is shown using PropBank relations. Eight inference axioms are obtained and their accuracy and productivity are evaluated. The model offers an unsupervised way of accurately extracting additional semantics from text.",
"pdf_parse": {
"paper_id": "W11-0106",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper presents a model to compose semantic relations. The model is independent of any particular set of relations and uses an extended definition for semantic relations. This extended definition includes restrictions on the domain and range of relations and utilizes semantic primitives to characterize them. Primitives capture elementary properties between the arguments of a relation. An algebra for composing semantic primitives is used to automatically identify the resulting relation of composing a pair of compatible relations. Inference axioms are obtained. Axioms take as input a pair of semantic relations and output a new, previously ignored relation. The usefulness of this proposed model is shown using PropBank relations. Eight inference axioms are obtained and their accuracy and productivity are evaluated. The model offers an unsupervised way of accurately extracting additional semantics from text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Semantic representation of text is an important step toward text understanding, performing inferences and reasoning. Potentially, it could dramatically improve the performance of several Natural Language Processing applications.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Semantic relations have been studied in linguistics for decades. They are unidirectional underlying connections between concepts. For example, the sentence The construction slowed down the traffic encodes a CAUSE and detecting it would help answer the question Why is traffic slower?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In Computational Linguistics, there have been several proposals to detect semantic relations. Current approaches focus on a particular set of relations and given a text they output relations. There have been competitions aiming at detecting semantic roles (i.e., relations between a verb and its arguments) (Carreras and M\u00e0rquez, 2005) , and between nominals (Girju et al., 2007; Hendrickx et al., 2009) .",
"cite_spans": [
{
"start": 307,
"end": 335,
"text": "(Carreras and M\u00e0rquez, 2005)",
"ref_id": "BIBREF1"
},
{
"start": 359,
"end": 379,
"text": "(Girju et al., 2007;",
"ref_id": "BIBREF6"
},
{
"start": 380,
"end": 403,
"text": "Hendrickx et al., 2009)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we propose a model to compose semantic relations to extract previously ignored relations. The model allows us to automatically obtain inference axioms given a set of relations and is not coupled to any particular set. Axioms take as their input semantic relations and yield a new semantic relation as their conclusion.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Consider the sentence John went to the shop to buy flowers. Figure 1 shows semantic role annotation with solid arrows. By composing this basic annotation with inference axioms, one can obtain the relations shown with discontinuous arrows: John had the intention to buy, the buying event took place at the shop and John and the flowers were at some point in the shop. ",
"cite_spans": [],
"ref_spans": [
{
"start": 60,
"end": 68,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Semantic relations are the underlying relations between concepts expressed by words or phrases. In other words, semantic relations are implicit associations between concepts in text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "In general, a semantic relation is defined by stating the kind of connection linking two concepts. For example, Hendrickx et al. (2009) loosely define ENTITY-ORIGIN as an entity is coming or is derived from an origin (e.g., position or material) and give one example: Earth is located in the Milky Way. We find this kind of definition weak and prone to confusion.",
"cite_spans": [
{
"start": 112,
"end": 135,
"text": "Hendrickx et al. (2009)",
"ref_id": "BIBREF9"
},
{
"start": 217,
"end": 245,
"text": "(e.g., position or material)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "Following Helbig (2005) , we propose an extended definition for semantic relations, including semantic restrictions for its domain and range. For example, DOMAIN(AGENT) must be an animate concrete object and RANGE(AGENT) must be a situation.",
"cite_spans": [
{
"start": 10,
"end": 23,
"text": "Helbig (2005)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "Moreover, we propose to characterize relations by semantic primitives. Primitives indicate if a certain property holds between the arguments of a relation. For example, the primitive temporal indicates if the first argument must happen before the second in order for the relation to hold. This primitive holds for CAUSE (a cause must precede its effect) and it does not apply to PART-WHOLE since the later relation does not consider time.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "Besides having a better understanding of each relation, this extended definition allows us to create a model that automatically obtains inference axioms for composing semantic relations. The model detects possible combinations of relations and identifies the conclusion of composing them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "Formally, we represent a relation R as R(x, y), where R is the relation type and x and y are the first and second argument respectively. R(x, y) should be read x is R of Y. DOMAIN(R) and RANGE(R) are the sorts of concepts that can be part of the first and second argument respectively. Any ontology can be used to define domains and ranges, e.g., Helbig (2005) defined one to define a set of 89 relations. Primitives are represented by an array P R of length n, where n is the number of primitives and P i R indicates the value R takes for the ith primitive.",
"cite_spans": [
{
"start": 347,
"end": 360,
"text": "Helbig (2005)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "The inverse of R is denoted R \u22121 and can be obtained by simply switching the arguments of R. Given R(x, y), R \u22121 (y, x) always holds. We can easily define R \u22121 given the definition for R: DOMAIN(R \u22121 ) = RANGE(R), RANGE(R \u22121 ) = DOMAIN(R), and P R \u22121 is defined according to the fourth column of Table 1 for each primitive, i.e., \u2200i \u2208 [1, n] :",
"cite_spans": [],
"ref_spans": [
{
"start": 296,
"end": 304,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "P i R \u22121 = Inverse(P i R ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Relations",
"sec_num": "2"
},
{
"text": "Relation primitives capture deep characteristics of relations. Huhns and Stephens (1989) define them as:",
"cite_spans": [
{
"start": 63,
"end": 88,
"text": "Huhns and Stephens (1989)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "They [primitives] are independently determinable for each relation and relatively self-explanatory. They specify a relationship between an element of the domain and an element of the range of the semantic relation being described.",
"cite_spans": [
{
"start": 5,
"end": 17,
"text": "[primitives]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "Relation primitives are fundamental properties that cannot be explained using other primitives; they are elemental. They specify basic attributes of a relation by stating if a particular property must hold by definition between the domain and range. Each relation takes a value for each primitive from the set V = {+, \u2212, 0}, where '+' indicates that the property holds, '\u2212' that it does not hold and '0' that it does not apply. For example, the primitive volitional indicates if a relation requires volition between domain and range. AGENT takes as value + for this primitive and PART-WHOLE takes 0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "Primitives complement the definition of a relation by stating if a particular property holds between its arguments. They help to understand the inter-relation differences and clustering relations. Primitives can be used as conditions to be fulfilled in order to determine if a potential relation holds. They are general enough to be determined for a relation, not a particular instantiation. In other words, they state properties that hold for all instances of a relation by definition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "Our set of primitives (Table 1) is inspired on previous work in Knowledge Bases (Huhns and Stephens, 1989) . We only select from them useful primitives for our purpose and add more primitives. The additional primitives are justified by the fact that we aim at combining relations capturing semantics ",
"cite_spans": [
{
"start": 80,
"end": 106,
"text": "(Huhns and Stephens, 1989)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 22,
"end": 31,
"text": "(Table 1)",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "R 2 R 1 -0 + - \u00d7 0 \u00d7 0 0 0 0 + \u00d7 0 + 2: Functional R 2 R 1 -0 + - -0 + 0 0 0 0 + + 0 + 3: Separable R 2 R 1 -0 + - --- 0 -0 + + -+ + 4: Temporal R 2 R 1 -0 + - - -\u00d7 0 -0 + + \u00d7 + + 5: Connected R 2 R 1 -0 + - --+ 0 -0 + + + + + 6: Intrinsic R 2 R 1 -0 + - -0 - 0 0 0 0 + -0 + 7: Volitional R 2 R 1 -0 + - -0 + 0 0 0 0 + + 0 + 8: F Impl. R 2 R 1 -0 + - -0 - 0 0 0 0 + -0 + 9: W Impl. R 2 R 1 -0 + - -0 - 0 0 0 0 + -0 +",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "P i R1 \u2022 P i R2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "from natural language. Whatever the set of chosen relations, it will describe the characteristics of events (who/when/where/how something happened), which elements were involved, connections between events (e.g. CAUSE, CORRELATION). Time (whether an argument is guaranteed to happen before than the other), space and volition (whether or not there must be volition between the arguments) also play an important role. The fourth column in Table 1 indicates the value of the primitive for the inverse relation. Same means the inverse relation takes the same value, opposite means it takes the opposite. The opposite of \u2212 is +, the opposite of \u2212 is +, and the opposite of 0 is 0.",
"cite_spans": [],
"ref_spans": [
{
"start": 438,
"end": 445,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "For example, P AGENT = {+, +, +, 0, \u2212, \u2212, +, 0, 0}, indicating that P 5 AGENT = \u2212 and P 7 AGENT = +, i.e., AGENT(x, y) does not require x and y to be connected and it requires volition between the arguments. Note that P AGENT \u22121 = P AGENT .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Primitives",
"sec_num": "2.1"
},
{
"text": "The key to automatically obtaining inference axioms is the ability to know beforehand the result of composing semantic primitives using an algebra. This way, one can identify prohibited combinations of relations and determine conclusions for the composition of valid combinations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "Given P i R 1 and P i R 2 , i.e., the values of R 1 and R 2 for a primitive p i , we define an algebra that indicates the result of composing them (i.e., P i",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "R 1 \u2022 P i R 2 )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": ". Composing two primitives can yield three values: +, \u2212 or 0, indicating if the primitive holds, does not hold or does not apply to the composition of R 1 and R 2 . Additionally, the composition can be prohibited, indicated with \u00d7. After composing all the primitives for R 1 and R 2 , we obtain the primitives values for the composition of R 1 and R 2 (i.e., P",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "R 1 \u2022 P R 2 ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "We define the values for the composition using a table for each primitive. Table 2 depicts the whole algebra. The ith table indicates the rules for composing the ith primitive. For example, regarding the intrinsic primitive, we have the following rules:",
"cite_spans": [],
"ref_spans": [
{
"start": 75,
"end": 82,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "\u2022 If both relations are intrinsic, the composition is intrinsic;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "\u2022 else if intrinsic does not apply to either relation, the primitive does not apply to the composition;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "\u2022 else, the composition is not intrinsic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "Other rules stated by the algebra are: (1) two relations shall not compose if they have different opposite values for the primitive temporal;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "(2) the composition of R 1 and R 2 is not separable if either relation is not separable; and (3) if either R 1 or R 2 are connected, then the composition is connected.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algebra for Composing Semantic Relations",
