A Mandarin to Taiwanese Min Nan Machine Translation System with Speech Synthesis of Taiwanese Min Nan

A Mandarin to Taiwanese Min Nan Machine Translation System

with Speech Synthesis of Taiwanese Min Nan

Chuan-Jie Lin and Hsin-Hsi Chen

Department of Computer Science and Information Engineering National Taiwan University

Taipei, TAIWAN, R.O.C.

E-mail: cjlin@nlg2.csie.ntu.edu.tw, hh_chen@csie.ntu.edu.tw

Abstract

This paper presents a design of a Mandarin to Taiwanese Min Nan (abbreviated as Taiwanese hereafter) machine translation system. It is the first machine translation system which focuses on these two languages. An input Mandarin sentence is segmented, tagged and translated word by word according to the part of speech of each word. The candidates come from a Mandarin-Taiwanese dictionary. If more than one candidate exists, an example base is consulted. When a Mandarin word is not found in the Mandarin-Taiwanese dictionary, it is translated according to a Single-Character dictionary. The output can be in terms of either speech or text. For speech output, we also deal with the tone sandhi problem in changing the tone of each Taiwanese syllable. Because the mapping between Taiwanese syllables and Chinese characters is still a subject of disagreement, and the phonetic spelling coding systems are not familiar to everybody, speech output is useful but is also a challenge.

Keywords: Machine Translation, Mandarin, Speech Synthesis, Taiwanese, Min Nan, Tone Sandhi.

1. Introduction

Mandarin and Min Nan are two languages commonly used around world. According to Ethnologue (Grimes, 1996), the populations1

of Mandarin- and Min Nan-speaking people are 885,000,000 and 49,000,000, respectively. They are ranked 1 and 21. In Taiwan, these two languages are also two of the major languages. This paper will study these two languages and present a Mandarin to Taiwanese Min Nan (abbreviated as Taiwanese hereafter) machine translation (MT) system, including a speech synthesizer of Taiwanese.

Taiwanese has 7 tones (i.e., upper even, rising, upper departing, upper entering, lower even, lower departing, and lower entering), 17 initials, and 75 finals. The major problem with Taiwanese processing is its representation. Several methods have been proposed for representing Taiwanese syllables. One is to relate syllables to Chinese characters. But for some frequently used words, we cannot write down their corresponding characters without any doubt. The following shows a typical example:

beh (want to, will) "Ú beh Pâ " (I am going to school)

Some people use "5" as its representation; some use ""; some even use "”". "5" and "" do represent the idea of "want to". However, "5" is read as "iau3" and "" is read as "iok8". It is hard to relate them to "beh". On the other hand, "”" does have a colloquial reading of "beh", but it means "fortune-telling". That is, it is irrelevant to "want to". Another way to represent Taiwanese is to spell out the syllables. Similarly, there exist many coding systems, e.g., the International Phonetic Alphabet, the Missionary Romanization, and systems modified from the Chinese Phonetic Alphabet. Different from character representation, these coding systems can be transformed into one another. Therefore, they are equivalent. This paper will adopt the Missionary Romanization system. Table 1 ({! , 1993) and Table 2 (ÄB¬ô» , 1988) list the initials, the finals, and the tones of Taiwanese.

This paper is organized as follows. Section 2 presents the architecture of our Mandarin-Taiwanese machine translation system. Section 3 focuses on the lexical selection problems. Two strategies are proposed. Section 4 evaluates the performance of lexical selection and discusses it. Section 5 deals with

Table 1. The Initials and the Finals of Taiwanese Initials p, ph, b, m, t, th, l, n, k, kh, g, ng, h, ch, chh, s, j

Finals a, ai, au, am, an, ang, e, eng, i, ia, iau, iam, ian, iang, io, iong, iu, im, in, ou, o, ong, oa, oai, oan, oe, u, ui, un, aN, aiN, eN, ouN, iN, iaN, iauN, iuN, oaN, oaiN, m, ng, ap, at, ak, ek, ok, iap, iat, iak, iok, ip, it, oat, ut, ah, auh, eh, ih, iah, iauh, ioh, iuh, ouh, oh, uh, oah, oeh, ahN, auhN, ehN, ihN, iahN, oaihN, mh, ngh

Table 2. A Representation of the Tones of Taiwanese

Tone number Tone Representation Modified Representation

1 upper even no mark, e.g. am am

2 rising ´, e.g. ám am2

3 upper departing `, e.g. àm am3

4 upper entering no mark, e.g. ap ap

5 lower even ^, e.g. âm am5

6 rising

7 lower departing ¯ , e.g. am am7

8 lower entering ', e.g. ap ap8

speech synthesis. The tone sandhi phenomena are touched on. Appendix lists some translation results obtained using our Mandarin-Taiwanese machine translation system. Section 6 demonstrates an experimental system used on the Web. Section 7 offers concluding the remarks.

2. Architecture of a Mandarin-Taiwanese Machine Translation System

Many different approaches, e.g., rule-based (Bennett and Slocum, 1985), statistics-based (Brown et al., 1990), example-statistics-based (Nagao, 1984), and knowledge-statistics-based (Mitamura et al., 1991; Baker et al., 1994) ones, have been proposed for MT design. They have both advantages and disadvantages (Chen and Chen, 1995; Chen and Chen 1996). No matter which kind of MT model is adopted, it has to capture the lexical and structural differences between the source and target languages. A typical transfer-based MT system is composed of a parser, a lexical transfer, a structural transfer and a generator. The parser analyzes the source sentences, and generates parsing trees. The lexical transfer selects the lexical items. The structural transfer captures the mapping of structures between source sentences and target sentences. The generator produces the target sentences.

