Popular in Psychology

Most Cited

The adapted mind

Dalhousie University Jerome Barkow

Psychology

Consumers’ pursuit of material and experiential purchases

Cornell University Tom Gilovich

Experimental Psychology

Time to investigate human visual shape perception

University of Montreal Martin Arguin

Most Cited

James S. Jackson and the program for research on Black Americans

Harvard University David R. Williams

Psychology

Violence and Mental Illness: Rethinking Risk Factors and Enhancing Public Safety

Duke University Eric Elbogen

Cognitive Psychology

Cognitive remediation – where are we now and what should we do next?

King's College London Til Wykes

Cognitive Psychology

Harnessing Plasticity in the Atypical Developing Brain using Neurostimulation

University of Surrey Roi Cohen Kadosh

Language-Specific Constraints on Conversation: Evidence from Danish and Norwegian

View more insights in Social Sciences - Psychology - Language and Memory

View more figures from Cornell University

View the page of Morten H. Christiansen

Read the Study

Next in Language and Memory

Metaphorical language processing and amygdala activation in non-native and native speakers

Francesca Citron | Lancaster University

The Acquisition of Noun and Verb Categories by Bootstrapping From a Few Known Words

Perrine Brusini | University of Liverpool

Do cross-linguistic differences in sound structure influence the use of conversational devices like backchannels, repair, and linguistic entrainment in everyday conversations?

Image courtesy of interviewee. January 7, 2025

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video discusses the nature of real-life conversations compared to scripted dialogue found in media. It emphasizes that everyday conversations are often chaotic, featuring incomplete sentences, rapid exchanges, and pauses, which can make understanding difficult. To navigate this complexity, people use conversational devices such as back channels (e.g., nodding, verbal affirmations), repairs (asking for clarification), and linguistic entrainment (reusing words and structures from one another).

Research indicates that this use of conversational devices varies with context. The study in question focuses on conversational differences between Danish and Norwegian speakers. Despite sharing cultural and linguistic similarities, the study found that Danes use linguistic entrainment more often than Norwegians, possibly due to the opaque sound structure of Danish, which complicates speech perception. In contrast, Norwegians request more clarifications, such as repeating information, reflecting their clearer speech communication.

The results highlight the importance of adapting communication strategies and technology interfaces to fit the linguistic characteristics and conversational behaviors of specific language speakers, suggesting that incorporating more linguistic entrainment in Danish-speaking interfaces could enhance clarity and reduce misunderstandings in interactions, whether human-to-human or human-to-computer. This video insight may indeed apply to the design of sophisticated communication technologies like chatbots or voice assistants.

Language:

Export as PDF

Export as PDF
Here are five key points from the video:

1. **Nature of Real Life Conversations**: Real-life conversations are often messy and chaotic compared to scripted dialogue in books or media. People frequently do not complete sentences, make false starts, and talk over one another, all of which contribute to the complexity of understanding each other.

2. **Conversational Devices**: To navigate these complexities, people employ various conversational devices, including back channels (like "mm-hmm" and "aha") to signal understanding, and "repair" mechanisms to clarify misunderstandings. Additionally, linguistic entrainment—or reusing words, syntax, and themes from the other person's speech—plays a significant role in enhancing mutual understanding.

3. **Cultural and Linguistic Differences**: The study highlights that conversational devices can vary by language and cultural convideo. For instance, the opaque phonetic structure of Danish leads to different conversational strategies when compared to Norwegian, which has clearer consonant pronunciations.

4. **Findings on Danish and Norwegian Conversations**: The research revealed that Danes and Norwegians have similar usage of back channels; however, Norwegians tend to ask for repairs more frequently, while Danes exhibit greater linguistic entrainment. This suggests that Danes utilize more redundancy in their speech to facilitate understanding due to the complexities of their language.

5. **Implications for Human-Computer Interfaces**: There are practical applications for this research in creating more effective human-computer interfaces. By incorporating linguistic entrainment relevant to specific languages, such as Danish, these systems can avoid misunderstandings and provide a more natural interaction for users, potentially improving the design of chatbots and other AI-driven conversation interfaces.

Export as PDF
Here are some key quotes extracted from the video:

1. **Nature of Real Conversations**: "Contrary to sort of the, the kind of scripted dialogue... real life conversations are actually very, very messy."

2. **Characteristics of Everyday Conversations**: "We don't complete our sentences, we do sort of, false starts... everyday conversations on top of that is really rapid fire interactions."

3. **Understanding in Conversations**: "It's actually close to a miracle that we can even try and that we can even understand one another."

4. **Conversational Devices**: "We employ a set of what we have called conversational devices."

5. **Back Channels**: "We use what's called back channels... in order to indicate to the other person... that we are kind of following what they're saying."

6. **Repair Mechanisms**: "Sometimes we miss what the other person is trying to say... we can ask them, for what's called a repair."

7. **Linguistic Entrainment**: "This is where we try to reuse part of what the other person was saying in our own response."

8. **Lexical, Syntactic, and Semantic Entrainment**: "We can use the same kind of syntactic combination of words... or we can continue talking about the same topic or theme."

