Modelling the Tip-Of-the-Tongue State in Guided
Propagation Networks
Dominique G. Béroule
and Michael Zock
LIMSI-CNRS, BP30, 91406 Orsay-cedex, France
LIF - Parc de Luminy, 163, Av. de Luminy Case 901, 13288 Marseille cedex 9, France
Abstract. We start by presenting the Tip-Of-the-Tongue (TOT) problem and
some theories accounting for it. We go then on to consider it within the frame-
work of a neurobiological, computational model: Guided Propagation Networks
(GPNs). “Selective facilitation”, a feature of the model, allows us to formulate
and test various hypotheses accounting for the impeded access to a target word.
We will illustrate and test them via a computer program simulating the TOT
state. A comparison with other “spreading activation” models is proposed in the
final discussion, as well as an evaluation of the proposed hypotheses. The ab-
stract should summarize the contents of the paper and should contain at least 70
and at most 150 words. It should be set in 9-point font size and should be inset
1.0 cm from the right and left margins. There should be two blank (10-point)
lines before and after the abstract.
1 Introduction
When the mental lexicon is contacted to access a specific word, the latter is retrieved
quickly and apparently without too much effort. This is the default case, i.e. the most
frequent situation. Unfortunately, there are cases when things go awry: one fails to
produce the needed target word. This kind of failure may be related to the speaker’s
ignorance of the target word, or, and this is the case we are interested in, to her/his
inability to access the right word, even though s/he knows it. This has often been
dubbed as tip-of-the-tongue (TOT) phenomenon.
Just as studying other dysfunctions of the human mind, analyzing TOT problems
may help to gain a better understanding of the structure of the mental lexicon, hence
inspire computational models allowing the simulation and support of Natural Lan-
guage Production (NLP).
A typical feature of the TOT phenomenon is that people who are experiencing it
have a strong feeling of knowing (FOK) that the retrieval of the target word is very
close [7], [28]. This feeling of imminence can be sustained by the fact that people feel
that they know the word, and that it is about to pop up, ready to cross their lips at any
moment. Indeed, next to conceptual knowledge, people seem to know a lot of other
things concerning the form of the target word [6]: number of syllables, beginning /
ending of the target word, part of speech (noun, verb, adjective, etc.), origin (Greek
or Latin), and even gender [30]. Strangely enough, the resolution of the problem
Béroule D. and Zock M. (2009).
Modelling the Tip-Of-the-Tongue State in Guided Propagation Networks.
In Proceedings of the 6th International Workshop on Natural Language Processing and Cognitive Science , pages 77-93
DOI: 10.5220/0002202300770093
occurs spontaneously, possibly at a time when deliberate retrieval attempts have been
abandoned [7], [25].
Existing models of word production provide assumptions concerning the origins
of the TOT. Following a brief presentation of these systems, we will consider the
TOT problem within the framework of Guided Propagation Networks (GPNs), a
special kind of computational model able to simulate word production. We present
then three hypotheses to model the TOT-phenomenon within the GPN framework:
low-frequency assumption, low-facilitation, and depressed internal flow. Finally, we
compare this work with other approaches.
2 Models of Lexical Production
Observations of naturally occurring TOTs, diary studies, as well as laboratory studies
suggest that at least some TOTs are caused by the blocking of the target word. Sever-
al hypotheses have been offered and tested, including the blocking theory (e.g., [28])
and the partial activation theory (e.g., [16]).
2.1 Psychological Models of the TOT
The blocking theory states that TOTs can occur when plausible, but incorrect answers
pop up in the speaker’s mind. As Woodworth has noted already 80 years ago [31]
similar words can cause this kind of blockage or inhibition of the target word.
The partial activation theory states that the TOT represents incomplete activation
of a target word, the latter failing to reach the necessary threshold. Research support-
ing this theory reveals that partial information about a target word, such as its first
letter or the number of syllables, are commonly available even though the word itself
rests eluded (e.g., [11]). Hence, definitions paired with phonologically related cue
words (for example, ‘abstract’ to prime ‘abdicate’) can free people from the TOT
state, helping them to access the phonological form of the target word (e.g., [16],
The production of a given word is usually viewed as starting with the activation of
semantic information, followed by the specification of a syntactic representation,
before the retrieval of the phonological components of the word (e.g., [5]). According
to this view, the TOT phenomenon is the result of a Transmission Deficit (TD), i.e. a
lack of strength of the connections transmitting activation to the phonological repre-
sentations [8], [21]. Retrieval thus fails because of the incomplete activation of repre-
sentations (see [7] for a review).
