4
$\begingroup$

For my master thesis I am working on a dialogue system that should be deployed in hospitals to administer simple questionnaires to patients. I already did literature research and I'm fine with what I found since I don't have to replicate something which has been already done, but I noticed that there are really few papers regarding this specific 'robot interviewer' topic.

Let me explain the task a bit more in details: in a real interview a human interviewer usually starts with greetings and with an explanation of the questionnaire to administer, and then he/she starts asking some more or less structured questions to the person to interview. The idea here is to replace a human interviewer with a dialogue system.

Now, at a first glimpse it seems like a task that can be easily hand coded, and indeed lot of real application simply use systems in which specific questions are stored in memory in association with some already made answers to choose from (here's an example), and the system simply show them (or read in the case of humanoid robots) to the people being interviewed, the system then wait for the answer and then move on with the next questions.

The point is that in real interviews the conversation flow is obviously much more smooth and natural. A human being can detect doubts in the voice of the interviewed person, which can also explicitly ask for explanations, a human being can also understand when an answer comes with emotional implications ('yes I feel sad every day') and we are able to automatically react to these hidden implications with emotional fillers ('I'm sorry to hear that'). All these aspects require to train some natural language understanding module in order to be replicated in an artificial agent (and this is actually what I'm currently working on), so I though I would have found more papers on this.

Now, despite having found a tons of paper related to open domain dialogue systems, affective and attentive systems, even systems able to reply with irony, I did not found many papers about dialogue systems for smooth interviews or questionnaire administration, which in my opinion sounds like an much easier task to tackle (especially if compared to open domain conversations). The only two papers that I found which truly focused on interviewer systems are:

So my question is: did I miss something, like some specific keywords? Or is there actually a gap in the literature with regard to the design of dialogue systems for interviews and questionnaire (or surveys) administration? I'm interested in any link or hint from anyone working on similar applicants, thank you in advance!

$\endgroup$
2
  • 1
    $\begingroup$ There are Scientific evidences that when a person is happy voice becomes higher pitched and probably lower pitched when sad. So I am sure there is a correlation between sound, words and emotion. I think emotion detection systems from sound might be the key. Google searched and found this sciencedirect.com/science/article/abs/pii/S0957417414001638 $\endgroup$
    – user9947
    Apr 8, 2020 at 16:03
  • $\begingroup$ Thank you for the link! I did found already lot of papers about using prosodic and other auditory features for emotion and doubt detection but every new link is useful. Actually the question refers more about papers that tries to develop entire systems created with the purpose of performing interviews or questionnaire administration, I'll edit it to make this point more explicit! $\endgroup$ Apr 8, 2020 at 16:09

1 Answer 1

4
$\begingroup$

[Disclaimer: I work for a company that provides a platform for developing conversational AI systems]

The platform used by the company I work for has a sentiment analysis component, so you can recognise if the user input expresses certain emotions. The dialogues are encoded in 'flows', which are graphs with an initial trigger consisting of output nodes and transitions that correspond to user input. In order to react to emotions expressed by the user you'd have a trigger that does not react to the user input text, but instead to a setting of the sentiment flags.

For example, in a call centre bot, if the user repeatedly uses angry sentiment in their utterances, the bot could detect this and trigger a flow that says "It seems you are not happy with my responses -- please wait while I transfer you to a human agent".

This all depends on the mechanics of organising the dialogues. If you have a ML model, this will probably not be as easy as if you have a manually designed graph structure (which gives you more control over such matters).

Being a commercial system, there are obviously no academic papers on this. But as it's so easy to do within the platform, it doesn't seem too much of a problem in the first place.

$\endgroup$
4
  • $\begingroup$ So, I guess that one of the keywords/terms that the OP was missing is "sentiment analysis". $\endgroup$
    – nbro
    Apr 8, 2020 at 18:09
  • $\begingroup$ Thanks for the answer! @nbro Actually I used the reference to sentiment analysis and auditory feature analysis just to show that these systems can be improved a lot compared to simple hand coded ones. Oliver: May I ask you if you work in the development of systems in which the robot has the lead of the conversation? That is actually the main key point in interviews from my perspective. Anyway you already gave me an hint: it could be that these type of dialogue systems are mostly researched in companies rather than academia, that might explain the rather low number of papers. $\endgroup$ Apr 8, 2020 at 18:22
  • 1
    $\begingroup$ @EdoardoGuerriero Yes, you can design the dialogue so that the bot takes the lead; in fact that is more likely (after the initial user utterance) as the user is typically asking the bot to do something, and the bot then requests required information etc. I'm not developing the system myself, though, but design dialogues. $\endgroup$ Apr 9, 2020 at 7:57
  • $\begingroup$ @OliverMason Thanks for the clarification! I'll wait.a couple of days to see if some else will answer with some reference otherwise I'll except the answer $\endgroup$ Apr 9, 2020 at 14:30

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .