‘O.V.R.’ Dialogue management

Presentatie over: "‘O.V.R.’ Dialogue management"— Transcript van de presentatie:

1 ‘O.V.R.’ Dialogue management
Alexandra Peters Augustus 2000

2 Openbaar Vervoer Reisinformatie (O.V.R.)
Verstrekt informatie per telefoon over reizen met openbaar vervoer Meer dan 400 informatrices 13 mln bellers per jaar 9 mln worden geholpen 4 mln wachtrij The dialogues from the corpus come from the dutch company OVR They are the largest provider of information by telephone concerning public transport in the Netherlands in utrecht main office 9 centres in the Netherlands have more than 400 operators who handle 13 million calls per year from which 4 mln calls are in waiting queue Waiting queues are not desirable for OVR because client’s appreciation will be lower questions like : At what time are the trains leaving from Delft? And What are the costs for travelling by train from Delft to Amsterdam? Augustus 2000

3 Onderzoeksdoelen Coderen van de dialogen Dialoog model ontwikkelen
Dialoog management bekeken vanuit de informatrice Modelleren van de gebruikte kennis van de informatrice Met de verkregen kennis ontwikkelen van een trainingsomgeving analysis implies these four points First, I will tell you more about the coding of the dialogues Data reduction which enables statistical processing coding labels the dialogues so that the structure of dialogues will be transparent Then the constructing of a dialogue model will be explained. with help of coding try to construct dialogue model as said before model can be used as script in ASP system thirdly, we will look at Dialogue management from the operator’s point of view Here will be explained how the operator manages the dialogue (to get an efficient and appreciative dialogue). The possible strategies and thoughts are explained and finally we will look at Modelling the operator’s knowledge. Several knowledge areas which the operator needs to have to be able to perform her tasks properly are explained. This knowledge has to be extracted from the operator and put in an ASP system to be able to simulate the operator’s side. But first I will give some general information concerning the used corpus. Augustus 2000

4 Beperkingen aan de codering van de dialogen
Verschillende coderingen voor de informatrice en de cliënt Een codering per zin/uiting Alleen op tekst gebaseerd To see the different roles for the operator and the client more clearly different dialogue acts are used for them One dialogue act per utterance because this way it is easier to construct a usable and transparent dialogue model. If more dialogue acts were used per utterance there would be too many possibilities and therefore the model to complicated To be able to use one dialogue act per utterance, some parts of the utterance have to be less important. We have chosen to see trivial remarks from client and operator as less important when they are combined with another dialogue act in an utterance, because they do not contain important information that is needed to provide the client with information. explanation prosody melody and intonation Prosody describes not the content of the utterance but the way the utterance is expressed. It involves the melody and sound of an utterance. Is it a question, is the utterance said angry etcetera Here not used because the dialogues were only available in plain text and the prosody of an utterance can not be derived without hearing it. So without prosody it can be difficult to interpret the utterance correctly; rules are necessary such as non-verbal utterances are discarded because there meaning can not be derived. Augustus 2000

5 Codering This are the dialogue acts used to code the dialogue acts. At the left side the dialogue acts for the operator can be seen and at the right side the client’s dialogue acts. As you can see here just a few action types possible for the operator and the client. They can greet, ask questions, answer questions and give remarks explanation that v,cl,ni,r are questions. With these questions the information that is needed to provide the information to the client is asked. This information is needed to fill the slots in the travel planner examples OO Good afternoon, travel information NI At what time do you want to leave ANI I want to leave at nine o’clock GC Thanks, goodbye Augustus 2000

