Language competence in NHPs

An assessment of the field in the light of a 'universal grammar'


"The Berlin wall is down, and so is the wall that separates man from chimpanzee."
(Elizabeth Bates)

"There is no debate, so I have no opinion."
(Noam Chomsky)


0 Introduction

The language competence of non-human primates is one of the most controversial issues in present-day linguistics, with disbelief ranging from bored indifference to vitriolic accusations of fraud. The present paper aims to assess the current state of debate from an open-minded, critical and detached perspective.

In a first part, a brief outline of earlier research in the language abilities of non-human primates - more precisely of apes (bonobos, urang-utangs, chimpanzees and gorillas) - is sketched. The second part focusses on the landmark studies published by Dr. Emily Sue Savage-Rumbaugh and her colleagues. A third section looks into the views of the Chomskyan field, leading up to the concluding section on the innateness debate.
 
 

1 Early research on non-human primates' capability for language
 
 

1.1 Attempts to teach NHPs to speak

The language capability of non-human primates has been a subject of research since the beginning of this century. In 1909 already did Witmer attempt to teach a chimpanzee to talk. He claims that the chimpanzee was capable of articulating the word ‘mama’. In 1916 Furness taught an orang-utan to say the words ‘papa’ and ‘cup’. After the unexpected death of this orang-utan, Kellogg and Kellogg wanted to follow up this work. In 1933 they obtained a young orang-utan – Gua – for the purpose of co-rearing it with their 9 1/2-month-old son Donald. By 12 1/2 months Gua evinced comprehension of 21 different requests while Donald, then 12 1/2 months old, was responding to 20 different requests. From that point on, Donald’s comprehension began to overtake Gua’s. It has to be said that the responses were always utterances which they heard frequently, were mostly action oriented and did not entail selection of a single object from a group of objects. Most of the time, they were uttered in a specific context which simplified interpretation. The last researchers to have tried to teach a non human primate to actually articulate spoken language are Hayes and Hayes in 1951. They raised Viki, a chimpanzee, in their home environment . After six and a half years of formal speech training – which was carefully documented – they say one was able to distinguish the words ‘papa’, ‘mama’, ‘cup’ and ‘up’. However, they do acknowledge that these words were very poorly articulated and doubt the referential function of these words to Viki. All this research shows that the belief that man is the only animal capable of communicating in language has long been questioned. Already as early as 1661 Samuel Pepys wrote in his diary :

I do believe it (a baboone) already understands much English ; and I am of the mind it might be taught to speak or make signs  (qtd. in Wallman 1992:11) 1.2 Projects in which NHPs are taught language

After it had been proven that non-human primates are physically incapable of articulating words because the shape of their vocal tracts disables them to articulate consonants (cp. infra), researchers tried different ways of teaching language to the ape. Some have used the American Sign Language (ASL) for the deaf and others have developed an artificial language.
 
 
 
 

1.2.1 The Washoe project

The first such project was the Washoe project, fielded by Gardner and Gardner. Washoe, a wild-born female chimpanzee, was acquired in June 1966. At that time, she was believed to be approximately 10 months old. She lived in a trailer in the Gardners’ backyard. This trailer was provided with furnishings and playthings so as to create an intellectually stimulating environment (like a child’s). The researchers’ aim was to teach Washoe ASL. In order to accomplish this, during her waking hours, Washoe was always accompanied by someone from the project personnel who mastered ASL. Aside from signing requests and comments, Washoe was also taught her signs through moulding and shaping. In moulding, the teacher physically guides the animal’s hand into the proper configuration, location, and/or movement for the sign. Shaping is the practice of rewarding successively closer approximations to a target behaviour, here execution of a certain sign. Except for these two techniques, the researchers hoped that Washoe would also learn ASL by way of imitation. A later study has shown that shaping and imitation are not efficient in teaching ASL to apes so moulding has become the dominant pedagogical instrument. Washoe’s utterances were recorded daily and by 36 months of age, she produced 85 different signs in the appropriate contextual situations and had even started to combine them as well. The Gardners’ initial conclusions that signs were much more readily acquired by apes than spoken language, that moulding greatly enhanced the rate of sign acquisition and that apes generalized their signs to exemplars not utilized during training, were supported by subsequent studies. It should be noted, however, that "the anatomy and motor control of the human hand is much superior to that of the ape hand " (Savage-Rumbaugh et al. 1998:14).
 
 

1.2.2 The Sarah project

In 1973, David Premack started teaching Sarah, a 5-year-old chimpanzee, an artificial language. She had to answer simple questions using plastic tokens instead of words. These tokens were metal backed and could be arranged on a magnetic board. The form-meaning relationship of these tokens was completely arbitrary (e.g. a blue triangle meant "apple"). Sarah was given elaborate step-by-step training in both definitions and sequence of tokens within a statement. Her task was to arrange the symbols on a magnetic board associating particular symbols with particular meanings. A shaping procedure was developed in which Sarah was rewarded for increases in the complexity of her performance. The major critique of the Sarah-project was made by Terrace(1979a), a fellow behavioral psychologist. First of all, he claims that

considering that almost all of her training, and testing sessions involved only a single paradigm, and that the alternatives being contrasted within a paradigm were almost always just two in number, it is quite possible that Sarah could attain her characteristic 80 percent level of correctness (Premack 1971) on a paradigm through memorization of that session’s construction and the correct, that is, rewarded, response for each. (qtd. in Wallman 1992:39) The transfer tests constitute a second point of criticism. These are intended to determine in how far the animal’s experience during training has become a general capacity. Premack has put the results of all the trials of these transfer tests into a single statistic and has thus disabled a solid interpretation of the transfer tests. Since the main aim of such a test is to see whether an animal has induced an abstract, context-independent solution for a problem type, only the results of the first presentation of each new problem are of importance, since performance on subsequent trials can be confused with the effect of new learning. The possibility exists that the ape might have acquired a way to quickly apprehend which behaviour is the rewarded one.

Terrace also criticizes the claimed ability of Sarah to comprehend complex sentences such as ‘Sarah banana pail apple dish insert’. He ascribes this success to the understanding of the hierarchical structure of such sentences, namely that the scope of application of "Sarah" and "insert" is the entire sentence.

To summarize the criticism one can say that Sarah has developed a skill in solving a number of problems devoid of linguistic significance. The narrow focus of each training session and the small number of alternatives might indicate that she could have memorized the correct responses. Some shortcomings of the project were that Sarah was not allowed to ‘speak’ spontaneously and that she only had the chips being used during each training paradigm available. Thus she could not have come to understand that her constructions with the chips or those of the trainer were communicative in nature.
 
 

1.2.3 The Lana project

The Lana (LANguage Analogue) project of Rumbaugh and colleagues began in 1972. The aim was to teach Lana, a two-year-old chimpanzee, a language created especially for the study. Yerkish, as the language is named, uses "a grammar of the correlational type, in which rules of sentence formation specify permissible combinations of word types, the word types being based on semantic attributes." (Wallman 1992:18) The vocabulary of this language consists of lexigrams. These are abstract symbols arrayed on a keyboard in Lana’s cage. Lana has a 24-hour access to the keyboard enabling her full access to everything the computer can provide such as food and drink, show slides and a movie, provide a view out of the window, and so on. The sole condition for her to obtain these things is that Lana first composes a grammatically acceptable lexigram sequence. The keys were regularly redistributed around her keyboard so as to ensure that Lana was learning a sequence of lexigrams rather than a sequence of positions on her keyboard. Above this keyboard are two rows of each seven projectors. The first row displays the lexigram corresponding to each key Lana pushes on her keyboard. The second row shows the communications from a trainer in an adjacent room.

This computersystem has many advantages. A first one is that it poses no articulary burden on the animal. A second advantage is that the durable image of the lexigram sequence on the projectors eliminates short-term memory problems. Another positive aspect is that the use of computers prevents the tainting of Lana’s linguistic performance and rules out possible cueing by the trainers(cp . Clever Hans). A last and very important advantage is the exhaustive and objective record of all linguistic transactions between Lana and the computer or Lana and her trainers.

