Towards to the Anticipation in Simultaneous Interpreting Eva Kiktová, Július Zimmermann and Mária Pal’ová Language Information and Communication Laboratory, Faculty of Arts, Pavol Jozef Šafarik University in Košice Košice, Slovakia {eva.kiktova, julius.zimmermann, maria.palova}@upjs.sk Abstract: This paper describes a very fine, supraseg- time as the speaker, or knows in advance the content of the mental linguistic feature – anticipation. An ’anticipation future statement [2], [4], [6], [10]. The ability to anticipate nucleus’ is a speech signal that refers to the uniquely is a key competency for professional interpreters. modulated part of an utterance that indicates continuation, This paper has the following structure, in Section 2 the it also allows for reduced semantic content in the utter- prosodic definition of anticipation nucleus is presented. ance. The extent that anticipation can be expressed relies Section 3 describes the research design, participants (in- on the phonetic, syntactic, and semantic capacity of a terpreters) and the used sound data. The statistical analy- language. In this work the special attention to the prosodic sis and its results are also presented. Section 4 focuses on structure of speech signal was paid regarding to this the detected anticipation moments; it describes the param- issue. The work provides information about performed eters extracted from the anticipation nuclei and discusses perceptual detection of anticipation nuclei, describes the the results. The final conclusion and discussion follows in measurement of key prosodic parameters and presents Section 5 obtained results of statistical analysis. 2 Prosodic definition of anticipation nucleus Keywords: Anticipation, nucleus, percipient, interpreting. To explain the essence of the phenomenon examined we 1 Introduction highlight that: If in the last one or last two syllables (their sonantic nu- Understanding what another is telling us is crucial to hu- clei) of the rhythmic group, the fundamental tone, F0 , rises, man civilization as we know it. Usually, we can grasp and is then followed by a pause filled with silence, a hes- meaning without great effort; however, some situations re- itation sound, or an intake of breath, it is likely that the quire us to focus and check that our understanding is as following rhythmic group contains important information. intended. In conversation, we control the importance that The last one or last two syllables of the rhythmic group we attribute to information, especially when it appears in- modulated in this way and followed by a pause, constitute consistent or unbelievable. In some situations, we could an anticipation nucleus. take over for a moment from the person speaking to us, The mention two realizations of anticipation nucleus is or finish the sentence that another has started. That such depicted in Fig. 1. The realization of anticipation nucleus situations are possible, indicates that we have the ability in the same syllables can occur in ’diftong’ or ’hijat’. to anticipate what another is about to say. Furthermore, if the speech utterance is consistent at all speech levels, the 3 Perceptual identification anticipation will be able to be performed without a signif- icant burden [5]. The aim of the perceptual identification was to obtain the In a foreign language, the ability to anticipate relies information about rhythmic groups, time of anticipation on an awareness of multiple levels of language, includ- and a type of anticipation event. ing non-verbal expressions. An interpreter will use all the Four interpreters participated in the present research; all information available to construct the resulting statement had a university degree in French, two also had a degree in in their mind [3]. The ability to anticipate in a foreign lan- linguistics. Three had been interpreters at conferences for guage is to some extent, a matter of learning interpretation more than 10 years, and one for more than five years; one strategies; these include intonation, semantic and syntactic was also an interpreter for the European Union. patterns, which indicate how a statement will continue [1], [8],[11]. 