Chao (1968) noted that the greatest degree of uniformity is found among all the dialects of the Chinese language in terms of grammar. Based on Chao's theory, we postulate that Mandarin and Taiwanese have similar structures. An example is shown as follows:

Mandarin: 4 Þú Ã 2 ó

Taiwanese: ² Þ2 Ã  ó

The above two sentences have the same word order. In our model, we focus on lexical selection. Source sentence analysis and the structure mapping between Mandarin and Taiwanese are neglected under the postulation. Figure 1 shows the overall architecture of our Text-to-Speech Mandarin-Taiwanese machine translation system.

Figure 1. Architecture of a Mandarin-Taiwanese Machine Translation System Input Sentences

Word Segmentation & POS Tagging

Word by Word Translation Unfound Word Translation Changing Tones Chinese Word Dictionary Mandarin-Taiwanese Dictionary Single Character Dictionary Tone Changing Rules Corpus Speech Output Text Output

An input sentence in Mandarin is first segmented, then tagged, and finally translated word by word according to the part of speech of each word. The candidates come from a Mandarin-Taiwanese dictionary. If more than one candidate exists, an example base is consulted. When a Mandarin word is not found in the Mandarin-Taiwanese dictionary, it is translated according to a Single-Character dictionary. The output can be in terms of either speech or text. For the speech output, we also deal with the tone sandhi problem in changing the tone of each Taiwanese syllable.

3. Lexical Selection

3.1 Mandarin-Taiwanese Dictionary

A sentence in Mandarin cannot be translated into one in Taiwanese character by character. For example, we do not say Õñ (today) as "kin-thiN" (Õñ) in Taiwanese, but "kin-a2-jit8" (Õ) ) instead. Thus, a word segmentation system is indispensable. We adopt the maximum matching criterion, 13 morphological rules selected from Lin’s work (Lin, Chiang, and Su, 1993) and a lexicon trained from CKIP Corpus (CKIP, 1995) to identify word boundaries. The reason why we do not use the Mandarin-Taiwan dictionary for word segmentation is that the dictionary we adopt is not quite complete. Table 4 shows that 109 of the 1,000 most frequently used Mandarin words are not collected in this Mandarin-Taiwanese Dictionary. To get better segmentation and be easier integration with a proper noun identifier, we adopt a well-established word segmentation system.

In our experiments, the Taiwanese-Mandarin Dictionary (WAðLYÛ, {! , personal communication), developed by Professor Robert L. Chang, was employed in the lexical selection. There are 38,287 Mandarin words and 39,865 Taiwanese words. In total, there are 53,500 entries. An entry contains the corresponding Mandarin and Taiwanese words in Chinese characters, the phonetic spelling of the Taiwanese word, the POSes of these two words, the frequency of the Mandarin word, and some notes about the accent and other information.

The dictionary is organized in 8 columns, from A to H. Column A contains some information about Taiwanese words. For example, ‘x’ denotes that the Taiwanese word in this entry is rarely seen. Column B shows the phonetic spelling of the Taiwanese word using a modified Missionary Romanization system. Column C is the Chinese character representation of the Taiwanese word. Column D shows

the corresponding Mandarin word, and column E its frequency. Columns F and G are the POSes of the Mandarin word and the Taiwanese word, respectively. The last column contains some other notes. For example, [2:û] means that it is from the accent spoken in Southern Taiwan, while [2:>] refers to from the accent of Northern Taiwan. The notes in columns A, G and H are still incomplete. Therefore, we use only B, C, D, and F columns. The following examples illustrate the format of an entry: B C D F {tek-pai5-a2% {.è2% d%.+% <Na {sau3-se% {ß)% 2d%.ßä% <N> {sau3-soe% {ß)% 2d%.ßä% <N> {kui-nng5-lit8% {՟% d%‹ú‰
% <D

In the dictionary, a Mandarin word may have more than one Taiwanese translation. For example, the word "A" may be translated as "e7", "e7-tang3", "e7-hiau2", "e7-tit-thang", or "hoe7". Some are synonyms (e.g., "e7-tang3" and "e7-tit-thang"), and some are not. This is so-called lexical disambiguation problem. Table 3 shows the distribution of Mandarin words and their possible Taiwanese translations in the Mandarin-Taiwanese dictionary.

Table 3. Translation Distribution of the Words in the Dictionary

# of Candidates 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 24 Dictionary 0 29465 5445 1906 791 340 149 68 53 30 12 6 5 6 2 5 3 1 Top-1000 in CKIP 109 418 179 91 78 49 33 10 10 9 4 1 2 2 2 1 1 1

The first row depicts the corresponding number of Taiwanese candidates the Mandarin words have. For example, a total of 29,465 Mandarin words have only one Taiwanese translation each; 5,445 Mandarin words have two translations and so on. The second row shows similar statistics, but it considers part-of-speech information. When part-of-speech is not considered, 76.06% of the words have only one candidate, and a Mandarin word has 1.4 Taiwanese translations on average. At first glance, lexical selection seems to be easy. But we further compute another statistic. Here, we consider the 1,000 most frequently used Mandarin words, which were retrieved from Academia Sinica Balanced Corpus (CKIP, 1995). The second row of Table 3 shows the distribution of these 1,000 words.

The average number of possible candidates increases to 2.49. If we further consider the frequency of these words used in the balanced corpus, i.e., if Formula (1) is used, the average number of candidates increases to 3.51. This number shows that lexical selection is not a trivial problem.