Lana was taught to start every utterance by "please". A majority of her utterances were based on stock sentences such as " Please machine give … ". Rumbaugh also claims that she showed signs of understanding grammatical structures. She responded differently to sentence beginnings provided by her trainer. She would push the ‘period’ key when given an incorrect beginning of a sentence. When she was given a correct beginning, she would finish the sentence.

Just as was the case in the Sarah-project, this project also had to deal with serious criticism. A good summary is given in Wallman (1992:37):

There is a paucity of evidence that the lexigrams in Lana’s strings had separable meaning for her and substantial indication that her sequences were rotely acquired wholes, a limited number of stock-phrases developed for each of a small number of contexts, with substitution of lexigrams in variable frames, depending upon what incentives and/or trainers were present. Further remarks are made concerning the claim that Lana would understand grammatical structures. Considering the fact that all of the keys on the keyboard have a specific colour and each colour designates a specific semantic category, the possibility exists that Lana might have cracked this colour code and thus rather than understanding syntactic structures, Lana might have memorized a chain of well-ingrained stimulus-response sequences.

Remarkable is the fact that this criticism was supported by the researchers of the project themselves. Rumbaugh and Savage-Rumbaugh have reassessed the Lana project, by way of contrasting it with their more recent work, which they feel avoids the shortcomings of the Lana project.

1.2.4 The Nim project

Another project was the one conducted by Terrace and colleagues on Nim. The full name of this chimpanzee is Nim Chimpsky referring to Noam Chomsky, who rejects the possibility that non-human primates might be capable of producing language. The main plan of the Nim project was to raise a chimpanzee in an environment approximating that of a child, including, especially, exposure to human language, and chart the animal’s acquisition, if any, of the language. The language that was to be taught was ASL although it was in many respects a reduced form of the true sign language. That is why it is called a pidgin sign language.

Nim was two weeks old when he was first exposed to sign language, but systematic language training did not start until he was nine months old. At that time, he lived with three student research assistants. A small room was prepared for the core training which took place for five hours every weekday. This room facilitated the observation and recording of Nim’s language performances, viz. the acquisition of 125 signs by the end of the project, when Nim was three years and eight months old. Although no controlled experiments on Nim’s vocabulary were conducted, informal tests and his general pattern of sign use demonstrated that Nim had developed consistent associations between signs and referents.

In many respects this project was compared to the earlier conducted Washoe project. Both Washoe and Nim lived in an intellectually stimulating environment in which they were exposed to ASL. Not only the way in which the projects were conducted was similar but also the criticism that both received. The main person to criticise the Nim project was Terrace himself. After having reviewed the videotapes of Nim’s signing sessions, he demonstrated that most of the multisign utterances were preceded by similar utterances on the part of his teacher. Terrace also analyzed segments of a film distiributed by the Gardners and came to the conclusion that the same phenomenon existed in Washoe’s behaviour. In conclusion, Terrace rejected the interpretations of Nim’s productions in favor of a nonsyntactic one. However, one might say that the system effectively taught Nim to imitate.
 
 

1.2.5 Koko and Francine Patterson

Another type of research was carried out by Francine Patterson with a female gorilla called Koko. The project started out in 1972 and is still continuing today, making it the longest running study of language abilities of an ape. The project is particularly unique because of three things: firstly, it remains the only linguistic study involving a gorilla. Secondly, Koko was brought up as a bilingual; she learnt both American Sign Language and spoken English, the two were even often used simultaneously. Thirdly, Koko’s linguistic achievements surpassed those of any of the other ASL apes (Table 1).

Table 1. Number of signs reported learned (for ASL apes) (Davies on Miles 1994: online)


APE Number of signs reported learned
Nim 125
Washoe 133
Loulis 70
Sarah 130
Chantek 150
Koko 224
(a deaf child of three years) 132

Patterson, who was inspired by the Gardners and their ‘Washoe project’ (cp. supra), insists that gorillas are "peace-loving vegetarians" who have suffered from "bad press", which has, undeservedly, branded gorillas "ferocious, stubborn, and stupid" (Patterson and Linden 1981: ch.3). However, she found out that the gorilla "is less aggressive, less excitable, and in some ways a good deal easier to work with than she had anticipated"(ibid.); therefore she decided to try and teach her gorilla the use of American Sign Language (ASL). At first, the animal easily matched some chimps’ vocabulary but then quickly outpaced them. At the age of three, Patterson reports, Koko signed about food most of the time. By age six, having passed infancy, Koko apparently started expressing her ideas about language and even began to call one of her female caretakers ‘bird’ and ‘nut’ in a pejorative way.

But more importantly, Koko has shown other feats which were only matched by a minority of the other non-human primates, according to Patterson. For instance, Koko quickly created new terms for unnamed objects (e.g. ‘eye hat’ for mask). Consequently, Koko showed an apparent ability to rhyme (use words that would rhyme in spoken English; e.g. "flower stink, fruit pink") and to jest about one of her caretakers (You dirty bad toilet). Furthermore, she manifested an ability to lie (Hawes 1995: online reports of an instance where Koko denied eating a crayon; she signed that she was applying lipstick). Finally, the gorilla seemed to have mastered the concepts of life and death, when referring to her demised pet kitten (P: Where do you go once dead? K: Comfortable hole bye). Patterson also tested Koko’s I.Q. and reported scores between 70 and 95 (human average being around 100).

As Patterson kept hauling up new exploits about metaphors, jokes, word games and Koko’s great creativity, many of the critics portrayed Koko’s abilities as "an elaborate hoax" (Tof 1996: online), not in the least because Patterson has constantly shrouded her research in secrecy. She hasn’t brought out any scientific publication about her project since 1981; nevertheless, the funding for the project has continued. She even admitted to lacking substantial, empirical proof to support her claims but nonetheless rated the anecdotal material, which had been collated, very highly.

An AOL chat session, during which people could ask Koko and Patterson questions, shows why some researchers have doubts about Koko’s reported abilities. The following excerpt displays heavy interpretation on behalf of Patterson.



Transcript of AOL chat-session (Koko online transcript 1998: online)

AOL: SickboyRE asks: Koko, have you taught other gorillas sign language, on your own?

PENNY: That's a good question. Have you taught other gorillas to sign?

KOKO: Myself lip.

PENNY: She taught herself. That's really true, too. That's very good and I think what part of that answer might be, is that she's taught us. In other words, 'myself lip' was her answer and 'lip' is her word for woman. So 'herself' has taught 'lips', perhaps. So there are a couple of interpretations there


Even Sue Savage-Rumbaugh, whom one thinks would be in the best position to judge, found it difficult to accept reports about Koko’s alleged puns, word play and knowledge of a concept like death; she said it "strained her credulity" (Savage-Rumbaugh and Lewin 1994: 46). Furthermore, she thought that Patterson "lacks the test data to rule out the impact of context-information" (o.c.:148).

Another major point of criticism was voiced by Terrace after he had evaluated a number of films of both Washoe and Koko. He concluded that both apes were as imitative as his own ape, Nim. Therefore, any evidence for grammatical patterning will be found inconclusive unless indications of imitation by the apes can be ruled out. Patterson responded by introducing blind tests, in which the subjects could not see the person giving them instructions.

Patterson’s research illustrates the dichotomy that has burgeoned in the scientific community. She belongs to the camp of the researchers who treat their apes more like children and focus on the accomplishments of their subjects and the similarities between ape and human language. Patterson believes that one needs a "true rapport" (Patterson and Linden 1981: 211) with the apes in order to "bring them to the limits of their abilities" (ibid.). For Koko, this rapport meant appreciating her representational drawings, talking about her feelings, laughing at her jokes and asking how her night was; in short, treating her like a person.

The researchers in the other camp treat their apes like experimental subjects; they focus on the failures of the apes and on the differences between ape and human language.