According to the TD model, a TOT would occur when the strength of the
connections among phonological nodes is too weak to transmit sufficient activity
towards all the phonemes of the target word. Recent and frequent activation of nodes
strengthens connections, therefore increasing transmission. Because connection
strength decreases over time if no activation has taken place, the TD model explains
why this phenomenon concerns low- rather than high-frequency names or words [8],
[14], [24].
2.2 Guided Propagation Networks
The ability to take instantaneous decisions based on the unconscious integration of
many pieces of information is one of the most striking features of our brain. This
works fine most of the time and without any apparent conscious effort, unless, some-
thing goes awry like in the case of the TOT problem. This is where Guided Propaga-
tion Networks (GPNs) become useful, as they are based on the detection of coinci-
dences between various flows of information across time. Their spontaneous flows
of activity allow them to respond even in the absence of stimulation, a main feature
for applications in both Artificial Intelligence and neuronal modelling. GPNs are also
supported at a physiological level by the spatio-temporal integration properties of the
Post-Synaptic Potentials for neuronal input [18].
A GPN is composed of several modules which work in parallel. The peripheral
ones code low-level information (letters, syllables), and stimulate deeper modules
containing more complex information (lexical, semantic, syntactic). Conversely,
deeper, or more central modules ease the activation of the peripheral modules by
temporary decreasing their propagation thresholds (Fig. 1).
For a given module, appropriate propagation of activity along memory pathways
relies on the coincidence of: (1) the spontaneous flow generated at the root of the
module, (2) incoming stimuli from more peripheral modules, and (3) the facilitation
provided by more central modules. Cells are meant to detect this coincidence. The
contribution of these three flows has to be well synchronized at the level of each cell
for the host module to run properly.
In a GPN module, the spontaneous flow is guided towards target output “event
detectors”. This is due to its coincidence with incoming stimuli and the facilitation
generated by other modules. In sum, the spontaneous flow will follow the pathway
made of a chain of cells with the lowest thresholds, stimulated across time by the
module’s input.
Fig. 1. GPN production modules. The white arrows pointing upwards stand for facilitation
flows. The grey arrow at the front comes from a neuromodulation-like module, facilitating
words of a certain emotional value.
2.3 Lexical GPN Production Modules
In mammals, proprioception is an essential 6th sense through which the brain perce-
ives what the body is doing [27]. In a GPN production module, the spontaneous flow
is sucked up along a specific pathway by the facilitation signals generated by deeper
modules. The internal proprioceptive stimuli are used to control the movement se-
quencing and to avoid Parkinson-like effects [29].
Word generation is performed in two steps:
Phase 1
: a wave of “threshold decrease” is instantaneously propagated backwards
a given lexical pathway.
Phase 2
: When this “backward facilitation” reaches the first cell of the pathway,
due to the priming by the spontaneous flow, a dam is opened: the activity flows to-
wards the next cells, including those whose threshold has just been decreased. The
spontaneous activity bypasses the latter a bit, but not enough to cross a third one. For
this to happen, the proprioceptive stimulus, signalling that the first syllable has just
been generated, is expected to provide its own contribution to the internal flow,
which will then cause an overflow of its contribution to the third cell. The process
continues along these lines until the end of the pathway is reached (Fig. 2).
For this procedure to run properly, three conditions must be met: (1) the backward
facilitation wave must be well regulated, (2) the spontaneous flow must be present at
the beginning of the pathway, and (3) proprioceptive stimuli must be produced by
more peripheral modules (i.e.: syllables). Otherwise, the internal flow may proceed
either too fast towards its target and be of too low intensity to produce the syllables in
the right order, or the activation is too low to reach its target, yielding problems akin
to the TOT phenomenon.
There are three factors likely to influence the propagation of the internal flow to-
wards the end of a pathway:
- the threshold resting level
- the intensity of the threshold offset
- the weight of the internal flow
Before getting more into details, a failure in the full activation of the lexical path-
way can be the result of an “excessively high” value of the threshold level prior to the
facilitation process. Conversely, the failure may also be the result of a lack of de-
crease of the threshold during facilitation. Third, it may be the result of a depressed
internal flow. The second possibility may arise when the value of the offset depends
on several factors, some of which being absent at the moment of trying to produce the
target-word. This holds only for the major lexical categories (nouns, verbs, adjective,
adverbs). In this model, ‘minor’ categories (determiners, preposi-tions, etc.) are only
triggered by a strong single syntactic input. They never cause a TOT-state.