"sec_num": "2.2"
},
{
"text": "In principle, one could define axioms for every single possible combination of relations. However, there are two necessary conditions in order to compose R 1 and R 2 :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Necessary Conditions for Composing Semantic Relations",
"sec_num": "3"
},
{
"text": "1. They have to be compatible. A pair of relations is compatible if it is possible, from a theoretical point of view, to compose them. Formally,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Necessary Conditions for Composing Semantic Relations",
"sec_num": "3"
},
{
"text": "R 1 and R 2 are compatible iff RANGE(R 1 ) \u2229 DOMAIN(R 2 ) = \u2205. 2. A third relation R 3 must fit as conclusion, that is, \u2203R 3 such that DOMAIN(R 3 ) \u2229 DOMAIN(R 1 ) = \u2205 and RANGE(R 3 ) \u2229 RANGE(R 2 ) = \u2205.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Necessary Conditions for Composing Semantic Relations",
"sec_num": "3"
},
{
"text": "Furthermore, P R 3 must be compatible with the result of composing P R 1 and P R 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Necessary Conditions for Composing Semantic Relations",
"sec_num": "3"
},
{
"text": "It is important to note that domain and range compatibility is not enough to compose two relations. For example, given KINSHIP(Mary, John) and AT-LOCATION(John, Dallas), no relation can be inferred between Mary and Dallas.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Necessary Conditions for Composing Semantic Relations",
"sec_num": "3"
},
{
"text": "An axiom is defined as a set of relations called premises and a conclusion. The composition operator \u2022 is the basic way of combining two relations to form an axiom. We denote an inference axiom as R 1 (x, y)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": "\u2022 R 2 (y, z) \u2192 R 3 (x, z)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": ", where R 1 and R 2 are the premises and R 3 the conclusion. In order to instantiate an axiom the premises must have an argument in common, y.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": "In general, for n relations there are n 2 = n(n\u22121) 2 different pairs. For each pair, taking into account the two relations and their inverses, there are 4 \u00d7 4 = 16 different possible combinations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": "We note that",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": "R 1 \u2022 R 2 = (R 2 \u22121 \u2022 R 1 \u22121 ) \u22121",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": ", reducing the total number of different combinations to 10. Out of these 10, (1) 4 combine R 1 , R 2 and their inverses (Table 3) ; (2) 3 combine R 1 and its inverse; and (3) 3 combine R 2 and its inverse. The most interesting combinations to use as premises for an axiom fall into category (1), since the other two can be resolved by the transitivity property of a relation and its inverse. Therefore, for n relations there are 2n 2 +n potential axioms: n 2 \u00d74+3n = 2\u00d7n(n\u22121)+3n = 2n 2 \u2212 2n + 3n = 2n 2 + n.",
"cite_spans": [],
"ref_spans": [
{
"start": 121,
"end": 130,
"text": "(Table 3)",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Inference Axioms",
"sec_num": "4"
},
{
"text": "Given a set of relations R defined using the extended definition, one can automatically obtain inference axioms using the following steps for each pair of relations R 1 \u2208 R and R 2 \u2208 R, where R 1 = R 2 : Repeat Steps 1, 2 and 3 for (",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "R 1 \u2022 R 2 R 1 \u22121 \u2022 R 2 R 2 \u2022 R 1 R 2 \u2022 R 1 \u22121 x R 1 / / R",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "+ + + 0 - - + 0 0 ARG1 + - + 0 - - - 0 0 ARG1 \u22121 + - + 0 - - - 0 0 MLOC + + 0 0 + - 0 0 0 MLOC \u22121 + + 0 0 + - 0 0 0 MCAU + + + + - + 0 + + MCAU \u22121 + + + - - + 0 - - MTMP + + 0 0 + - 0 0 0 MTMP \u22121 + + 0 0 + - 0 0 0 MPNC + - + - - - - 0 - MPNC \u22121 + - + + - - - 0 + MMNR + - + 0 - - + 0 0 MMNR \u22121 + - + 0 - - + 0 0",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "R i , R j ) \u2208 [(R 1 , R 2 ), (R 1 \u22121 , R 2 ), (R 2 , R 1 ), (R 2 , R 1 \u22121 )]:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "1. Domain and range compatibility If",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "RANGE(R i ) \u2229 DOMAIN(R j ) = \u2205, break 2. Primitives composition",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "Using the algebra for composing semantic primitives, calculate",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "P R i \u2022 P R j 3. Conclusion match Repeat for R 3 \u2208 R If DOMAIN(R 3 ) \u2229 DOMAIN(R i ) = \u2205 and RANGE(R 3 ) \u2229 RANGE(R j ) = \u2205 and consistent(P R 3 , P R i \u2022 P R j ), then inf erence axioms += R i (x, y) \u2022 R j (y, z) \u2192 R 3 (x, z)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "The method consistent(P 1 , P 2 ) is a simple procedure that compares the values assigned to each primitive one by one. Two values for the same primitive are compatible unless they have different opposites or either value is '\u00d7' (i.e., prohibited).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An Algorithm for Obtaining Inference Axioms",
"sec_num": "4.1"
},
{
"text": "PropBank (Palmer et al., 2005 ) adds a layer of predicate-argument information, or semantic role labels, on top of the syntactic trees provided by the Penn TreeBank. Along with FrameNet, it is the resource most widely used for semantic role annotation.",
"cite_spans": [
{
"start": 9,
"end": 29,
"text": "(Palmer et al., 2005",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Case Study: PropBank",
"sec_num": "5"
},
{
"text": "PropBank uses a series of numeric core roles (ARG0 -ARG5) and a set of more general roles, ARGMs (e.g. MTMP, MLOC, MMNR). The interpretation of the numeric roles is determined by a verb-specific framesets, although ARG0 and ARG1 usually correspond to the prototypical AGENT and THEME. On the other hand, the meaning of AGRMs generalize across verbs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Case Study: PropBank",
