Chapter 2 Unsupervised Joint Prosody Labeling and
2.5 Analyses of the Labeled Breaks and Prosodic
2.5.1 Analyses of the Labeled Break Types
Since the purpose of announcers’ broadcasting is to propagate information accurately to the audience relying exclusively on their audio perception, our well-trained informant skillfully manipulated as many segmental and prosodic cues as possible, such as clear and precise articulation, strategic variations in the fundamental frequency, volume, syllable length, and types of breaks. These prosodic information carried in the utterance speech, in turn, reflects the informant’s mental grammar, his/her Mandarin linguistic competence that determines when to form a semantically appropriate word chunk, a prosodic phrase, or a larger unit, and hence where and how long a break in an utterance should be so that the informant’s speech would sound natural, informative, and attention attracting to the audience.
As a research based on our informant’s speech data rich in the Mandarin prosodic cues, our break-type-labeling model also can generate appropriate break types consistent with native speakers’ psychological reality. To verify this point, we examined the relationship between some special groups of words/morphemes and their concurring break types that both our break-type-labeling model and the ordinary Mandarin native speakers would consistently produce. These special groups of words/morphemes include (1) affix morpheme; (2) DE; (3) Ng, Di, and T; (4) VE; (5) Caa and Cb; and (6) P [72]. The results are discussed in more detail as follows.
1. Set of Affix Morpheme
It is well-known that prefixes and suffixes are bound morphemes that attach to their preceding or following heads to form units of complex words. Since the resultant form after combining the head and the affix is a unit, it is reasonable to predict that
the breaks at the boundaries between the head and the affix tend to fall in B0 or B1 types. These phenomena were observed in our corpus. We found that some Mandarin Chinese mono-syllabic prefixes, such as bu- “不 not, un-, dis-, in-,” ke- “可 -able,”
wu- “無 not, -less, un-,without,” etc. [67,73], tend to join the following roots to form legitimate words as in bu-li “不利 unfavorable,” bu-fang-bian “不方便 inconvenient,”
ke-wu “ 可 惡 detestable,” ke-sing “ 可 行 feasible,” wu-sian “ 無 限 limitless,”
wu-shuang “無雙 unparalleled.” Similarly, by attaching mono-syllabic suffixes, such as -bian “邊 side,” -zhe “者 -er, -or,” -hua “化 -ize,” etc., to the preceding roots, we can derive complex words as in lu-bian “路邊 roadside, curb,” he-bian “河邊 riverside,” zuo-zhe “ 作 者 author, writer,” sing-zhe “ 行 者 religious practicer,”
gung-yie-hua “工業化 industrialize,” min-zhu-hua “民主化 democratize.”
Table 2.2 lists the statistics of the break types labeled for the syllable boundaries of 121 prefixes and 195 suffixes. It can be seen from the table that 79.6% of the post-syllable boundaries of these 121 prefixes and 98.5% of the pre-syllable boundaries of these 195 suffixes were labeled as B0 or B1. These prosodic findings reflect the fact that morphologically the combination of head and affix generates a lexical unit, and thus the break between them is determined to be the break type of intra-PW category by our method. The results were also consistent with some rules found in Refs. 64, 65, and 68.
Table 2.2: Statistics of break types labeled for 121 prefixes and 195 suffixes.
Labeled break type B0 B1 B2-1 B2-2 B3 B4 Total count
Pre-boundary 94 1289 460 545 193 5 2586
Prefix
Post-boundary 584 1475 344 178 5 0 2586
Pre-boundary 1046 2466 31 20 3 0 3566
Suffix
Post-boundary 307 1479 272 482 568 458 3566
2. Word Set of DE
The words in the DE set particularly refer to de, zhe, and di, which serve multi-functions including a possessive marker, an adjective marker, and an adverbial marker [72]. They are characterized by the fact that a DE word can combine with a wide range of preceding syntactic constituents to form a possessive adjective as in an noun phrase (NP)-de structure: xue-sheng-de quan-li “學生的權力 students’ right,”
“思鄉之情 nostalgia,” or to function as an adverbial phrase as in a DM-de structure:
ke-ren yi-bo-bo-di yong-jin-dien-lai “客人一波波地湧進店來 guest were flocking to the shop.” Despite the variety of the preceding constituent, a DE word, similar to a suffix, builds closer connection with its preceding constituent to form a larger syntactic unit; consequently, it is predictable that the break at the DE words’
pre-boundary position tends to fall into B0 and B1, which means a pause is hardly to be perceived at this juncture. It is also reasonable to infer that due to a looser connection between the DE words and the following constituent, less B0 and B1 would occur at the post-boundary position.