( )

( )

( )

∑

∑

where Q is the set of 1,000 most frequently used Mandarin words collected in the Mandarin-Taiwanese dictionary, CD(w) denotes the number of candidates of w, and fCKIP(w) is the number of occurrences of w in the CKIP

Corpus.

3.2 Part of Speech

We adopt the following two criteria to select a suitable Taiwanese lexical item from the candidates.

(a) Part-of-speech;

(b) Corpus-based Method.

Table 3 shows that the most frequently used Mandarin words have 3.51 candidates in the Mandarin-Taiwanese dictionary on average. Thus, the first step is to reduce the number of candidates. Part of speech information is useful. For example, the word "V" has several POSes: Caa (and), VC (to be with), Na (heel), etc. They correspond to different Taiwanese words. When it means "and" (Caa), its translation is "kah"; when it means "to be with" (VC), it is translated into "toe3"; and into "kin" when it means "the heel" (Na). In our dictionary, a column records the parts of speech of Mandarin words. We employ this information to reduce the number of candidates. Table 4 verifies this key point. The average number of candidates is reduced to 1.98 words when parts of speech are considered, and to 2.27 when Formula 1 is adopted. That is better than the word-information-only-method (2.49). We further proceed an investigation. Here, the top 1000 <Word, POS>-pairs are examined. Table 5 shows the statistics.

Table 4. Translation Distribution of <Word, POS>-Pairs

# of Candidates 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 20 Dictionary 0 30964 5453 1806 714 295 117 58 38 18 9 3 4 2 1 4 1 1 Top-1000 in CKIP 1076 590 249 98 61 31 23 10 6 6 3 0 0 0 0 1 0 0

Table 5. The Number of Candidates of the 1,000 Most Frequently Used <Word, POS>-Pairs

# of Candidates 0 1 2 3 4 5 6 7 8 9 10 15

# of <Word,POS>-Pairs 239 402 186 69 44 24 16 9 3 4 3 1

The average number of candidates becomes 2.02 and 2.28, respectively.

Some translation rules are based on parts of speech. For example, the Mandarin word "+" has two meanings, i.e., he and his, and their corresponding Taiwanese words are different ("i" vs. "in"). The following examples clarify this point:

(1) 4 2 /k (He is happy.) i chin hoaN-hi2

(2) 4  2 /k (His father is happy.) in a-pa3 chin hoaN-hi2

A translation rule is shown below:

IF Mandarin word="4" and its POS="Nh", THEN IF the POS of the following word is "Na" or "Nc",

THEN the corresponding Taiwanese is "in"

ELSE the corresponding Taiwanese is "i"

3.3 Corpus-based Method

Part of speech information eliminates some of the candidates. However, it is still possible that more than one candidate has the same part of speech. In other words, we have to disambiguate further. Context information is necessary. The co-occurrence relationships between Taiwanese words are important. We can count how many times a word appears with other words and use the statistics to disambiguate the word usage. That is the so-called Markov model. However, a large scale of segmented Taiwanese corpus is needed to compute reliable statistics. Unfortunately, this is unavailable at present. We simplify the high-gram model to a unigram. That is, only the frequency of a word is employed. We always select high frequency words.

The statistics were computed based on the book "f֛‹MYîNWÖðÿ YPB" by Professor Robert L. Cheng ({!2, 1989). This book originated from" S->֍Y" ()à, 1980), which lists frequently used Mandarin words and example sentences. Professor Cheng adds the most likely corresponding Taiwanese translations of each Mandarin word, and the corresponding Taiwanese sentences of

the Mandarin example sentences. Our approach is as follow: if a Taiwanese word appears in most of the example sentences of a Mandarin word, then the Taiwanese word is regarded as the correct translation of the Mandarin word. Of course, part-of-speech information is also considered. We segment and tag the Mandarin sentences first, and then extract the information we need. After the information is extracted, we rearrange the order of the entries in the dictionary. Since we prefer such a Taiwanese to be the translation of the Mandarin word with this part of speech, we simply set this Taiwanese word in the first entry of this Mandarin <Word, POS>-pair. During translation, we just choose the first entry of the matched <Word, POS>-pair. 3.4 Translation of Unfound Words

Table 4 shows that even the 1000 most frequently used words may not all be found in the Mandarin-Taiwanese dictionary, not to mention proper names such as those of persons, companies, and organizations. In order to deal with these unfound words, we have developed a Single-Character dictionary. It contains 5,680 single Chinese characters and their readings in Taiwanese. For example, "N" is read as "chu" in Taiwanese. These characters contain the 5,401 most frequently used characters as proposed by The Ministry of Education (é ô, 1979) as well as those characters which may be used in Chinese names and transliterated foreign names (Chen and Lee, 1996). The readings were selected from "fWÖO" (oÐ6, 1992). Each character may have literary readings and colloquial readings. For Example,

Literary Reading Colloquial Reading

I chhian3 chhiaN3

 sui2 chui2

4. Evaluation of Lexical Selection 4.1 Test Results

The test data came from the Evening News Broadcast of the Chinese Television Station (CTS) on October 16, 1996. The CTS Evening News was broadcast in both Mandarin and Taiwanese simultaneously. There were 789 Mandarin total sentences and 6,484 total words in our test data, and 758 Taiwanese sentences and 5,916 words in the Taiwanese version. A total of 31 Mandarin sentences did not have any corresponding Taiwanese sentences. Thus, these sentences were neglected in the tests.