9. **Variation in Conversational Convideos**: "Most previous research... have tended to treat all conversations as if they're essentially the same."

10. **Observations on Language Use**: "Even speakers in Denmark and Norway actually have different combinations... in how they use these conversational devices."

11. **Danish Language Characteristics**: "Danish as a language is rather unusual in that it has a rather opaque sound structure."

12. **Challenges in Language Acquisition**: "This opaque sound structure is challenging for Danish children learning their native language."

13. **Comparative Language Study**: "Norwegians tend to pronounce their consonants, where Danish oftentimes don’t, producing a much more opaque speech signal."

14. **Findings on Back Channels and Repairs**: "Danes and Norwegians both use back channels to the same degree... Norwegians actually asked for more repairs."

15. **Linguistic Entrainment Differences**: "Danes actually use more linguistic entrainment than Norwegians."

16. **Preemptive Communication**: "Danes do a lot of work upfront in order to make sure that people better understand what they’re saying."

17. **Applications in Technology**: "It would be helpful to counter potential misunderstanding by using more linguistic entrainment."

These quotes encapsulate the major themes and findings discussed in the video regarding conversational dynamics, language characteristics, and implications for communication technology.

X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:90000

1
00:00:00.365 --> 00:00:03.245
Contrary to sort of the, the kind of scripted dialogue

2
00:00:03.245 --> 00:00:05.005
that we might come across in books

3
00:00:05.265 --> 00:00:07.885
or listen to on tele, in television shows

4
00:00:07.985 --> 00:00:10.285
or in movies, real life conversations

5
00:00:10.345 --> 00:00:11.805
are actually very, very messy.

6
00:00:12.545 --> 00:00:14.485
And oftentimes what happens when we are talking,

7
00:00:14.505 --> 00:00:17.965
we don't complete our sentences, we do sort of,

8
00:00:18.455 --> 00:00:20.925
false starts where we sort of end up rephrasing

9
00:00:21.195 --> 00:00:23.245
what we are saying halfway through a sentence.

10
00:00:23.825 --> 00:00:26.405
And, and so everyday conversations on top of

11
00:00:26.405 --> 00:00:28.885
that is really rapid fire interactions

12
00:00:28.935 --> 00:00:31.325
where we produce very short utterances

13
00:00:31.325 --> 00:00:33.765
that are only about two to four seconds long.

14
00:00:34.025 --> 00:00:37.005
And these are interspersed by incredibly short pauses

15
00:00:37.145 --> 00:00:39.285
of about a quarter of a second.

16
00:00:39.905 --> 00:00:42.165
And on top of that, it's actually quite common for us

17
00:00:42.185 --> 00:00:43.885
to talk over one another as well.

18
00:00:44.465 --> 00:00:47.485
So it's actually close to a miracle that we can even try

19
00:00:47.485 --> 00:00:49.485
and that we can even understand one another.

20
00:00:49.785 --> 00:00:52.125
So what we do in order to maintain

21
00:00:52.265 --> 00:00:54.285
and establish a common understanding of

22
00:00:54.355 --> 00:00:57.405
what we are talking about, we employ a set of

23
00:00:57.435 --> 00:01:00.285
what we have called conversational, devices.

24
00:01:00.865 --> 00:01:03.485
So for example, we use what's called back channels,

25
00:01:03.815 --> 00:01:06.125
which means that we are doing things like nodding.

26
00:01:06.425 --> 00:01:07.965
We are saying mm-hmm, aha

27
00:01:08.145 --> 00:01:10.125
and so on in order to indicate to the other person

28
00:01:10.125 --> 00:01:12.005
that we are talking to that we are kind

29
00:01:12.005 --> 00:01:13.485
of following what they're saying.

30
00:01:13.745 --> 00:01:17.565
We also, sometimes what might happen is that we,

31
00:01:17.925 --> 00:01:20.725
sometime, we sometime, let me start again.

32
00:01:22.285 --> 00:01:25.065
Of course, sometimes we miss what the other person is trying

33
00:01:25.065 --> 00:01:28.785
to say, and then what we do is that we can ask them,

34
00:01:29.005 --> 00:01:31.545
for what's called a repair, meaning

35
00:01:31.545 --> 00:01:33.505
that we can simply go like, huh,

36
00:01:33.645 --> 00:01:35.945
if we didn't understand anything of what they're saying,

37
00:01:36.205 --> 00:01:39.105
or we can ask for specific information such as,

38
00:01:39.255 --> 00:01:40.945
John date, what,

39
00:01:41.445 --> 00:01:44.225
and so that we can use to sort of get,

40
00:01:44.225 --> 00:01:46.265
missing information and get the conversation

41
00:01:46.265 --> 00:01:47.425
back on track again.

42
00:01:47.925 --> 00:01:51.185
Now, a, a third kind of these conversational devices

43
00:01:51.765 --> 00:01:54.705
is what's called linguistic entrainment.

44
00:01:55.325 --> 00:01:57.185
And this is where we,

45
00:01:57.245 --> 00:01:59.865
and this is sometimes referred to as, alignment

46
00:02:00.085 --> 00:02:01.145
or linguistic alignment.