As one can see by looking at the architecture depicted in Fig.1, every meaningful
word stays under the influence of several facilitating factors: syntactic, semantic,
emotional. For example, the syntactic facilitation flow may touch nouns, while words
of a certain semantic frame are facilitated at the same time; words of the current emo-
tional value may also be focused on. The “best” candidate among the words to be
generated has the greatest intersection of the different subsets.
1/ 2/
Fig. 2. Well-regulated activity of a production pathway. At the left-hand side, you can see a
pathway starting by the internal flow generator (cell n°0) ; from top to bottom, the following
cell n°3 is aimed at producing a first syllable (i.e.: MI); cell n°5 generates the 2
(LAN), and the square cell n°6 represents the related full word (which is MILAN). The horizon-
tal axes stand for the time dimension in the diagrams of activity (vertical axes); the cells’ re-
sponse thresholds are represented by dotted lines. The backward facilitation generated by a
deeper module (long, grey arrow pointing upwards) results in a threshold decrease. While this
threshold decrease does not affect the inactive cells (5 and 6), it does induce the response of the
first cell of the pathway (n°3), because of the preactivation due to the internal flow (top dia-
gram). The cell n°3 reacts to this by sending an activation signal to its offspring (n°5), and a
facilitating signal to the syllable-effector n°7. Once this syllable is uttered, its effector stimu-
lates cell n°3 with a “proprioceptive” feedback, which allows the internal flow to move for-
ward. The spontaneous flow eventually reaches the end of the path (bottom diagram), with cell
n°6 sending a “proprioceptive” stimulus to the deeper modules.
According to this view, two mechanisms are used to produce the appropriate word:
1/ decision thresholds associated with each word can be regulated in such a way
that only words having received all the energy become fully active.
2/ mutual inhibition between the remaining candidates simulate a competition
where only the most activated candidate, i.e. word, wins.
The labels t
, t
and t
of the figures here below correspond respectively to four
time slices represented along the horizontal axis of Fig. 2.
3 Exploring Three Hypotheses
Human language production would not work if all possible words for a given slot
were considered serially before being selected. Parallelism is an efficient means to
choose instantaneously one among the possible candidates, as long as the involved
mechanisms work properly.
Fig. 3.a. “Word” production module fed by a “Syllable” module whose effectors appear at the
top (row of square cells). Before being aspirated in a specific direction by a facilitation signal,
the spontaneous activation of the root unit (n° 0) can potentially feed in parallel every pathway
of the module. If the word effector n°6 is facilitated by deeper modules (not represented here),
a threshold decrease proceeds upwards (white and grey arrows) towards the root, touching its
constituent cells.
Fig. 3.b. The response threshold of cell n°3 has decreased and can now be reached by the
spontaneous flow generated by the root cell (n°0). Cell n°3 sends a slight activation signal
towards cells n°4 and n°5, and a facilitation signal up to the syllable effector n°7. The pathway
to the latter yields now the production of its associated syllable-pattern of activity.
Fig. 3.c. When the syllable is generated, the corresponding effector (n°7) gets fully activated
by the contextual flow of its module, spreading then activation towards the cells 3 and 8. Cell 3
has been waiting for this feedback to re-activate cell 5 (as well as cell 4 which has not been
facilitated). The same process applies again, facilitating the pathway of the syllable-effector 9.
Fig. 3.d. At the end of this process, the word effector n°6 gets fully activated, indicating via
activation of deeper modules that the associated word production is completed. The next word
can be triggered in the same way by deeper modules (white and grey arrow at the bottom-
3.1 The Low-frequency Assumption
The TOT state can be experimentally induced by presenting an image or a definition
of a low-frequency object [15]. The frequency (or familiarity) factor seems to be quite
important in the case of temporary lexical amnesia characterizing the TOT. It may be
considered together with the “never-ending learning” schema that allows GPN mod-
ules to grow from scratch [3].
In this theory, the occurrence of an unexpected series of internal stimuli results in
the sprouting of a related pathway inside a GPN module. In agreement with recent
data showing that new neurons are continually born throughout adulthood [23], new
GPN cells are used for pathways to autonomously grow whenever needed in the
course of the “system life”. The following selective consolidation and general extinc-
tion, which both apply to lexical pathways, provide an efficient way for separating
meaningful items from noise. Thanks to this continuous process, a snapshot of the
lexical landscape may contain the most frequently used paths or “highways”, living
together with newborn “tracks”. While being beneficial for the access of very-
frequent words, this kind of selective process may also impede the retrieval of rare
items. Hence, the TOT phenomenon could be considered as a simple side-effect of a
more general learning principle.
