Proceedings of STeP'96. Jarmo Alander, Timo Honkela and Matti Jakobsson (eds.),
Publications of the Finnish Artificial Intelligence Society, pp. 79-88.

FinnPro: A Tool for the Simulation of Connectionist Models of Language Production

Anneli Tikkala
Department of Computer Science and Applied Mathematics,
University of Kuopio,
P.O.Box 1627, 70211 Kuopio, Finland

Hans-Jürgen Eikmeyer
Faculty of Linguistics and Literature,
University of Bielefeld, Bielefeld, Germany

Matti Laine
Finnish Academy and Department of Neurology,
University of Turku, Turku, Finland

Jussi Niemi
Department of General Linguistics,
University of Joensuu, Joensuu, Finland

Abstract

FinnPro is a tool for the simulation of local connectionist models of language production. The tool defines an abstract network description language which is translated into PASCAL data-structures. The sequential order of linguistic units such as phonemes, syllables, morphemes and words in language output is a challenge to any parallel neural network. For this purpose FinnPro uses a sequentialization machinery which allows for the formulation of control node chains/networks implementing linear constraints found in natural language utterances (Eikmeyer & Schade 1991; Schade 1992). The tool is here used for the simulation of a production model for Finnish nouns. Finnish is a morphologically complex language where nouns can be marked for number, case and possession, as well as with several enclitic suffixes. The tool simulates normal language production as well as slips of the tongue, i.e., errors brought about by adding noise to the system. In this paper, the behavior of the system in the case of two simultaneously activated words is presented and the results of the simulation tests are compared to data from a corpus of Finnish slips.

Introduction

Cognitive models are constructed to account for empirical data about human behavior. The models are made to perform in a manner that is similar to human performance in some respects. The purpose of such modelling is to obtain information on how that behavior is produced (Bechtel & Abrahamsen 1990). Although connectionist models in many respects act quite neuron-like, they are not seen as models of the neural level. Rather, they are to be regarded as (neuro)psychological models.

Since the 1980s researchers have focused on connectionist models in order to model and simulate language production. Several kinds of connectionist architectures have been proposed for modelling natural language processing, child language development and language disorders (Wright 1995). After Dell introduced his interactive activation model for language production (Dell 1985, 1986, 1988), much research work has been done in that area (Stemberger 1985; McKay 1987; Eikmeyer & Schade 1991, 1993; Schade 1992).

In our recent papers (Tikkala 1995; Tikkala et al.; Tikkala, submitted), we have introduced an interactive activation modelling tool FinnPro. Its basic psycholinguistic assumptions follow those of the SAID model (Laine et al. 1994; Niemi et al. 1994).

Fig. 1. The structure of FinnPro

FinnPro: an overall picture of the tool

FinnPro is a tool for simulating the retrieval of Finnish nounforms. Information about several nouns in different forms may be included in a FinnPro model. In the present version of our tool, each simulation run produces one wordform. In an undisturbed simulation, FinnPro outputs the correct wordform; the production of slips of the tongue may be simulated by adding noise to the system.

FinnPro (Fig. 1) has two kinds of input: (i) a textual network description for the creation of the network and (ii) information about the simulation run to be made (target word stem and form, and some parameter values). The output of FinnPro is the (correct or violated) wordform produced by the tool.

FinnPro is based on the principles of spreading activation (McClelland & Rumelhart 1981; Rumelhart & McClelland 1986). During the simulation process activation flows from the nodes representing the input of the network. After a few simulation steps, the most highly activated phoneme nodes are selected as output nodes. The correct sequence of the output phonemes (as well as syllables) is controlled by a sequentialization mechanism which follows the ideas of a corresponding tool used for German (Schade 1992; Eikmeyer & Schade 1991, 1993). Furthermore, FinnPro includes a control structure to maintain Finnish vowel harmony (Tikkala, submitted).

The topology of FinnPro networks

A FinnPro network consists of two subnetworks. The selection network (Fig. 2) includes nodes labeled by the constituent parts (i.e., morphemes, syllables and phonemes) of the words and by word form specifications. These labels are mnemonic only. The control network (Fig. 3) serves for sequentialization and maintenance of vowel harmony.

The selection network consists of hierarchically arranged layers of nodes. Within and between these layers, the nodes representing a linguistic entity and its constituent parts are linked by connections as presented in the input network description. Connections within a (sub)layer are bidirectional and inhibitory (lateral inhibition, not shown in Fig. 2). Connections between nodes in the different layers are excitatory and (mainly) bidirectional.

The control network consists of two control chains with a hierarchical organization. The hierarchically higher chain represents the structure of a noun and the lower chain the structure of a syllable. Furthermore, the control network includes control nodes for vowel harmony. The control chains are composed of two kinds of nodes: control and valve nodes. Each control node represents a linguistic constituent, i.e., a syllable of the output wordform or a phoneme of a syllable. The valve nodes (the black circles in Fig. 3) control the selection of the control nodes.

Each of the control nodes of the higher chain (i.e., nodes representing a certain syllable location, e.g., the first syllable) are connected to the start node of the lower chain (representing any syllable). The end node of the lower chain is linked back to the higher chain nodes. These links (not shown in Fig. 3) enable activation flow between the control chains.