"sec_num": "5"
},
{
"text": "An example of PropBank annotation is the following: [Winston] Palmer et al. (2005) discuss the creation of PropBank. For more information about the semantics of each role, we refer the reader to the annotation guidelines 1 .",
"cite_spans": [
{
"start": 52,
"end": 61,
"text": "[Winston]",
"ref_id": null
},
{
"start": 62,
"end": 82,
"text": "Palmer et al. (2005)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Case Study: PropBank",
"sec_num": "5"
},
{
"text": "Since ARG2, AGR3, ARG4 and ARG5 do not have a common meaning across verbs, they become not composable. For example, ARG2 is used for INSTRUMENT in the frameset kick.01 and for BENEFACTIVE in the frameset call.02. Table 5 : Results after applying the steps depicted in Section 4.1 using PropBank semantic roles. A letter indicates an inference axiom R 1 \u2022R 2 \u2192 R 3 by indicating the conclusion R 3 . '-' indicates that the combination is not prohibited but a relation compatible with P R1 \u2022 P R2 could not be found; '=' indicates that the cell corresponds to a relation and its inverse.",
"cite_spans": [],
"ref_spans": [
{
"start": 213,
"end": 220,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Case Study: PropBank",
"sec_num": "5"
},
{
"text": "The remaining labels (ARG0, ARG1 and all ARGMs) do generalize in meaning across verbs. Roles MEXT, MDIS, MADV, MNEG, MMOD, MDIR, are not composable because they encode a very narrow semantic connection. Manual examination of several examples leads to this conclusion. Table 4 depicts the primitives for the roles which are composable and their inverses. Note that for any two relations their primitives are different.",
"cite_spans": [],
"ref_spans": [
{
"start": 268,
"end": 275,
"text": "Table 4",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Case Study: PropBank",
"sec_num": "5"
},
{
"text": "PropBank does not provide domains and ranges for its roles, although we can specify our own. We do so by using the ontology defined by Helbig (2005) . All relations in PropBank are denoted as R(x, y), where x is an argument of y, and y is a verb. The range of all relations is a situation. The domain of AGR0 and ARG1 are objects, the domain of MLOC and MTMP local and temporal descriptors respectively, the domain of MMNR qualities or states, and the domain of MPNC and MCAU are situations.",
"cite_spans": [
{
"start": 135,
"end": 148,
"text": "Helbig (2005)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Case Study: PropBank",
"sec_num": "5"
},
{
"text": "Out of the four possible axioms between any pair of relations (Table 3) , the only way to compose two relations from PropBank is by using as common argument y a verb. This restriction is due to the fact that PropBank exclusively annotates relations between a verb and its arguments. Thus, the only possible axiom for any pair of roles",
"cite_spans": [],
"ref_spans": [
{
"start": 62,
"end": 71,
"text": "(Table 3)",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": "R 1 and R 2 is R 1 (x, y) \u2022 R 2 \u22121 (y, z) \u2192 R 3 (x, z)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": ", where y is a verb. Table 5 shows the eight inference axioms obtained after following the steps depicted in Section 4.1. Note that the matrix is symmetric as stated by the property",
"cite_spans": [],
"ref_spans": [
{
"start": 21,
"end": 28,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": "R 1 \u2022 R 2 = (R 2 \u22121 \u2022 R 1 \u22121 ) \u22121",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": ". Some of the axioms obtained are:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": "\u2022 MCAU \u2022 MLOC \u22121 \u2192 MLOC \u22121",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": ", the location of a cause is the same than the location of its effect.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": "\u2022 MPNC \u2022 ARG0 \u22121 \u2192 ARG0 \u22121 , the agent of an action is inherited by its purpose.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": "\u2022 MPNC \u2022 MMNR \u22121 \u2192 MMNR \u22121",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": ", the manner of an action is inherited by its purpose.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference Axioms from PropBank",
"sec_num": "5.1"
},
{
"text": "First, we evaluated all the instantiations of axiom MPNC \u2022 MMNR \u22121 \u2192 MMNR \u22121 . This axiom can be instantiated 237 times using PropBank annotation, yielding 189 new MANNER not present in PropBank. The overall accuracy is 0.797, superior to state-of-the art semantic role labelers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "5.2"
},
{
"text": "Second, we have evaluated the accuracy of the eight inference axioms (Table 5 ). Since PropBank is a large corpus, the amount of instantiations found for all axioms is too large to be checked by hand. We have manually evaluated the first 1,000 sentences that are an instantiation of any axiom. Since a sentence may instantiate several axioms, we have actually evaluated 1,412 instantiations. The first 1,000 sentences which are an instantiation of any axiom are found within the first 31,450 sentences in PropBank. Table 6 shows the number of roles PropBank annotates for these sentences. Table 7 : Axioms used during evaluation, number of instances, accuracy and productivity. Results are reported both using and not using the heuristic. Productivity refers to the number of relations added by the axiom in relative terms. Table 7 depicts the total number of instantiations for each axiom and its accuracy (columns 3 and 4). Accuracies range from 0.12 to 0.90, showing that the plausibility of an axiom depends on the axiom. The average accuracy for axioms involving MCAU is 0.53 and for axioms involving MPNC is 0.54.",
"cite_spans": [],
"ref_spans": [
{
"start": 69,
"end": 77,
"text": "(Table 5",
"ref_id": null
},
{
"start": 515,
"end": 523,
"text": "Table 6",
"ref_id": "TABREF7"
},
{
"start": 590,
"end": 597,
"text": "Table 7",
"ref_id": null
},
{
"start": 825,
"end": 832,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "5.2"
},
{
