We extracted the lexical entries from the dictionaries. The numbers of the words are shown in Table 1.

We compared the coverage by checking the returnability of E-J-E matches, i.e., translating English to Japanese, and then back to English again (see Figure 1). We enumerated each Japanese translation word (J1) for the English index word (E1) in the English-Japanese dictionary, and looked up the corresponding entry in the Japanese-English dictionary. Then, we compared the English words of both end results (E1, E2), and analyzed whether E1 returns to E2 when we search for the same word as E1 among plural translations.

We also carried out the J-E-J match and made a comparison with E-J-E. However, it is redundant to describe the analysis about both E-J-E and J-E-J. So that, we mainly describe about E-J-E as long as we do not annotate specially.

Source word Matching word Terminal word

We classified the result of the correspondence into four types, as given below, and the number of each type is shown in Tables 2 and 3. Note that if several translation words are described per index word in dictionary 1, they are counted separately.

(a) returns exactly (entry J1 exists in J-E and E1 exists as a translation)

(b) returns to a part of the morphemes (entry J1 exists in J-E and E1+x exists as a translation)

(c) entry of the matching word exists in dictionary 2, but the source word and the terminal word differ (entry J1 exists in J-E but E1 and E2 differ)

(d) entry of the matching word does not exist in dictionary 2 (entry J1 does not exist in J-E)

Table 3 : Classification of J-E-J matching result The following sections show some examples of each type.

Generally speaking, only a limited number of Japanese translation words are shown in the commercial English-Japanese dictionary which are sufficient to understand the English word. For this reason, we can only get a few words by simply gathering the Japanese words for the translation word for each index word.

The Japanese-English dictionary shows a more marked tendency of the limitation of translation words than the English-Japanese dictionary. As seen in Table 1, the J-E dictionary not only has considerably fewer index words, but also considerably fewer translation words per index word (B/A) than the E-J dictionary. Therefore, the number of English translation words in the J-E dictionary seems to be too small. This is derived from an editorial policy in which the translation words were reduced intentionally and the example sentences were increased in order to prompt the users to avoid literal English translations, because the dictionary was published for human-reading. However, it is not suitable for obtaining a large vocabulary of English translation words. For example, Shirai & Yamamoto (2001) carried out matching between a Korean-English dictionary and a Japanese-English dictionary and reported a problem that different words are used to describe the same meaning, such as "occupation" and "business", and this caused unmatchings between some entries in J-E and K-E.

Thereupon we gained a large vocabulary in English by gathering the English index words out of the E-J dictionary, which can be grouped into paraphrasing word groups by matching the Japanese translation words. If we use the paraphrasing word groups, the problem in Shirai & Yamamoto (2001) can be resolved and it is fruitful to apply them into other matchings, such as sentence-matching.

We created the English paraphrasing word groups by gathering and grouping the English index words in the E-J dictionary according to the existence of the same Japanese word as a translation word. Consequently, we could obtain a greater variety of English words than we could by simply gathering the English translation words in the J-E dictionary. For example, in E-J-E matching, as shown in Figure 2, in the case where the Japanese matching word is "''shokugyou'"" if we simply gather E in J-E, we get only four words: "occupation", "profession", "trade" and "job". However, by adding E in E-J, we can get more words: "business", "calling", "career", etc.

On the other hand, when the Japanese matching word has a wide meaning, incompatible English words are classified in one group. Therefore, we excluded such groups from extraction into paraphrasing words. We defined the judgment for excluding or not as follows: if the English translation description in the J-E dictionary is classified into plural subsections, we regard the Japanese index word as having a wide meaning. For example, the translation of "saisef in J-E dictionary is classified into three subsections:

1: (produce newly) grow again 2: (use again) recycle 3: (playback music etc.) playback

Therefore, we exclude the matching of the word "saisei". As a result, we can omit the group that includes incompatible words such as "rebirth"and "replay".

In the same way, we also extracted Japanese paraphrasing word groups by J-E-J matching.

shokugyou => occupation, profession, trade, job

Paraphrasing words

playback => saisei rebirth => saisei regeneration => saisei replay => saisei reproduction => saisei saisei => 1:grow again saisei => 2:recycle saisei => 3:playback

Table 10 shows the amount of the extracted English paraphrasing words by E-J-E matching and Table 11 shows the amount of the extracted Japanese paraphrasing words by J-E-J matching. The number 6676 of the extracted groups in Table 10 seems quite large as against 1257 in Table 11. This is because the E-J dictionary has more entries than the J-E dictionary and because the entries in the E-J dictionary tend to be classified into more subsections than the J-E dictionary.

