About express dates ( speed-dates
) on Habré already written
. In short, a group of participants gathers, they talk in pairs for about 5 minutes each, and then they exchange interlocutors. Each participant about each of their partners notes how much he liked him; if the sympathy turns out to be mutual, the organizers give each other contacts.
A group of researchers from Stanford University is engaged in analyzing human dialogues, trying to recognize both the
and the perception of
speech by the listener. The discrepancy between the implied and the perceived is a natural property of natural speech. For the analysis, transcripts from express dates were used, on which each side evaluated the partner's "flirting", and noted her own. The constructed system of automatic recognition of flirting was able to correctly determine the intentions of the speaker in 71.5% of cases; this surpassed the accuracy of the express interview participants themselves. As it turned out, people are more projecting their own feelings to the interlocutor than analyzing his speech.
In 2005, volunteer students conducted three express-dating sessions, during which each participant held an included voice recorder. Records about 1100 four-minute meetings were collected. Since the meetings took place in a natural setting, the recordings included a strong background noise that made automatic recognition of the spoken text impossible. The text was manually recognized, and for each sentence, the time of its beginning and end was recorded on the audio recordings, so that the tempo and timbre of speech could be automatically determined.
In addition to sound recordings, each participant evaluated on a 10-point scale (1 = never, 10 = constantly): “How often did the interviewee behave this way?” (Flirting, awkward, funny, energetic, etc.) and “ How often did you yourself behave in this way? ”In a paper published last year, the researchers analyzed only estimates of flirting. Separately, the expression of their flirtation was considered by the participants and participants, and separately - the perception by the participants and participants of the flirtation of their interlocutor. Thus, four speech analyzers were built, using both the sound recording of the conversation and its transcript.
The analyzers determined several dozen numerical parameters that characterize the intonation of speech, dialogue, and the vocabulary used. The intonation parameters include the height (fundamental frequency, F0) of speech, its volume (RMS), and tempo (words / min). Dialogue parameters were obtained from the transcript by specially composed regular expressions. The following events were counted:
- Acknowledgments: Aha; Okay; Well...
- Approvals: Wow; Yep, exactly; Great
- Interrogations: Wait? In the sense of?
- Completion of phrases for the interlocutor
- Disagreeing replies beginning with Well
- Filling pauses: Uh ...
- Cliffs unfinished phrases
- Interrupting one speaker
The vocabulary used was analyzed by belonging to ten categories, compiled specifically for this study:
- You: yours; you; you
- We: ours; us; us; come on
- I: mine; me; to me
- Approval: yes; okay; cool; fine; I agree
- Abuse: hell; horseradish; pancake; shit; see
- Guess: think; to feel; believe; understand; represent
- Anger: hate; ridiculous; stupid; kill fucking spiked
- Dissatisfaction: bad; problem; heavy; strange; boring; sad
- Sex: love; adore; fuck; virgin
- Food: eat; drink; water; wine; coffee; bar; dinner; dish
It also counted the number of past tense uses (according to the researchers' assumption, its use in the dialogue means the transition of one of the interlocutors to a monologue), and the number of references to the experiment being conducted ( meeting; questionnaire; form; research
The selected parameters may not be sufficiently expressed in a short conversation time. The authors constructed a statistical model linking the words used in the conversation with each of the parameters, and performed a “smoothing” of the observed values of each parameter, given its statistically most probable value for the words used in the conversation.
|Flirting||Perception of flirting|
|with explicit parameters||61.5%||70.0%||77.0%||59.5%|
Interestingly, the proposed system equally accurately recognizes the flirting of young people and girls, while significantly better recognizing the reaction of girls than young people to flirting with the interlocutor. Another interesting result is that only 23 parameters were statistically significant for young people, compared with 31 for girls. Thus, the girls were predictable, but richer in expressive means.
As it turned out, flirting students ask more questions, use you
more often, and we
laugh more, use more words from the categories of “anger”, “discontent”, “sex”. Their speech is faster and higher in tone, but quieter. Reactions of the interlocutor, testifying to the flirting of her partner: laughter, abuse, sexual vocabulary, increased tempo and tone.
Students, when they flirt, use a wider range of pitch; laugh I
say more often; often ask, but do not ask other questions; use more sexual vocabulary, less endorsements and endorsements; their sentences are getting longer and rarer. The interlocutor uses you
more often, asks more questions, speaks quieter and faster.
In addition, we managed to find out when a person believes that his interlocutor is flirting. Misunderstanding between people is often caused by the fact that the speaker uses some means of expression, and the listener draws attention to others. So, although the parameters that indicate “he / a flirts” and “decide about him / her that he / a flirts,” are largely the same, there are important differences. The girls believe that their interlocutor flirts when he uses approvals less often and interrupts less often - although statistically these parameters do not indicate the young man’s intentions. In addition, girls exaggerate the value of the questions asked by the young man, and his accelerated speech. Young people, on the other hand, do not attach enough importance to the interlocutor's laughter, her questioning, and long sentences.
It is interesting to compare the behavior of flirting students and female students. The general characteristic is that they laugh more, speak faster, and raise the tone of voice. There are differences: when young people play around, they ask more questions than usual; when girls are less than usual; but girls start asking again more often, but young people do not. Girls more often say I
and less often we
; young people - more often we
. Girls reduce the use of endorsements, while young people do not.
"It's not about you, but about me."
How did it happen that the automatic recognition of flirting significantly exceeded the accuracy of the living participants in the experiment? To illustrate the essence of the problem, the researchers present the questionnaire of participant No. 101 and participant No. 127:
|I'm flirting||The interviewee flirts|
|Participant number 101||eight||7|
|Participant number 127||one||one|
The impressions of this couple about their dating are completely opposite; but at the same time, everyone appreciated the intentions of his partner almost the same as his own. This is not an isolated example: it turned out that the correlation (“I flirt,” “I think my interlocutor flirts”) is 0.73, while the correlation (“I flirt,” “my interlocutor actually flirts”) is only 0.15. Similar results were obtained for other criteria for assessing the past meeting:
|(self-assessment, assessment of the interlocutor)||(self-assessment, evaluation by the interlocutor)|
Although the accuracy with which people recognize the intentions of a partner is low in all cases, the flirting is better recognized than other characteristics. Researchers attribute this to the fact that in terms of express-date participants are more focused on the partner’s flirting, and less analyze other elements of his behavior. Another important observation is that for the “attracting” behavioral traits (flirting, benevolence), the participants more closely associated self-esteem with the interlocutor's assessment than for the “repulsive” (awkwardness, pressure).
Of course, to recognize the intentions of a partner in four short minutes is not easy. But there is no doubt that each participant evaluated not so much the interlocutor's behavior, but his own impressions and intentions, assuming that, by default, there is reciprocity. The automatic system is devoid of such “blind optimism”, and recording a four-minute meeting is enough for it to make a more accurate assessment than the meeting participants themselves, even though the participants also had facial expressions and gestures, but they were not available to the system.
As applications of this work, we see both the discovery of key data in an array of text (for example, effective indexing of correspondence, conversations in social networks, minutes of meetings and interviews) and the implementation of more sophisticated automatic interlocutors, including for dating sites.