Table 6. Precision and Recall Rates in Lexical Selection (MAT: Number of Words Matched; R: Recall; P: Precision)

Answer Correctly Seg-Tagged Input Plain Text Input

Set A Set B MAT R P Set C MAT R P

No POS 5916 6323 2790 47.16 44.12 6639 2619 44.27 39.45 POS 5916 6323 2738 46.28 43.30 6639 2443 41.29 36.80 POS + Corpus 5916 6323 3204 54.16 50.67 6639 2848 48.14 42.90

The Chang-Chou Accent and Chuan-Chou Accent are two of the accents of Taiwanese. In our experiment, we set the accent option to Chuan-Chou first. The experimental results are listed in Table 6. A denotes the correct answer in the test data. B and C denote the translation results of our system. To show the influence of a word segmentation system and a part-of-speech tagging system, B received input which was correctly segmented and tagged by hand, and C received input from our word segmentation and part-of-speech tagging systems. The first row depicts the precision rate and recall rate when POS was not considered. We randomly chose a translation from all the candidates. The results listed in the second row are those obtained when we employed POS information but randomly chose a translation from all the candidates of the same POS. The third row shows the results obtained when both POS and corpus-based knowledge were considered.

Table 6 shows that the correctness of the word segmentation and POS tagging systems indeed affected the lexical selection performance. The precisions of the segmentation and tagging systems were 88.0% and 78.69%, respectively. The performance of the segmentation system was not perfect, but this did not affect the translation performance so much as we expected. This is because that proper noun identification had not yet been integrated, and these proper nouns were not easily translated. This will be discussed later in Section 4.2.3.

Surprisingly, the recall rate and the precision rate dropped 1-3% when POS information was considered. In both cases B and C, integrating POS information and corpus-based knowledge worked better than did the other two methods. When we analyzed why the POS approach was a little worse than the No-POS approach, we found that the POS information of in our Mandarin-Taiwanese dictionary was incomplete. Thus, we did an investigation. This time, we focused only on those Mandarin words whose <Word, POS>-pairs appeared in the dictionary. Table 7 shows the results.

Table 7. Examination of 4,609 Pairs Found In the Dictionary number of words which had no correct candidates 2837 number of words which had only one candidate 1641 number of words which had multiple candidates 1196 number of correctly selected translations in multiple candidates 1009

correct rate 84.36%

During translation, a total of 4,609 <Word, POS>-pairs were found in the Mandarin-Taiwanese dictionary, but only 2,837 appeared in the answer set. Among these 2,837 pairs, 1,641 had only one candidate (lexical selection is trivial in this case), and 1,196 had more than one candidate. Among these 1,196 pairs, 1,009 were correct, so the correct rate was 84.36%. This finding supports our point.

4.2 Discussion

Besides the propagation errors from segmentation and tagging, we found several other kinds of errors. They will be discussed in the following subsections.

4.2.1 Errors from the Test Data

Our test data were selected from the CTS Evening News. It is a live broadcast, so both Mandarin and Taiwanese broadcasters speak at the same time. It is interesting that the Taiwanese broadcaster always spoke more slowly and it took her a little more time to speak out the same sentence which Mandarin broadcaster said. Thus, the Taiwanese broadcaster tended to eliminate words and sentences to keep up with the Mandarin broadcaster. The following shows an example. The symbol (C) denotes the correct answer, and (D) denotes our translation result.

(3) Ìx 1O ýÐ ù- í Q L| Ý m ì½ ån ‘ K

(C) ia2-m7-ko3 heng5-cheng3 pou7-bun5 * * * * iu5-oan5 si7 chhi5-tioh8 siong-tong po2-liu5 e5 thai7-tou7

(D) m7-koh heng5-cheng3 pou7-bun5 tui3-u5 che ko3 kian3-gi7 iu5-goan5chhi5 siong-tong po2-liu5 tek thai7-tou7

The phrase "ðäN+I" did not appear in the Taiwanese sentence. That lowered the precision rate of our system. In contrast with the above situation, some phrases were added in the Taiwanese sentences but had no corresponding phrases in the Mandarin sentences. For example,

(4) ¢ñ (ãÝ ç `Gq (© LQ û_) ×c ¬ï ½Ï ¥ ‘ c

(C) cha2-khi2 chip8-hap8 ti7-leh tai5-pak-chhi7 se3-bou7 kong2-tiuN5 thau5-cheng5 chiap-siu7 kau-thong kiok8-tiuN2 ho7-tan5-tan3 e5 kiam2-iat8

(D) cha2-khi2 chiap-siu7 kau-thong kiok8-tiuN2 ho7 tin5 tan3 e5 kiam2-iat8

The phrase "ÚÔÞW>h m0ò8" was eliminated by the Mandarin broadcaster. That lowered our recall rate.

Table 8. Comparison between the Chang-Chou and the Chuan-Chou Accents CHUAN-CHOU CHANG-CHOU EXAMPLE (CHUAN-CHOU,CHANG-CHOU)

l j Î (liat8, jiat8); Ï (lin5, jin5)

e oe o (ke3, koe3); · (be2, boe2)

oe e ž (boe2, be2); Q (koe, ke)

iN eN j (piN5, peN5); À (siN3, seN3)

ui oe 7 (mui5, moe5); ? (hui3, hoe3)

u i ­ (lu2, li2); T (tu7, ti7)

un in ¿ (kun7, kin7);  (kun, kin)

Furthermore, the accents also lowered the precision rate and recall rate. Recall that we adopted the Chuan-Chou accent in the preliminary test, but the reporter still read some words using the Chang-Chou accent. Table 8 shows comparisons between these two accents.