47
00:02:01.285 --> 00:02:03.865
And this is where we try to reuse part of

48
00:02:03.865 --> 00:02:07.145
what the other person was saying in our own response.

49
00:02:07.485 --> 00:02:11.345
So we can do this by, reusing the same words

50
00:02:11.445 --> 00:02:13.065
as the other person is, is using.

51
00:02:13.245 --> 00:02:14.825
That's called lexical entrainment.

52
00:02:15.285 --> 00:02:17.905
Or we can use the same kind of syntactic combination

53
00:02:17.905 --> 00:02:19.505
of words that the other person is doing

54
00:02:19.505 --> 00:02:21.185
that's called syntactic entrainment.

55
00:02:21.765 --> 00:02:23.745
Or we can talk about,

56
00:02:24.145 --> 00:02:25.545
continue talking about the same topic

57
00:02:25.605 --> 00:02:28.025
or theme that's called semantic entrainment.

58
00:02:28.725 --> 00:02:33.345
Now, this, this way of reusing some of the same words,

59
00:02:33.485 --> 00:02:35.545
the same syntax, the same meaning

60
00:02:35.885 --> 00:02:39.665
or semantics that the other person was just saying,

61
00:02:39.695 --> 00:02:43.145
that makes it easier for us to understand what that,

62
00:02:43.405 --> 00:02:45.345
it makes it easier for the other person we are talking to,

63
00:02:45.345 --> 00:02:47.505
to understand what we are saying,

64
00:02:47.505 --> 00:02:50.945
because they have just used the same, the very same words,

65
00:02:51.015 --> 00:02:54.105
syntactic instructions and, and touched on a similar topic.

66
00:02:54.765 --> 00:02:56.825
So what has happened is that most previous

67
00:02:57.705 --> 00:02:58.785
research on how we

68
00:02:58.805 --> 00:03:02.165
Use these kind of conversational devices to, to keep,

69
00:03:02.285 --> 00:03:03.925
a conversation on track, have tended

70
00:03:03.925 --> 00:03:05.205
to treat all conversations

71
00:03:05.345 --> 00:03:08.835
as if they're essentially the same, no matter matter

72
00:03:08.835 --> 00:03:10.155
what the local context is.

73
00:03:10.155 --> 00:03:12.355
Like, whether people are trying to solve a task together,

74
00:03:12.505 --> 00:03:15.555
whether people are just sort of shooting debris, sort

75
00:03:15.555 --> 00:03:17.715
of hanging out socially,

76
00:03:18.175 --> 00:03:21.275
or even if considering whether the language

77
00:03:21.305 --> 00:03:24.675
that the speakers use in their normal day-to-day

78
00:03:24.675 --> 00:03:26.715
interactions are different from one another.

79
00:03:27.255 --> 00:03:29.515
And so what we found, well, so,

80
00:03:30.215 --> 00:03:32.075
so the underlying a assumption here is

81
00:03:32.075 --> 00:03:33.555
that the conversational context

82
00:03:33.895 --> 00:03:36.755
or the native language of the speaker doesn't really matter.

83
00:03:36.895 --> 00:03:39.995
But what we found in our study was that even speakers

84
00:03:40.215 --> 00:03:42.955
of two closely related languages like Danish

85
00:03:42.955 --> 00:03:47.075
and Norwegians actually have different combinations,

86
00:03:47.135 --> 00:03:48.955
quantitatively speaking in

87
00:03:49.055 --> 00:03:51.395
how they use these conversational devices when

88
00:03:51.395 --> 00:03:52.475
they're talking to one another.

89
00:03:52.475 --> 00:03:54.995
That is when Danes are talking to other Danes

90
00:03:54.995 --> 00:03:57.315
and Norwegians to other Norwegians.

91
00:03:57.455 --> 00:04:01.155
So Danish as a language is rather unusual in

92
00:04:01.385 --> 00:04:04.275
that it has a rather opaque sound structure.

93
00:04:04.735 --> 00:04:07.555
And we've been stu, we've been starting this phenomena,

94
00:04:07.815 --> 00:04:08.835
in what we call the puzzle

95
00:04:08.835 --> 00:04:11.355
of Danish project at old university in Denmark.

96
00:04:11.855 --> 00:04:14.395
So Danish as a language has one

97
00:04:14.395 --> 00:04:16.675
of the largest vial inventories,

98
00:04:16.695 --> 00:04:19.995
of any language in the world, about 40 different vials,

99
00:04:19.995 --> 00:04:23.955
roughly in total compared to English, which has,

100
00:04:24.085 --> 00:04:25.315
about half of that,

101
00:04:25.495 --> 00:04:27.515
or actually less than half of that, about 13

102
00:04:27.515 --> 00:04:30.595
to 15 vials depending on the dialect that that you speak.

103
00:04:31.095 --> 00:04:34.035
So on top of that, Danes also have a tendency

104
00:04:34.035 --> 00:04:38.595
to turn consonants into v like sounds when they speak.