Num. of extracted groups_

Num. of words per group (E in J-E) Num. of extracted groups

Num. of words per group (J in E-J)

Table 11 : Amount of extracted Japanese paraphrasing words by J-E-J

Next, in order to see the quality of the extracted result, we chose 100 groups at random out of the English paraphrasing word groups and conducted an evaluation with a native speaker of English. In each group, we chose one English word (X) at random that came from J-E, chose one English word (Y) at random that entered into the group newly by adding words in E-J, and compared X and Y for meaning and part-of-speech. We evaluated the meaning of the X and Y pairs by classifying them into five types as follows:

A: The meanings of X and Y are exactly the same.

B: Y has a wider meaning than X.

C: Y has a narrower meaning than X.

D: The meanings of X and Y overlap partly.

E: Nonsense.

The evaluator was not familiar with the part-of-speech system. Therefore, we asked him/her to recall some sentences using X, replace X into Y, and judge whether the sentences become ungrammatical or not as an evaluation of part-of-speech. We said that he/she need not consider whether the meaning changes or not in evaluating part-of-speech.

O: Still grammatical.

X: Becomes ungrammatical.

?: Both cases are possible.

Note that we hid the Japanese matching words and let him/her evaluate only with English words.

In the same way, we also conducted an evaluation of Japanese paraphrasing word groups with a native speaker of Japanese.

Table 12 shows the evaluation of English paraphrasing words. Almost half of the pairs were classified into type A, and the sum of type A to D amounts to 80%. Ten Xes in type A seems to be a lot. This is because some of them had a different part-of-speech in English but were the same word in Japanese with a change in the inflection by adding functional words:

frivolous / levity : keihaku(+na)

Yet some of them seem to be misjudgments because the evaluator could not recall all usages written in the dictionary when he/she only saw the English words. For example, the next example was evaluated as "AX" but the correct answer is "A?". This is because the evaluator could recall only "surrender" as a verb in spite of the existence of a noun in the dictionary:

surrender / capitulation : koufuku

We also evaluated the excluded words (see the example of "saisef in Figure 2) in order to see the performance of the exclusion (see Table 13). There are 20 pairs of type A. This is because the pairs are not always incompatible even if they are excluded. In fact, the words in the excluded groups should be subdivided into further groups to get proper paraphrasing words. Therefore, if we can subdivide them in some way, we will obtain more paraphrasing words. However, we must devise something for subdivision because we cannot know which subsection in J-E the J in E-J corresponds to.

We also evaluated Japanese. Table 14 shows the evaluation of Japanese paraphrasing words. Table 15 shows the evaluation of the excluded words in order to see the performance of the exclusion.

There are still 20 pairs of type E in Table 12. We further investigated the dictionary entries of the word Y in each pair to see the cause, i.e., E-J of E-J-E matching.

If we want to obtain paraphrasing words accurately rather than widely, we can use subsection numbers to improve the accuracy. Generally, the usage with high frequency is placed first in a dictionary, and the usage with low frequency is placed rearward. Therefore, we can omit the uncommon usages by cutting off the rearward subsections. To see the effect of this cutting off, we classified the result in Table 12 into subsection numbers (see Table 16). The smaller the subsection number, the better the evaluation. Furthermore, if we want to get type A only, it is covered by subsections 1 to 3 in this example, i.e., 34+8+2=44.

We compared our paraphrasing word with WordNet, which is an English thesaurus, using the matching word "shokugyou" as an example (see Table 17). We picked some entries that have the meaning of "shokugyou" from WordNet. The words are divided into small hierarchies in WordNet, while the words are widely contained in the same group in our proposed method. In particular, the group of the proposed method contains words for which it is difficult for non-native speakers of English to distinguish whether there is a difference or not, such as "profession" and "work". Therefore, the proposed method is especially useful for sentence matching for non-native speakers of English.

On the other hand, words of large ambiguity (e.g., "walk", "way") and words of small ambiguity (e.g., "business") are contained in the same group. This means that the proposed method can be used to replace a word of large ambiguity with a word of small ambiguity.