Another inconsistency was that a word could be spoken differently in the broadcast. For example,

(5) Jì 9y óÿC Þú Z âÁ *õ ' ¤y Fó ç

`G *J

(C) chong2-thong2-hu2 chu-cheng3 ng5-siau3-kok kin-a2-jit8 thau3-cha2 in-ui7 kam2-mou7 in2-hoat hi3-iam7 ti7 tai5-pak eng5-chong2 koe3-sin

(6) óÿC “ - ‡o û œ‹

(C) ng5-siau2-kok si7 ti7 bin5-kok cheng5 chap8-it-ni5 chhut-si3 ti7 The name "êö:" was read in two different ways. All the above problems show that a high quality bilingual evaluation is hard to prepare. In the experiments, we used a strict measure, so the test data influenced the precision rate and the recall rate severely.

4.2.2 Errors from Dictionary

Some errors resulted from dictionary. Consider the following example:

(7) ?- š Š ‘ J“ ȳ   Nh (âº)

(C) chi3-u5 chin2-sou2 * * i-seng hou7-su7 si7 m7-si7 u7 ke3-sit chek-jim7

(D) chi3-u5 chin2-sou2 lai7 e5 sian-siN hou7-su7 u7-but8-u7 ke3-sit Here, "kŠ" was translated as siN" in our system. The Taiwanese word "sian-siN" is an old-fashioned word, a respectful way to address a doctor. The inappropriate translation dragged down our performance.

4.2.3 Proper Noun Translation

Proper nouns are not easy to translate. Each character has literal reading and colloquial reading. It is not clear when to use which reading. The most famous case is "¬". "<¬>" is read as "lim5-toa7-hai2", but "¬ @Ž" is read as "tai7-kiat8 li2-heng5-sia7".

Table 9 shows our results for names of persons, locations, and companies. Most of the errors were due to incorrectly reading selection. Some of them were eliminated by the Taiwanese broadcaster (as the case discussed in Section 4.2.1). Other errors included mistakes in the Single Character Dictionary, accent, word-structure, error-reading or inconsistent-reading of the broadcaster. Location names were much easier to translate because many of them were collected in the bilingual dictionary.

Table 9. Investigation on Proper Noun Translation Incorrect

Names of Correct

Selection Elimination Others

Persons 67 45 17 38

Locations 147 15 13 10

Companies 28 16 2 0

4.2.4 Quantifier Problem in Translation

The translation of quantifiers is a common problem in machine translation. So far, we have not dealt with this problem. The following shows an example:

(8) Mandarin: ‹ â Ø&

Taiwanese: ‹ Ø&

The quantifier for an umbrella in Taiwanese is "" instead of "Ù". But "‚ Ù ö Ò" is still translated as "‚ Ù öÒ".

4.2.5 Other Problems

Word-by-word Translation is not always acceptable. Consider the following two examples.

(9) º_VC ‘_Di ë_Nc æ6_Nc D~_Na ½·_Na (C) liah8 tioh8 lam5-tau5 te7-khu siu2-ok hun7-chu2 (D) liah8 liau2 lam5-tau5 te7-khu siu2-ok le2

(10)<*_Cbb \_P d_Neu Q_Nf “_Na Õ_Nd º_VC (C) ji7-chhiaN2 chiong si3 e5 lang5 tong5-si5 liah8=khi0-lai0 (D) ji7-chhiaN2 chiong su3 ko3-jin5 siang5-si5 liah8

In example (9), "±ù" is translated as "liah8". In example (10), "±ù" should be translated as " liah8=khi0-lai0". This is because if we place "liah8" after the target, we have to say, e.g., "±ù +" (to arrest him), "chiong i liah8=khi0-lai0" (‰ + ± ù) or "kah8 i liah8=khi0-lai0" (Ù + ±ù).

5. Speech Synthesis 5.1 Tone Sandhi

Tone sandhi is a common problem in tonal languages like Chinese. In Mandarin, when two or more third-tone characters are put together, the preceding third-tone characters is read using the second tone. In Taiwanese, there is also a tone sandhi problem. However, it is quite different from the one in Mandarin.

The tone used to read a character individually is called its original tone. When more than one character is spoken, we always change the tones of the preceding characters. Consider the following example:

(11) Mandarin: 4 Þú Ã 2 ó

Taiwanese: ² Þ2 Ã  ó

Original tones i kin-a2-jit8 sim-cheng5 chin ho2 Changed tones i7 kin7 a jit8 sim7 cheng5 chin7 ho2

he today mood very is-good

In the above example, the last characters of "ÿº" and "Õ) " retain their original tones. The word "ê" is at the end of the sentence, so its tone remains unchanged, too. The tones of the other characters are changed.

In the Mandarin-Taiwanese dictionary, the phonetic spelling of a word follows the original tones. During speech synthesis, we have to change the tones in order to

speak fluently. When and how the tones should be changed depends on situation and accent. Section 5.2 will discuss this topic in detail.

5.2 The Rules 5.2.1 General Rules

Figure 2 shows general rules for the Chang-Chou accent and Chuan-Chou accent ({ !2, 1993). The tone numbers are shown in Table 2. P, t, and k are stop endings, and h is a glottal stop ending. The rules for the Chang-Chou accent and Chuan-Chou only differ a little. In the Chang-Chou accent, the lower even tone (the 5th tone) changes to the lower departing tone (the 7th tone). In contrast, the lower even tone (the 5th tone) changes to the upper departing tone (the 3rd tone) in the Chuan-Chou accent.