6 Dialoog voorbeeld Codering Dialoog OO
informatrice : goedemiddag reisinformatie cliënt : goedemiddag ik wil graag weten wanneer de trein van Delft naar Amsterdam centraal vertrekt informatrice : wanneer wilt u aankomen cliënt : rond elf uur informatrice : u kunt vertrekken om twee voor tien cliënt : twee voor tien informatrice : en dan komt u om vijf voor elf aan in Amsterdam centraal cliënt : dank u wel dag informatrice : graag gedaan dag Codering OO OC NI ANI GA RC GC GO 200 dialogues used to code topics restricted to travelling by train dialogues are recorded and transliterated One dialogue act per utterance in this way the dialogues were transliterated; no punctuation simply plain text Explanation per utterance this is the opening of the operator where she greets the client then client greets and asks his opening question nine fifty eight is an example where it is not clear if it is a remark or a verification question. Because the difference between nine… and nine …. can only be heard . Augustus 2000

7 Codeer tool Augustus 2000

8 Bepalende factoren Prosodie Context Classificatie Codeerniveau Inhoud
Subjectieve codering Geen betekenis non-verbale uitingen Context Classificatie Codeerniveau Inhoud Relaties tussen uitingen Coding can be difficult because several factors influence the coding Without prosody the meaning of a sentence is not always clear, as I illustrated in the last example. And sometimes it can not be solved with the help of the next utterance and then you try to interpret it and this leads to subjective coding. Non-verbal utterances can also have a meaning but without prosody this meaning is not clear. Also the history and the future of the utterance are needed in coding the utterance because there are relations between utterances. For example a question is followed by an answer most of the time. These relations determine what is meant by these utterances. An utterance has no stand-alone interpretation. The only exception to this is the first utterance of the dialogue. For example the utterance ‘central station’ uttered by the client can be an answer, a verification question or a remark. It depends on the location in the dialogue and the previous and next utterance how this utterance is coded. Augustus 2000

9 Voorbeeld “centraal station” antwoord verificatie vraag
“Van welk station in Amsterdam vertrekt u?” verificatie vraag “Ja, Amsterdam centraal station” opmerking/parafrasering “En dan komt u om twee uur aan in Utrecht” Augustus 2000

10 Onderzoeksdoelen Coderen van de dialogen Dialoog model ontwikkelen
Dialoog management bekeken vanuit de informatrice Modelleren van de kennis van de informatrice Met de verkregen kennis ontwikkelen van een trainingsomgeving The second step is constructing a dialogue model. I will this explain in more detail. Augustus 2000

11 Top niveau dialoog model
1 2 3 Analysis has shown that a dialogue can be divided in four phases explanation phases Explain skipping second phase explain another question greeting: in this phase the operator and the client greet each other and often introduce themselves; the client with his or her name and the operator with travel information. This is followed by the opening question of the client that includes information about the trip to be made. If the client has provided enough information for the operator to give an answer immediately the second phase is skipped. In this second phase the operator asks questions to get the necessary information to be able to answer the client’s question. The operator asks for further information if there are still slots to be filled. During the third phase the operator retrieves information from the travel planner and presents this information to the client. After the travel scheme is presented it is possible that the client has another question which is indicated with this line. Because the operator might need to ask question again to answer this question and therefore returns to the query. In the last phase, the participants thank each other and say goodbye. And the dialogue ends. 4 Augustus 2000

12 Dialoog voorbeeld informatrice : goedemiddag reisinformatie
cliënt : goedemiddag ik wil graag weten wanneer de trein van Delft naar Amsterdam centraal vertrekt informatrice : wanneer wilt u aankomen cliënt : rond elf uur informatrice : u kunt vertrekken om twee voor tien cliënt : twee voor tien informatrice : en dan komt u om vijf voor elf aan in Amsterdam centraal cliënt : dank u wel dag informatrice : graag gedaan dag begroeting of opening dialoog vraagstelling of vullen van de slots informatie presentatie of antwoord geven Indication of four phases Description dialogue Here the four phases are indicated in a dialogue example. The operator and the client greet each other and the client states his opening question. Then the operator asks a question because she still needs to fill a slot. She asks for the arrival time. The client answers. Then the operator gives her answer and they say goodbye. afscheid of einde dialoog Augustus 2000

