Mining the Web for Hyponymy Relations based on Property Inheritance Shun Hattori, Hiroaki Ohshima, Satoshi Oyama, and Katsumi Tanaka Department of Social Informatics, Graduate School of Informatics, Kyoto University Yoshida-Honmachi, Sakyo, Kyoto 606-8501, Japan {hattori, ohshima, oyama, tanaka}@dl.kuis.kyoto-u.ac.jp Abstract. Concept hierarchies, such as hyponymy and meronymy relations, are very important for various natural language processing systems. Many researchers have tackled how to mine very large corpora of documents such as the Web for them not manually but automatically. However, their methods are mostly based on lexico-syntactic patterns as not necessary but sufficient conditions of concept hierarchies, so they can achieve high precision but low recall when using stricter patterns or they can achieve high recall but low precision when using looser patterns. In this paper, property inheritance from a concept to its hyponyms is assumed to be necessary and sufficient conditions of hyponymy relations to achieve high recall and not low precision, and we propose a method to acquire hyponymy relations from the Web based on property inheritance.
1
Introduction
Conceptual hierarchies, such as hyponymy and meronymy relations, are very important for various natural language processing systems. For example, query expansion in information retrieval [1, 2], WebQA [3], machine translation, object information extraction by text mining [4], and so forth. While WordNet [5, 6] and Wikipedia [7, 8] are being manually constructed and maintained as lexical ontologies, many researchers have tackled how to extract concept hierarchies from very large corpora of text documents such as the Web not manually but automatically [9–15]. However, their methods are mostly based on lexico-syntactic patterns, e.g., “x such as y” and “y is a/an x”, as sufficient but not necessary conditions of concept hierarchies. Therefore, they can achieve high precision but low recall when using stricter patterns or they can achieve high recall but low precision when using looser patterns. To achieve high recall and not low precision, this paper proposes a method to extract hyponyms and hypernyms of a target concept from the Web based on property inheritance. Our method assumes property inheritance from a concept to its hyponyms to be necessary and sufficient conditions of hyponymy relations. The remainder of the paper is organized as follows. First, Section 2 introduces one assumption on the relationship between hyponymy and property inheritance. Next, Section 3 and 4 describe our method of mining the Web for hyponyms and hypernyms based on it, respectively. Section 5 shows some experimental results to validate our method. Finally, we conclude the paper in Section 6.
2
Hyponymy and Property Inheritance
In this section, we introduce and discuss one assumption on the relationship between hyponymy and property inheritance before proposing our method of mining the Web for hyponymy relations (hyponyms and hypernyms of a target concept) based on property inheritance in the following sections. Assumption 1. Let C be the universal set of concepts (words). We assume that if and only if a concept x ∈ C is a superordinate (hypernym) of y ∈ C, in other words, a concept y is a subordinate (hyponym) of a concept x, then the set of properties that a concept y has, P (y), completely includes the set of properties that a concept x has, P (x): { 1 if P (x) ⊆ P (y) and x ̸= y, hyponym(y, x) = hypernym(x, y) = 0 otherwise. P (c) = {pi ∈ P | has(pi , c) = 1}, where P stands for the universal set of N properties, P = {p1 , p2 , ..., pN }, and has(pi , c) indicates whether or not a concept c ∈ C has a property pi ∈ P , { 1 if a concept c has a property pi , has(pi , c) = 0 otherwise. In other words,
∑ ∑ 1 if has(pi , x) · has(pi , y) = has(pi , x) · has(pi , x), p ∈P p ∈P i i (i.e., P (x) · P (y) = P∑ (x) · P (x) ) ∑ hyponym(y, x) = has(pi , x) · has(pi , y) < has(pi , x) · has(pi , x), 0 if pi ∈P pi ∈P (i.e., P (x) · P (y) < P (x) · P (x) ) P (c) = (has(p1 , c), has(p2 , c), ..., has(pN , c)).