Figure 2. Tone-Changing Rules

In our study, the upper entering tone did not change to the lower entering tone. The changed tone was most like a rising tone except that it was a checked sound while the lower entering tone shared the same tone value with the lower departing tone. For better speech output, we added another tone number 9, to indicate that the tone changed from the upper entering tone. The cause might have been the accents again because in some accents their tones 8 and 9 (defined by us) are the same (i.e., the 8th tones in different accents are different).

5.2.2 Tone-Changing Rules before "a2" (ò'JÌõ)

The character "a2" ()) is a special character in Taiwanese. It is a suffix and is always combined with other characters to form a word. For example,

(12) Á “i2” Á2 “i2-a2” i a2

(13) : “koa-hi3” :2 “koa-a2-hi3” koa7 a hi3 Chang-Chou Accent Chuan-Chou Accent 1 2 3 5 7 4h 8h 4p,t,k 8p,t,k 1 2 3 5 7 4h 8h 4p,t,k 8p,t,k

The tone of the character before "a2" also changes, but the rules are different from the general rules mentioned in the previous section. We adopted the analyses (o¨, 1991) shown in Table 10.

Table 10. Tone-Changing Rules before "a2"

Original 1 2 3 4 5 7 8

Chang-Chou 7 2 2 9 (ptk) 2 (h) 7 7 8 (ptk) 7 (h) Chuan-Chou 7 2 2 9 (ptk) 2 (h) 3 7 8 (ptk) 7 (h)

5.2.3 Tone-Changing Rules for Neutral Tone (ïOJ'JÌõ)

If a character is followed by another character of which tone is neutral, it retains its original tone. However, the neutral tone will change according to the following two tone-changing rules (¢cÏ, 1996).

(a) Solid tone value (§ÊJ): Some neutral-tone characters will retain a solid tone value. This tone value is a little lower than the upper departing tone. Characters such as "C" (lai0), "G" (khi0), and "ºŠ" (sian0-siN0) determine the tone-change, for example, "oC" ("koe3=lai0"), "ûG" ("bo5=khi0"), and "ºŠ" ("tan5=sian0-siN0"). The characters "o", "û", and "" retain their original tones.

(b) Change according to the preceding tone (¶ò'J): The tones of some neutral-tone characters will change according to the tones preceding them. The rules are shown below:

if the preceding tone is the upper even tone,

then the tone of the neutral-tone character change to the upper

even tone

else if the preceding tone is one of the upper tones, then it change to the upper departing tone else it change to the lower departing tone.

ˆ"e5", "ah0" and the pronouns are such characters. Examples are:

(14) A[ “sin=e5” “sin e”

(15) ÒÚ “hou7=goa2” hou7 goa7”

However, if a neutral tone character of this kind is not in the end of a syntactic structure, or if, as in (15), the pronoun is indeed combined with the following word, the tone will change according to the general tone-changing rules.

5.2.4 Triple-Character Adjectives (`ûÅÉY'JÌõ)

In Taiwanese, we sometimes will triple a one-character adjective (e.g., sour, "ÿ") to emphasize the situation (in this case, "ÿÿÿ"). This makes the sound more interesting. The tone of the second character of such a word will change according to the general tone-changing rules while the tone of the first character will change according to Table 11 (o¨, 1991).

Table 11. Tone-Changing Rules for the First Character of Triple-Character Adjectives

Original 1 2 3 4 5 7 8

Changed Tone 5 1 2 9(ptk)2(h) 5 5 *(ptk)5(h)

5.3 Evaluation

The same test data were used to evaluate the Tone Sandhi problem. To avoid other sources of error, we only considered correctly translated words. There were 5,576 such characters. The tones of 4,602 characters changed correctly. That is, the correct rate was 82.53%. After we examined the experimental results, we found the following types of errors:

(a) Within a phrase

Because the syntactic structures were not used, the tones of many nouns in noun phrases did not change while many verbs, in contrast, did change. An example is shown below:

(14)`Í_Nc qæ_Na YÄ_VC _Na (C) tai7 tiong7 chhi3 te7 tiong7 oe3 pe2 an3 (D) tai3 tiong chhi3 te7 tiong3 oe3 pe2 an3 (b) Object was a clause

The objects of some verbs were clauses. Since the clause behind such a verb was very long, its last tone did not change. For example,

(15)Æ_D ¾6_Cbb >_VF œn_VD _Na [_DE æM_Na  _Caa o[_Na 9_Na

(C) chia chin2 chit pou7 iau7 kiu5 the3 chhut9 ki7 bit seng3 e7 kong7 bu7 ia3 si3 kong kok9 hong5 e7 chu7 liau7

(D) iu3 ko2 chin2 li3 iau7 kiu3 the3 kiong7 ki7 bit siN3 e3 kong7 bu7 hek kok9 hong5 chu7 liau7

When a long clause was followed by the verb "5ú", it retained its last tone in the correct answer.

(c) Nominalized verbs like "Ðj", "ݛ", and "P»" retained their last tones.

(d) In predicate-object structure words like "|[" and "Ö\", the second character was like an object, so it tended to retain its tone.

6. An Experimental System on the Web

An experimental system is demonstrated on the web site:

http://nlg3.csie.ntu.edu.tw/group/cjlin/MTMT.html

Figure 3 illustrates the home page of this system. After giving an input sentence, users can choose which accent is preferred. Besides, if the input sentence is selected from classical literatures or poems, it is preferable in the literary readings. The option "Z" provides a selection "9". In the normal case, "\@" is chosen.