Fig. 2. Part of a FinnPro network lintu - mikro: the selection subnetwork.

The selection network and the control network are linked to each other with connections (the small numbered circles in Figs. 2 and 3) to influence each other's behavior. Each of the control nodes, when activated, sends an activation jolt to all the nodes in its control area, e.g. to all possible first syllables, or to all coda phonemes. The selection nodes, in turn, send activation to the valve nodes to direct the appropriate route selections in the control chains. The nodes representing vowels in the selection chains activate the corresponding front and back vowel harmony control nodes. The vowel harmony nodes, in turn, send back excitatory and inhibitory activation jolts to the vowel nodes along the linguistic vowel harmony constraints of Finnish.

The dynamics of FinnPro

The simulation of word production proceeds in discrete time steps. At each step, new activation values are computed for each of the nodes. This makes the activation spreading process a parallel phenomenon. The amount of steps needed for the production of one word varies as a function of the number of the syllables and the phonemes in a wordform, as well as the number of steps used to process each of them (this is a parameter of the system). In our tests, simulations take less than a hundred steps.

Fig.3. Part of a FinnPro network lintu - mikro: the control subnetwork

The simulation begins by activating the start node in the highest control chain (i.e., noun), and ends when the activation reaches the end of this chain. In the simulations, usually, one stem node representing the target noun and one form specification node representing the target form are initialized to full activation (i.e., 1). The initial activation values of other nodes are zero.

The activation values of the selection nodes are computed using formulae (1)-(3). Value n(i,t) is the activation change of node i in time step t+1. In formula (1) a(j,t) and a(k,t) are the activation values (greater than possible thresholds) of nodes j and k in time step t. Symbols j and k represent selection, control or vowel harmony nodes connected to node i excitatorily and inhibitorily and and represent the excitatory and inhibitory connection strengths between nodes j and i, or k and i, correspondingly. New activation values are calculated using formulae (2) and (3).

(1)
(2)
(3)

The activation values of the valve nodes and the vowel harmony nodes are calculated using formulae (1)-(3), symbols j and k represent selection nodes and (in case of the latter nodes, also) vowel harmony nodes. A control node is fully activated when selected.

Every mth step (m is a parameter) is a selection step. During that step the following algorithm is performed:

  1. Compute new activation values to all selection nodes, valve nodes and vowel harmony nodes.

  2. If one of the phoneme or syllable control nodes is activated, find the winning selection node and, in the case of phonemes, output its label.

  3. Select the route forward in the control chain and let activation flow along that route one node further.

In the process of producing the output string phoneme by phoneme, any winner phoneme is inhibited after its selection. Further, any winner syllable is inhibited after its selection. These inhibitions make it possible for other constituents to proceed with high activation values.

When the network is disturbed by noise, after using formulae (1) to (3) to compute the activation values for the selection nodes, calculations of formulae (4) and (5) are performed. Noise is a random value taken from normal distribution with the standard deviation value given as a parameter.

(4)
(5)

Simulation tests

In normal simulations one of the word stems in the model is activated for each simulation run. The tool is capable of producing correctly each of the forms defined in the model. To simulate slips of the tongue, this simulation process may be disturbed by noise.

In our present tests, we chose the most robust phonological frame constraint in speech errors, the initialness effect, for closer scrutiny. The initialness effect means that in slips of the tongue, particularly the initial segments are prone to errors (e.g., mait a winute < wait a minute).

We sought to test whether the behavior of FinnPro would follow the lines suggested by Eikmeyer & Schade (1993). They suggested that the initialness effect could rise from the dynamics of the production model without the assumption of a special word-onset consonant slot. Specifically, Eikmeyer & Schade proposed that at the moment of the selection of the initial segment, the activation coming from the target word node is still on the rise. Relatively weak activation together with some noise in the network would thus increase chances for the mis-selection of the initial consonant of a competing word. In non-initial positions, chances for mis-selection would be lower as by the time they are selected, the target word is giving stronger activation and the competing word(s) and its segments are dampened through lateral inhibition.

For the purposes of the present analysis, we devised FinnPro networks with two competing words: lintu 'bird' - mikro 'micro' and lintu - metso 'capercaillie' in inessive case form. In both cases the models included two first and two second syllables and one third syllable. At the beginning of a simulation run, instead of only one stem, both word stems were getting activation. Each member of the two word pairs served as a target in the simulation runs, and the to-be-produced form was either in inessive case (e.g., linnu+ssa 'bird+in') or in inessive case+possessive (2nd person) (e.g., linnu+ssa+si 'bird+in+your').

The initial activation value of the target word was 0.55 and the competing word was activated to the value 0.45. The number of steps between the selection phases was 5. The standard deviation of noise was 0.01.