Rare words, or those that have not been used much recently, would then have a
high propagation threshold resting level. The dynamic threshold decrease which trig-
gers a lexical production would fail to aspirate the spontaneous flow.
3.2 The Low-facilitation Assumption
As mentioned already, the thresholds resting levels of cells which form a word-
pathway are heavily decreased for the word about to be generated. As shown in Fig-
ure 1, there are at least three factors accounting for the decrease of these thresholds,
hence, its generation: syntactic, semantic and emotional facilitations. Only the con-
nection between syntactic and lexical dimensions seems to be well preserved in TOT.
Emotion may interfere in several ways with the generation of a word:
- Emotional guiding. According to newer theories of the brain architecture, part of
the cortex stores the emotional value of our experience to promote behaviour which
proved to be rewarding [10]. Applying this strategy to word production may result in
blocking words associated with negative events.
- Mismatch between emotional values. Among the dimensions determining the
word to be generated at a certain moment, the emotional value associated through
learning with full-fledged words can be compared with the “general mood” of the
current generation process. Generation of words experienced as rather ‘positive’ may
for instance be inhibited in case of negative mood.
3.3 The Depressed Spontaneous-flow Assumption
Given a possible correlation between the strength of the spontaneous flow and the
Serotonin neuromodulator [29], a lack of serotonin, as found when being in a depres-
sive state, may impede the generation of a target-word. This view is compatible with
studies showing a higher rate of TOT states for depressed people [1].
3.4 Relevance of Learning
The TOT-state seems to be only a retrieval problem, since the internal representation
of the target word exists, but cannot be fully accessed in time. The word may be re-
trieved later on, incidentally, without any extra learning.
However, a mismatch between the cumulative knowledge associated with a target
word and the current situation may result in a TOT effect. For instance, an accidental
event occurring during the initial acquisition of a word may impede later on its pro-
duction in normal conditions. Learning by imitation is allowed by the homogeneity of
GPN modules: the perception of a new word generates a new pathway to be used later
on in the production mode before an autonomous “production” pathway is created
[4]. A possible implication of this unsupervised strategy and the TOT has not been
considered here, as it falls outside the scope of this paper.
4 Computer Simulation
A full simulation of the TOT state would require a complete lexicon, with words of
various types and emotional values. Right now, we are interested in the low-level
mechanisms causing or accounting for the failure to produce a given target-word
represented in a GPN pathway. This first experiment does not account for the possi-
ble mutual inhibition between candidate words, neither does it include emotional
mismatches like those that may occur in real life.
The GPN software used here is a prototype of a GPN translator, where French and
English are respectively the source and target language [4]. The “production” mod-
ules are shown in Fig.1. “Perception” modules are organized according to a similar
architecture. They send facilitating signals to their “production” counterpart in the
course of processing. In this way, the production of a given word depends on the
multiple facilitations received from the various sources: - a French word which feeds
all its possible translations in English, - a semantic frame pathway feeding several
words of the same meaning, - a syntactic pathway facilitating words of a certain class.
The target-word stands at the intersection of all these facilitated sub-lexicons.
As always with GPN software, the dynamic links between the cells are
represented by pointers attached to a source-cell data-structure, signaling the ad-
dresses in memory of other tables of data respectively associated with destination-
cells. Among other information, each pointer of this kind has a weight tending to
decrease slightly at each time-step of the program. Although the instantaneous value
of this weight, as well as the one of other parameters (thresholds, time-delays, signal
durations) is worked out by the learning algorithm, it can be adapted “manually” to
achieve a given effect in this type of “local” representation. In the network’s graphi-
cal interface, the user can also click on a given cell (or pathway), to display its activi-
ty over time, as shown in the next sections. A user interface has been added to this
software, allowing the propagation parameters of the GPN modules, i.e. the “lexical
production”, to be fixed to any value, not just the theoretical one. The resting level of
the thresholds can be set to the minimum value, thus simulating the case of a recently
acquired (or low-frequency) word. The temporary decrease of a pathway’s thresholds
that triggers the production of a word can also be reduced. In the same way, the
weight of the internal spontaneous flow can be decreased to simulate a depressed
activity in the module.