Figure 3. The Home Page of the Mandarin-Taiwanese MT System

There are three kinds of ways to display the output results. That is, Chinese characters, Missionary Romanization and tone-changed form. Figure 4 gives an example, which is in classical literature reading. It is the first three sentences of the poem "ñ<\" written by a famous Chinese poet la. We read it character by character. If this sentence is in colloquial reading, some characters will be translated into multi-character words. In that case, some will be translated wrongly. For example, "¬" is read as "toa7" instead of "tai7". Both examples in Figures 5 and 6 are in colloquial readings but in different accents. Figure 5 shows the example of Chang-Chou accent and Figure 6 is in Chuan-Chou accent. The characters "Q" and "·" have different readings, and the tones of "," and "ˆ" are changed differently.

7. Concluding Remarks

This paper has presented the first Mandarin-Taiwanese machine translation system in the world. We adopted an integrated approach in designing our MT system. In our system, a rule-based segmentation system divides a Chinese sentence into a sequence of tokens. A statistics-based tagger assigns a part-of-speech to each token. A lexical selection system chooses target lexical items. A speech synthesizer generates the speech output. The major problems we have tackled are the unknown word problem, lexical disambiguation problem, and tone sandhi problem.

Dictionary is one of the major knowledge resources in developing such a system. The coverage of the Mandarin-Taiwanese dictionary and the inconsistencies in the tagging result are major errors in lexical selection. Setting up a large-scale dictionary will be indispensable for future research. Because a large-scale segmented and tagged Taiwanese corpus is not available, we have adopted parts of speech, and preferences in selecting suitable lexical items. The association of lexical items has not been considered in this paper. However, context information is important for making lexical choices. Further experiments are needed to integrate context knowledge. The tone sandhi phenomena in Taiwanese are complex. Although this study employed several rules for capturing such phenomena, there were still many errors. Knowledge of syntactic structure does help. How to incorporate the parsing results will be investigated further.

Figure 5. An Example in the Chang-Chou Accent

Acknowledgements

We are very thankful to Professor Robert L. Cheng for providing us with the Taiwanese-Mandarin Dictionary.

References

Baker, K. et al. (1994), "Coping with Ambiguity in a Large-Scale Machine Translation System," Proceedings of COLING-94, Kyoto, Japan, 1994, pp. 90-94. Bennett, W. and Slocum, J. (1985), "The LRC Machine Translation System,"

Computational Linguistics, Vol. 11, No. 2-3, 1985, pp. 111-119.

Brown, P. et al. (1990), "A Statistical Approach to Machine Translation,"

Computational Linguistics, Vol. 16, No. 2, 1990, pp. 79-85.

Chao, Y.R. (1968), A Grammar of Spoken Chinese, University of California Press, 1968, pp. 13-14.

Chen, H.H. and Lee, J.C. (1996), "Identification and Classification of Proper Nouns in Chinese Texts," Proceedings of 16th International Conference on Computational

Linguistics, Copenhagen, Denmark, August 5-9, 1996, pp. 222-229.

Chen, K.H. and Chen, H.H. (1995), "Machine Translation: An Integrated Approach,"

Proceedings of the Sixth International Conference on Theoretical and Methodological Issues in Machine Translation, Leuven, Belgium, July 5-7, 1995,

pp. 287-294.

Chen, K.H. and Chen, H.H. (1996), "A Hybrid Approach to Machine Translation System Design," International Journal of Computational Linguistics & Chinese

Language Processing, 1(1), 1996, pp. 159-182.

CKIP (YÛ²‹) (1995), "ë)ÖۈÚÉîÜ+," ÄYŒôÛ²‹Õà K‚#95-02, Ä^ë), 1995.

Grimes, B.F. (1996), Ethnologue: Languages of the World, Thirteenth Edition, Summer Institute of Linguistics, Inc., Dallas, Texas, 1996.

Lin, M.Y., Chiang, T.H., and Su, K.Y. (1993), "A Preliminary Study on Unknown Word Problem in Chinese Word Segmentation," Proceedings of ROCLING-93, pp. 119-141, 1993.

Mitamura, T., Nyberg, E. and Carbonell, J. (1991), "An Efficient Interlingua Translation System for Multilingual Document Production," Proceedings of

Nagao, M. (1984), "A Framework of Mechanical Translation between Japanese and English by Analogy Principle," Artificial and Human Intelligence (Elithorn, A. Eds.), 1984, pp. 173-180. Äf¬$ô» (1988), Ö9î, Äf¬$ô» 7wŽ, >^, 1988, pp. 227-228. )à (1980), S->֍$Y, ODÈ Ë, >^, 1980. ¢™Ï (1996), "¬IûÖïy'æ­Ö[ˆæ," ¬IûÖnL¨Û N3, 1996, pp. 330-382. é2ô (1979), ›‹fî.», |Ä ´, W>, 1979. o'¨ (1991), ¬IûÖÖ[ñ, ¬ó7wŽ, W>, 1991. oÐ6 (1992), fWÖy,  R7wŽ, W>, 1992. {32 (1989), f֛‹w‰îxWÖðÿ‰€f, á7w 4ÝP, 1989. {32 (1993), ‰éWÖÜ$î•­ÌJ, %Ž, W>, 1993.

Appendix

Some of the results are included here to demonstrate the translation performance. Each set has one sentence in Mandarin, a corresponding Taiwanese sentence in the test data, and a translation output by our system (which received input from the word segmentation system and POS tagging system).