To conclude, it is important to see the need for a happy medium and not lose track of the central question, which is not about the moral issues surrounding the non-human primate language debate. Patterson has unquestionably become very attached to Koko in the 25 years she has worked with Koko and can therefore not be counted as a valid judge to interpret her language abilities. Savage-Rumbaugh circumvented this problem by using a number of ‘independent’ observers (cp. infra). However, it must be said that ASL will always ‘lend itself more difficult to interpret’ than lexigram use and comprehension tests Savage-Rumbaugh used.

On the other hand, human language can not be used as a sole standard to measure ape language by and differences between ape and human language must not be endlessly exaggerated. Furthermore, some rapport between the caretaker and the subject is needed as a catalyst to ensure some kind of understanding (cp. Nim Chimpsky project: a possible reason for the failure is that Nim had 60 different trainers, obviously too many for a decent rapport).
 
 
 
 

1.2.6 The Primatech Project

This project differs completely from previous projects in that it does not exclusively focus on non-human primate language. The very novel aim of the Primatech project is to create an interactive simulation of a signing orang-utan.

Central in this project is Chantek, an orang-utan nearing age 22 who thinks like a 4-year old human, according to his caretakers. Lyn Miles, an anthropologist from the University of Tennessee who ‘brought up’ Chantek, says the orang-utan’s skills resemble those of other primates like Koko. This research follows her general creed that "there is extraterrestrial intelligence – right here on earth, in the orang-utan" (Hawes 1995: online). As opposed to the education of other apes, Lyn Miles adopted a "developmental perspective that seeks to identify the cognitive and communicative processes that might underlie language development" (Miles 1990: 512).

Chantek has apparently mastered at least 150 words in ASL and has learned to clean his room, memorize the way to McDonald’s, where he spends his allowance, and lie ("He'd tell me he had to go to the bathroom and then go in there just to play with the knobs on the toilet" Lyn Miles: online article). However, Miles insisted on calling Chantek’s utterances ‘Pidgin Sign English’, because the "sentences are composed of signs combined in English word order but are devoid of articles and lack most of the grammatical morphemes of English" (Wallman 1992: 28). On the whole, Chantek’s achievements do not differ remarkably from those in the other projects. However, he did overcome some of the errors that were found in the Nim project.

Table 2. Statistics Nim vs. Chantek(Davies on Miles 1990a: online)


  NIM CHANTEK
imitation rate 38% 3%
spontaneity rate 4% 49-88%
interruptions of caregiver 35% 8%
interruptions of ape ~18% 9%

Like Francine Patterson with Koko (cp. supra), Lyn Miles reported a number of amazing achievements for Chantek. The orang-utan too can create novel words (e.g. ‘eye drink’ for contact lens solution) and lie (Miles calculated that Chantek lied an average of three times a week). Furthermore, Chantek seemed to have invented some of his own grammatical rules; he would always place an action in front of a visible object (e.g. ‘play ball’) and put an invisible object in front of an action (e.g. ‘ball, play).

Miles’s strongest claims were about reference. Comparing Chantek’s development to that of a human child, she found evidence that Chantek was using words and signs referentially. Like human children, Chantek showed different stages in his referential communication; he went from pointing to an object, to "labelling and semantic overgeneralization or undergeneralization" (Miles 1990:524) and ended up with representational naming.

As orang-utans mostly live a solitary life in the wild, the education of an orang-utan partly rescinded earlier scientific claims that a primate’s perceived language ability rests on his need for social interactions with other individuals. Offering another account for the language ability, Miles called on other, general cognitive capabilities of apes as underlying their apparent capacity to learn human language. First of all, primates seem to be able to "remember over the course of several years the location of various reliable fruiting trees within a vast territory" (Hawes 1995: online). Subsequently, primates show a remarkable expertise with tools (Chantek used 22 tools in a sequence once).

However, Miles says the ‘social living theory’ must not totally be written off; orang-utans may well have been more ‘social’ previously and were driven to a more solitary way of life due to environmental degradation. Human disturbance brought along a move to new habitats, where food was scarcer, thus increasing the competition within the groups.

Where previous projects concentrated on the ape and his ability to both sign and understand signing, the Primatech Project incorporates the ape’s abilities in a larger frame. Jim Davies, from the college of computing at Georgia Tech, who is co-ordinating the project, says the aim of Primatech is to "create an installation that simulates communicating with an orang-utan" (Davies 1999: online). Davies acknowledges that ape language research is a controversial subject and suggests that the best approach to modelling ape language is to "rely on basic cognition before introducing higher-level language structures" (Davies, Primatech Abstract 1999: online).

If all goes well, members of the public will ultimately be able to communicate with a virtual orang-utan, whose movements, signing, and language abilities are modelled on those of Chantek. Figure 1 schematizes these processes.

Figure 1. Schematic representation of the Primatech project

The research will cover three domains. Firstly, Animation will be of great importance for the virtual orang-utan. Based on a 3D model of Chantek, the graphical model will move in predefined ways, which correspond to actual orang-utan anatomy. Existing theories of animal behaviour will also figure prominently in the shaping of the graphical model.

Secondly, Intelligence, and, more specifically, artificial intelligence will be used to both recognise and generate language like a non-human primate would. Especially this computational modelling of primate language use is state of the art in primate language research and will certainly shed a new light on the issue.

A third important aspect of the Primatech Project is voice and sign language recognition. Once again, the installation will be modelled on Chantek in that it will recognise a subset of spoken English and ASL. On a first level, utterances will be recorded by both a camera and a microphone; then they will be analysed by a recognizer which will send them to the intelligence module (cp. supra). The system should be able to handle a number of different grammars and vocabularies.

In the end, the simulation will be placed in the Atlanta Zoo, where visitors will be able to communicate with the virtual orang-utan.
 
 

1.2.7 The Sherman and Austin project

So far, none of the projects have dealt with the question whether the signs or keys that are used by the apes have a referential function to them. The focus has been on the capacity to master a certain degree of syntax. What makes the Sherman and Austin project so important is its investigation of the cognitive competencies that underlie symbolic processes.

The Sherman and Austin project started in 1975, at which time Sherman was two and a half and Austin was one and a half years old. This study was performed by Savage-Rumbaugh and Rumbaugh. They were interested in the animals' ability to comprehend and to communicate, they wanted Sherman and Austin, two chimpanzees, to use symbols referentially. The medium to do so was a modified version of the computerized keyboard used in the Lana project. Symbols that represent objects or activities are arbitrary geometric forms based on the Yerkish language. The keyboard eventually grew to 92 symbols. So as to prevent that the chimps would learn keyboard positions, the keys were randomly rearranged after every keyboard use. Another medium used to get the chimps to do something was speech. After much research on how to teach a chimp vocabulary, they used the ‘request task’, meaning "I show you X, you select the lexigram that goes with X, and then I give you X". In this case, the ape selects the symbol or makes the sign on the basis of what it anticipates will happen afterward. From these experiences flowed a means of communicating wishes, which Sherman and Austin unexpectedly learned on their own. The ability to request is just one of three major abilities in communication. The other two are (i) the ability to name objects and (ii) a comprehension of symbols as referring to objects. In order to teach the chimps the naming of objects without receiving them afterwards, a technique called ‘fading’ was applied. This refers to the fact that the size of the food item given to the chimp gradually diminishes, while the size of the item he is being shown does not change.