The statistics in Table 2.3 indicates that the distribution of the break types labeled by our model just conformed with our anticipation; while 92.3% pre-boundary breaks of the DE words were B0 and B1, only 65% post-syllable boundaries of the DE words fell into the same types, which suggests that for the DE words, the majority of the neighboring breaks are unperceivable, and in most cases only at the post-boundary position can perceivable breaks be sensed. This result also matched the findings in Refs. 64, 65, and 68. Two examples are given below for illustration:
Ex.1: …。因為(because, Cbaa) 女性(women, Nad) 在(in, P21) 社會(society, Nac) B1 的(DE) B1 地位(position, Nad) 提高(raise, VC2)，…
(… Because women’s social status has been improved, …)
Ex.2: …。目前(now, Nddc) 我(I, Nhaa) 是 (am, V_11) 三十一歲(thirty one year old, DM) B1 的(DE) B2-2 單身(single, VH11) 女郎(woman, Nab)，
(…. Now, I am a thirty-one-year-old single woman.)
Table 2.3: Statistics of break types labeled for the DE words.
Labeled break type B0 B1 B2-1 B2-2 B3 B4 Total count
Pre-boundary 168 1600 146 1 0 0 1915
Post-boundary 210 1035 331 294 41 4 1915
3. Word Sets of Ng, Di, and T
Ng, Di, and T represent the word sets of Mandarin Chinese localizers, aspectual adverbs, and particles [72], respectively. The distinctive shared feature of these sets of words is that almost all the words are no longer than two syllables in length and that when combining with other syntactic constituent to form a larger phrase, they are all positioned at the end of the derived phrase, such as san-tian hoNg “三天後 three days
latter,” kai-hui dang-zhungDi “開會當中 while the meeting is being held,” and bu-qu le-maT “不去了嗎? Not going?.” Due to the characteristic of being post-positioned in a phrase, these words are inclined to be incorporated with their preceding constituents, and predictably barely any pauses can be perceived at the pre-boundary position.
The statistic results listed in Table 2.4 indicate that our model’s break-labeling performance just exactly met our expectation. As high as 94%, 93%, and 87% of the pre-syllable boundaries of the words in this category were labeled as B0 or B1. Three examples are given below:
Ex.3: …，牠(it, Nhaa) 跟(and, Caa) 婦人(woman, Nab) 和(and, Caa) 相命(tell one's fortune, VB11) 者(person, Nab) B1 之間(between, Ng) B3 並(Dbb) 無(no, VJ) 勾串(conspire, Nv1) 作弊(cheat, Nv4)，…
(…. There is no conspiracy to cheat among it, the woman and the fortune teller.) Ex.4: …，忘記(forget, VK1) B0 了(Di) B1 婚後(after marriage, Ndc) 現實(realistic, VH11) 的(DE) 環境(environment, Nac)。Be
(… have forgotten the realistic environment after marriage.)
Ex.5: …，因為(because, Cbaa) 我(I, Nhaa), 永遠(always, Dd) 有(have, V_2) 看 (read, VC2) 不完(have no limit, VC2) 的(DE) 書(book, Nab) B1 啊(“ah”, Tc) B4！
(… for I always have books to read!)
On the other hand, it is also interesting to find that for the breaks at the post-syllable boundaries, 67% and 96% of them were labeled as B3 /B4 especially for the Ng-set and T-set words, respectively. Further investigation reveals that most of the longer breaks were caused by a following PM, an index representing the occurrence of a detectable pause. Besides, because the T-set words are phrasal or sentential final particles and hence are highly likely to be followed by a PM, a much higher ratio of B3/B4 could be found. Two examples are given in the following:
Ex.6: Bb 對(to, P31)人類(human, Naeb) B0 來說(in some sense, Ng) B3，…
(…To human beings, …)
Ex.7: …，想(think, VE2) 辦法(idea, Nac) 解決(to solve, VC2) B0 而已(nothing but, Tb) B4！…
(…just figured out the solution …)
Table 2.4: Statistics of break types labeled for the word sets of Ng, Di, and T.