It is very essential for hyponymy extraction based on the above assumption to calculate the binary value has(pi , c) ∈ {0, 1} of any concept and property accurately. However, it is not easy, and we can only use the continuous value has(pi , c) ∈ [0, 1] in this paper. Therefore, we suppose that the ratio of the number of properties that a concept ∑ y inherits from a concept x to the number has(pi ,x)·has(pi ,y) of properties that a concept x has, ∑pi ∈P , can simulate how has (pi ,x)·has(pi ,x) pi ∈P suitable a concept y is for a hyponym of a concept x, as an approximation of whether or not a concept y is a hyponym of a concept x, hyponym(y, x). That is to say, we consider a concept y as a hyponym of a concept x when the ratio is enough near to one, while we consider that a concept y is not a hyponym of a concept x when the ratio is not near to one.
3
Hyponym Extraction based on Property Inheritance
In this section, we propose a method to mine such a very large corpus of documents as the Web for hyponyms of a target concept, by using not only lexicosyntactic patterns as sufficient (but not necessary) conditions of hyponymy, but also property inheritance as its necessary and sufficient condition. When a target concept x ∈ C is given, our method executes the following three steps to extract its hyponyms from the Web. First, a set of candidates for its hyponyms, C(x), is collected. Second, some weight of each pair of a property p ∈ P and a concept c (the target concept x or its hyponym candidate y ∈ C(x)), has(p, c) is calculated by using Web search engine indices. That is, the property vector of each concept c, P (c), is obtained. Last, some weight of each pair of the target concept x and its hyponym candidate y ∈ C(x), hyponym(y, x), is calculated based on property inheritance, and then a set of its top k hyponym candidates ordered by their weight, Ck (x), would be outputted to the users. Step 1. Listing of candidates for hyponyms of a target concept Our method has to collect a set of candidates for hyponyms of a target concept x, C(x), as exhaustively as possible and enough precisely. If C(x) is set to the universal set of concepts, C, its recall always equals to 1.0 (the highest value) but its precision nearly equals to 0.0 (too low value). Meanwhile, if y ∈ C(x) is collected from some sort of corpus of documents by using too strict syntactic pattern such as “y is a kind of x”, its precision is enough high but its recall is too low in most cases. Therefore, our method collects a set of candidates for hyponyms of a target concept x, C(x), from the Web by using not too strict but enough strict syntactic pattern such as “y is a/an x”. In the after-mentioned experiments, any 2-gram noun phrase y where the syntactic pattern “y is a/an x” exists at least once in the title or summary text of the top 1000 retrieval results by submitting a phrase “is a/an x” as a query to Yahoo! Web search API [16] is inserted into C(x) as a candidate for hyponyms of the target concept x.
x: target concept y: hyponym candidate P(c): property vector
P(x)
x
hyponym (y,x) = ? ß inherit ?
P(y) ・・・
y
・・・
Fig. 1. Property Inheritance (PI) based hyponym extraction
Step 2. Extracting of typical properties of each concept Our method has to obtain the property vector of each concept c (a target concept x ∈ C or its hyponym candidate y ∈ C(x)), P (c). The weight of each pair of a property p ∈ P and a concept c, has(p, c), is defined as follows: has(p, c) :=
df(“c’s p”) , df(“c’s”) + 1
where df(q) stands for the number of documents that meet a query condition q in such a corpus as the Web. In the after-mentioned experiments, we calculate it by submitting each query to Yahoo! Web search API [16]. Note that has(p, c) is not a binary value {0, 1} but a continuous value [0, 1], so has(p, c) cannot indicate whether or not a concept c has a property p but it would indicate how typical a property p is of a concept c. Step 3. Weighting of candidates for hyponyms of a target concept In our hyponym extraction based on property inheritance, hyponym candidates of a target concept x, C(x), that are collected in Step 1, have to be ordered based on how many properties a hyponym candidate y ∈ C(x) inherits from the target concept x. The weight of a candidate y for hyponyms of a target concept x is defined as follows: ∑ has(pi , x) · has(pi , y) I hyponymP n (y, x) :=
pi ∈Pn (x)
∑
has(pi , x) · has(pi , x)
,
pi ∈Pn (x)
where Pn (c) stands for a set of the top n typical properties of a concept c ordered by their weight has(p, c) that is calculated in Step 2. Note that if n = N , i.e., Pn (x) = Pn (y) = P , we cannot decide which x or y is I PI subordinate to the other because of hyponymP n (y, x) = hyponymn (x, y). In the after-mentioned experiments, we set n to 3 or 10 fixedly. Meanwhile, in an existing hyponym extraction based on such a syntactic pattern as “y is a/an x”, it is defined in such a simple manner as follows; hyponymSP (y, x) := df(“y is a/an x”). In the after-mentioned experiments, we use hyponymSP (y, x) as a baseline to validate our hyponym extraction based on property inheritance. Last, we define a hybrid of the above two kinds of weights as follows: SP I (y, x) + 1) · (hyponymP hyponymHB n (y, x) + µ), n (y, x) := log2 (hyponym
where µ stands for a constant holding the smallest positive nonzero value to I prevent hyponymSP (y, x) from being ineffective when hyponymP n (y, x) = 0. −324 In the after-mentioned experiments, µ is set to 4.9 · 10 .