For example, if the food we wanted them to name was "sweet potato", we gave them a smaller and smaller piece of sweet potato each time they did so, along with a large piece of some other food, to indicate we were still pleased. (Savage-Rumbaugh and Lewin 1994:68-69) After having learned this for three food items, tests were held to see whether they could generalize the concept to the other symbols in their vocabulary. They did so easily. By having learned this, Sherman and Austin had learned to make a distinction between communication for its own ends and communication about the properties of the units of the system. The next step was to teach them receptive skills, comprehension. To do so, the trainer went into their room with a container with food in it. The chimps knew that when she entered with food in her hands, they were allowed to ask for the food with their keyboard. Since they did not know the content of the container, they had to listen to the trainer in order to find out what was in it. After the trainer had told them what was in it, they went to their keyboards and asked for the food. This shows that they had been listening to what the trainer was saying. The next step is to make them tell each other what is in the container. One of the chimps (the informer) was allowed to see what was being put inside the container. When he came back in his room, he pushed the correct key thus informing the other chimp of the contents of the container. Only if both had correctly requested the food in the container were the contents of the container shared. Through this procedure Sherman and Austin had learned both interindividual communication and cooperation. Thus, they even learned to share their food which is highly unusual for chimpanzees. With surprising ease, Sherman and Austin learned to attend to each other, attend to the requests made, and proffer the specified food. The experiment was taken one step further : the food was hidden in sites in their room. In order to obtain the food, the chimp needed one of six tools (key, money, stick, straw, sponge, and wrench). The chimpanzees were put into two rooms separated from each other by a window. One chimp would have access to the tools while the other would be in the room where food was put in a site. They immediately understood how to obtain the food. The one in the room containing food would use his keyboard to tell the other which tool he needed. The tool-supplier would then give the appropriate tool and in the end they shared the obtained food. This experiment proves that the chimps not only comprehend but also show intentionality.

Although these results were extremely pertinent to the ape-language debate, little attention was paid to them since the supposed primacy of syntax still held most researchers in its thrall. The little response that was received was from the behaviorist camp that claimed that these experiments could be compared to previous experiments with pigeons. Savage-Rumbaugh and Rumbaugh replied to this assault by saying that Sherman and Austin were exhibiting conscious intentionality during their communication. To prove this, it was important to demonstrate beyond reasonable doubt not only that Sherman and Austin could communicate with each other, but also that they indeed knew that they were communicating. To do so, another experiment was set up in which one of the chimps would see what food was put in a container, was then brought back to his room where the keyboard had been turned off and had to try and tell the other which food was in the container with the help of food labels that were scattered on the floor. The chimps had seen the food labels before when being around the refrigerator. This experiment entailed many presuppositions: would they feel the need to tell the other what was in the container when the keyboard was turned off? would they recognise the label? and would they think of using the labels as a means of communication? But surprisingly, Sherman and Austin had no difficulties with the situation. The test was continued with 30 different food labels all of which were correctly recognised by Sherman and Austin. Another experiment that was carried out was to introduce the idea of generic symbols. This experiment was split up into three stages: first, they were taught the categories ‘food’ and ‘tool’ after which they were shown food items and tools that had not been part of the training regime. Then the same process was repeated with plastic-covered photographs of food items and tools. Finally, they were shown symbols for food items and tools. This third part of the experiment was crucial in that it could prove the representational value of Sherman and Austin’s symbols. The results of the tests were very positive; both chimps scored successes on all three tasks.

In conclusion,

Sherman and Austin did not merely imitate their teachers : they were economic and informative with their symbol production; they did take turns in communication, both with each other and with their teachers; and finally, although much of their symbol product was request oriented, they moved beyond this sage to spontaneous symbol production either to announce or comment on their action. (Savage-Rumbaugh and Lewin 1994:91) 2 Kanzi and beyond
 
 

2.1 Introduction

Even though the Washoe project yielded interesting results and received ample public attention, the Gardners were unable to refute Terrace's claims. Rather, they replicated their work with other chimpanzee subjects, but again they failed to address the crucial issues, those of imitation and inadvertent cueing by researchers (Savage-Rumbaugh et al. 1993:8). The report of Fouts et al. (1989) of signing between apes without humans present, failed to show that "these gestures differed in function from the many nonverbal gestures and 'body language' that the chimpanzees employ as well" (Savage-Rumbaugh et al. 1993:8). After the ground-breaking work with Sherman and Austin, reported above, it was finally the work with Kanzi that provided a major break-through in ape language research, as this research would also reveal syntactic competence in addition to (decontextualized) symbol comprehension.
 
 

2.2 Introducing Kanzi

The bonobo (Pan Paniscus) Kanzi was born in 1980 and he picked up language by observing his mother's language training. Kanzi was not taught the use of lexigrams actively, but appeared to have mastered their use on his own, even though he had never seemed to be particularly interested in his mother's language course. Indeed, at two and a half years of age, he correctly employed nearly all of the ten lexigrams that were on his mother's keyboard at that time, and what is more, he also understood the spoken English words which the lexigrams represented (Kanzi 1998: online). Whereas Sherman and Austin seemed to understand spoken language, carefully controlled tests showed that true speech comprehension was lacking (ibid.). Kanzi, however, is truly capable of symbol-to-symbol transfer: he can listen to a word and select the written or printed symbol for it. But Kanzi can do more than that, as the subsequent paragraphs will demonstrate.
 
 
 
 

2.3 Kanzi's comprehensive abilities compared to those of a human child
 
 

2.3.1 Basic tenets

A detailed study of Kanzi's comprehensive competence, in comparison to that of a human child named Alia, is offered in Savage-Rumbaugh et al. 1993. In accordance with recent findings in research on child language, it is one of the authors' basic tenets that comprehension precedes production when language is not trained but acquired naturally through observation (cp. infra). After all, in basic language production all that is required is the 'wording' of the speaker's own ideas or mental states, whereas even the most elementary comprehension precludes foreknowledge: the other's intentions cannot be known to the hearer. As both these data, and data from child language research (as summarized in Savage-Rumbaugh et. al. 1993:16-23), have shown, comprehension abilities emerge earlier and rise to more impressive levels earlier than productive skills. As Savage-Rumbaugh et al. note:

Children do not learn language by talking; they learn it by listening. (o.c.:24) When training emphasizes comprehension, production will occur as a "spontaneous by-product" of such training (o.c.:19). Comprehension does not require reinforcement: an infant will naturally strive to "understand" in order to "predict what the speaker is going to do as a consequence of having produced a particular utterance" (ibid.).

Another basic tenet of this (and subsequent) research by the team led by Dr. Sue Savage-Rumbaugh is that reward-based rote learning has to be abandoned. The failure of Matata, Kanzi's mother, to learn the use of lexigrams with any amount of ease, and the subsequent success of Kanzi himself, have convincingly shown that, as with children, NHPs do not need special training to acquire language. The only condition for observational learning is that they are exposed to language ("immersion") from an early age. Indeed, for NHPs as for children, there appears to be a "critical period" for language, due to the higher brain plasticity of infants. It would seem that brains of language-competent NHPs outweigh those of nonlanguage-competent ones (Savage-Rumbaugh et al. 1998:213).
 
 

2.3.2 Reference: learning how to mean

It remains to a large extent a mystery - often referred to as Quine's dilemma - how infants determine where words begin and end, and how they "come to understand that words refer to particular objects, events, emotions, etc." (o.c.:13), irrespective of the presence or absence of these objects (cp. "Kanzi, don't bite the ball", "Where's your ball?", "He's having a ball", ibid.). A "process-oriented perspective of reference" (o.c.:15) is proposed, in which the use of words is seen as an attempt to "bring about a certain behavior or set of beliefs in the listener" (o.c.:14), based on previous perceived utterances of a word and the various events surrounding its previous use.

Rather than learning words to NHPs through associative training in "naming", a successful approach is that of interindividual routines. When an ape learns to name a "banana", it will not be able to respond appropriately to messages regarding the state, absence, location, etc. of the bananas (o.c.:15). Indeed, it is precisely the decontextualized use of symbols that points at true comprehension:

It is easy to train an ape to say apple in order to get an apple but difficult to teach it to use apple to describe a food that it is not allowed to eat, a food that it sees someone else eating, a food that it does not like, a food that is in a particular location, etc. (ibid.) Once the stage of decontextualization has been reached, infants are able to assign a new word to a new referent.

As already noted in passing, the language acquisition model based on observational learning describes infants' symbol acquisition in terms of the learning of interindividual routines, as observed by Savage-Rumbaugh et al. (1993) in four apes and two human children.