3.1 The sound database Anticipation in simultaneous interpreting means that an interpreter is able to complete the statement at the same Analysed data include speeches of European Parliament Copyright c 2019 for this paper by its authors. Use permitted un- members delivered in French (74 speakers; males and fe- der Creative Commons License Attribution 4.0 International (CC BY males). From a total of 7366 sentences consisting of 4.0). rhythmic groups, a representative sample of 200 (100 + groups that comprised a compound sentence (Fig. 3). In the remaining 30% of cases, they differed on one (B), two (C), or more (D) of the rhythmic groups. The majority of disagreements occurred in compound sentences, which have more than five rhythmic groups; it should be noted that these sentences are particularly interesting for inter- preters, as their high number of anticipatory moments cre- ate many opportunities for new speech utterances [4]. Next, statistical evaluation focused on the extent to which the interpreters could identify syllables with antici- pation nuclei (Fig. 4). The degree of discord is higher than in the previous evaluation. At only 9% of the time were all four participants in agreement on when the anticipation nucleus occurred; in other cases, one participant (B), two participants (C) or more (D), disagreed on the point of the Figure 1: Example of two different realizations of antic- anticipation nucleus. ipation nuclei. First, a two-syllable occurance; second, a Given that each anticipation nucleus is not clearly con- one-syllable occurance. firmed by agreement in the perception tests, results in- dicate that anticipation perception is largely individual. The presence of the anticipation nucleus is not a logi- 100) sentences was randomly selected. From MP4 for- cal value but rather the opposite. The anticipation phe- mat, sound data in mono mode at 44100 Hz sampling fre- nomenon could be better described as a grade of antici- quency, 16-bit were extracted. pation. This corresponds to the reality, that not each an- ticipation nucleus is confirmed clearly by the perceptions 3.2 Description of perceptual identification tests. The anticipation moment detected only one has the low grade of anticipation, or in other words it has a low Each sentence was thoroughly assessed by the four in- probability of anticipation. Conversely, the anticipation terpreters who recorded the number of rhythmic groups, moment which was confirmed several times through per- the order number of the rhythmic group with the antici- ception identifications, had a higher probability of antici- patory nucleus, the time of the anticipation nucleus and pation. In both evaluations, the stochastic character of the the anticipation events (type). The rhythmic group refers anticipation phenomenon was confirmed. to the melodically characteristic part of a compound sen- In the next phase, standard deviations [13] of time real- tence [9]. It divides sentence into shorter parts according ization of the anticipation nucleus were calculated for the to the specific tone and time modulation, see Fig 2. An first hundred sentences and for the second hundred sen- example of data collected are depicted in Tab.1. tences. Results are depicted in Fig. 5. A key point to The aim of recording interpreters’ perceptions was to note is reduced spread of standard deviation in the second define the anticipation nucleus on the oscillogram (Fig. 2, hundred sentences. As this round of perception tests were upper part); this refers to defining the saturation point at conducted approximately two weeks after the first round, which the participant has detected an anticipatory hint, and reduced spread reflects more precise guidelines given to can anticipate one or several possible trajectories that the participants for how to determine an anticipation nucleus. following speech could address. In the first hundred, one participant strictly adhered to Interpreters did not receive any advanced notice of the the criteria of increased intonation and the subsequent anticipation events. They described in their own words pause, as an exclusive criterion to identify the rhythmic what information they anticipated and their expectations group with anticipation; however, the remaining partici- for the continuation of the sentence (anticipatory trajec- pant consciously combined this criterion with anticipatory tory), see Tab.1.. Finally, in studies [5], [7] 12 different moments provided by other linguistic factors, this was es- anticipation events such as a core of utterance, change of pecially the case in ‘flat’ sentences that lacked intonation. theme, determinative syntagm (3), emphasis (9) etc., were Therefore, in the analysis described below, only the sec- defined. ond hundred sentences were included. 