4
Hypernym Extraction based on Property Inheritance
In this section, we propose a method to mine such a very large corpus of documents as the Web for hypernyms of a target concept, by using not only lexicosyntactic patterns as sufficient (but not necessary) conditions of hyponymy, but also property inheritance as its necessary and sufficient condition. When a target concept x ∈ C is given, our method executes the following three steps to extract its hypernyms from the Web. First, a set of candidates for its hypernyms, C(x), is collected. Second, some weight of each pair of a property p ∈ P and a concept c (the target concept x or its hypernym candidate y ∈ C(x)), has(p, c) is calculated by using Web search engine indices. That is, the property vector of each concept c, P (c), is obtained. Last, some weight of each pair of the target concept x and its hypernym candidate y ∈ C(x), hypernym(y, x), is calculated based on property inheritance, and then a set of its top k hypernym candidates ordered by their weight, Ck (x), would be outputted to the users. Step 1. Listing of candidates for hyponyms of a target concept Our method has to collect a set of candidates for hypernyms of a target concept x, C(x), as exhaustively as possible and enough precisely. On the same score as described in Step 1 of Section 3, we don’t use the universal set of concepts, C, as C(x), and we don’t use too strict syntactic pattern such as “x is a kind of y” to collect it from the Web by text mining. Therefore, our method collects a set of candidates for hypernyms of a target concept x, C(x), from the Web by using not too strict but enough strict syntactic pattern such as “x is a/an y”, to achieve enough high precision and not too low recall. In the after-mentioned experiments, any 1-gram noun phrase y where the syntactic pattern “x is a/an y” exists at least once in the title or summary text of at most 1000 retrieval results by submitting a phrase “x is a/an” as a query to Yahoo! Web search API [16] is inserted into C(x) as a candidate for hypernyms of the target concept x.
P(y) ・・・
y
・・・
hypernym (y,x) = ? ß inherit ?
x: target concept y: hypernym candidate P(c): property vector
P(x)
x
Fig. 2. Property Inheritance (PI) based hypernym extraction
Step 2. Extracting of typical properties of each concept The same as described in Step 2 of Section 3. Our method has to obtain the property vector of each concept c (a target concept x ∈ C or its hypernym candidate y ∈ C(x)), P (c). The weight of each pair of a property p ∈ P and a concept c, has(p, c), is defined as follows: has(p, c) :=
df(“c’s p”) , df(“c’s”) + 1
where df(q) stands for the number of documents that meet a query condition q in such a corpus as the Web. In the after-mentioned experiments, we calculate it by submitting each query to Yahoo! Web search API [16]. Note that has(p, c) is not a binary value {0, 1} but a continuous value [0, 1], so has(p, c) cannot indicate whether or not a concept c has a property p but it would indicate how typical a property p is of a concept c. Step 3. Weighting of candidates for hyponyms of a target concept In our hypernym extraction based on property inheritance, hypernym candidates of a target concept x, C(x), that are collected in Step 1, have to be ordered based on how many properties the target concept x inherits from a hypernym candidate y ∈ C(x). The weight of a candidate y for hypernyms of a target concept x is defined as follows: ∑ has(pi , y) · has(pi , x) I hypernymP n (y, x) :=
pi ∈Pn (y)
∑
has(pi , y) · has(pi , y)
,
pi ∈Pn (y)
where Pn (c) stands for a set of the top n typical properties of a target concept c ordered by their weight has(p, c) that is calculated in Step 2. Note I PI that hypernymP n (y, x) is equivalent to hyponymn (x, y) defined in Step 3 of Section 3. In the after-mentioned experiments, we set n to 3 or 10. Meanwhile, in a traditional hyponym extraction based on such a syntactic pattern as “x is a/an y”, it is defined in such a simple manner as follows; hypernymSP (y, x) := df(“x is a/an y”). In the after-mentioned experiments, we use hypernymSP (y, x) as a baseline to validate our hypernym extraction based on property inheritance. Last, we define a hybrid of the above two kinds of weights as follows: SP I (y, x) + 1) · (hypernymP hypernymHB n (y, x) + µ), n (y, x) := log2 (hypernym
where µ stands for a constant holding the smallest positive nonzero value to I prevent hypernymSP (y, x) from being invalid when hypernymP n (y, x) = 0. −324 In the after-mentioned experiments, µ is set to 4.9 · 10 .