The term interindividual routine is defined as a "more or less regularly sequenced set of interindividual interactions that occur in a relatively similar manner on different occasions" (o.c.:25). Apart from verbal markers, postural and gestural markers are equally important in engaging in such an interaction. As is observed regularly in publications by Savage-Rumbaugh and her team, such additional markers are important to signal the intentionality of the communication taking place. Once a routine (such as 'changing diapers' or 'playing with bubbles') and its markers - verbal and others - are understood, the infant can manipulate the course of events by using these markers to initiate the routine. (Note again how comprehension precedes production.) These initiations grow gradually less context-dependent: instead of picking up the bottle of bubbles, for instance, infants will say "bubbles" or point to the "bubbles" lexigram, thus assuming the role of a "symbolic communicator capable of announcing their intentions to another party" (o.c.:30). The ultimate driving force behind language acquisition, according to this model, is then

not the caretaker; rather, it is the desire of the child or ape accurately to predict what is going to happen to him or her next that motivates the attention toward the acquisition of vocal and gestural markers. (o.c.:32) The knowledge of specific referents comes "from a group of intermeshed routines that have overlapping markers" (o.c.:33): from routines such as "playing with the bubbles", "helping find the bubbles", "putting the bubbles in the bath water" and "hide the bubbles", it appears that the item common to all these routines is called "(bottle of) bubbles".
 
 

2.3.3 Methodology of the comparative study in Savage-Rumbaugh et al. 1993

The rearing histories of Kanzi and Alia were very similar. They were both exposed to both spoken English and lexigrams from infancy, but they were not "trained" to talk. They shared a prime caregiver, Jeannine Murphy, who spent forenoons with Kanzi and afternoons with her daughter Alia, in a mobile home with approximately the same indoor environment. Both subjects were separated from the experimenter by a mirror during the tests.

During a number of nonblind trials, the subjects were "accustomed to responding to requests produced by a disembodied voice" (o.c.:50). The blind trials started after 244 trials for Kanzi, and 180 for Alia. "For purposes of data analysis, only trials subsequent to the 244th are treated as 'blind' trials for both subjects." (o.c.:51) In the trials, the subjects were asked to perform one or other action. This required them to understand novel, and often very unusual, sentences and to act accordingly, without any further contextual cues. During these tests, several objects were in front of the subjects, so they had to select the correct one to perform an action on it (table 3). Moreover, several sentences were given for each array of objects:

For example, the sentences "Go vacuum Liz," " Go put some soap on Liz," "Rose is gonna chase Kanzi," "Put on the monster mask and scare Linda," "I want Kanzi to grab Rose," "Take the mushrooms to Matata," and "Kanzi is going to chase Rose" were all presented while the array in front of Kanzi consisted of the following items: monster mask, ball, bunny puppet, sweet potato, melon, soap, umbrella, straw, toothbrush, toy gorilla, hose, vacuum, mushrooms, two televisions, and a shoe. The persons in the room were Liz, Linda, Rose, and Kelly, and the available locations were tool room, colony room, microwave, refrigerator, outdoors, and potty. (o.c. 53)
Table 3. Mean number of objects and agents in the arrays facing Kanzi and Alia

(Savage-Rumbaugh et al. 1993:51)

 
KANZI
ALIA
TRIALS
Objects
Agents
Objects
Agents
1-100
18
3
5
2
101-200
13
2
4
2
201-300
7
3
6
2
301-400
8
3
7
3
401-500
9
3
7
3
501-600
12
3
8
3
601-end
12
3
8
3

As mentioned earlier, the sentences were also quite unusual. Kanzi and Alia were familiar with "dogs" and "snakes" as real animals, whereas in these tests, they referred to toy animals, and (toy) snakes appear to be able to bite (toy) dogs. Similarly, neither subject could know from previous experience what sentences such as "Take the lettuce out of the microwave" or "Wash the hotdogs" mean, as lettuce is not normally kept in microwaves, and hotdogs are not normally washed (o.c. 53).

The subjects were confronted with 660 different sentences, for which the entire test environment was recorded (cp. previous block quote). For the blind trials, extra information was also taken down. Moreover, all tests were videotaped.

The sentences were divided into 7 main types with a number of subtypes, yielding 13 types in total:
 
Table 4. Sentence types (Savage-Rumbaugh et al. 1993:54-67)
Type Description Example
1A Put object X in/on transportable object Y Put the ball on the pine needles
1B Put object X in nontransportable object Y Put the ice water in the potty
2A Give (or show) object X to animate A Give the lighter to Rose
2B Give object X and object Y to animate A Give the peas and the sweet potatoes to Kelly
2C (Do) action A on animate A Give Rose a hug
2D (Do) action A on animate A with object X Get Rose with the snake
3 (Do) action A on object X (with object Y) Knife the sweet potatoe
4 Announce information The surprise is hiding in the dishwasher
5A Take object X to location Y Take the snake outdoors
5B Go to location Y and get object X Go to the refrigerator and get a banana
5C Go get object X that's in location Y Go get the carrot that's in the microwave.
6 Make pretend animate A do action A on recipient Y Make the doggie bite the snake
7 All other sentences Take the potato outdoors and get the apple

Some clarifications are in order. The difference between 1A and 1B was introduced because requests involving a nontransportable object are presumably simpler, as such an object cannot be moved. Hence, in sentence type 1B, inversal of the object-object relation is precluded, whereas in 1A, it might occur.

In sentence types 2A-2D, the 'animates' are persons or toy animals. It is the subjects who transferred the terms from the real animals to the inanimate representations of them. 2B serves to demonstrate that the subjects do not simply perform a random action involving object X and Y (the 'key words'), but indeed understand the relation implied in the verb. A number of different verbs was used in type 2C (bite, chase, groom, hammer, hide, hug, play, scare, slap, tickle, wash, vacuum) and 2D (brush, feed, get, hide, hit, put, scare, take, tickle, throw, wash). The added complexity of type 2D requires the subject to grasp the syntax of the complete sentence, as the object-recipient relation is not inherent in its semantics. For example, "Hit the dog with the stick" could result either in hitting the stick with the (toy) dog, or in hitting the dog with the stick.

Similarly, requests of sentence type 3 cannot be executed correctly simply by responding to key words: the sweet potato of the example in table 4 could act on the knife just as well. A basic syntactic understanding is again required.

The sentences of type 4 were not included to investigate syntactic features, as they simply provide the subject with information (viz. where an object was hidden, or that another party was about to engage in a game with the subject). This sentence type was included so as to determine whether the subjects could make the correct inferences from a statement (or 'non-request').

Sentence type 5A is roughly synonymous to type 1B, but distinct spatial entities are used as 'locations', rather than objects as 'recipients'. For sentence type 5B, between six and nine different words indicating locations were employed during each session, and three or more objects were placed in each of these locations, so that the subject had to remember which object to retrieve. Moreover, these objects were not usually found at that place. In a control set of sentences, "the item to be retrieved from the distal location was duplicated in the central array" (o.c.:64). However, errors in carrying out these requests could be due to the semantic ambiguity inherent in 5B: "the subject could interpret the sentence as a request to go to a location, then return and "get" the object in the central array" (o.c.:64-65). Type 5C sentences preclude this possibility by using an embedded phrasal modifier, thus introducing an element of recursion in the sentence structure.

Initially, the responses received one of three codes: correct, partially correct or incorrect - by Sue Savage-Rumbaugh for Kanzi, and by Jeannine Murphy for Alia. A third person scored both subjects' responses on videotape. A fourth observer coded 326 randomly selected trials from videotape of Alia. Statistical calculations were made to assess judgement agreement:

Percentage agreement on the judgment of correct, partially correct, or incorrect response was .98 for Kanzi and .89 for Alia. Percentage agreement on judgments that behaviors on the part of the experimenter may have helped the subject was .83 for Kanzi and .64 for Alia. (o.c.:71) Subsequently, Sue Savage-Rumbaugh reviewed the tapes to see what led to discrepancies in judgements. This review revealed that Savage-Rumbaugh and Murphy made inferences "as to whether Kanzi and Alia were cooperating, attending, purposefully ignoring them, etc." (ibid.). Indeed, it is important to take this 'intimate' knowledge of the researcher working with a subject most of the time, into account. Hence, the coding system was substantially extended, as represented in table 5. More statistical analysis determined coding agreement at .72 for Kanzi and .83 for Alia. On top of that, 38 randomly selected requests were presented twice (with week- or month-long intervals) to Kanzi so as to provide a measure of validity. On 79% of these second-presentation trials, Kanzi did as well or better. The test-retest reliability was not expected from Alia, due to her rapid development.
 