3.3 Statistical evaluation of perceptual identification 4 Measurement of identified anticipation Data gathered from the interpreters’ perceptions were sta- nuclei tistically evaluated. Findings indicate how compound sen- tences and anticipation moments are perceived. According to the data obtained from four linguists- Results revealed that in 70% of cases (A), linguist- interpreters, measurements of time, fundamental tone and interpreters were in agreement on the number of rhythmic intensity were performed in Speech Analyzer 3.0.1. First, rhythmic group (n) rhythmic group (n+1) Figure 2: Example generated using Speech Analyzer 3.0.1. displays raw waveform, spectrogram, relative intensity [dB] and F0 [Hz]. Also it shows a part of the one sentence with an anticipation nucleus and background; and a decomposition of a sentence on rhythmic groups. Table 1: Comparison of perceptions for the same file. Num. of Num. of Rhythmic group Anticipation Supposed type File percipient rhythmic group with anticipation nucleus time of anticipation 1. 6-29m 3 1,2 0:00.700; 0:02.146 9,3 2. 6-29m 3 2 0:02.487 3 3. 6-29m 2 2 0.02:510 3 4. 6-29m 2 1 0:02.524 3 Figure 3: Results of perceptions: number of rhythmic groups. Figure 5: Results of perceptions: variance of time when anticipation nucleus is presented. the sentences were de-noised and normalized, then accord- ing to the time indicated by interpreters, the anticipation Figure 4: Results of perceptions: time of the anticipation moments were identified. nucleus. At the penultimate syllable, the extraction of time T 1, Table 2: Examples of measured values. File T1 [s] F0 1 [Hz] I1 [dB] T2[s] F0 2 [Hz] I2 [dB] T3 [s] T4 [s] F0 2 - F0 1 [s] 6-29m 1.895 131 -2.6 2.165 177 -7.1 0.419 2.731 46 138-28f 3.220 222 -5.2 3.397 361 -2.4 0.335 3.844 139 31-13m 1.572 152 -16.9 1.797 137 -11.7 1.659 3.556 -15 Table 3: Location and scatter statistics of 134 anticipation nuclei. Mean Confidence interval Min Max Standard Standard x̄ -95% +95% deviation error Fundamental tone differences 35.98 31.84 40.11 -17 91 24.20 2.09 Differences of the effective value of intensity 1.69 0.83 2.54 -10.90 14.60 5.00 0.43 Pause duration value 0.39 0.36 0.42 0.03 0.89 0.18 0.01 fundamental tone F0 1 and intensity I1 were performed. F0 2-F0 1 is greater than 31.84 Hz and less than 40.11 Hz The same three parameter were extracted in the ultimate (Tab. 3). syllable ( T 2, F0 2, I2). The measurement on the one sylla- Differences in the effective value of the intensity (mea- ble was performed in the case of the anticipation nucleus sured in dB) between the last two syllables: I2-I1 were being identified on the same syllable (diftong, hijat). The calculated. Results show that at the 5% significance level, pause duration (T 3) and the time of next utterance begin- the mean value of the selective distribution of the effective ning (T 4) were also recorded. Tab.2 shows an example of value of intensity I2-I1 is more than 0.83 dB and less than extracted parameters from several sentences; F0 difference 2.54 dB (Tab 3). (F0 2 - F0 1) is calculated in the last column. Pauses are measured in seconds. The results presented The first example (6-29m) in Tab.2 represents a com- in Tab. 3 show that at the 5% significance level, the aver- mon linguistic setting to identify an anticipation nucleus age pause length of the sample is more than 360 millisec- (similar to Fig. 2); the F0 increase from the sonantic nu- onds and less than 420 milliseconds. The pause can be cleus of penultimate to ultimate syllable is in the normal filled with silence, hesitation sounds or an intake of breath. range. The second example (138-28f) represents the hy- Pauses of longer duration (approximately more than one perprosody [5], when the base F0 values and difference F0 second) indicate the interruption of the anticipation pro- values are too high; hyperprosody undermines the antici- cess. pation process due to frequent emphasis on selected parts Of the above mentioned suprasegmental parameters, of an utterance without new information. Conversely (31- fundamental tone is the most relevant to anticipation nu- 13m), when speech is too flat, referred to as hypoprosody cleus detection, followed by pause duration, and finally, [5], grasping key information and anticipating future con- intensity change. tent is also difficult. 