105
00:04:39.055 --> 00:04:42.435
So for example, in Norwegian, the word for street is God.

106
00:04:43.055 --> 00:04:44.675
That's a nice T sound in there.

107
00:04:44.775 --> 00:04:46.515
The, the word in Danish is very similar,

108
00:04:46.775 --> 00:04:49.355
but there is a, what we call a soft D in Danish.

109
00:04:49.355 --> 00:04:51.275
So we say Gil,

110
00:04:51.295 --> 00:04:53.035
and that makes it harder for people to

111
00:04:53.555 --> 00:04:55.275
perceive the consonant sound there

112
00:04:55.275 --> 00:04:56.795
because it's more vial light.

113
00:04:57.135 --> 00:05:00.635
And finally, Danish people often swallow the answer words,

114
00:05:00.885 --> 00:05:03.275
which means that they would admit on, on average,

115
00:05:03.405 --> 00:05:06.435
about a quarter of all the syllables that,

116
00:05:06.775 --> 00:05:10.035
you might write down on a, sort of in an email or,

117
00:05:10.055 --> 00:05:11.755
or in a, in a text more generally.

118
00:05:12.255 --> 00:05:13.275
And that's true even

119
00:05:13.375 --> 00:05:16.235
for radio broadcasters when they read their scripts aloud.

120
00:05:17.075 --> 00:05:19.535
So in our past work, we have found

121
00:05:19.535 --> 00:05:22.135
that this opaque sound structure is challenging

122
00:05:22.135 --> 00:05:25.615
for Danish children learning their native language Danish.

123
00:05:25.675 --> 00:05:29.215
And this actually caused delays in vocabulary acquisition

124
00:05:29.355 --> 00:05:32.535
and in the acquisition of the past tense system of Danish.

125
00:05:33.235 --> 00:05:34.375
But we were also interesting

126
00:05:34.375 --> 00:05:38.095
to find out if there are any effects in adults when they use

127
00:05:38.335 --> 00:05:39.655
language to communicate,

128
00:05:39.715 --> 00:05:41.895
or rather Danish to communicate with one another.

129
00:05:42.555 --> 00:05:46.695
Now. So given the ambiguous speech signal, we ha we

130
00:05:47.535 --> 00:05:50.575
hypothesized that Danes may use conversational devices

131
00:05:50.575 --> 00:05:52.855
differently than speakers of other languages.

132
00:05:53.475 --> 00:05:57.095
Uh, and here Norwegian provides an almost perfect natural

133
00:05:57.255 --> 00:06:01.445
experiment for us to look at, to test this hypothesis

134
00:06:01.445 --> 00:06:02.565
because Denmark

135
00:06:02.565 --> 00:06:05.045
and Norway are very similar in most respects.

136
00:06:05.065 --> 00:06:06.845
So they're closely related.

137
00:06:07.365 --> 00:06:10.365
Historically, culturally, they have similar,

138
00:06:10.505 --> 00:06:12.085
well-developed welfare systems,

139
00:06:12.115 --> 00:06:13.845
educational systems, and so on.

140
00:06:14.145 --> 00:06:16.005
The two languages, Danish

141
00:06:16.005 --> 00:06:18.485
and Norwegian are also quite closely related.

142
00:06:18.945 --> 00:06:21.845
Uh, they have very similar grammar, similar vocabulary,

143
00:06:21.875 --> 00:06:24.045
similar past tense systems and so on.

144
00:06:24.465 --> 00:06:27.045
But there is a key difference between these two languages,

145
00:06:27.065 --> 00:06:29.045
and that is that Norwegians tend

146
00:06:29.045 --> 00:06:30.365
to pronounce their consonants,

147
00:06:30.365 --> 00:06:32.965
where Danish oftentimes don't,

148
00:06:32.965 --> 00:06:37.045
producing this much more opaque, speech signal.

149
00:06:37.345 --> 00:06:39.805
So because of this opaque Danish speech signal,

150
00:06:40.345 --> 00:06:43.885
we expected Danes to rely more on conversational devices

151
00:06:44.065 --> 00:06:46.205
to avoid misunderstandings,

152
00:06:46.235 --> 00:06:47.285
sort of throughout the board.

153
00:06:47.285 --> 00:06:49.205
We were expecting them to use more back channels,

154
00:06:49.235 --> 00:06:53.005
more repairs, more alignment, essentially.

155
00:06:53.665 --> 00:06:58.295
Um, this is not what we found.

156
00:06:58.845 --> 00:07:03.295
What we found was that, first of all, Danes

157
00:07:03.295 --> 00:07:06.415
and Norwegians both use back channels to the same degree.

158
00:07:06.435 --> 00:07:09.895
So they're both, using sort of things like mm-hmm, aha

159
00:07:10.115 --> 00:07:11.375
and so on equally often.

160
00:07:11.635 --> 00:07:14.135
So that's no difference there between the two languages.

161
00:07:14.715 --> 00:07:18.535
Um, second, and this is, was surprising to us,

162
00:07:18.705 --> 00:07:21.255
Norwegians actually asked for more repairs.