5
Experiment
In this section, we show some experimental results to validate our method of mining the Web for hyponymy relations (hyponyms and hypernyms of a target concept) based on property inheritance, by comparing it with a traditional syntactic pattern based hyponymy extraction and the hybrid of them. 5.1
Experiment on Mining the Web for Hyponyms
We applied our method of mining the Web for hyponyms based on property inheritance to five kinds of concepts: “amphibian”, “bird”, “fish”, “mammal”, and “reptile”. All of them are subordinate of “animal”. Table 1 shows the top 10 typical properties with their weight has(p, c) for each of five target concepts and “animal” (which is one of their common hypernyms). Note that their weights were calculated by using Yahoo! Web search API [16] in October of 2007. It shows that these subordinate concepts of “animal” inherit such a typical property of “animal” as “body”, “head” and “life”. The results seem to be also good for a meronymy extraction. For example, “beak” and “wings” seem to be characteristic components for “bird”, “gills” is for “fish”, and “milk” is for “mammal”. Table 1. Top 10 typical properties with their weight for each concept k 1 2 3 4 5 6 7 8 9 10 k 1 2 3 4 5 6 7 8 9 10
amphibian skin (0.01039) internal (0.00825) chest (0.00796) life (0.00782) body (0.00653) habitat (0.00604) eye (0.00467) assault (0.00291) descent (0.00277) heart (0.00242) mammal body (0.02596) notebook (0.00846) milk (0.00688) blood (0.00558) heart (0.00524) life (0.00411) skeleton (0.00408) immune (0.00363) head (0.00357) brain (0.00311)
k 1 2 3 4 5 6 7 8 9 10 k 1 2 3 4 5 6 7 8 9 10
bird eye (0.44815) nest (0.09907) foot (0.01787) head (0.01352) beak (0.01157) wings (0.00821) song (0.00797) body (0.00701) nests (0.00700) life (0.00647) reptile habitat (0.01304) head (0.01304) lair (0.01170) body (0.01141) skin (0.00822) eyes (0.00785) mouth (0.00763) tail (0.00563) hide (0.00549) cage (0.00459)
k 1 2 3 4 5 6 7 8 9 10 k 1 2 3 4 5 6 7 8 9 10
fish mouth (0.05737) body (0.02953) tail (0.02673) head (0.02240) belly (0.01971) eye (0.01626) ability (0.01035) life (0.01029) gills (0.01006) skin (0.00819) animal body (0.04674) life (0.03774) head (0.02414) health (0.02280) behavior (0.01665) skin (0.01467) people (0.01345) ability (0.01253) owner (0.01245) neck (0.01077)
Table 2. Top 5 hyponym candidates for each concept by Syntactic Pattern (SP) based vs. Property Inheritance (PI) based vs. Hybrid hyponym extraction (Note that |C(x)| stands for the number of candidates for hypernyms of a concept x, and |S(x)| stands for the number of acceptable ones in C(x).) amphibian (|C(x)| = 255, |S(x)| = 44) PI-based (n=3) PI-based (n=10) Hybrid (n=3) mink frog bird painted frog painted frog chubby frog man sp100 mink frog mink frog man crested newt frog crested newt vertebrate animal bird (|C(x)| = 434, |S(x)| = 140) k SP-based PI-based (n=3) PI-based (n=10) Hybrid (n=3) 1 love snake-bird snake-bird mind 2 mind mind mind ostrich 3 chicken ostrich ostrich snake-bird 4 eagle crested crane crested crane cuckoo 5 ambition cuckoo cuckoo eagle fish (|C(x)| = 492, |S(x)| = 114) k SP-based PI-based (n=3) PI-based (n=10) Hybrid (n=3) 1 fish tycoon horse horse horse 2 tycoon intermediate host natural fish electric eel 3 mother electric eel chilazon lingcod 4 school chilazon intermediate host whale 5 mccarthy school lingcod sea horse intermediate host mammal (|C(x)| = 350, |S(x)| = 169) k SP-based PI-based (n=3) PI-based (n=10) Hybrid (n=3) 1 cat brazilian tapir brazilian tapir brazilian tapir 2 whale pilot whale pilot whale cow 3 dog goat goat raccoon 4 leo cow cow vertebrate 5 panthera leo animal species vertebrate goat reptile (|C(x)| = 283, |S(x)| = 65) k SP-based PI-based (n=3) PI-based (n=10) Hybrid (n=3) 1 snake dragon horse snake 2 maher serpent dragon dragon 3 bill maher snake snake serpent 4 pet zoo viper alligator alligator 5 zoo alligator serpent turtle k 1 2 3 4 5