The simulation results are given in Table 1. Error rates varied between 104 (10.4%) and 163 (16.3%) in the 1000 runs for each of the targets. The error patterns show a strong preponderance on word-initial consonant slips, as was observed in the empirical data (an unpublished compilation of Finnish slip corpora: 255 phonological slips with 21 onset-to-onset exchange errors and no coda-to-coda errors in word-initial syllables and 10 consonant exchanges in the non-initial syllables). This effect is due to the relatively lower activation boost that the initial consonant is receiving, as suggested by Eikmeyer & Schade (1993). Other types of slips were also observed, including coda errors that were not observed in normals' exchanges. However, their rate was very low in the simulations.

Examples of production errors: in the test with the network lintu - metso the following wordforms were produced with target wordform linnussa (compare to Table 1).

minnussa92word-initial consonant
metnussa2whole initial syllable
mennussa2word-initial consonant+vowel
lennussa2initial syllable nucleus
metnunsa1other error
letnussa1other error

Table 1. Simulation of the initialness effect. Each error distribution is based on the first 100 errors produced.

The activated word pairTarget wordformType of exchange of errorsNumber
LINTU - MIKROLINNU+SSAword-initial consonant
whole initial syllable
initial syllable coda
other error
96
2
1
1
MIKRO+SSAword-initial consonant
whole initial syllable
initial syllable coda
other error
97
1
1
1
LINTU - METSOLINNU+SSAword-initial consonant
whole initial syllable
word-initial consonant+vowel
initial syllable nucleus
other error
92
2
2
2
2
METSO+SSAword-initial consonant
word-initial consonant+vowel
initial syllable nucleus+coda
92
6
2
LINTU - MIKROLINNU+SSA+SIword-initial consonant
whole initial syllable
initial syllable coda
other error
95
1
1
2
MIKRO+SSA+SIword-initial consonant
whole initial syllable
initial syllable coda
other error
94
3
2
1
LINTU - METSOLINNU+SSA+SIword-initial consonant
whole initial syllable
word-initial consonant+vowel
initial syllable nucleus
initial syllable nucleus+coda
other error
86
1
5
5
1
2
METSO+SSA+SIword-initial consonant
whole initial syllable
word-initial consonant+vowel
initial syllable nucleus
initial syllable nucleus+coda
other error
86
2
5
4
2
1

Conclusions and future work

A FinnPro model is the first proposal of a spreading activation based word production model seriously dealing with the morphological structure of a natural language which is quite complex. Our simulations so far have shown the ability of our tool to produce error patterns near to those of normals both in the vowel harmony behavior (Tikkala, submitted) and the initialness effect. In the future, these simulations will allow for the formulation of hypotheses concerning more details of the word production process.

References

Bechtel, W. & Abrahamsen, A. (1990). Connectionism and the mind. An introduction to parallel processing in networks. Oxford: Basil Blackwell.

Dell, G.S. (1985). Positive feedback in hierarchical connectionist models. Cognitive Science, 9, 3-23.

Dell, G.S. (1986). A spreading activation theory of retrieval in language production. Psychological Review, 93, 283-321.

Dell, G.S. (1988). The retrieval of phonological forms in production: Tests of predictions from a connectionist model. Journal of Memory and Language, 27, 124-142.

Eikmeyer, H.-J. & Schade, U. (1991). Sequentialization in connectionist language-production models. Cognitive Systems, 3-2, 128-138.

Eikmeyer, H.-J. & Schade, U. (1993). The role of computer simulation in neurolinguistics. Nordic Journal of Linguistics, 16, 153-169.

Laine, M., Niemi, J., Koivuselkä-Sallinen, P. & Hyönä, J. (1994). A neurolinguistic analysis of morphological deficits in a Finnish-Swedish bilingual aphasic. Clinical Linguistics & Phonetics, 8, 177-200.

McClelland, J.L. & Rumelhart, D.E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88. 375-407.

McKay, D.G. (1987). The Organization of Perception and Action. A Theory for Language and Other Cognitive Skills. New York, Springer-Verlag.

Niemi, J., Laine, M. & Tuominen, J. (1994). Cognitive morphology in Finnish: foundations of a new model. Language and Cognitive Processes, 9, 423-446.

Rumelhart, D.E. & McClelland, J.L. (1986). Parallel Distributed Processing: Explorations in the Microstucture of Cognition, Vol 1: Foundations. Cambridge, MA: MIT Press.

Schade, U. (1992). Konnektionismus: zur Modellierung der Sprachproduction. Opladen: Westdeutscher Verlag. Germany.

Stemberger, J.P. (1985). An interactive activation model of language production. In Ellis, A.W. (Ed.) Progress in the Psychology of Language. London: Erlbaum. Vol 1, 143-186.

Tikkala, A. (1995). MORPHO: A connectionist language production tool for Finnish Nouns. Report A/1995/1. Dept. of Computer Science and Applied Mathematics. University of Kuopio. Finland.

Tikkala, A. (submitted to Neural Computer & Applications). A connectionist word production tool for Finnish nouns with a model for vowel harmony restrictions.

Tikkala, A., Eikmeyer, H.-J., Niemi, J. & Laine. M. (in preparation) On the Production of Finnish Nouns: A Psycholinguistically Motivated Connectionist Model.

Wright, J.F. (1995) Connectionist Architectures for Language Disorder Simulation. Dept. of Mathematical and Computing Sciences, University of Surrey, Guilford, UK