"text": "Axiom MCAU \u2022 ARG0 \u22121 \u2192 ARG0 \u22121 adds 201 relations, which corresponds to 0.89% in relative terms. Its accuracy is low, 0.40. Other axioms are less productive overall, but have a greater relative impact and accuracy. For example, axiom MPNC \u2022 MMNR \u22121 \u2192 MMNR \u22121 , only yields 71 new MMNR, and yet it is adding 3.21% in relative terms with an accuracy of 0.82.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "5.2"
},
{
"text": "It is worth noting that overall, applying the eight axioms used during evaluation adds 1,412 relations on top of the ones already present (2.26% in relative terms) with an accuracy of 0.54.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "5.2"
},
{
"text": "Because of the low accuracy of axioms 1 and 5, an error analysis was performed. We found that unlike other axioms, these axioms often yield a relation type that is already present in the semantic representation. Specifically, axioms 1 and 5 often yield R(x, z) when R(x', z) is already known.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Error Analysis",
"sec_num": "5.3"
},
{
"text": "An example can be found in Figure 4 , where axiom 5 yields ARG1(orders, to buy) when the relation ARG1(the basket, to buy) is already present. We use the following heuristic in order to improve the accuracy of axioms 1 and 5: do not instantiate an axiom",
"cite_spans": [],
"ref_spans": [
{
"start": 27,
"end": 35,
"text": "Figure 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Error Analysis",
"sec_num": "5.3"
},
{
"text": "R 1 (x, y) \u2022 R 2 (y, z) \u2192 R 3 (x, z) if a relation of the form R 3 (x', z) is already known.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Error Analysis",
"sec_num": "5.3"
},
{
"text": "This simple heuristic allows us to augment the accuracy of the inferences at the cost of lowering their productivity. The last three columns in Table 7 show results when using the heuristic. The eight axioms add 840 relations (1.35% in relative terms) with an accuracy of 0.86.",
"cite_spans": [],
"ref_spans": [
{
"start": 144,
"end": 151,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Error Analysis",
"sec_num": "5.3"
},
{
"text": "In this section we present several examples of instantiations. We provide the full text of each example, but only the relevant semantic annotation for instantiating axioms. For all examples, solid arrows indicate semantic role annotation from PropBank, and discontinuous arrows inferred relations. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Examples",
"sec_num": "5.4"
},
{
"text": "] \\ [ [ Z Y X W V U T S S came h h MLOC g g MTMP",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Examples",
"sec_num": "5.4"
},
{
"text": "bef ore the House . . . yesterday to talk about Figure 4 , indicated with *. Using the heuristic, all inferences are correct.",
"cite_spans": [],
"ref_spans": [
{
"start": 48,
"end": 56,
"text": "Figure 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Examples",
"sec_num": "5.4"
},
{
"text": "MPNC g g ' ' MLOC P R T V X [ ] a c f h j l ( ( MTMP U Z d i",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Examples",
"sec_num": "5.4"
},
{
"text": "There have been abundant proposals to detect semantic relations without taking into account composition of relations. All these approaches, regardless of their particular details, take as their input text and output the relations found in it. In contrast, the framework proposed in this article obtains axioms that take as their input relations found in text and output more relations previously ignored.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Previous Work",
"sec_num": "6"
},
{
"text": "Generally, efforts to extract semantic relations have concentrated on particular sets of relations or a single relation, e.g. CAUSE (Bethard and Martin, 2008; Chang and Choi, 2006) and PART-WHOLE (Girju et al., 2006) . Automatic detection of semantic roles has received a lot of attention lately (M\u00e0rquez et al., 2008; Carreras and M\u00e0rquez, 2005) . The SemEval-2007 Task 04 (Girju et al., 2007) and SemEval-2010 Task 08 (Hendrickx et al., 2009) aimed at relations between nominals. There has been work on detecting relations within noun phrases (Moldovan et al., 2004; Nulty, 2007) , clauses (Szpakowicz et al., 1995) and syntax-based comma resolution (Srikumar et al., 2008) .",
"cite_spans": [
{
"start": 132,
"end": 158,
"text": "(Bethard and Martin, 2008;",
"ref_id": "BIBREF0"
},
{
"start": 159,
"end": 180,
"text": "Chang and Choi, 2006)",
"ref_id": "BIBREF3"
},
{
"start": 196,
"end": 216,
"text": "(Girju et al., 2006)",
"ref_id": "BIBREF5"
},
{
"start": 296,
"end": 318,
"text": "(M\u00e0rquez et al., 2008;",
"ref_id": "BIBREF11"
},
{
"start": 319,
"end": 346,
"text": "Carreras and M\u00e0rquez, 2005)",
"ref_id": "BIBREF1"
},
{
"start": 374,
"end": 394,
"text": "(Girju et al., 2007)",
"ref_id": "BIBREF6"
},
{
"start": 420,
"end": 444,
"text": "(Hendrickx et al., 2009)",
"ref_id": "BIBREF9"
},
{
"start": 545,
"end": 568,
"text": "(Moldovan et al., 2004;",
"ref_id": "BIBREF12"
},
{
"start": 569,
"end": 581,
"text": "Nulty, 2007)",
"ref_id": "BIBREF13"
},
{
"start": 592,
"end": 617,
"text": "(Szpakowicz et al., 1995)",
"ref_id": "BIBREF16"
},
{
"start": 652,
"end": 675,
"text": "(Srikumar et al., 2008)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Previous Work",
"sec_num": "6"
},
{
"text": "Previous research has exploited the idea of using semantic primitives to define and classify semantic relations under different names. Among others, the literature uses relation elements, deep structure, aspects and primitives. To the best of our knowledge, the first effort on describing semantic relations using primitives was made by Chaffin and Herrmann (1987) . They introduce Relation Element Theory, and differentiate relations by relation elements. The authors describe a set of 31 relations clustered in five groups (CONTRAST, SIMILARS, CLASS INCLUSION, CASE-RELATIONS, PART-WHOLE), and distinguish each relation by its relations elements and not just a definition and examples. Their 30 relation elements are clustered into five groups (elements of intensional force, dimension elements, elements of agreement, propositional elements, elements of part-whole inclusion). They only use the elements to define relations, not to compose relations. Winston et al. (1987) work with six subtypes of PART-WHOLE and uses 3 relation elements (functional, homeomerous and separable) to distinguish the subtypes. Cohen and Losielle (1988) introduce the notion of deep structure and characterize it using two aspects: hierarchical and temporal. Huhns and Stephens (1989) extend previous works by considering an extended set of 10 primitives.",