4.1 Simulation of Low-frequency
In the GPN formalism, two parameters determine the cell’s propagation behaviour: its
excitability, i.e. its potential to reach its propagation threshold, and the ratio between
the internal flow and the stimuli (weight of the spontaneous flow). New cells are born
with low excitability. As the cell is used, this parameter gradually increases, correlat-
ing with the familiarity of the item that the cell codes for. In order to simulate the low
frequency of a word, all the cells of its associated pathways are set to the minimal
level of excitability (E = 1.).
Fig. 4. Pattern of propagation along a generation pathway not frequently used (low-frequency
assumption). The same diagrams as the ones shown in Fig.2 are superposed in grey, in order to
allow for comparison of this basic situation with the TOT state. Here, the decision thresholds
(black dotted lines) of the cells displayed at the left-hand side have been shifted to their maxi-
mum value (excitability: 1).
The decrease of the threshold is kept similar to the basic situation. The first syllable of the word
(cell n°7) is activated at a lower level, hence with less intense feedback towards cell 3, whose
propagated activity does not reach the threshold of cell 5. The activation remains stuck at the
same sub-threshold level, and the internal flow does not spread anymore along the pathway,
whatever its length (the word’s number of syllables).
4.2 Simulation of Low-facilitation
The most robust value for the threshold offset provoking generation is the one at the
middle of a permitted interval. For this second experiment, excitability has been set to
1.5, and the offset to 1/3 of the interval.
Fig. 5. Pattern of propagation along a generation pathway in which the threshold offset is
depressed. Either because of an emotional mismatch, a weak connection with the current se-
mantic context, or because of the inhibition due to a more frequent word-candidate, the thre-
shold decrease being half its normal value. Combining low-frequency and depressed links
would simply prevent the first syllable from being activated.
4.3 Simulation of Low Spontaneous Flow
The weight of the internal flow usually follows an arithmetic progression for the
intensity of the output to give a matching score of a series of stimuli and its internal
representation. By setting this weight to unity all along a pathway, without updating
the thresholds calculation, the spontaneous flow decreases in intensity as it proceeds
towards the end of the pathway.
Fig. 6. Pattern of propagation along a pathway fed by a weak internal flow. The two first syl-
lables are produced with less intensity, hence the flow cannot cross the next cell’s threshold
5 Discussion
5.1 Comparison with other Approaches
While not addressing strictly speaking the TOT problem, several computational
models can be compared with GPNs: in particular those applying the “spreading
activation” scheme to information retrieval.
The development of GPNs started in the 80’s, while a new tide of learning algo-
rithms was reaching the shore of Artificial Neural Networks (ANNs). Working on a
pre-wired net of formal neurons – either layered or fully connected -, the supervised
procedures modify internal parameters (connection weights and propagation thre-
sholds) so that a selection of input patterns of activity - or vectors – be trained to
generate another given set of output vectors.
Apart from the root-cell that initiates the internal flow, GPN modules are initially
empty, instead. New cells are chained in the course of processing, when the module
anticipations are not confirmed by internal stimuli. No supervision is needed. A new
pathway is automatically created from the cell currently activated, provided that
propagation thresholds of the concerned module are set up to their maximum value.
As a matter of fact, GPN parameters only determine the status of the internal repre-
sentations: flexible Recognition, on-line Learning, and flexible Production. The poss-
ible values of the related two parameters are not learnt, but are taken in relatively
small areas [4]; they are dynamically modified either by another module or by the
general algorithm standing for a central Control Unit.
The activation implemented in multilayered ANNs also spread towards lexical
models in psycholinguistics. Two structures are then involved: high-
level/semantic/content-related, and low-level/ phonological/form-related.
Indeed, studying performance, [13] observed that people tend to make two kinds
of errors: meaning-based substitutions (left instead of right) or substitutions based on
similarity of form (historical instead of hysterical). Given this fact, and given the
little evidence for form choices to interfere with the semantic choices, led them to
claim that lexical access is a two-step process, whose components are serially related:
meaning choices taking place before form-related computations (but see [2] and [9]).
The process is supposed to take place in the following way: given some information
(semantic, conceptual), a lemma is retrieved, triggering the activation (or computa-
tion) of a lexeme, the lemma’s corresponding phonological-, or graphemic-form.
Please note, unlike for computational linguists, a lemma is not a concrete dictionary
entry, but an abstract structure containing semantic and syntactic information (part of
speech), while the lexeme contains the phonological form (syllabic structure, pho-
nemes, intonation curve, etc.).