Labeled break type B0 B1 B2-1 B2-2 B3 B4 Total count
Pre-boundary 97 420 19 12 0 2 550
Ng Post-boundary 26 81 17 58 245 123 550
Pre-boundary 107 83 12 1 0 2 205
Di Post-boundary 30 68 36 41 11 19 205
Pre-boundary 89 84 14 11 0 0 198
T Post-boundary 0 5 1 2 22 168 198
4. Word Set of VE
VE represents a class of transitive verbs that take a sentence as the object, such as ren-wei “認為 to suppose/think/believe (that),” gan-dao “感到 to feel (that),”
biao-she “表示 to show/indicate/mean/suggest (that),” etc [72]. It is evident that since the message carried in a sentential object, compared to a NP object for example, demands longer time to process mentally before being accurately expressed, a longer pause is reasonably anticipated to occur after a VE verb for information operation.
Based on the statistic results listed in Table 2.5, on the whole 72% post-word boundaries of the VE verbs were labeled as breaks with distinctly audible pauses, namely B2-1, B2-2, B3, or even B4, another quite favorable evidence that the break types labeled by our model were consistent with the pause duration people usually take in their utterances. A typical example is given in Ex. 8.
Table 2.5: Statistics of break types labeled for word set of VE.
Labeled break
type B0 B1 B2-1 B2-2 B3 B4 Total
count
Post-boundary 63 177 99 108 159 234 840
Ex.8: …。 當局(government , Nad) B1 說(proclaim, VE2) B3 他(he, Nhaa) 違反 (violate, VJ1) 市政府(city government, Ncb) 一戶(one, DM) 人家(family, Nab) 不 能(can not, Dbab) 儲存(store, VC33) 二千五百雙(2,500, DM) 鞋子(shoes, Nab) 的(DE) 規定(rule, Nac)，…
( …City government proclaimed that he violated the rule that no more than 2,500 pairs of shoes were allowed to be stored in a family…. )
However, it cannot be neglected that no less than 28% post-word boundaries of the VE verbs were labeled as B0 or B1, implying that seemingly our model still
generated quite a few unexpected break types for the VE verbs. Further observation of the data, nevertheless, found two main reasons to account for this discrepancy of labeling. First, besides a sentential object, part of the VE verbs could also take a NP object, so the breaks occurring before a NP object were predictably shorter than before a sentential object. A typical example is given below:
Ex.9: …，因(because, Cbaa) 你(you, Nhaa) 可(can, Dbab) 從(from, P19) 過去(past, Ndda) 經驗(experience, Nac) 中(in the course of, Ng) B2-2 檢討(examine, VE2) B1 成敗(success and failure, Nad)，…
(…because you can examine your success and failure based on the past experience…) The other reason for the occurrence of B0/B1 after a VE verb is that to express attitudinal, temporal, spatial, or manner information about a VE verb, a small word from the DE, Di, Ng, or T sets (such as de, zhe, le, guo, etc.) was attached to the verb, and this attachment and the close connection between the small word and the VE verb caused no need to pause at the juncture. However, the originally expected long pause (B3/B4) after the VE verb did not actually disappear; it was retained and only lagged behind to occur after the VE verb, for instance:
Ex.10: …，以(with, P11) 行動(action, Nad) 說明(prove, VE2) B1 了(Di) B3 他(he, Nhaa) 在乎(care, VK1) 妳(you, Nhaa) 的(DE) 感覺(feeling, Nac) 與(and, Caa) 期望(expectation, Nac) 。…
(…his actions have already proved that he cares about your feeling and expectation….)
5. Word Sets of Caa and Cb
Caa and Cb are two subcategories of Mandarin conjunctions, representing conjunctive conjunctions and correlative conjunctions [72], respectively. In the case of Caa, the arguments linked by the Caa conjunctions are words or phrases of identical syntactic categories and are usually associated in their meaning as in fengN
heCaa yuN “風和雨 wind and rain,” reVH hia-shiCaa lengVH “熱還是冷 hot or cold,”
siNeu zhiCaa shiNeu suiNf “四至十歲 from four to ten years old,” and the like. Upon observation, we found that people usually tend to take a longer pause at pre-word boundary than at the post-word context, forming a sensible rhythmic variation and
The statistics of the labeling results in Table 2.6 informs us that 90% of the Caa pre-boundary breaks were not shorter than B2-2, while, on the contrary, 98% of the post-word breaks were not longer than B2-2, a labeling outcome verifying our observation of the Caa words’ neighboring breaks; that is, longer pauses tended to occur at the boundary between the preceding argument and the conjunction. The results matched some findings in Ref. 40. One example is given below for illustration:
Ex.11: …。生活(life, Nad) 緊張(stress, VH21) B3 與(and, Caa) B1 結婚(marriage, Nv4) 延後(put off, VC2) 的(DE) 問題(problem, Nac)，…
(…the problems of stressful life and postponed marriage….)