"cite_spans": [
{
"start": 337,
"end": 364,
"text": "Chaffin and Herrmann (1987)",
"ref_id": "BIBREF2"
},
{
"start": 954,
"end": 975,
"text": "Winston et al. (1987)",
"ref_id": "BIBREF17"
},
{
"start": 1111,
"end": 1136,
"text": "Cohen and Losielle (1988)",
"ref_id": "BIBREF4"
},
{
"start": 1242,
"end": 1267,
"text": "Huhns and Stephens (1989)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Previous Work",
"sec_num": "6"
},
{
"text": "In Computational Linguistics there have been previous proposals to combine semantic relations. Harabagiu and Moldovan (1998) manually extract plausible inference axioms using WordNet relations. Helbig (2005) transforms chains of relations into theoretical axioms. On the other hand, the model presented in this paper extracts inference axioms automatically.",
"cite_spans": [
{
"start": 95,
"end": 124,
"text": "Harabagiu and Moldovan (1998)",
"ref_id": "BIBREF7"
},
{
"start": 194,
"end": 207,
"text": "Helbig (2005)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Previous Work",
"sec_num": "6"
},
{
"text": "Composing relations has been proposed before in the more general field of Artificial Intelligence, in particular in the context of Knowledge Bases. Cohen and Losielle (1988) point out that two relations shall combine if and only if they do not have contradictory values for the aspect hierarchical or temporal. They work with a set of nine specific relations (CAUSES, COMPONENT-OF, FOCUS-OF, MECHANISM-OF, PRODUCT-OF, PURPOSE-OF, SETTING-OF, SUBJECT-OF and SUBFIELD-OF) and their inverses. Huhns and Stephens (1989) are the first to propose an algebra for composing semantic primitives. Unlike ours, their set of relations is not linguistically motivated; ten of them map to some sort of PART-WHOLE (e.g.",
"cite_spans": [
{
"start": 148,
"end": 173,
"text": "Cohen and Losielle (1988)",
"ref_id": "BIBREF4"
},
{
"start": 490,
"end": 515,
"text": "Huhns and Stephens (1989)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Previous Work",
"sec_num": "6"
},
{
"text": "In this paper, we have presented a model to compose semantic relations. The model is independent of any particular set of relations and is able to obtain inference axioms. These axioms take as their input two semantic relations and yield a previously ignored relation as conclusion.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "The model is based on an extended definition of semantic relations, including restrictions on domains and ranges and values for a set of semantic primitives. We have defined an algebra for composing semantic primitives. This algebra is the key to automatically identify the resulting relation of composing a pair of compatible relations and to form an axiom.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "The proposed algorithm to compose semantic relations identifies eight inference axioms using Prop-Bank relations. When instantiated in a subset of PropBank, these axioms add 2.26% of annotation in relative terms with an accuracy of 0.54. We believe these results are worthwhile for a completely unsupervised approach to obtain semantic relations. Adding a simple heuristic improves the accuracy to 0.86, lowering the productivity in relative terms to 1.35%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "The model has limitations and is not always correct. First, relations are defined manually and mistakes could be made when assigning values to their primitives. Second, the algebra for composing primitives is also manually defined.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "We find the first problem easy to overcome. Whatever the set of relations one might use, we believe thinking in terms of primitives helps to understand the nature of the relations and their differences. An issue might be that the proposed set of primitives is not enough for a particular set, but more primitives could be added to solve this eventuality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "A further issue with the algebra is the fact that primitives are composed orthogonally. This is a simplification, but we have shown that this simplified algebra works.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "Even though different sets of semantic relations may call for different ontologies to define domains and ranges, and possibly an extended set of primitives, we believe the model presented in this paper is applicable to any set. As far as we are concerned, this is a novel way to compose semantic relations in the field of Computational Linguistics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "http://verbs.colorado.edu/\u02dcm palmer/projects/ace/PBguidelines.pdf",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Learning Semantic Links from a Corpus of Parallel Temporal and Causal Relations",
"authors": [
{
"first": "S",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "J",
"middle": [
"H"
],
"last": "Martin",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08: HLT, Short Papers",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bethard, S. and J. H. Martin (2008). Learning Semantic Links from a Corpus of Parallel Temporal and Causal Relations. In Proceedings of ACL-08: HLT, Short Papers, Columbus, OH.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Introduction to the CoNLL-2005 shared task: semantic role labeling",
"authors": [
{
"first": "X",
"middle": [],
"last": "Carreras",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "M\u00e0rquez",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. of the 9th Conf. on Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Carreras, X. and L. M\u00e0rquez (2005). Introduction to the CoNLL-2005 shared task: semantic role label- ing. In Proc. of the 9th Conf. on Computational Natural Language Learning, Morristown, NJ.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Relation Element Theory: A New Account of the Representation and Processing of Semantic Relations",
"authors": [
{
"first": "R",
"middle": [],
"last": "Chaffin",
"suffix": ""
},
{
"first": "D",
"middle": [
"J"
],
"last": "Herrmann",
"suffix": ""
}
],
"year": 1987,