1. ½ÿ_VK [_Nc ¿s_D æL_VC v_Neu N_Nf _Na ½_D É· _VH ë_VAC

hi-bong7 hek8-su3 ian5-au7 nng7 ko3 geh8 khai-kang

hi-bong7 hek-su3 e7-tit ian5-au7 nng7 ko3 geh8 chai3 khai-kang ë

2. 'Ø_Cbb ²ø_Na ÇË_Ng ¿s_D j‹_VC ä_Nep v_Neu N_Nf  _Na ˆ_DE Ë_Na ½_D ˆN_VE ì_VAC

i2-pian7 tiau5-ia2 e7-tang3 li7-iong7 che nng7 ko3 geh8 e5 si5-kan ko2-chai3 pheng5-kou2

i2-pian7 tiau5-ia2 chi-kan e7-tang3 li7-iong7 che nng7 ko3 geh8 tek si5-kan chai3 pheng5-kou2 ì

3. @p_Nc ‰_D A_D œr_D  _P í%ô_Nc g_D ë)_VE yx_D H'_D æL_VC [_Nc É·_VH ˆ_DE ±_Na í_VAC

heng5-cheng3-iN7 chiong e7 sui5-si5 kah keng-che3-pou7 lai5 gian2-kiu3 si7-m7-si7 e7-tang3 ian5-au7 hek8-su3 khai-kang jit8-ki5

heng5-cheng3-oan7 chiong e7-hiau2 lip8-chit kah Keng-che3-pou7 lai5 gian2-kiu3 si7-m7-si7 e7-tang3 ian5-au7 hek-su3 khai-kang tek jit8-ki5 í

4. W^_Nc ‰_D e_D U__VJ ¬_VH 4_Nf ˆ_DE TäÅ_Na í_Na tai5-tian7 chiong sun2-sit toa7 pit e5 poe5-siong2-kim

Tai5-tian7 chiong e7-hiau2 sun2-sit toa7-pit e5 pe5-siong5-kim í

5. ßü_Cbb E_VF €Ý_VA ˆ_DE œ¸_Na ¤áÙ_Nb 2_Cab Š_Na ظ_VE k„_Na ÿ7_D €Ü_VA Ü+_VE ,_Na

sui-jian5 sin-chhiaN2 sek-hian3 e5 lip8-ui2 tiuN-chun3-hiong5 in jin7-ui5 chong2-thong2 eng3-kai ai3 kau3-teng5 soat-beng5

sui-jian5 sin-chheng2 sek-hian3 e5 lip8-ui2 tiuN-chun3-hiong5 teng2 lang5 lin7-ui5 chong2-thong2 eng3-kai kau3-teng5 seh-beng5 ,

6. 9y_Cbb k„ã_Nc ‚_Nh Ôü_D y_SHI '_P (p_Nc [Dô _Nc -»_VK 7Ù_VC í_VAC

tan7-si7 chong2-thong2-hu2 iu5-goan5 si7 i2 heng5-cheng3-iN7 hoat-bu7-pou7 tai7-piau2 chhut-sek8

tan7-si7 Chong2-thong2-hu2 it hong iu5-goan5 si7-i2 heng5-cheng3-oan7 Hoat-bu7-pou7 tai7-piau2 chhut-sek8 í

7. Þ_P [Ñ_Nc  _Caa pÑ_Nc þ_VC ˆ_Di P¬_VH uI_Na ˆ

_DE Bk„_Na yN_D H'_D ‚±_VG (pÆ_Na ˆ_Na YY

_Na ,_Na

ti7 hoat-kai3 kah cheng3-kai3 in2-hoat tiong7-tai7 cheng-gi7 e5 hu3-chong2-thong2 si7-m7-si7 e7-eng7=e kiam-jim7 heng5-cheng3-iN7-tiuN2 e5 bun7-toe5 chai7 hoat-kai3 kah cheng3-kai3 in2-hoat ah0 tiong7-tai7 cheng-gi7 e5 hu3-chong2-thong2 si7-m7-si7 e7-eng7=chit kiam-jim7 heng5-cheng3-oan7-tiong5 tek bun7-te5 ,

8. Ž_P œ¸_Na ¤áÙ_Nb 2_Cab Š_Na E_VF ¬[Ë_Na AI_Na 5P_VE L_Ng ,_Na

iu5 lip8-ui2 tiuN-chun3-hiong5 chiah-e5 lang5 sin-chheng2 tai7-hoat-koaN hoe7-gi7 kai2-soe7 liau2-au7

iu5 lip8-ui2 tiuN-chun3-hiong5 teng2 lang5 sin-chheng2 toa7-hoat-koaN hoe7-gi7 kai2-sek au7 ,

9. E_VF PÝ_VA ˆ_DE -»_Na èÊf_Nb “7_VC CY_VE ,_Na sin-chheng2 sek-hian3 e5 tai7-piau2 pang5-teng7-kok the5-chhut sun5-bun7 sin-chheng2 sek-hian3 e5 tai7-piau2 Pang5 teng7 kok te5-chhut sun5-bun7 , 10.[DôÆ_Na ú|ã_Nb €ä_D y_SHI -»_VC k„ã_Nc Ây_Caa

-»_VC (p_Nc ,_Na

hoat-bu7-pou7-tiuN2 liau7-cheng3-ho5 tau3-toe2 si7 tai7-piau2 chong2-thong2-hu2 ia2-si7 tai7-piau2 heng5-cheng3-iN7

hoat-bu7-pou7-tiong5 liau7-cheng3-ho5 tau3-te2 si7 tai7-piau2 Chong2-thong2-hu2 hoan5-si7 tai7-piau2 heng5-cheng3-oan7 ,