This view has led to various models. Indeed the major psycholinguistic theories of
word production are all activation-based, multilayered network models. Most of them
are implemented, and their focus lies on modelling human performance: speech errors
or the time course (latencies) as observed during the access of the mental lexicon. The
two best-known models are those of Dell [12] and Levelt [19] [20], which take op-
posing views concerning conceptual input (conceptual primitives vs. holistic lexica-
lized concepts) and activation flow (one-directional vs. bi-directional).
The Dell model is an interactive-activation-based theory that, starting from a set
of features, generates a string of phonemes. Information flow is bi-directional, that is,
lower level units can feed back to higher-level components, which may lead to errors.
For example, the system might produce rat instead of the intended cat. Indeed, both
words share certain components. Hence, both of them are prone to be activated. At
the conceptual level (from the top) they share the feature animal, while at the phono-
logical level (from the bottom) they share two phonemes. When the word node for cat
is active, any of the following segments /k/, /æ/, and /t/ is co-activated. The latter two
phonemes may feed back, leading to rat, which may already be primed and be above
baseline due to some information coming from a higher-level component. The model
can account for various other kinds of speech errors like preservations (e.g., beef
needle soup), anticipations (e.g., cuff of coffee), etc.
Based on the distribution of word errors, Dell argues that some aspects of speech
generation rely on retrieval (phrases, phonological features, etc), while many others
(word/phoneme and possibly morpheme combinations) rely on synthesis. Since gen-
eration is a productive task, it is prone to swapping or reuse of elements.
WEAVER++ (Word Encoding by Activation and VERification) is also a computa-
tional model. It has been designed to explain how speakers plan and control the pro-
duction of spoken words [26]. The model is "hybrid" as it combines a declarative
associative network and procedural rules with spreading activation and activation-
based rule triggering. Words are synthesized by weaving together various kinds of
WEAVER++ is also activation-based, information flow is only one-
directional, top-down. Processing is staged in a strictly feed-forward fashion. Starting
from lexicalized concepts (concepts for which a language has words) it proceeds
sequentially to lemma selection, morphological, phonological and phonetic encoding,
to finish off with a motor plan, necessary for articulation. Unlike the previous model,
WEAVER++ accounts primarily for reaction time data. Actually, it was developed on
the basis of such data collected during the task of picture naming. However, more
recently the program managed to parallel a number of findings obtained in psycholin-
guistics where other techniques (chronometry, eye tracking, electrophysiological and
neuro-imaging) have been used.
Again, as mentioned already, these models deal with word access in general and
not with modelling the TOT problem.
Also, just as Pattern Recognition must deal with variations of real-world temporal
patterns, the actual generation of syllables undergoes variations of rhythm. Propri-
oception, as viewed in GPNs, provides a solution to this problem, by sending a feed-
back to the lexical motor plan as soon as a phonetic unit has been successfully ut-
tered. Whereas ANNs as well as psychological models implement a single type of
signal (activation) for conveying both low and high levels information, GPNs provide
a dynamic “threshold decrease” for high-level information to influence lower levels,
while the opposite influence is handled by activation signals.
As shown in the previous sections, and in agreement with the two-step process
model, the production of a word is triggered by a meaning-related signal. However,
the corresponding threshold decrease is not effective without a basic activation of the
phonological level (spontaneous activation), nor the occurrence of a series of propri-
oceptive stimuli. The syntactic module itself waits for lexical feedback before trigger-
ing the next word class to be instantiated by the next word in the string.
To summarize, in the field of language production, the possible contributions of
GPNs to the “spreading activation” approach is twofold:
- A dialog between two types of signals originating respectively from low-level
and high-level information.
- A never-ending learning procedure that incorporates new pieces of knowledge
“in context”, and therefore without necessary supervision; this involves a central
unit that allows a given module to temporarily switch to the learning mode
through an increase of thresholds.
5.2 Comparison of the Three Hypotheses
The aforementioned hypotheses may together impede the production of a target word.
If the model is correct and if you’d like to avoid the unpleasant TOT-experience, then
you’d better not be depressed (H3), only utter words that you are familiar with in
general (H1), including in the current situation (H2).
Not all these explanations have the same status, though. The following may clari-
fy the part played by each one of them, by underlining that basically only two coinci-
dent factors determine access to a target word:
- Activations
: general spontaneous activity fed by proprioceptive stimuli (syl-
lables) along the path.