"venue": "Memory and Learning. The Ebbinghaus Centennial Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chaffin, R. and D. J. Herrmann (1987). Relation Element Theory: A New Account of the Representation and Processing of Semantic Relations. In D. S. Gorfein and R. R. Hoffman (Eds.), Memory and Learning. The Ebbinghaus Centennial Conference.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Incremental cue phrase learning and bootstrapping method for causality extraction using cue phrase and word pair probabilities",
"authors": [
{
"first": "D",
"middle": [
"S"
],
"last": "Chang",
"suffix": ""
},
{
"first": "K",
"middle": [
"S"
],
"last": "Choi",
"suffix": ""
}
],
"year": 2006,
"venue": "Information Processing & Management",
"volume": "42",
"issue": "3",
"pages": "662--678",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chang, D. S. and K. S. Choi (2006). Incremental cue phrase learning and bootstrapping method for causality extraction using cue phrase and word pair probabilities. Information Processing & Manage- ment 42(3), 662-678.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Beyond ISA: Structures for Plausible Inference in Semantic Networks",
"authors": [
{
"first": "P",
"middle": [
"R"
],
"last": "Cohen",
"suffix": ""
},
{
"first": "C",
"middle": [
"L"
],
"last": "Losielle",
"suffix": ""
}
],
"year": 1988,
"venue": "Proceedings of the Seventh National conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cohen, P. R. and C. L. Losielle (1988). Beyond ISA: Structures for Plausible Inference in Semantic Networks. In Proceedings of the Seventh National conference on Artificial Intelligence, St. Paul, MN.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Automatic Discovery of Part-Whole Relations",
"authors": [
{
"first": "R",
"middle": [],
"last": "Girju",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Badulescu",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Moldovan",
"suffix": ""
}
],
"year": 2006,
"venue": "Computational Linguistics",
"volume": "32",
"issue": "1",
"pages": "83--135",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Girju, R., A. Badulescu, and D. Moldovan (2006). Automatic Discovery of Part-Whole Relations. Com- putational Linguistics 32(1), 83-135.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "SemEval-2007 Task 04: Classification of Semantic Relations between Nominals",
"authors": [
{
"first": "R",
"middle": [],
"last": "Girju",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Nakov",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Nastase",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Szpakowicz",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Turney",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Yuret",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the Fourth International Workshop on Semantic Evaluations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Girju, R., P. Nakov, V. Nastase, S. Szpakowicz, P. Turney, and D. Yuret (2007). SemEval-2007 Task 04: Classification of Semantic Relations between Nominals. In Proceedings of the Fourth International Workshop on Semantic Evaluations, Prague, Czech Republic.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "WordNet: An Electronic Lexical Database and Some of its Applications",
"authors": [
{
"first": "S",
"middle": [],
"last": "Harabagiu",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Moldovan",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "684--714",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Harabagiu, S. and D. Moldovan (1998). Knowledge Processing on an Extended WordNet. In C. Fellbaum (Ed.), WordNet: An Electronic Lexical Database and Some of its Applications, Chapter 17, pp. 684- 714. The MIT Press.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Knowledge Representation and the Semantics of Natural Language",
"authors": [
{
"first": "H",
"middle": [],
"last": "Helbig",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Helbig, H. (2005). Knowledge Representation and the Semantics of Natural Language (1st ed.). Springer.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "SemEval-2010 Task 8: Multi-Way Classification of Semantic Relations Between Pairs of Nominals",
"authors": [
{
"first": "I",
"middle": [],
"last": "Hendrickx",
"suffix": ""
},
{
"first": "S",
"middle": [
"N"
],
"last": "Kim",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Kozareva",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Nakov",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Diarmuid",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Pad\u00f3",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Pennacchiotti",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Romano",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Szpakowicz",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Workshop on Semantic Evaluations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hendrickx, I., S. N. Kim, Z. Kozareva, P. Nakov, Diarmuid, S. Pad\u00f3, M. Pennacchiotti, L. Romano, and S. Szpakowicz (2009). SemEval-2010 Task 8: Multi-Way Classification of Semantic Relations Between Pairs of Nominals. In Proceedings of the Workshop on Semantic Evaluations, Boulder, CO.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Plausible Inferencing Using Extended Composition",
"authors": [
{
"first": "M",
"middle": [
"N"
],
"last": "Huhns",
"suffix": ""
},
{
"first": "L",
"middle": [
"M"
],
"last": "Stephens",
"suffix": ""
}
],
"year": 1989,
"venue": "IJCAI'89: Proceedings of the 11th international joint conference on AI",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Huhns, M. N. and L. M. Stephens (1989). Plausible Inferencing Using Extended Composition. In IJCAI'89: Proceedings of the 11th international joint conference on AI, San Francisco, CA.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Semantic Role Labeling: An Introduction to the Special Issue",
"authors": [
{
"first": "L",
"middle": [],
"last": "M\u00e0rquez",
"suffix": ""
},
{
"first": "X",
"middle": [],
"last": "Carreras",
"suffix": ""
},
{
"first": "K",
"middle": [
"C"
],
"last": "Litkowski",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Stevenson",
"suffix": ""
}
],
"year": 2008,
"venue": "Computational Linguistics",
"volume": "34",
"issue": "2",