13 Lager niveau dialoog model
Each split into more details Following model derived from transitions Can not read this picture But shows there can be structure and can be more complicated than you think 80% of dialogues are included next two phases on next slide Now each phase is split into more details and the following lower level model is derived from the transitions between the dialogue acts. I do not expect you can read this picture because it is to large. But this shows that there can be an underlying structure and that it can be more complicated than you think 80% of the corpus dialogues are included in model. Because this model is not clear this way, the first two phases are shown on the next slide. Augustus 2000

14 Lager niveau dialoog model (1)
Description of every step in model same line as in top level model skipping the second phase Augustus 2000

15 Lager niveau dialoog model (2)
Augustus 2000

16 Onderzoeksdoelen Coderen van de dialogen Dialoog model ontwikkelen
Dialoog management bekeken vanuit de informatrice Modelleren van de kennis van de informatrice Met de verkregen kennis ontwikkelen van een trainingsomgeving Next step to explain is dialogue management from the operator’s point of view Augustus 2000

17 Directief dialoog fragment
informatrice : goedemiddag van welk station naar welk station wilt u reizen cliënt : goedemiddag ik wil van Delft naar Utrecht informatrice : wanneer wilt u vertrekken of aankomen cliënt : ik wil vanavond om acht uur vertrekken informatrice : ik kijk even een moment cliënt : ja informatrice : De verwachte aankomst in Utrecht centraal is om vijf voor negen en u moet dan in Rotterdam centraal overstappen cliënt : …. Augustus 2000

18 Non-directief dialoog fragment
informatrice : goedemiddag reisinformatie cliënt : goedemiddag ik wil naar Utrecht [ehm]. informatrice : ja... [noise] cliënt : en ik vertrek van Delft nee ….. van Rotterdam. informatrice : van Rotterdam naar Utrecht ja… cliënt : ik wil voor de lunch in Utrecht zijn informatrice : ieder uur rijdt er een trein en de eerste trein van Utrecht naar Rotterdam vertrekt om tien over zes cliënt : …. Augustus 2000

19 Dialoog stijlen Augustus 2000

20 De gedachtengang van de informatrice
Wat is de beste reactie op de prompt van de cliënt? Wat wil de cliënt weten? Is het begrijpelijk wat de cliënt zegt? Is het duidelijk wat de cliënt zegt? Augustus 2000

21 Mogelijke hypotheses Augustus 2000
Goedemiddag, ik wil om tien uur in Amsterdam aankomen en ik vertrek vanuit Utrecht. Hij wil waarschijnlijk om tien uur ‘s avonds aankomen omdat het nu al elf uur ‘s morgens is. Hij wil waarschijnlijk weten wanneer hij kan vertrekken. Hij wil waarschijnlijk vertrekken van Utrecht centraal. Hij wil waarschijnlijk vandaag reizen Hij wil waarschijnlijk naar Amsterdam centraal. Hij wil waarschijnlijk met de trein reizen. Augustus 2000

22 Mogelijke prompts Augustus 2000

23 Voorbeelden goede/foute prompt
Ik wil naar Amsterdam.. Ja... Amsterdam centraal? Ik wil graag morgen om negen uur in Delft zijn. Wanneer kan ik dan vanuit Utrecht vertrekken? Vertrekt u van Utrecht centraal? Ja…. Here you see two examples in which the same operator’s reaction can be seen. The same utterance can be a good one at one moment and a bad one at another moment. It is good if it gets the dialogue closer to the goal, that is providing the client with an answer. It also means that the operator can fill a slot. In the first example the operator notices that the client is not finished with question and she therefore encourages him to go on In the second example all the information is given so reacting with an alignment is not that sensible because the client probably gives no more information Augustus 2000