SP-based frog turtle newt salamander toad
Hybrid (n=10) bird frog animal crested newt man Hybrid (n=10) mind ostrich snake-bird cuckoo eagle Hybrid (n=10) horse whale sea horse lingcod electric eel Hybrid (n=10) cow brazilian tapir vertebrate raccoon goat Hybrid (n=10) snake dragon alligator serpent animal
Table 2 compares the top 5 hyponym candidates for each target concept by syntactic pattern based, property inheritance based (n = 3, 10), or hybrid (n = 3, 10) hyponym extraction. It shows that our hyponym extraction based on property inheritance is superior to a traditional hyponym extraction based on such a syntactic pattern as “is a/an”, and the hybrid of them seems to be somewhat superior to both of them. Figure 3 shows precision-recall graphs for each target concept and the 21point interpolated average for all of five target concepts by syntactic pattern based, property inheritance based (n = 3, 10), or hybrid (n = 3, 10) hyponym extraction. Our hyponym extraction based on property inheritance and the hybrid are completely superior to a traditional hyponym extraction based on such a syntactic pattern as “is a/an”. It also shows that the bigger number of typical properties for the calculation of property inheritance from a target concept to its hyponym candidate, n, the more precise result, and that the hybrid (n = 10) is superior to the others in the middle recall range (about from 0.2 to 0.9).
for "amphibian"
for "bird"
1
SP-based PI-based (n=3)
0.9
0.9
PI-based (n=10)
0.8
0.8
Hybrid (n=3) Hybrid (n=10)
0.7
0.7
0.6
Precision
Precision
1
0.5
0.6 0.5
0.4
0.4
0.3
0.3
0.2
0.2
SP-based PI-based (n=3) PI-based (n=10)
0.1
0.1
Hybrid (n=3) Hybrid (n=10)
0
0 0
0.1
0.2
0.3
0.4
0.5 0.6 Recall
0.7
0.8
0.9
1
0
0.1
0.2
0.5 0.6 Recall
0.7
0.8
0.9
1
0.7
0.8
0.9
1
1
SP-based PI-based (n=3)
0.9
0.9
PI-based (n=10)
0.8
0.8
Hybrid (n=3) Hybrid (n=10)
0.7
0.7
0.6
Precision
Precision
0.4
for "mammal"
for "fish" 1
0.5
0.6 0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
SP-based PI-based (n=3) PI-based (n=10) Hybrid (n=3) Hybrid (n=10)
0
0 0
0.1
0.2
0.3
0.4
0.5 0.6 Recall
0.7
0.8
0.9
0
1
0.1
0.2
for "reptile"
0.3
0.4
0.5 0.6 Recall
21-point average
1
1
SP-based PI-based (n=3) PI-based (n=10)
0.9 0.8
SP-based PI-based (n=3) PI-based (n=10)
0.9 0.8
Hybrid (n=3) Hybrid (n=10)
0.7
Hybrid (n=3) Hybrid (n=10)
0.7
0.6
Precision
Precision
0.3
0.5
0.6 0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0 0
0.1
0.2
0.3
0.4
0.5 0.6 Recall
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5 0.6 Recall
0.7
0.8
0.9
1
Fig. 3. Precision-Recall graph for each concept or 21-point average by Syntactic Pattern (SP) based vs. Property Inheritance (PI) based vs. Hybrid hyponym extraction
5.2
Experiment on Mining the Web for Hypernyms
We applied our method of mining the Web for hypernyms based on property inheritance to five kinds of concepts: “amphibian”, “bird”, “fish”, “mammal”, and “reptile”. All of them are subordinate of “animal”, “creature”, “physical entity” and so forth in such a large lexical database of English as WordNet [5]. Table 3 shows the acceptable candidates with their rank for each target concept by syntactic pattern based, property inheritance based (n = 3, 10), or hybrid (n = 3, 10) hypernym extraction. Figure 4 shows the 21-point interpolated average precision-recall graph for all of five target concepts by syntactic pattern based, property inheritance based (n = 3, 10), or hybrid (n = 3, 10) hypernym extraction. Our hypernym extraction based on property inheritance is inferior in precision to a traditional hypernym extraction based on such a syntactic pattern as “is a/an” at low recall (0 to about 0.2), but is completely superior at from middle to high recall. The hybrid of them is fundamentally superior to each of them. Especially, the hybrid (n = 3) is always superior to the other hypernym extractions. Possibly, we have the hybrid (the other n) that is better than these in this paper. Therefore, it is one of our future works to invent how to find the optimal number n.