"pages": "145--159",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M\u00e0rquez, L., X. Carreras, K. C. Litkowski, and S. Stevenson (2008, June). Semantic Role Labeling: An Introduction to the Special Issue. Computational Linguistics 34(2), 145-159.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Models for the Semantic Classification of Noun Phrases",
"authors": [
{
"first": "D",
"middle": [],
"last": "Moldovan",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Badulescu",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Tatu",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Antohe",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Girju",
"suffix": ""
}
],
"year": 2004,
"venue": "HLT-NAACL 2004: Workshop on Computational Lexical Semantics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Moldovan, D., A. Badulescu, M. Tatu, D. Antohe, and R. Girju (2004). Models for the Semantic Classi- fication of Noun Phrases. In HLT-NAACL 2004: Workshop on Computational Lexical Semantics.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Semantic Classification of Noun Phrases Using Web Counts and Learning Algorithms",
"authors": [
{
"first": "P",
"middle": [],
"last": "Nulty",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the ACL 2007 Student Research Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nulty, P. (2007). Semantic Classification of Noun Phrases Using Web Counts and Learning Algorithms. In Proceedings of the ACL 2007 Student Research Workshop, Prague, Czech Republic.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "The Proposition Bank: An Annotated Corpus of Semantic Roles",
"authors": [
{
"first": "M",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Gildea",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Kingsbury",
"suffix": ""
}
],
"year": 2005,
"venue": "Computational Linguistics",
"volume": "31",
"issue": "1",
"pages": "71--106",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Palmer, M., D. Gildea, and P. Kingsbury (2005). The Proposition Bank: An Annotated Corpus of Semantic Roles. Computational Linguistics 31(1), 71-106.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Extraction of Entailed Semantic Relations Through Syntax-Based Comma Resolution",
"authors": [
{
"first": "V",
"middle": [],
"last": "Srikumar",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Reichart",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Sammons",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Rappoport",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Roth",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08: HLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Srikumar, V., R. Reichart, M. Sammons, A. Rappoport, and D. Roth (2008). Extraction of Entailed Semantic Relations Through Syntax-Based Comma Resolution. In Proceedings of ACL-08: HLT, Columbus, OH.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Interactive semantic analysis of Clause-Level Relationships",
"authors": [
{
"first": "B",
"middle": [],
"last": "Szpakowicz",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Barker",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Szpakowicz",
"suffix": ""
}
],
"year": 1995,
"venue": "Proc. of the 2nd Conference of the Pacific Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Szpakowicz, B., K. Barker, and S. Szpakowicz (1995). Interactive semantic analysis of Clause-Level Relationships. In Proc. of the 2nd Conference of the Pacific Association for Computational Linguistics.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "A Taxonomy of Part-Whole Relations",
"authors": [
{
"first": "M",
"middle": [
"E"
],
"last": "Winston",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Chaffin",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Herrmann",
"suffix": ""
}
],
"year": 1987,
"venue": "Cognitive Science",
"volume": "11",
"issue": "4",
"pages": "417--444",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Winston, M. E., R. Chaffin, and D. Herrmann (1987). A Taxonomy of Part-Whole Relations. Cognitive Science 11(4), 417-444.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "Semantic representation of the sentence John went to the shop to buy flowers.",
"uris": null,
"type_str": "figure"
},
"FIGREF3": {
"num": null,
"text": "In the fibers division, profit remains weak, largely because of persistent overcapacity. (wsj 0552, 28). First Tennessee National Corp. said it would take a $4 million charge in the fourth quarter, as a result of plans to expand its systems operation. (wsj 0621, 0). When it occurs, the traders place orders via computers to buy the basket of stocks . . . in whichever market is cheaper and sell them in the more expensive market; . . . (wsj 0118, 48).A man . . .",
"uris": null,
"type_str": "figure"
},
"FIGREF5": {
"num": null,
"text": "A man from the Bush administration came before the House Agriculture Committee yesterday to talk about . . . (wsj 0134, 0). Figures 2 and 3 instantiate axioms 1, 2 and 3. For these examples, all inferences are correct. Figures 4 and 5 instantiate the rest of axioms. Not using the heuristic leads to a wrong inference in the example shown in",
"uris": null,
"type_str": "figure"
},
"TABREF1": {
"type_str": "table",
"content": "",
"text": "Primitives for characterizing semantic relations, values for the inverse relation and references.",
"html": null,
"num": null
},
"TABREF2": {
"type_str": "table",
"content": "",
"text": "",
"html": null,
"num": null
},
"TABREF4": {
"type_str": "table",
"content": "| Role | | | | Primitive | | | | Role | | | | Primitive | |
| Composable | Functional | Separable | Temporal | Connected | Intrinsic | Volitional | Fully Impl. | Weakly Impl. | | Composable | Functional | Separable | Temporal | Connected | Intrinsic | Volitional | Fully Impl. | Weakly Impl. |
| ARG0 | + | + | + | 0 | - | - | + | 0 | 0 | ARG0 | \u22121 | | | | | | |
",
"text": "The four unique axioms taking as premises R 1 and R 2 . R 3 indicates the conclusion.",
"html": null,
"num": null
},
"TABREF5": {
"type_str": "table",
"content": "",
"text": "",
"html": null,
"num": null
},
"TABREF6": {
"type_str": "table",
"content": "",
"text": "ARG0 [procrastinated] rel [a lot] MADV [due to his nervous demeanor] MCAU .",
"html": null,
"num": null
},
"TABREF7": {
"type_str": "table",
"content": "| no heuristic | with heuristic |
",
"text": "Number of relations in PropBank for the first 31,450 sentences.",
"html": null,
"num": null
}
}
}
}