 
 
Table 5. Criteria for coding subjects' responses to requests (Savage-Rumbaugh et al. 1993:73)
Code and type of response Response
Correct:
C

C1-C5

S carries out the request immediately and correctly 

S first hesitates or engages in a tangential activity, then ...

C1

C2

C3

C4
 

C5

S carries out the request correctly

E repeats the request; S carries it out promptly

E rewords and may repeat the request; S carries it out promptly

S's involvement in the tangential activity interferes with his/her ability to attend; E insists the activity stop and then repeats the request; S carries it out promptly

Asked to retrieve a distal object, S attends to its duplicate in the immediate array; E redirects attention to the distal object; S retrieves it promptly

Partially correct:
PC

OE

I

S is partially correct in carrying out the request

S retrieves more objects than requested

S carries out the act in inverse order but is correct with regard to all other components of the request

Incorrect:
W

NR

M

S is incorrect with regard to all aspects of the request

S does not respond or refuses to respond

Mistrial (an item mentioned in the request was unavailable)

Note: S= Subject; E= Experimenter.

 

2.3.4 Results

The results are quite impressive: Kanzi was correct on 72 % of all trials, and 74 % of the blind ones, and Alia on 66 and 65 % respectively. The overall scores are given in table 6; table 7 lists the results per sentence type. (In table 7, only the 'correct' codes are included in the 'percentage correct', hence, both partially correct and incorrect responses were counted as errors. Many of the partially correct responses showed an understanding of the sentence structure, but not of all the words, as the subjects performed the correct activity with the incorrect object; o.c. 80.)
 
 
 
Table 6. Proportions of different response codes obtained by Kanzi and Alia (Savage-Rumbaugh et al. 1993:77)
 
KANZI
ALIA
 
All trials
Blind trials
All trials
Blind trials
Response
N
%
N
%
N
%
N
%
C

C1

C2

C3

C4

C5

369

11

36

34

8

9

57

2

6

5

1

1

246

10

16

21

7

7

59

2

4

5

2

2

319

20

32

13

7

0

54

3

6

2

1

0

220

18

16

9

4

0

54

4

4

2

1

0

Total correct
467
72
307
74
391
66
267
65
PC

OE

I

W

NR

153

6

9

8

10

23

1

1

1

2

87

5

8

4

4

21

1

2

1

1

124

20

9

32

11

21

3

2

6

2

84

17

7

24

8

21

4

2

6

2

Total wrong
186
28
108
26
196
34
140
35
Overall total
653
100
415
100
587
100
407
100

 
 
 
 
Table 7. Proportion of responses to different sentence types coded correct for Kanzi and Alia(Savage-Rumbaugh et al. 1993:78)
 
KANZI
ALIA
Sentence type
Count
% Correct
Adjusted res.
Count
% Correct
Adjusted res.
All trials:
1A

1B

2A

2B

2C

2D

3

4

5A

5B

5C

6

7

80/126

36/49

56/69

7/21

16/18

61/86

56/80

10/16

64/85

40/47

27/35

7/10

7/11

64

74

81

33

89

71

70

63

75

85

77

70

64

-1.4

.1

.0

-1.6

.2

1.5

.6

-1.3

.9

3.5

2.2

.5

1.5

88/123

31/43

47/58

12/21

13/15

45/75

41/63

11/13

55/80

23/45

16/31

6/10

3/10

72

72

81

57

87

60

65

85

69

51

52

60

30

1.4

-.1

-.0

1.6

-.2

-1.5

-.6

1.3

-.9

-3.5

-2.2

-.5

-1.5

Blind trials only:
1A

1B

2A

2B

2C

2D

3

4

5A

5B

5C

6

7

39/62

13/17

36/46

7/19

10/11

37/49

40/49

8/12

45/58

32/39

27/35

6/9

7/9

63

77

78

37

91

76

82

67

78

82

77

67

78

-1.3

.4

-.7

-1.3

.0

1.6

2.0

-.9

.8

3.4

2.2

.5

1.9

47/64

12/17

36/43

12/21

10/11

31/51

25/40

10/12

42/59

18/40

16/31

5/9

3/9

73

71

84

57

91

61

63

83

71

45

52

56

33

1.3

-.4

.7

1.3

.0

-1.6

-2.0

.9

-.8

-3.4

-2.2

-.5

-1.9

Note: adjusted res.= adjusted residuals (indication of better performance compared to the other subject; the numbers for Kanzi and Alia are the same, but the + or - sign is the opposite each time)

The mere statistical probability that the simple object-location sentences would be responded to correctly, is calculated to be 2.4 %, and for other sentence types, it would be even less (o.c. 76). We may safely conclude that true comprehension was ascertained in both subjects. Let us consider the results of some of the sentence types in more detail.

Type 1A and B elicited a large number of correct responses from both subjects, and Kanzi performed 10 % better on type 1B requests, in which the two objects are not readily reversible due to the non-transportable nature of one of them (although some "non-transportable" objects proved to be quite transportable once Kanzi decided to move them…). Type 2B appeared to be very difficult for both, but significantly more difficult for Kanzi than for Alia, as it required the subjects to memorize two unrelated objects and perform the same action on both. "Kanzi's difficulty was perhaps due more to short-term memory limitations on the overall amount of information than to processing limitations on the information available to him" (o.c. 85). One must after all bear in mind that a human brain is considerably larger than an ape's… Indeed, Kanzi performed better on a sentence such as "Feed the dog some milk", in which the information can be chunked meaningfully, than on a type 2B request such as "Show me the milk and the dog".

Some of the errors on type 2D and 3 requests were not due to comprehension errors: for type 2D, a hesitancy of the subject to act directly on another person with an object prompted errors, and for type 3, the decidedly odd nature of requests (e.g. "Bite the picture of the oil", o.c. 86) may have caused some errors.

Kanzi's performance on type 5B and 5C requests was significantly better than Alia's. As noted earlier, type 5B sentences were ambiguous, whereas in 5C sentences, the ambiguity was resolved by using a phrasal modifier ("Go get object Y that's in location X"). Indeed, the subjects performed far better on control trials - in which the item to be retrieved was present both in the immediate array and in the distal location - for 5C than for 5B: Kanzi, for example, acted on the object in the near array on only 9 % of these control trials, as contrasted with 50 % of the type 5B control trials. Similar observations hold for Alia (o.c. 89-90). The understanding of a phrasal modifier clarifying the object of reference indicates that both subjects could interpret the syntactic device of recursion.

The research team led by Dr. Savage-Rumbaugh regrouped the obtained data into three groups in order to see whether the subjects comprehended reversal of word order: (A) verb plus word order changes, and appropriate response differs (e.g. "Could you take the pine needles outdoors?"/"Go outdoors and get the pine needles"), (B) word order remains constant, but appropriate response differs ("Take the rock outdoors"/"Go get the rock that's outdoors") and (C) word order changes, and appropriate response changes ("Put the juice in the egg"/"Put the egg in the juice") (o.c. 92). Results are listed in table 8, indicating that these sentences were a difficult challenge to both, and that Kanzi performed significantly better than Alia, who tended to return from various locations with more than one object.
 
