Fig 1.
Baddeley's model of the working memory.
Fig 2.
Localization of the Memory (yellow), Unification (cyan) and Control (grey) components of the MUC model proposed by Hagoort. Brodmann's areas are marked by numbers.
Fig 3.
Type of neural gating mechanism used in the ANNABELL model.
(adapted from Ref. [39]).
Fig 4.
Schematic diagram of the ANNABELL system main components.
Fig 5.
Schematic diagram of the main system architecture.
Each rectangle represents a subnetwork, which is composed by interconnected artificial neurons. Only the main subnetworks are represented in this diagram. The arrows that join the rectangles represent directional connections among neurons of different subnetworks.
Table 1.
Sentences of the people dataset.
The social environment described in this dataset includes twenty persons. In the second column, <person> can be “Mum”, “Dad”, or the name of one of the other eighteen persons, <relationship> can be “father”, “mother”, “sister”, “friend”, “cousin”, “Grandma”, “Grandpa”, “aunt” or “uncle”. <number> can be a number or, in row 21, also “some” or “many”. The “(s)” denotes the possibility of a plural form. In row 15, <verb> and <complement> describe the profession in terms understandable for a preschool child, e.g. “the journalist writes in the newspaper”. The sentences in row 24 use the present progressive, as in “Susan is reading a book”. The sentences in row 25 (how-to sentences) are verbal prescriptions, expressed through the natural language, that are used to instruct the system on how to perform specific tasks in language processing.
Table 2.
Sentences of the parts-of-the-body dataset.
Thirty-three body parts are included in the dataset. In the first column, <part> is the name of a body part. The “(s)” refers to a possible plural form.
Fig 6.
Map of the virtual house used to build the sentences for the test set of the virtual-environment dataset.
Table 3.
Number of declarative sentences, number of interrogative sentences used for training, number of interrogative sentences used for the test and number of output sentences in the five learning sessions.
The conversational sentences that do not expect a turn taking from the system are treated as declarative sentences, while the ones that expect a turn taking are treated as interrogative sentences. The output sentences in the virtual environment (*) are actually commands issued by the system to perform actions in the environment itself.
Fig 7.
Distribution of the questions used for training (a) and for test (b) among the question categories: alternative (choice) questions, polar (yes/no) questions, what/how questions, where/when questions, who questions, question-like imperative sentences (e.g. tell me), conversational sentences that expect a turn taking from the system.
Fig 8.
Distribution of the number of words and of the word classes in the input and output sentences.
Distribution of the number of words in the input sentences (a) and in the output sentences (b); distribution of the words used in the input and output sentences among different word classes (c,d); percentage of word classes in the input and output sentences (e,f).
Table 4.
Number of correct answers over the total number of expected answers in the test stage of the cross-validation rounds for the first three datasets.
An answer to an interrogative sentence is considered valid only if it is correct both syntactically and semantically.
Table 5.
Number of tasks that are performed correctly by the system on the virtual environment dataset over the total number of assigned tasks in the test stage of the cross validation rounds, as a function of the number of training examples used in the training stage.
Fig 9.
Extract of a side-by-side comparison between the human/ANNABELL-system dialogue on one side and the mother/real-child dialogue on the other side, based on the Warren-Leubecker corpus from the CHILDES database.
The right side is a transcription of a conversation between a 5-years-and-10-months old child and his mother, extracted from the file “david.cha” of the CHILDES database. Note that the human/ANNABELL dialogue system does not use punctuation, which has been added here for clarity.
Fig 10.
Comparison between some distributions related to the output sentences produced by the system in the communicative interaction test, based on the Warren-Leubecker corpus from the CHILDES database, and to the utterances of the real child for the same part of the corpus: distribution of the number of words in the output sentences (a,b); distribution of the words used in the output sentences among different word classes (c,d); percentage of word classes in the output sentences (e,f).
Fig 11.
Average time that the system needs to answer a question as a function of the number of allocated connections in the state-action-association subnetwork, evaluated on a CPU system based on a Xeon 2.50 GHz dual-processor-quad-core 16GB RAM, and on a system equipped with a high-performance GPU NVIDIA Kepler GK104 having 1536 cores.
Table 6.
Percentage of correct answers after removal of the connections from the STM components to the central executive, and after complete removal of the components.
The values are averaged over the three datasets people, parts of the body and categorization, and over the four rounds of the cross validation. The percentage of correct answers with all components and with all connections is 86.5%.
Fig 12.
Percentage of correct answers over the total number of expected answers, evaluated on the first three learning sessions: [(a) and (b)] as functions of the weight-saturation value Wmax for the main components of the system; (c) as a function of the parameter k used for the k-winner-take-all algorithm in the state-action-association subnetwork.
Table 7.
Number of correct answers produced by the four instances of the system over the number of interrogative sentences for the three extended datasets people, parts of the body and categorization.
The four instances of the system are the ones obtained by training the system on the original datasets during the four rounds of the cross-validation, respectively.
Table 8.
Group of words that compose the sentences of the database used for this experiment.
The X represents a closed-class word. In case of ambiguities, the system is trained to use the largest groups.
Table 9.
Meaning errors and sentence errors on the ten rounds of the ten-fold cross validation.
Meaning error is the percentage of incorrect thematic role assignment. Sentence error is the percentage of sentences in which there is at least one wrong thematic role assignment.