Table 3. Ranks of hypernyms for each concept by Syntactic Pattern (SP) based vs. Property Inheritance (PI) based vs. Hybrid hypernym extraction (Note that |C(x)| stands for the number of candidates for hypernyms of a concept x, and |S(x)| stands for the number of acceptable ones in C(x).) amphibian (|C(x)| = 72, |S(x)| = 6) hypernym SP-based PI-based (n=3) PI-based (n=10) Hybrid animal 1st 7th 9th creature 7th 13th 15th organism 37th 23rd 31st organisms 70th 18th 27th lifeform 36th 19th 13th aeroplane 31st 42nd 46th bird (|C(x)| = 144, |S(x)| = 4) hypernym SP-based PI-based (n=3) PI-based (n=10) Hybrid animal 12th 15th 21st creature 38th 17th 20th creation 101st 71st 44th instrument 2nd 52nd 55th fish (|C(x)| = 123, |S(x)| = 3) hypernym SP-based PI-based (n=3) PI-based (n=10) Hybrid animal 7th 6th 5th matter 33rd 79th 60th person 35th 8th 12th mammal (|C(x)| = 142, |S(x)| = 2) hypernym SP-based PI-based (n=3) PI-based (n=10) Hybrid animal 2nd 1st 1st creature 34th 3rd 4th reptile (|C(x)| = 108, |S(x)| = 2) hypernym SP-based PI-based (n=3) PI-based (n=10) Hybrid animal 6th 7th 9th organism 39th 34th 38th
(n=3) Hybrid (n=10) 1st 2nd 7th 9th 24th 31st 23rd 32nd 19th 17th 40th 42nd (n=3) Hybrid (n=10) 11th 16th 17th 22nd 72nd 49th 37th 40th (n=3) Hybrid (n=10) 4th 2th 76th 57th 8th 9th (n=3) Hybrid (n=10) 1st 1st 3rd 4th (n=3) Hybrid (n=10) 2nd 3rd 36th 37th
21-point average 1
SP-based
0.9
PI-based (n=3) PI-based (n=10) Hybrid (n=3) Hybrid (n=10)
0.8
Precision
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0
0.1
0.2
0.3
0.4
0.5 0.6 Recall
0.7
0.8
0.9
1
Fig. 4. Average precision-recall graph for five concepts by Syntactic Pattern (SP) based vs. Property Inheritance (PI) based vs. Hybrid hypernym extraction
6
Conclusion and Future Work
In this paper, to achieve high recall and not low precision in automatic acquisition of hyponymy relations, we assumed property inheritance from a concept to its hyponyms to be necessary and sufficient conditions of hyponymy relations, and then we proposed a method to mine the Web for hyponyms and hypernyms of a target concept based on property inheritance. The experimental results show that our hyponymy extraction based on property inheritance is superior to a commonly-used hyponymy extraction based on such a lexico-syntactic pattern as “is a/an”, and the hybrid of them is more superior to each of them. In the future, we plan to refine a method to extract typical properties for each concept, to invent how to find the optimal number of typical properties for the calculation of property inheritance from a target concept to its hyponym candidate, and moreover, to utilize coordinate terms [17] for more robust method. Acknowledgments. This work was supported in part by (1) a Grant-in-Aid for JSPS (Japan Society for the Promotion of Science) Fellows (Grant#: 1955301); (2) a MEXT (Ministry of Education, Culture, Sports, Science and Technology) Grant-in-Aid for Scientific Research on Priority Areas: “Cyber Infrastructure for the Information-explosion Era,” Planning Research on “Contents Fusion and Seamless Search for Information Explosion” (Project Leader: Katsumi Tanaka, A01-00-02, Grant#: 18049041); and (3) the Kyoto University Global COE (Center of Excellence) Program: Informatics Education and Research Center for Knowledge-Circulating Society (Program Leader: Katsumi Tanaka).