24 Dialoog strategieën Min-Max principe Verminderen van dubbelzinnigheid
Nieuwe informatie verkrijgen Informatie volledig maken Vermindering van fouten how to choose the best prompt; the following strategies apply min max principle  with minimal effort maximal results great efficiency if you think several steps ahead you know which prompt is getting you nearer to the goal; goal maximal filling of slots as soon as possible, so the information can be provided the information can be incomplete so she can ask a clarification question (central station), or the operator is insecure about the information; she can verify the information, or she misses some information; she can ask for the missing information Augustus 2000

25 Computational model Augustus 2000
Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

26 Computational model Augustus 2000 Informatrice:
goedemiddag reisinformatie Cliënt: goedemiddag, ik wil naar Utrecht Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

27 Computational model Augustus 2000 Informatrice: ja Cliënt:
en ik vertrek van Rotterdam Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

28 Computational model Augustus 2000 Informatrice : wanneer wilt u reizen
Cliënt : overmorgen Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

29 Computational model Augustus 2000 Informatrice :
u bedoelt hemelvaartsdag Cliënt : ja Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

30 Computational model Augustus 2000 Informatrice :
hoe laat wilt u aankomen Cliënt : ik wil rond acht uur vertrekken Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

31 Computational model Augustus 2000 Informatrice :
‘s morgens of ‘s avonds Cliënt : ‘s avonds Thoughts of operator are visualised in slots. Every circle corresponds with a slot the arrows correspond with an utterance of the operator and an utterance of the client explanation with dialogue parts, what is the operator thinking Augustus 2000

32 Onderzoeksdoelen Coderen van de dialogen Dialoog model ontwikkelen
Dialoog management bekeken vanuit de informatrice Modelleren van de kennis van de informatrice Met de verkregen kennis ontwikkelen van een trainingsomgeving And the last point of dialogue analysis I will explain further is modelling the operator’s knowledge Augustus 2000

33 Benodigde kennis Domeinkennis ‘Gezond verstand’ Ervaringskennis
Geografische kennis Dialoog management kennis Gebruik reisplanner To engage in a dialogue the operator has to have the following knowledge she has to know what the several specific public transport terms mean, such as Thalys, intercity she has to know how to interpret the time ; if the client is saying around two o’clock; can she give a travel option that is a quarter to or maybe only five minutes to two or five past she has to know how to handle the timetables so that she can provide the information of these time tables to the client common sense she has to use her common sense, but I think that is what everyone does for example; if a client calls at two o’clock and wants to travel at eight today it is clear to the operator that the client means eight o’clock in the evening dialogue management the operator has to know how she has to participate in a dialoguedat houdt in dat ze weet hoe ze moet sturen maar ook over de inhoud Augustus 2000

34 O.V.R. reisplanner Augustus 2000
operators use travel planner to look up information slots needed to provide information; explanation slots departure city arrival city date departure and arrival time city and location or station needed because they not only provide information about trains but also about busses, subways, operators need not only the city but also the location in the city operator only has to fill in the information; By choosing from the list in the combo boxes ;she does not have to know this info. Augustus 2000

35 Onderzoeksdoelen Coderen van de dialogen Dialoog model ontwikkelen
Dialoog management bekeken vanuit de informatrice Modelleren van de kennis van de informatrice Met de verkregen kennis ontwikkelen van een trainingsomgeving Augustus 2000

36 Demo Dialogue Training Environment Augustus 2000

37 Conclusies & Resultaten
Dialogen zijn succesvol te modelleren Dialogen hebben onderliggend model Er is een codeertool ontwikkeld Er is een (computational) model ontwikkeld voor dialoog management Domeinkennis uit dialogen geexpliciteerd Prototype voor dialoog training is ontwikkeld Augustus 2000

38 Aanbevelingen Lange termijn Nieuwe rol informatrice Korte termijn
Onderwerp van dialogen uitbreiden Verdere implementatie prototype Author tool Gebruik van spraakherkenning Augustus 2000

39 Augustus 2000

40 Augustus 2000

41 Augustus 2000

42 Augustus 2000

43 Augustus 2000

44 Augustus 2000