163
00:07:21.255 --> 00:07:22.375
That is, they asked for more.

164
00:07:22.755 --> 00:07:25.615
Uh, either repetitions had questions about what was said,

165
00:07:25.875 --> 00:07:27.135
so this was puzzling to us.

166
00:07:27.475 --> 00:07:31.455
But then what we found was that, thirdly, that

167
00:07:32.305 --> 00:07:34.415
Danes actually use more linguistic

168
00:07:34.445 --> 00:07:36.015
entrainment than Norwegians.

169
00:07:36.395 --> 00:07:39.575
And so our interpretation of this was that,

170
00:07:39.645 --> 00:07:41.735
because of the opaque speech signal,

171
00:07:42.265 --> 00:07:44.135
Danes would use more entrainment.

172
00:07:44.135 --> 00:07:46.935
That is, there would be more likely to reuse previous words,

173
00:07:47.375 --> 00:07:48.615
previous and tactic constructions,

174
00:07:48.995 --> 00:07:51.775
or focused on previous semantic themes

175
00:07:52.325 --> 00:07:53.975
because,

176
00:07:55.615 --> 00:07:57.505
that would help the other person understand

177
00:07:57.505 --> 00:08:00.825
what they're saying upfront in a sense, preempting the need

178
00:08:01.125 --> 00:08:02.705
for having to do repair.

179
00:08:02.805 --> 00:08:04.705
So in other words, the d dangers do a lot

180
00:08:04.705 --> 00:08:07.785
of work upfront in order to make sure that,

181
00:08:07.805 --> 00:08:09.865
people better understand what they're saying

182
00:08:10.205 --> 00:08:13.385
and leaving less need for actually doing repair later on.

183
00:08:13.385 --> 00:08:15.705
Whereas Norwegians, they don't need to do this

184
00:08:15.705 --> 00:08:17.745
because their speech signal is clearer.

185
00:08:18.125 --> 00:08:20.425
And so instead, what they have to do every now

186
00:08:20.425 --> 00:08:23.025
and then is to fall back on repair when some sort

187
00:08:23.025 --> 00:08:24.545
of misunderstanding arises.

188
00:08:24.545 --> 00:08:26.985
There's a number and different ways in which people have

189
00:08:26.985 --> 00:08:29.145
been looking at this, in the literature.

190
00:08:29.165 --> 00:08:33.025
So what we did was for lexical entrainment, that is sort

191
00:08:33.025 --> 00:08:34.745
of the reuse of words, we simply

192
00:08:35.585 --> 00:08:37.145
measured the degree to which speakers.

193
00:08:37.165 --> 00:08:39.785
We used the same words as the person,

194
00:08:39.785 --> 00:08:41.185
that spoke previously.

195
00:08:41.405 --> 00:08:45.785
And when it comes to synap entrainment, we,

196
00:08:46.685 --> 00:08:49.705
used a simplified version where essentially

197
00:08:49.705 --> 00:08:53.465
what we did was, so we took each of the word in a, we turned

198
00:08:53.465 --> 00:08:57.165
that into what's called parts of speech, meaning that,

199
00:08:57.465 --> 00:09:01.205
for example, if we take the sentence the cat runs,

200
00:09:01.675 --> 00:09:04.085
that would be turned into a sequence of

201
00:09:04.685 --> 00:09:06.725
determiner, noun verb.

202
00:09:06.745 --> 00:09:10.845
And that's because there is a determiner, cat is a noun,

203
00:09:11.145 --> 00:09:12.565
and runs is a verb.

204
00:09:12.705 --> 00:09:14.845
So we got a sequence of these,

205
00:09:14.995 --> 00:09:16.325
so-called parts of speech.

206
00:09:16.715 --> 00:09:19.845
Then we do, we determine the degrees of overlap

207
00:09:20.635 --> 00:09:22.685
between these parts of speech sequences

208
00:09:23.385 --> 00:09:26.765
across different entrances that the speakers were producing.

209
00:09:27.505 --> 00:09:30.765
Um, and then that allow us to, look at the degree

210
00:09:30.865 --> 00:09:35.725
of synthetic overlap between, turns in the conversation.

211
00:09:36.145 --> 00:09:38.725
And finally, for semantic entrainment,

212
00:09:38.825 --> 00:09:42.045
we created what's called distributional, meaning,

213
00:09:42.045 --> 00:09:44.725
vectors for each word in a sense and in a sentence,

214
00:09:44.725 --> 00:09:47.605
and compared them, across speaker terms.

215
00:09:47.705 --> 00:09:51.805
Now, distributional meaning vectors are, are computed,

216
00:09:51.805 --> 00:09:53.325
using a computational model

217
00:09:53.675 --> 00:09:57.685
that looks at the way in which word could occur in a text.

218
00:09:58.265 --> 00:10:01.445
So, for example, the basic idea is that words that tend

219
00:10:01.445 --> 00:10:02.805
to occur in the same

220
00:10:03.465 --> 00:10:07.245
or in similar, centenial context have similar meaning.