- Decision thresholds
: initial resting level of the pathway cells, decreased by sig-
nals conveying the semantic and syntactic current context.
Starting from their initial state, these two factors are modified in the course of
processing. Once the thresholds have been lowered along the target pathway, the
proprioceptive stimuli occur one after the other to guide the spontaneous propagation
until the end of the path. One may note that there is another possible hypothesis to
account for the TOT phenomenon: the lack of proprioception, not considered here
since it would concern the entire lexicon rather than just the target-word. Likewise,
the depressed spontaneous flow is not restricted to one particular target word.
The main question then is: how to deal with or how to compensate for the TOT
A lack of internal spontaneous flow may find a pharmacological answer, not specific
to the production of a target word. This would in addition require specific training.
Once retrieved, repetitions of a target word would ‘decrease its basic response thre-
sholds’. Using this word in various contexts and moods would extend the number of
links between its ‘associated pathway’ and conceptual representations, thus increas-
ing its availability for subsequent context-driven, two-step retrievals.
6 Conclusions and Perspective
One of the most amazing problems people encounter in language production is their
failure to access a word they are certain to know. That they have memorized the word
can easily be shown (they produced the word in the past, they are able to recognize it
without mistake and immediately in a list of words, or they may produce it later dur-
ing the day); yet, what accounts for their frustration is the fact that they cannot access
the word when needed mostly, now.
While the metaphor of the Tip-of-the-Tongue seems self-explanatory and attrac-
tive, it is also misleading, if not a misnomer. Indeed, those being in this state have the
impression that the word they are looking for has made it all the way from the mind
to the mouth, or close to it, since it got stuck (or blocked) at the very last moment,
when it reached the tip of the tongue. Of course, what happens in reality is different
in kind, occurring at a different, deeper and far more central place, the brain.
As other dysfunctions of the brain machinery, the TOT effect may reveal basic
mechanisms and inner organization of the involved structures. Sometimes considered
as a type of “speech error”, this phenomenon may better be described as a kind of
word blockage, impeding fluency or access to the word to be inserted in a particular
point of the chain, i.e. host sentence.
The reason generally given to explain why a lexeme could not be uttered at the
right moment is lack of energy of the word’s internal representation: the needed acti-
vation threshold could not be reached, as it may happen for a rare word. We offer
other explanations in order to account for a subset of TOT symptoms. More precisely,
we suggested two specific impairments of the GPN production mechanisms, namely:
weak decrease of propagation thresholds and depressed spontaneous flow of activity.
These features remain to be studied with the same computer simulation, in expe-
riments involving a large set of words of known frequencies and emotional values.
Comparisons with other models addressing the TOT question, such as the “two-
stage” lexical access [17], may be carried out.
Another TOT feature that we plan to investigate is: “delayed retrieval”. The target
word pops up, but long after having been consciously searched for, and too late to be
inserted in the specific slot reserved for it. For this goal to be met, a serial retrieval
process is to be implemented, decreasing slightly, one after the other, the response
thresholds of the lexical candidates. The word pathway whose thresholds has been
decreased the most – but possibly not enough to make it at a given moment – would
then get a better chance to be retrieved, finally.
1. Andersen, J., Gardner, J., George, J., Fulton., A.: The Effects of Depression on Tip of the
Tongue Experiences, Proceeds of the 77th Annual WPA Convention, Seattle, April 25,
2. Badecker, W., Miozzo, M., & Zanuttini, R.: The two-stage model of lexical retrieval: Evi-
dence from a case of anomia with selective preservation of grammatical gender. Cognition,
57, (1995) 193-216.
3. Béroule, D.: The Never-ending Learning, in: Neural Computers, R.Eckmiller, C.v.d. Mals-
burg (Eds.), Springer-Verlag (1988) 219-230.
4. Béroule., D.: An instance of Coincidence Detection Architecture relying on Temporal
Coding, in IEEE Trans. On Neural Networks, Special Issue on Temporal Coding for Neural
Information Processing, DeLiang Wang & al (Eds.), Vol. 15, N°5 (2004) 963-979.
5. Bock, K., and Levelt, W.: Language Production: Grammatical Encoding. Gernsbacher, M.
(Ed.) Handbook of Psycholinguistics, Orlando, FL, Academic Press (1994) 945-984
6. Brown, R., and McNeill, D.: The tip of the tongue phenomenon. In Journal of Verbal
Learning and Verbal Behaviour, 5 (1996) 325-337.