References 1. Mandala, R., Takenobu, T., Hozumi, T.: The use of WordNet in information retrieval. In: Proceedings of the COLING-ACL Workshop on Useage of WordNet in Natural Language Processing Systems. (1998) 31–37 2. Hattori, S., Tezuka, T., Tanaka, K.: Activity-based query refinement for contextaware information retrieval. In: Proceedings of the 9th International Conference on Asian Digital Libraries (ICADL’06), LNCS Vol.4312. (2006) 474–477 3. Fleischman, M., Hovy, E., Echihabi, A.: Offline strategies for online question answering: Answering questions before they are asked. In: Proceedings of 41st Annual Meeting of the Association for Computational Linguistics (ACL’03). (2003) 1–7 4. Hattori, S., Tezuka, T., Tanaka, K.: Mining the web for appearance description. In: Proceedings of the 18th International Conference on Database and Expert Systems Applications (DEXA’07), LNCS Vol.4653. (2007) 790–800 5. WordNet: http://wordnet.princeton.edu/perl/webwn (2007) 6. Miller, G., Beckwith, R., Fellbaum, C., Gross, D., Miller, K.: Introduction to wordnet: An online lexical database. International Journal of Lexicography 3(4) (1990) 235–312 7. Wikipedia: http://www.wikipedia.org/ (2007) 8. V¨ olkel, M., Kr¨ otzsch, M., Vrandecic, D., Haller, H., Studer, R.: Semantic wikipedia. In: Proceedings of the 15th International Conference on World Wide Web (WWW’06). (2006) 585–594 9. Hearst, M.A.: Automatic acquisition of hyponyms from large text corpora. In: Proceedings of the 14th International Conference on Computational Linguistics (COLING’92). (1992) 539–545 10. Caraballo, S.A.: Automatic construction of a hypernym-labeled noun hierarchy from text. In: Proceedings of the 37th annual meeting of the Association for Computational Linguistics on Computational Linguistics (ACL’99). (1999) 120– 126 11. Sanderson, M., Croft, B.: Deriving concept hierarchies from text. In: Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. (1999) 206–213 12. Morin, E., Jacquemin, C.: Automatic acquisition and expansion of hypernym links. Computers and the Humanities 38(4) (2004) 363–396 13. Shinzato, K., Torisawa, K.: Automatic acquisition of hyponymy relations from HTML documents. Japanese Journal of Natural Language Processing 12(1) (2005) 125–150 14. Kim, H., Kim, H., Choi, I., Kim, M.: Finding relations from a large corpus using generalized patterns. International Journal of Information Technology 12(7) (2006) 22–29 15. Ruiz-Casado, M., Alfonseca, E., Castells, P.: Automatising the learning of lexical patterns: An application to the enrichment of wordnet by extracting semantic relationships from wikipedia. Data & Knowledge Engineering 61(3) (2007) 484– 499 16. Yahoo!: http://api.search.yahoo.co.jp/WebSearchService/V1/webSearch (2007) 17. Ohshima, H., Oyama, S., Tanaka, K.: Searching coordinate terms with their context from the web. In: Proceedings of the 7th International Conference on Web Information Systems Engineering (WISE’06), LNCS Vol.4255. (2006) 40–47