5 Conclusion and discussion 4.1 Statistical evaluation of anticipated nuclei In the next phase of investigating the anticipation phe- This paper contributes useful information to the topic of nomenon, statistical evaluation of identified anticipatory anticipation, particularly as it is used in simultaneous in- nuclei was performed. From a total of 140 detected antici- terpreting. It describes a process of perceptual identifica- pation nuclei, six were excluded owing to extreme values. tion of anticipation nuclei, evaluation of perception results, The Shapiro-Wilk test [12], [13] was used to investigate and then reports the extraction of relevant prosodic param- the data’s normality. eters (fundamental tone, intensity, time). It additionally First, the difference in the fundamental tone F0 [Hz] of contributes statistical evaluation parameters for measuring the last or last two syllables: F0 2-F0 1 was analyzed. Re- anticipation nuclei. sults show that at a significance level of 5%, the mean Although interpreting requires multiple skills and is value of the selected difference of the fundamental tone largely individual, the results of the current study indicate that interpreters have the ability to identify the same antic- [10] Ondas, S., Juhar, J.: Analysis of turn-taking in the Slo- ipatory moment at which they can estimate the same con- vak interview corpus. Proc. ICETA 2018, IEEE International tent of future statements (anticipation type). Conference on Emerging eLearning Technologies and Ap- This research demonstrates a research model for collect- plications, p. 411-416, 2018. ing data for future studies on simultaneous interpreting. [11] Allwood, J., Nivre, J., Ahlsn, E.: On the semantic and prag- For such phonetics and linguistics research, data acquisi- matics of linguistic feedback. In: Semantics, vol. 9, no. 1, tion depends on the results of perception tests in which 1992. a group of people indicate their perceptions. This study [12] Kiktova, E., Zimmermann, J.: Detection of anticipation nu- therefore contributes to the body of literature on antic- cleus using HMM and fuzzy based approaches. Proc. DISA 2018, World Symposium on Digital Intelligence for Systems ipation nuclei, which is currently being prepared in the and Machines, p. 355-360, 2018. French language covering 1000 sentences, which will also [13] Statistica. StatSoft, Inc. p. 1878, 1998, ISBN 1-884233-42- be perceptually evaluated by the same four linguist – inter- 2. preters. Their perceptions represent the probability of an- ticipation being used in analyzed sound data in the future and their contribution will be the foundation for further measurements of anticipation nuclei. This analysis of key suprasegmental parameters indi- cates how anticipation nuclei are detected. In the future the parameters extracted here could be used to create a software tool able to detect the presence of anticipation nuclei. Acknowledgement This work was supported by the Slovak Research and De- velopment Agency under the contracts No. APVV-15- 0492 and No. APVV-15-0307. References [1] K. G. Seeber, Intonation and anticipation in simultaneous interpreting. Ecole de Traduction et d’Interprétation, Uni- versité de Genève. [2] K. G. Seeber, Cognitive load in simultaneous interpreting: Model meets data. International journal of bilingualism, vol. 16, no. 2, p. 228-242, 2011. [3] Adamowicz, A.: The role of anticipation in discourse: Text processing in simultaneous interpreting. In: Polish Psycho- logical Bulletin, vol. 20, no. 2, p. 153-160, 1989. [4] Bartłomiejczyk, M.: Strategies of simultaneous interpreting and directionality. In: Interpreting, vol. 8, no. 2, p. 149-179, 2006. [5] Palova, M., Kiktova, E.: Prosodic anticipatory clues and ref- erence activation in simultaneous interpretation. In: XLin- guae, vol. 12/01XL/02, p. 13-22, 2019. [6] Anticipation in simultaneous interpreting. Available: https://www.languageconnections.com/2015/05/anticipation- in-simultaneous-interpreting/ [7] Pal’ová, M., Zeleňáková, M.: On prosodic anticipatory hint. In: XLinguae, vol. 12/1, p. 165-180, 2019. [8] M. V. Butz, O. Sigaut, G. Pezzulo and G. Baldassarre, Antici- patory Behavior in Adaptive Learning Systems. From Brains to Individual and Social Behavior. Springer-Verlag Berlin Heidelberg, p.378, 2007, ISBN 13-978-3-540-74261-6. [9] G. Kleiber and R. Sock, Ce + N + Relative : Sémantique et Prosodie. Lingvisticae Investigationes 29: 2, John Ben- jamins Publishing Company, p. 251-273, 2006.