 
 
Table 8. Comparison of Kanzi and Alia's performance on reversed sentences over three subtypes of reversals: summary statistics (Savage-Rumbaugh et al. 1993:92)
 
KANZI
ALIA
C/N
%
C/N
%
Subtype A  
  Sentences

Pairs

38/46

17/23

83

74

26/44

8/21

59

38

Subtype B  
  Sentences

Pairs

22/28

8/14

79

57

18/27

5/13

67

38

Subtype C  
  Sentences

Pairs

33/42

12/21

79

57

27/39

7/18

69

39

C = number of correct responses (C, C1-5 in table 3); N = total number of sentences or sentence pairs given to the subject

It is important to note that real inversion errors were scarce: semantic errors predominated (e.g. putting the melon in the water when requested to "put the melon in the tomatoes", o.c. 96). Indeed, both subjects' overall performance indicates that they were "sensitive to word order as well as to the semantic and syntactic cues that signaled when to ignore word order [e.g. "Go get the carrot that's outdoors"] and when to attend to it [e.g. "Make the doggie bite the snake"]" (o.c. 97).

In conclusion, it should be clear by now that both Kanzi and Alia were able to understand the semantics and the syntactic structure of unusual English sentences, even though neither of them was as yet a fluent speaker.

The lack of contingent reward, the novel nature of the requests, the absence of previous training to perform these specific requests, and the unique nature of each trial countermand simple explanations that depend on the conditioning of responses independently of semantic and syntactic comprehension. Both subjects clearly demonstrated a capacity to process the semantic and syntactic information in the sentences presented to them. Moreover, the manner in which they did so revealed that they did not interpret the words contained in sentences as randomly juxtaposed events, to be acted on independently. Instead, they invariably attempted to carry out a complex set of related actions that reflected their interpretation of the semantic and syntactic features of each novel utterance. Thus, for example, Kanzi’s solution to "Put the water on the carrot" was to toss it out into the rain. Such innovative actions revealed a sophisticated processing of the speaker’s intent (in this case, to get the carrot wet) rather than a rote, unthinking solution. Even when the subjects failed, they virtually never did so in a way that would suggest that they were assigning key words randomly. (o.c. 98-99) In addition, it should be noted that this study was carried out with an amazing sense for detail and precision, as shown in the elaborate coding system applied by various coders, the statistical analyses applied to the data at various stages, the recurrent methodological reflections, and the very detailed information provided in the appendix (o.c. 111-210).
 
 

2.4 The importance of a language-rich environment

In order to demonstrate once more that the good results obtained with Kanzi and with another language-competent bonobo, Panbanisha, are not due to subtle cues provided by the context, Panbanisha's results were confronted with those for another bonobo, Tamuli, who was only exposed to a 'language-rich' environment at 3.5 years of age. The sentence constructions under consideration employed A(ctions), O(bjects), R(ecipients) and L(ocations) in the combinations AO, AL, ARO, AOR, AOL, yielding a total of 145 sentences (examples: table 9). As in the previous study, many alternatives were available in the context. The results are listed in table 10.
 
Table 9. Examples of sentence types (Williams et al. 1997:308-311)
type example
Action-Object (AO) Hammer the soap
Action-Location (AL) Go to the potty
Action-Recipient-Object (ARO) Let's give doggies a rock
Action-Object-Recipient (AOR) Take the hot dog to Tamuli
Action-Object-Location (AOL) Put the straw in the ball

 
 
Table 10. Summary of sentence comprehension data for Panbanisha and Tamuli (Williams et al. 1997:307)
 
PANBANISHA
TAMULI
KANZI
Sentence type
(#C/F)
%
(#C/F)
%
(#C/F)
%
Action-object

Action-location

Subtotal

Action-recipient-object

Action-object-recipient

Action-object-location

Subtotal

Other

Grand total

45/50

9/12

54/62

19/24

6/07

25/42

50/73

8/10

112/145

90

75

87

79

86

60

68

80

77

5/50

2/12

7/62

1/24

0/07

1/42

2/73

0/10

9/145

10

17

11

04

00

02

03

00

06

105/107

23/23

128/130

3/04

17/19

33/36

53/59

65/74

246/263

98

100

98

75

89

92

90

88

94

Number correct (#C), frequency (F), percentage correct (%). Kanzi's data is from Savage-Rumbaugh 1987. 

Whereas "Tamuli does possess the communicative skills necessary to function successfully in the ape world" (o.c.:313), it appears from these data that Tamuli's comprehension of spoken English and lexigrams is nearly non-existent: she responded correctly on 6% of the sentences, as contrasted to Panbanisha's score of 77 %. Moreover, Panbanisha was never completely incorrect; whereas Tamuli was given an 'incorrect' score on 63% of her responses. She also refused to respond more often than did Panbanisha (15 times as opposed to twice). Her overall score of 6 % cannot readily be accounted for by single word comprehension, as appeared from an additional test. Rather, contextual information of some kind may have helped her to respond appropriately, and even chance circumstances served to elicit a correct response:

When asked to ‘Find the book’, Tamuli ran to look out the window and picked up a book lying in her path. Two other sentences to which Tamuli responded appropriately involved a plastic bag, an object which she found highly attractive and which she often incorporated in her incorrect responses as well. (o.c.:309-311) 2.5 Kanzi's grammar

Let us go on to consider more specifically the issue of grammar. Greenfield and Savage-Rumbaugh (1990) define grammar as "a set of formal rules for marking relations between categories of semiotic elements" (o.c. 541). The comprehension of various syntactic construals by Kanzi has already been demonstrated in our discussion of Savage-Rumbaugh et al. (1993). In order to ascertain to what extent Kanzi also uses (and in fact, partly invents) a (proto)grammar in his production of lexigram-lexigram and lexigram-gesture combinations, Greenfield and Savage-Rumbaugh (1990:541) adopt five criteria (table 11), and show how these are all met by Kanzi's performance (o.c. 553ff.). From the onset, it is stressed that evidence of "the invention of grammatical rules by apes would be stronger evidence of evolutionary continuity than merely learning of rules" (o.c. 543). Indeed, early protohumans did not have any models to rely on: at one or other stage, language as we know it was "invented".
 
Table 11. Five criteria necessary for a grammatical rule (Greenfield and Savage-Rumbaugh 1990:541, cp. Savage-Rumbaugh and Lewin 1994:158-159)
1
Each component of a combination must have independent symbolic status.
2
The relationship between symbols must be reliable and meaningful (semantic).
3
A rule must specify relations between categories of symbols across combinations, not merely a relation between individual symbols.
4
Some formal device, such as statistically reliable order, must be used to relate symbol categories across combinations.
5
The rule must be productive: a wide variety of spontaneous combinations must be generated.

The corpus resulting from five months of observation yielded 1,422 combinations of two or more elements (out of 13,691 utterances, or 10.39 %, o.c. 550). Half of this material was disregarded, mainly because there was no second observer present to record the context for these tokens. (Apart from that, partial or complete imitations - 2.67 % of total combinations - and utterances that were solicited in any way were not included in the final corpus.) The resulting 723 two-symbol utterances nevertheless constitute a much larger corpus than the usual child language corpus for a similar stage of development (o.c. 551). Lexigrams clearly are no icons - they do not bear any resemblance to their referent - and they can transcend indexical usage to function as a true 'symbol' (in Peirce's terminology, o.c. 552ff.). Gestures on the other hand, often are iconic and/or indexical. This does not invalidate their status as semiotic elements, however: human grammar uses a number of indexical signs as well (e.g. demonstrative pronouns), and interestingly, pointing gestures are also important signs for deaf children of hearing parents (which, like Kanzi, lack a pre-established model of language). Apart from 'demonstrative gestures', Kanzi used clearly distinct gestures for "go", "come", "bite", "tickle", "chase", "yes", "open", and "bad" (o.c.564n). In addition, it should be noted that the lexigram system does not offer the possibility of morphological inflection, so Kanzi's grammar can only be expected to use word order as a formal device.

The two-symbol combinations were analyzed in terms of semantic roles similar to those used in child language studies. Seven out of the eight major relations in children's utterances at the two-word stage were found in Kanzi's data, viz. agent-action, action-object, agent-object, entity-demonstrative, goal-action, location-entity and entity-attribute. Possession is the 'missing' relation; conjoined action (e.g. 'tickle bite', cp. English 'go get') was missing in children's utterances.