7. Brown, A. S.: A review of the tip-of-the-tongue experience. Psychological Bulletin, 109
(1991) 204-233.
8. Burke, D. M. & MacKay, D.G., Worthley, J.S. and Wade, E.: On the Tip of the Tongue:
What Causes Word Finding Failures in Young and Older Adults? Journal of Memory and
Language 30 (1991) 542-579.
9. Caramazza, A.: How many levels of processing are there in lexical access? Cognitive
Neuropsychology, 14 (1997) 177-208.
10. Cohen, M.X., Heller, A.S., Ranganath, C.: Functional connectivity with anterior cingulated
and orbitofrontal cortices during decision-making, Cognitive Brain Res. 23 (2004) 61-70.
11. Cohen, G. & Faulkner, D.: Memory for proper names: Age differences in retrieval. British
Journal of Developmental Psychology, 4 (1986) 187-197.
12. Dell, G.S. (1986): A spreading-activation theory of retrieval in sentence production. Psy-
chological Review, 93, 283-321.
13. Fay D. & Cutler A.: Malapropisms and the structure of the mental lexicon. Linguistic
Inquiry, 8 (1977) 505-520.
14. Harley, T. and Bown, H.: What causes a tip-of-the-tongue state? Evidence for lexical
neighbourhood effects in speech production. British Journal of Psych., 89 (1977) 151-174.
15. Heine, M. K., Ober, B. A. & Shenaut, G. K.: Naturally occurring and experimentally in-
duced tip-of-the-tongue experiences in three adult age groups. Psychology and Aging, 14
(1977) 445-457.
16. James, L. E. & Burke, D. M.: Phonological priming effects on word retrieval and tip-of-
the-tongue experiences. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 26 (2000) 1378-1391.
17. Kempen, G., & Huijbers, P.: The lexicalization process in sentence production and nam-
ing: Indirect election of words. Cognition, 14 (1983) 185–209.
18. Laget., P.: Relations synaptiques et non-synaptiques entre les éléments nerveux, Ed. Mas-
son (1970)
19. Levelt, W.: Accessing Words in Speech Production: Stages, Processes and Representa-
tions. Cognition 42 (1992) 1-22
20. Levelt, W.J.M., Roelofs, A., & Meyer, A.S.: A theory of lexical access in speech produc-
tion. Behavioral and Brain Sciences, 22 (1999) 1-75.
21. MacKay, D. and Burke, D.: Cognition and aging: A theory of new learning and the use of
old connections. In T. Hess (Ed.), Aging and cognition: Knowledge organization and utili-
zation, Amsterdam: North-Holland (1990) 213-263.
22. Meyer, A., and Bock, K.: The tip-of-the-tongue phenomenon: Blocking or partial activa-
tion? Memory and Cognition, 20 (1992) 715-726.
23. Rakic, P.: Neurogenesis in adult primate neocortex: an evaluation of the evidence, Nat Rev
Neurosci. 3 (1) (2002) 65–71
24. Rastle, K., and Burke, D.: Priming the tip of the tongue: Effects of prior processing on
word retrieval in young and older adults. Journal of Memory & Lang., 35 (1996) 586-605.
25. Reason, J. & Lucas., D.: Using cognitive diaries to investigate naturally occurring memory
blocks. In J. E. Harris & P. E. Morris (Eds.), Everyday memory: Actions and absentmind-
edness. London: Academic Press. (1983) 53-69.
26. Roelofs, A.: A spreading-activation theory of lemma retrieval in speaking, Levelt, W. (ed.)
Special issue on the lexicon, Cognition, 42 (1992) 107-142
27. Sacks, O.: The Disembodied Lady, in "The Man Who Mistook His Wife for a Hat and
Other Clinical Tales, Summit Books (1985).
28. Smith., S.M.: Frustrated feelings of imminence: On the tip of the tongue. In J. Metcalfe &
A. Shimamura (Eds.), Metacognition: Knowing about knowing. Cambridge, MA: MIT
Press, (1994) 27-45.
29. Toffano-Nioche, C., Béroule, D., Tassin, J.P.: A Functional Model of some Parkinson's
Disease Symptoms using a Guided Propagation Network, Artificial Intelligence in Mede-
cine, N°14 (1998) 237-258.
30. Vigliocco, G., Antonini, T. and Garrett, M.: Grammatical gender is on the tip of Italian
tongues. Psychological Science, 8 (1997) 314-317.
31. Woodworth, R.: Psychology. Oxford, England: Holt (1929).