221
00:10:07.345 --> 00:10:09.725
So for example, if we take a word such as street

222
00:10:09.745 --> 00:10:12.685
and road, they will oftentimes occur in very similar

223
00:10:12.685 --> 00:10:16.085
context, indicating that they have similar meaning,

224
00:10:16.085 --> 00:10:18.125
similar meaning, and we can pick that up in terms

225
00:10:18.125 --> 00:10:19.645
of these computational analysis.

226
00:10:19.645 --> 00:10:24.125
So that's how we, compared lexical entrainment syntax

227
00:10:24.185 --> 00:10:28.645
and Symantec, entrainment was especially, in the use

228
00:10:28.645 --> 00:10:30.045
of linguistic entrainment

229
00:10:30.045 --> 00:10:33.725
that Dane showed less variation across, different tasks,

230
00:10:34.025 --> 00:10:37.365
whether they were sort of just, talking socially

231
00:10:37.385 --> 00:10:39.765
or trying to solve some sort of, task.

232
00:10:40.585 --> 00:10:43.565
Um, and in, in specifically what we found was

233
00:10:43.565 --> 00:10:47.325
that they showed less decrease in linguistic entrainment,

234
00:10:47.985 --> 00:10:51.685
compared to Norwegians when it came to these,

235
00:10:51.715 --> 00:10:53.125
task oriented conversation.

236
00:10:53.125 --> 00:10:54.325
That is when we're trying to solve

237
00:10:54.325 --> 00:10:55.405
some sort of task together.

238
00:10:55.865 --> 00:10:58.205
And we believe that's because the dames,

239
00:10:58.415 --> 00:11:00.125
might need a higher level of sort

240
00:11:00.125 --> 00:11:02.565
of baseline linguistic entrainment in,

241
00:11:02.565 --> 00:11:06.685
because of this sort of o ness of the speech signal in order

242
00:11:06.825 --> 00:11:09.845
to, make sure they understand one another.

243
00:11:09.905 --> 00:11:14.005
So in other words, daes need to preempt potential,

244
00:11:14.205 --> 00:11:15.685
misunderstanding misunderstandings

245
00:11:15.685 --> 00:11:18.125
by incorporating additional redundancy in

246
00:11:18.125 --> 00:11:19.685
what they say in order

247
00:11:19.685 --> 00:11:23.285
to facilitate understanding human computer interfaces become

248
00:11:23.525 --> 00:11:26.645
more sophisticated, perhaps sort of moving in the direction

249
00:11:26.645 --> 00:11:29.765
of the kind of spoken language interfaces we see in science

250
00:11:29.765 --> 00:11:32.525
fiction movies such as Star Trek or what have you.

251
00:11:33.105 --> 00:11:36.405
Um, it's probably worthwhile for the designers

252
00:11:36.405 --> 00:11:40.805
of these system to, adapt them to the specific patterns

253
00:11:40.805 --> 00:11:45.605
of conversational device use that are, relevant

254
00:11:45.665 --> 00:11:47.765
for particular languages like Danish

255
00:11:47.765 --> 00:11:49.405
or Norwegian or English.

256
00:11:49.585 --> 00:11:50.805
So for example, if you wanted

257
00:11:50.805 --> 00:11:54.405
to create a more naturally sounding human computer interface

258
00:11:54.425 --> 00:11:57.645
for Danish speakers, it would be, helpful

259
00:11:57.865 --> 00:11:59.725
to counter potential misunderstanding

260
00:11:59.865 --> 00:12:03.205
by using more linguistic, entrainment, for example,

261
00:12:03.765 --> 00:12:08.165
compared to if you're, developing an English, based,

262
00:12:08.625 --> 00:12:12.685
human computer interface with spoken, language.

263
00:12:13.505 --> 00:12:15.685
Now, this might actually also be relevant

264
00:12:15.685 --> 00:12:19.005
for thinking about, text interfaces as well, such as

265
00:12:19.005 --> 00:12:22.925
what we might use for interacting with, current,

266
00:12:22.925 --> 00:12:25.645
systems like chat GBT or something like that.

267
00:12:26.265 --> 00:12:28.445
Um, now in this case, the Danish system,

268
00:12:28.705 --> 00:12:32.565
the Danish language interface could be tuned, fine tuned

269
00:12:32.665 --> 00:12:35.485
to essentially incorporate more linguistic entrainment

270
00:12:35.485 --> 00:12:36.605
than for other languages.

271
00:12:36.605 --> 00:12:39.365
Again, sort of that would, that would make it be,

272
00:12:39.915 --> 00:12:43.205
that would make it sort of appear more natural sounding to,

273
00:12:43.545 --> 00:12:45.085
the person using it.

274
00:12:45.785 --> 00:12:47.975
Um, and more generally

275
00:12:48.495 --> 00:12:51.175
actually adding more linguistic entrainment,

276
00:12:51.275 --> 00:12:54.535
to chatbots more generally might actually, be useful

277
00:12:54.535 --> 00:12:56.735
to pre, to prevent misunderstandings.

278
00:12:57.075 --> 00:12:59.695
And indeed, we've actually been doing some,

279
00:12:59.695 --> 00:13:02.935
preliminary analysis of chat GBT conversations,

280
00:13:03.435 --> 00:13:05.175
and what we found was actually they tend

281
00:13:05.175 --> 00:13:09.415
to involve more entrainment, than when we look at human

282
00:13:09.715 --> 00:13:11.015
to human interactions.

283
00:13:11.275 --> 00:13:13.415
So this actually might already have been something

284
00:13:13.415 --> 00:13:15.415
that either have been built into the system

285
00:13:15.775 --> 00:13:17.095
explicitly or implicitly.

286
00:13:17.515 --> 00:13:18.735
Uh, we don't really know,

287
00:13:18.735 --> 00:13:21.255
but that does seem to be already a, a move in, in

288
00:13:21.255 --> 00:13:21.815
that direction.

Understanding the Complexity of Everyday Conversations

00:00:00 - Contrary to sort of the, the kind of scripted dialogue that we might come across in books or listen to on tele, in television shows or in movies, real life conversations are actually very, very messy. And oftentimes what happens when we are talking, we don't complete our sentences, we do sort of, false starts where we sort of end up rephrasing what we are saying halfway through a sentence. And, and so everyday conversations on top of that is really rapid fire interactions where we produce very short utterances

00:00:34 - And these are interspersed by incredibly short pauses of about a quarter of a second. And on top of that, it's actually quite common for us to talk over one another as well. So it's actually close to a miracle that we can even try and that we can even understand one another. So what we do in order to maintain and establish a common understanding of what we are talking about, we employ a set of what we have called conversational, devices. So for example, we use what's called back channels,

00:01:06 - We are saying mm-hmm, aha and so on in order to indicate to the other person that we are talking to that we are kind of following what they're saying. We also, sometimes what might happen is that we, sometime, we sometime, let me start again. Of course, sometimes we miss what the other person is trying to say, and then what we do is that we can ask them, for what's called a repair, meaning that we can simply go like, huh, if we didn't understand anything of what they're saying,

Linguistic Entrainment in Conversation

00:01:39 - John date, what, and so that we can use to sort of get, missing information and get the conversation back on track again. Now, a, a third kind of these conversational devices is what's called linguistic entrainment. And this is where we, and this is sometimes referred to as, alignment or linguistic alignment. And this is where we try to reuse part of what the other person was saying in our own response.

00:02:11 - as the other person is, is using. That's called lexical entrainment. Or we can use the same kind of syntactic combination of words that the other person is doing that's called syntactic entrainment. Or we can talk about, continue talking about the same topic or theme that's called semantic entrainment. Now, this, this way of reusing some of the same words, the same syntax, the same meaning or semantics that the other person was just saying,

00:02:43 - it makes it easier for the other person we are talking to, to understand what we are saying, because they have just used the same, the very same words, syntactic instructions and, and touched on a similar topic. So what has happened is that most previous research on how we Use these kind of conversational devices to, to keep, a conversation on track, have tended to treat all conversations as if they're essentially the same, no matter matter what the local context is.

Conversational Dynamics: Language Use Among Danes and Norwegians

00:03:12 - whether people are just sort of shooting debris, sort of hanging out socially, or even if considering whether the language that the speakers use in their normal day-to-day interactions are different from one another. And so what we found, well, so, so the underlying a assumption here is that the conversational context or the native language of the speaker doesn't really matter. But what we found in our study was that even speakers

00:03:42 - and Norwegians actually have different combinations, quantitatively speaking in how they use these conversational devices when they're talking to one another. That is when Danes are talking to other Danes and Norwegians to other Norwegians. So Danish as a language is rather unusual in that it has a rather opaque sound structure. And we've been stu, we've been starting this phenomena, in what we call the puzzle of Danish project at old university in Denmark.

00:04:14 - of the largest vial inventories, of any language in the world, about 40 different vials, roughly in total compared to English, which has, about half of that, or actually less than half of that, about 13 to 15 vials depending on the dialect that that you speak. So on top of that, Danes also have a tendency to turn consonants into v like sounds when they speak. So for example, in Norwegian, the word for street is God.

Challenges of Danish Language Perception

00:04:44 - The, the word in Danish is very similar, but there is a, what we call a soft D in Danish. So we say Gil, and that makes it harder for people to perceive the consonant sound there because it's more vial light. And finally, Danish people often swallow the answer words, which means that they would admit on, on average, about a quarter of all the syllables that, you might write down on a, sort of in an email or, or in a, in a text more generally. And that's true even

00:05:17 - So in our past work, we have found that this opaque sound structure is challenging for Danish children learning their native language Danish. And this actually caused delays in vocabulary acquisition and in the acquisition of the past tense system of Danish. But we were also interesting to find out if there are any effects in adults when they use language to communicate, or rather Danish to communicate with one another. Now. So given the ambiguous speech signal, we ha we

00:05:50 - differently than speakers of other languages. and here Norwegian provides an almost perfect natural experiment for us to look at, to test this hypothesis because Denmark and Norway are very similar in most respects. So they're closely related. Historically, culturally, they have similar, well-developed welfare systems, educational systems, and so on. The two languages, Danish and Norwegian are also quite closely related.

