Two-picture matching game with robots  (e.g. Chen 2011)

In Chen (2011), a picture-matching game was used to elicit naturally spoken SVO sentences with either the topic–tail–focus structure or the focus–tail–topic structure in a controlled but interactive setting. The population were monolingual native speakers of Dutch, between the ages of 4;5 and 5;7. This description is adapted from that article in the Journal of Child Language.


In the game, the experimenter showed a child a picture each time and asked a question about the picture in order to find the matching picture. The question could be either a who-question (1) or a what-question (2):


(1) EXPERIMENTER:       Kijk! Een biet. Wie eet de biet?

ÔLook! A beet. Who is eating the beet ?'


PARTICIPANT :          [De poetsvrouw]focus eet [de biet]topic.

ÔThe cleaning-lady is eating the beet.'


(2) EXPERIMENTER:       Kijk! Een poetsvrouw. Wat pakt de poetsvrouw?

ÔLook! A cleaning-lady. What is the cleaning-lady picking up?'


PARTICIPANT :          [De poetsvrouw]topic pakt [een vaas]focus.

ÔThe cleaning-lady is picking up a vase.'


To ensure consistency in choice of word and word order in children's answers to the questions, we introduced a few virtual robots to the game as children's informants. The robots provided children with the Ôraw' answer to each question using the intended words in SVO word order via a headphone set. Crucially, the robots' sentences lacked normal sentence-level prosody such that they were hardly imitable: The words that made up the robots' answer sentences were first recorded in a randomized wordlist by a female native speaker of Dutch. The words in each sentence were then spliced together with a 200 ms pause in between to form the sentence. This way, no sentence-level intonational or rhythmical properties were present in the robots' sentences. In addition, the original pitch pattern was erased and the pitch level was set at 200 Hz using Praat (Boersma, 2001) such that the sentences sounded similar in pitch level across trials. A method similar to ours was used in MŸller et al. (2006) to create experimental materials.


 Children were supposed to reconstruct the robots' answers in their own intonation. Previous studies have shown that children's reproduction and comprehension were not affected by abnormal sentence-level prosody, although their reproduction could be affected by abnormal word order (e.g. Lahey, 1974). We therefore expected children to reconstruct the robots' sentences with no difficulty. However, to rule out misunderstandings on the task, we introduced up-front the rules of the game, whereby children should use the robots' words to answer the questions but should speak like they normally did (see below for the precise instructions given to the children). Further, in the light of Garrod and Pickering's (2004) alignment model for conversations, we expected children to have no difficulty with using the robots' words and word order to answer the questions. In addition, our informal survey of how children answered adult interlocutor's questions in a popular TV talk show with four- to seven-year-olds as the guests (ÔPraatjesmakers') revealed that children tended to answer both wh-questions and yes/no questions with full sentences or constituents larger than the words conveying the required information. This suggested that answering questions in full SVO sentences in our game should be natural for children.


Thirty-six pairs of pictures were included in the game and spread over thirty-six trials. Eighteen sentences with sentence-initial topic and sentence-final focus and eighteen sentences with sentence-initial focus and sentence-final topic were to be elicited from each participant. Each subject noun occurred in both the topic condition and the focus condition. The same held for each object noun. But each subject noun and each object noun occurred only once together in an answer sentence in the game.


Instructions to Children

I've got two boxes full of pictures here. The pictures are about a fantasy world where anything goes. For example, animals can speak like humans and tables can kick objects with their legs. A picture from one box goes together with a picture from the other box. But the pictures have got mixed up and I need your help to sort the pictures out. Every time, I take a picture from one box and ask you a question about it. Of course you don't know the answer either. But you will get help from a robot. The robot will tell you the answer via a headphone set so that only you can hear what the robot says. You will then know the answer and be able to answer my question. There will be four robots to help you in the game. The robots always know the answers but they speak in a very funny way. You will use the robots' words to answer my questions but you should speak like you normally do. These are the rules of the game. With the help of your answers, I can then find the matching pictures.

Trial procedure

1) the experimenter took a picture (e.g. a picture of a cleaning-lady) from one box, drew the child's attention to the picture, and established what the picture was by saying Kijk! Een poetsvrouw! ÔLook! A cleaninglady!' with a high-pitch stretch or a weak rise on the attention-getter Kijk, and a rise on the noun to indicate that more was to come following the naming. These intonation patterns are common in comparable conversational situations in Dutch. Often the child chatted a bit about the picture with the experimenter. If the child did not do so but the word was identified as a potentially difficult word for four- to five-year-olds in our pilot tests, the experimenter took the initiative to talk a bit about the picture. This was done to make sure that the entity on the picture was referentially given to the child by the time a question about the entity was asked.


2) The experimenter asked a question about the already introduced referent (e.g. In the picture, the cleaning-lady seemed to be picking up something; the question was then Wat pakt de poetsvrouw? ÔWhat is the cleaning-lady picking up?'), with a high-pitch stretch or a weak rise on the wh-word, a fall or a downstepped fall on the verb, an optional downstepped fall on the noun, and a low tone at the end of the wh-question. An informal survey of wh-questions uttered as information-seeking questions in the Spoken Dutch Corpus (Oostdijk, 2000) showed that these patterns are commonly used in wh-questions in spoken Dutch.


3) Following the experimenter's question, the child turned to a robot for help by clicking on a picture of the robot displayed on his or her computer screen. The child received the answer from the robot via a headphone set. Four robots were used in the game; they were randomly assigned to the trials.


4) Having heard what the robot said, the child answered the experimenter's question using the same words as the robot but in his or her own intonation (e.g. De poetsvrouw pakt een vaas. ÔThe cleaning-lady is picking up a vase.').


5) The experimenter looked for the matching picture (e.g. the picture of a vase) in the other box and handed over both pictures to the child.