On the other hand, the Cb conjunctions function to join two clauses – a syntactic unit much larger than Caa’s arguments – into a compound sentence, and therefore have higher potential to be preceded or followed by a PM in written texts to delimit the domain of a clause or a sentence; in read speech the occurrence of a PM elicits the announcer to take a longer pause to index a message transition or a piece of new message is coming. Our statistic results show that in the case of Cb conjunctions 80% of the pre-word boundaries and 20% of the post-word boundaries were labeled as B3/B4, which means much more PMs occurred before Cb conjunctions than afterward. One typical example is given below for demonstration:
Ex.12: …，B4 因為(because, Cbaa) B2-1 學歷(academic credential, Nad) 並(Dbb) 非(not, VG2) 擇偶(choose spouse, VA4) 的(DE) 絕對(absolute, A) 條件(condition, Nac)。
(…, because the academic credentials are not absolute conditions of choosing spouse.)
Table 2.6: Statistics of break types labeled for word sets of Caa and Cb.
Labeled break type B0 B1 B2-1 B2-2 B3 B4 Total count
Pre-boundary 5 32 1 127 214 26 405
Caa Post-boundary 52 104 157 85 7 0 405
Pre-boundary 61 46 23 39 168 512 849
Cb Post-boundary 135 284 166 95 150 19 849
6. Word set of P
P represents the class of Chinese prepositions, which precede a required argument and together play several semantic roles and indicate various relationships
such as time, location, tool, purpose, etc. Although Chinese Knowledge Information Processing (CKIP) categorizes prepositions into 65 types [73], only 13 types are active in the Sinica Treebank corpus. As for the adjacent pause of a preposition, it is reasonable to expect that due to the close connection of a preposition and its following argument, the pause at the post-word boundary tends to be short. For convenience of illustration, only ba/jiang “把,將” (labeled as P07) and zai “在”
(labeled as P21), two typical and most frequently used prepositions, are selected out as the representative examples for discussion.
The statistic results in Table 2.7 show that on the whole for both ba/jiang “把, 將” and zai “在” about 90% of the post-word boundaries were labeled as breaks no longer than B2-1 (a break type caused by a pitch jump instead of lengthened pause duration), which indicates that the pauses at this juncture were either unperceivable or tending to be very short, again another confirmation of our model’s sound labeling job. Besides, a closer look at the distribution of break type percentages reveals that as high as 49% and 69% of the post-word breaks were B2-1 for ba/jiang “把,將” and zai
“在”, respectively. This statistics reflected our informant’s idiosyncratic style of articulating prepositional phrase; namely, besides leaving no pauses (Ex. 13), she often made a pitch jump between a preposition and the following argument to cause a sensible short pause (Ex. 14).
Ex.13: ，B4 把(P07) B1 孩子(children, Nab) 當做(regard as, VG1) 一塊(a piece of, DM) 璞玉(uncut jade, Nab)，
(….treat children as unpolished jade….)
Ex.14: 曾(at one time, Dd) 探詢(investigate, VE2) 他(he, Nhaa) B3 在(among , P21) B2-1 過去(past, Ndda) 眾多(a large number of, VH11) 的(DE) 著作(works, Nab) 中(among, Ng)
(…has investigated that among a large number of works he wrote…)
On the other hand, as far as the labeling at the pre-word boundary is concerned, most labels were either B1 or B3/B4; that is, 46% and 41% of the labels were B3/B4 and 44% and 49% of them were B1 for ba/jiang “把,將” and zai “在,” respectively, which suggests that our informant either took quite a long pause or just no pause at the pre-word position. To explain this phenomenon, further examination on the data
before a preposition were contributed by a left PM (Ex. 15), and in the remained cases she usually took no pause at this position (Ex. 16).
Ex.15: …，B4 將(P07) B2-1 卷證(document, Nab) 移送(transfer, VC32) 桃園 (Taoyan, Nca) 地院(district court, Ncb) 審理(process, VC2)。
( …transfer the documents to Taoyan District Court to process….)
Ex.16: ，協會(association, Nac) 就(an auxiliary confirming, Dd) 設(establish, VC33) B1 在(at, P21) B1 他(his, Nhaa) 家(home, Ncb) 。
(….The association is established at his home….)
Table 2.7: Statistics of break types labeled for word sets of P07 and P21.
Labeled break type B0 B1 B2-1 B2-2 B3 B4 Total count
Pre-boundary 0 39 0 9 32 9 89
P07 Post-boundary 8 38 34 9 0 0 89
Pre-boundary 1 168 12 24 88 53 346
P21 Post-boundary 27 79 208 28 4 0 346