Conversational Devices in Danish and Norwegian Speech

00:06:21 - similar past tense systems and so on. But there is a key difference between these two languages, and that is that Norwegians tend to pronounce their consonants, where Danish oftentimes don't, producing this much more opaque, speech signal. So because of this opaque Danish speech signal, we expected Danes to rely more on conversational devices to avoid misunderstandings, sort of throughout the board. We were expecting them to use more back channels,

00:06:53 - this is not what we found. What we found was that, first of all, Danes and Norwegians both use back channels to the same degree. So they're both, using sort of things like mm-hmm, aha and so on equally often. So that's no difference there between the two languages. second, and this is, was surprising to us, Norwegians actually asked for more repairs.

00:07:22 - either repetitions had questions about what was said, so this was puzzling to us. But then what we found was that, thirdly, that Danes actually use more linguistic entrainment than Norwegians. And so our interpretation of this was that, because of the opaque speech signal, Danes would use more entrainment. That is, there would be more likely to reuse previous words, previous and tactic constructions, or focused on previous semantic themes

Understanding Conversational Clarity: Danish vs. Norwegian Communication

00:07:55 - that would help the other person understand what they're saying upfront in a sense, preempting the need for having to do repair. So in other words, the d dangers do a lot of work upfront in order to make sure that, people better understand what they're saying and leaving less need for actually doing repair later on. Whereas Norwegians, they don't need to do this because their speech signal is clearer. And so instead, what they have to do every now and then is to fall back on repair when some sort

00:08:24 - There's a number and different ways in which people have been looking at this, in the literature. So what we did was for lexical entrainment, that is sort of the reuse of words, we simply measured the degree to which speakers. We used the same words as the person, that spoke previously. And when it comes to synap entrainment, we, used a simplified version where essentially what we did was, so we took each of the word in a, we turned

00:08:57 - for example, if we take the sentence the cat runs, that would be turned into a sequence of determiner, noun verb. And that's because there is a determiner, cat is a noun, and runs is a verb. So we got a sequence of these, so-called parts of speech. Then we do, we determine the degrees of overlap between these parts of speech sequences across different entrances that the speakers were producing.

Exploring Linguistic Entrainment and Semantic Overlap

00:09:30 - of synthetic overlap between, turns in the conversation. And finally, for semantic entrainment, we created what's called distributional, meaning, vectors for each word in a sense and in a sentence, and compared them, across speaker terms. Now, distributional meaning vectors are, are computed, using a computational model that looks at the way in which word could occur in a text.

00:10:01 - to occur in the same or in similar, centenial context have similar meaning. So for example, if we take a word such as street and road, they will oftentimes occur in very similar context, indicating that they have similar meaning, similar meaning, and we can pick that up in terms of these computational analysis. So that's how we, compared lexical entrainment syntax and Symantec, entrainment was especially, in the use

00:10:30 - that Dane showed less variation across, different tasks, whether they were sort of just, talking socially or trying to solve some sort of, task. and in, in specifically what we found was that they showed less decrease in linguistic entrainment, compared to Norwegians when it came to these, task oriented conversation. That is when we're trying to solve some sort of task together. And we believe that's because the dames, might need a higher level of sort

Improving Human-Computer Interaction through Linguistic Entrainment

00:11:02 - because of this sort of o ness of the speech signal in order to, make sure they understand one another. So in other words, daes need to preempt potential, misunderstanding misunderstandings by incorporating additional redundancy in what they say in order to facilitate understanding human computer interfaces become more sophisticated, perhaps sort of moving in the direction of the kind of spoken language interfaces we see in science fiction movies such as Star Trek or what have you.

00:11:36 - of these system to, adapt them to the specific patterns of conversational device use that are, relevant for particular languages like Danish or Norwegian or English. So for example, if you wanted to create a more naturally sounding human computer interface for Danish speakers, it would be, helpful to counter potential misunderstanding by using more linguistic, entrainment, for example,

00:12:08 - human computer interface with spoken, language. Now, this might actually also be relevant for thinking about, text interfaces as well, such as what we might use for interacting with, current, systems like chat GBT or something like that. now in this case, the Danish system, the Danish language interface could be tuned, fine tuned to essentially incorporate more linguistic entrainment

Enhancing Chatbot Communication Through Linguistic Entrainment

00:12:36 - Again, sort of that would, that would make it be, that would make it sort of appear more natural sounding to, the person using it. and more generally actually adding more linguistic entrainment, to chatbots more generally might actually, be useful to pre, to prevent misunderstandings. And indeed, we've actually been doing some, preliminary analysis of chat GBT conversations, and what we found was actually they tend to involve more entrainment, than when we look at human

00:13:21 - So this actually might already have been something that either have been built into the system explicitly or implicitly. we don't really know, but that does seem to be already a, a move in, in that direction.

Faculti AI can make mistakes

Research

Policy

Analysis

Data

Originals

Explainers

Podcasts