In general, it appears that the first criterion is met: individual lexigrams were only regarded as distinct semiotic elements if they occurred spontaneously on 9 of 10 occasions and if Kanzi "demonstrated behavioral concordance on 9 of 10 of the subsequent occasions" (o.c. 553). The relation between symbols was clearly reliable (criterion 2): Kanzi "rarely paraphrased or repeated himself or formed combinations that were semantically unrelated" (o.c. 556): only 6 out of 723 utterances showed no 'direct relation' (e.g. Kanzi commented "potato oil" after a researcher had put oil on him while he was eating a potato, o.c. 557). Since all partial or complete imitations of researchers' utterances (only 2.67 %) were omitted from the corpus, a degree of creative spontaneity (criterion 5) is clearly present.

One rule that Kanzi learned from his caregivers is 'action precedes object', an order he began to use to a statistically significant degree after about a month into the observation period. This rule remained "undisturbed by countertrends of individual lexical items" (o.c. 559; criterion 4), unlike with Nim, and clearly relates categories of symbols rather than individual symbols (cp. table 20.2 in o.c. 558-559).

An 'arbitrary' (not functionally motivated) rule which Kanzi coined himself is 'place gesture after lexigram'. This is the exact opposite of the human caregivers' order, who invariably use a lexigram first. The arbitrary nature of this rule appears from this anecdote:

At one point Kanzi was observed to move away from a person he would later indicate as agent, go to the board (where he indicated an action lexigram), and then return to the person (using the gesture to designate her as agent). In that situation, the rule Kanzi had invented demanded extra motor steps and therefore seemed purely arbitrary. (o.c. 560) The rule pertains to four relations (creative productivity): agent-action, demonstrative-entity, goal-action, and agent-object. In the case of goal-action, the rule implied relations between two categories (lexigrams, and on the other hand, action gestures such as 'open', 'go', 'come'; criterion 3). In the other cases, however, only a demonstrative gesture was used.

Another rule invented by Kanzi is directly related to Kanzi's interests (games such as chasing and tickling, for instance): preferred orderings for conjoined actions. "Chase" and "tickle" tended to appear in the first position, whereas "hide", "slap" and "bite" usually appeared in the second position (criterion 3 and 4). "Grab" and "hug" showed no statistically significant position preference. The semantic reason (criterion 2) for this categorization is suggested to be that first-position lexigrams function as "invitations to play", whereas second-position lexigrams represent "the play content that follows" (Kuroda pers.comm. qtd. in Greenfield and Savage-Rumbaugh 1990:566). A later suggestion by Patricia Greenfield is that Kanzi "tended to place the actions that required a greater distance between the two parties in the first position and the action requiring closer contact between the two parties in the last position" (reported in Savage-Rumbaugh and Lewin 1994:162). This might iconically reflect primates' tendency to move from distal to proximal actions in their play (ibid.). A similar rule-governed verb order appears in human languages - even to a limited extent in English ('go get'). Interestingly, Kanzi's productive rule - 10 of 16 predicted combinations were attested - was imitated by his caregivers (whose communicative acts were studied on hours of videotape). The conjoined action rule is relevant to an evolutionary approach to language origins, according to which language evolved as "an instrument to plan coordinated action" (Greenfield and Savage-Rumbaugh 1990:568).

Difference in symbol order could signal difference in meaning: animate agents tended to be placed in the first position in lexigram-lexigram combinations ("Matata bite"), whereas animate beings functioning as objects of action were placed in the second position ("grab Matata").

Of Kanzi's three-word utterances (which, unlike Nim's, were not redundant), only action-action-agent was sufficiently productive during the observation period (e.g. "chase bite (person: demonstrative gesture)").

It would appear that Kanzi even produced the "rudiments of an ergative system in the face of the accusative model presented by his English-speaking caregivers" (o.c. 570) as he placed intransitive agents and transitive objects after the action symbol.

Kanzi's grammatical development was, of course, much slower than children's. For one thing, an ape's brains are only a third of the size of human brains (cp. infra). The fact that the amount of requests was far larger than statements (96 %) probably reflects at least in part

a bias stemming from the fact that in captivity, a chimpanzee's behaviour and environment are under the control of humans, from whom he must request activities or objects. In the wild, a given animal might, for example, state his planned activity, rather than requesting it. This difference is quantitative rather than qualitative and may reflect a tendency on the part of the researchers to code chimpanzee statements as requests in the interests of conservative interpretation. (o.c. 568) 2.6 Rearing or species variable?

In order to determine whether bonobos were privileged in their language competence, a comparative study between a bonobo (Panbanisha) and a chimpanzee (Panpanzee) was set up. That they were co-reared, is extremely significant: from previous studies with chimpanzees one may have got the impression that chimpanzees were not capable of anything near the level of language comprehension apparent in Kanzi, Panbanisha and another bonobo, Mulika. However, in these previous studies, the chimpanzee subjects were older, and were taught on the basis of trial and reward, rather than observationally.

Both subjects learned to comprehend speech and lexigram use, but "the bonobo did so more extensively than the chimpanzee" (Brakke and Savage-Rumbaugh 1996:363). Indeed, Panbanisha acquired symbols earlier, more rapidly and in greater numbers than Panpanzee, she combined them in more novel and abstract ways and used gestures more often to stress her communicative intent (Panbanisha and Panpanzee: online). This is not to suggest that Panpanzee was not intelligent: she was considerably better at maze learning and puzzle working than both Kanzi and Panbanisha. She did also reach an important level of linguistic competency, though not as impressive as that of Kanzi or Panbanisha.

An illustration of Panpanzee's lag in language development when compared to Panbanisha in the realm of language production is offered in figure 2. The rate of early language acquisition is comparable to that of human infants. After this point, however, children's vocabularies start to increase very rapidly. That this is not the case for apes may be due to their smaller brain size, but also (in part) to the low number of symbols available on the lexigram keyboard (viz. 256).
 
Figure 2. Number of different lexigrams used non-imitatively at least once per month (Brakke and Savage-Rumbaugh 1996:368)

In conclusion, it appears that chimpanzees are also able to learn symbols spontaneously given the correct linguistic input (viz. early immersion in a language-rich environment). Indeed, the "rearing variable" is far more important than the "species variable" (Panbanisha and Panpanzee 1998: online). However, the lower scores for the chimpanzee when compared to the bonobo, suggest that they could be situated at the left on a continuum - rather than an 'all-or-none' division - as represented in figure 3:
 
Figure 3. Continuum of language competence


 
 

  2.7 Sex, lies and videotapes: a Theory of Mind

As is argued throughout a number of publications by Savage-Rumbaugh and her team (and in short also above), comprehension precedes production and is a more relevant pointer of language competence: listening requires "the assumption of the perspective of others", and the assumption "that others have something to say that the listener does not know. In brief, listening requires what is called a Theory of the Mind" (Savage-Rumbaugh 1997:50). Language is not just about expressing oneself - which is roughly what Washoe and Lana could do - but about how individuals structure their interactions with one another, which entails the capacity to act in response to the expressions of other individuals. The results obtained with Sherman and Austin, Kanzi and Panbanisha showed that apes are capable of this. The latter two of these were also raised from infanthood in a language-rich environment, and developed impressive language skills, as reported above.

Such truly "language-competent" apes, as Kanzi and Panbanisha, can "easily participate in three- and four-way conversations - conversations that deal with the intentions and actions of multiple parties and with the states of minds of the parties" (o.c. 57). Savage-Rumbaugh reports an elucidating example which shows that Panbanisha can answer questions about what it is that someone else thinks is hidden in a box (o.c. 64). In a Theory of Mind test, Liz asks Sue for candy in Panbanisha's presence. Sue agrees to put some candy in a box for Liz. Then Liz leaves, and in Liz's absence Panbanisha watches as Sue replaces the candy with a bug. When Liz returns, she pretends to have difficulty opening the box, at which point Sue asks Panbanisha what Liz is looking for. Panbanisha responds by pointing to the lexigram for 'candy'. In addition, Panbanisha was able to comm