Table 1:
CSWL models in literature.
| Paper | Input | Formalism | Key features | Constraints/Biases | Captures | Implications | Limitations |
|---|---|---|---|---|---|---|---|
| Hypothesis Testing Models | |||||||
| Siskind (1996) | Symbolic | Discrete rule-based inference; incremental | Pre-defined rules detect noise and homonymy; Heuristic functions disambiguate word senses under homonymy. | Mutual exclusivity and coverage (to narrow down the set of meanings for a word); composition |
Data: Artificially generated corpus Behaviour: Learns under variable vocabulary size and degree of referential uncertainty; fast mapping; bootstrapping from partial knowledge |
Incremental systems of CSWL and mutual exclusivity (ME) can solve lexical acquisition problems like of children | Not possible to revise the meaning of a word once it is considered learned; sensitive to noise and missing data |
| Frank, Goodman & Tenenbaum (2009) | Sub-symbolic audiovisual | Bayesian | Batch processing; Uses speaker’s intention to derive mappings | Speaker’s intent is known |
Data: Small corpus of mother-infant interactions CHILDES dataset Behaviour: ME; Fast map; Generalizes from social cues |
Some language-specific constraints such as ME may not be necessary | Learns small lexicon; arbitrary representation of speaker’s intention |
| Truesell et al. (2013) | Symbolic | Mathematical | Incremental; Retains one referent per word at a time; minimal free parameters | Some degree of failure to recall |
Data:
Trueswell et al. (2013) Behaviour: Captures participant’s knowledge (or lack) of referents from preceding trials |
Adults retain only one hypothesis about a word’s meaning at each learning instance | Arbitrary assumptions of recall probability |
| Sadegi, Scheutz & Krause (2017) | Sub-symbolic audiovisual | Bayesian, Embodied | Incremental; Uses speaker’s referential intentions; Robotic implementation; Robust to noise | Speaker’s Intent; ME; Limited memory |
Data: Simple utterances from a human to the robot Behaviour: Learns under referential uncertainty |
Incremental models help avoid a need for excessive memory | Tested on a very limited data set and an artificial experiment only |
| Najnini & Banerjee (2018) | Symbolic | Connectionist | Integrates socio-pragmatic theory; Batch processing; Deep reinforcement learning; Uses four reinforcement algorithms | Novel-Noun Novel-Category (N3C); Attentional, prosodic cues |
Data: Artificial experiments on two transcribed video clips of mother-infant interaction from CHILDES corpus Behaviour: Referential uncertainty; N3C bias |
Reinforcement learning models are well-suited for word-learning | Learns one-to-one mappings only; No modelling of empirical experiments |
| Associative Learning Models | |||||||
| Yu & Ballard (2007) | Sub-symbolic Audio-symbolic visual | Probabilistic | Batch processing; Uses expectation maximization; Adds speaker’s visual attention and social cues in speech | Visual Attentional and Social cues |
Data: 600 mother utterances from CHILDES database corpus Behaviour: Referential uncertainty; Role of prosodic and visual cues |
Speakers’ attentional and prosodic cues guide CSWL learning | No modelling of empirical results |
| Fazly, Alishahi & Stevenson (2010) | Sub-symbolic audio; Symbolic visual | Probabilistic | Incremental; Calculates and accumulates probability for each word-object pair | Prior knowledge bias |
Data: Artificial experiments on the CHILDES database corpus Behaviour: Referential uncertainty; ME bias |
Inbuilt biases such as ME not necessary; Primarily, input shapes development | Basic CSWL; no modelling of empirical results |
| Yu & Smith (2011) | Eye-tracking data | Mathematical | Incremental; Moment-by-moment modelling; Uses eye fixations to build associations | Selective visual attention |
Data:
Yu & Smith (2011) Behaviour: Learning under referential uncertainty in infants; Selective attention |
Visual attention drives learning; Learners actively select word–object mappings to store | Mathematical treatment of infant gaze data; Not a model of audiovisual input |
| Nematzadeh, Fazly & Stevenson (2012) | Sub-symbolic audio Symbolic visual | Probabilistic | Extension from Fazly et.al. (2010) forgetting and attention to novelty | Prior knowledge; Attention to novelty; Memory |
Data : Artificial experiments on a small corpus from CHILDES database Behaviour: Referential uncertainty; spacing effect |
Memory and attention processes underlie spacing effect behaviours | No modelling of empirical results |
| Kachergis, Yu & Shiffrin (2012, 2013, 2017) | Symbolic | Mathematical | Incremental; Learned associations and novel items compete for attention; Associations decay; WM supports successive repeated associations | Familiarity/prior knowledge; Novelty/entropy for attentional shifting |
Data:
Kachergis, Yu & Shiffrin (2012, 2013, 2017), Behaviour: ME as well as its relaxation; Sensitivity to variance in input frequency and contextual diversity |
ME can arise in associative mechanisms through attentional shifting and memory decay | Bias to associate uncertain words with uncertain objects similar to ME; Unexplained parametric variations |
| Yurovsky, Fricker, Yu, & Smith (2014) | Symbolic | Mathematical, Bayesian | Compares role of full and partial knowledge in generating mutual exclusivity behaviour | Prior knowledge bias |
Data:
Yurovsky et al (2014) Behaviour: ME; Bootstrapping from partial knowledge |
Partial knowledge can help disambiguate word meanings | Specific to analysis of the role of prior knowledge reported in this work |
| Rasanen & Rasilo (2015) | Sub-symbolic audiovisual | Probabilistic | Transition probability-based; Joint learning of word segmentation and word-object mappings from continuous speech | Transition probability (TP) analysis |
Data : Yu and Smith (2007), Yurovsky, Yu, & Smith (2013) Caregiver Y2 UK corpus; Behaviour: ME, Sensitivity to varying degrees of referential uncertainty |
CSWL can aid bootstrapping of speech segmentation and vice versa | Hard allocation of TPs into disjoint referential contexts; No experiments on development |
| Bassani & Araujo (2019) | Sub-symbolic audiovisual | Modular connectionist | Incremental trial-by-trial learning; Raw images of objects, streams of phonemes as input data | Time-Varying Self-Organizing Maps |
Data:
Yurovsky et al. (2013), Yu and Smith (2007), Trueswell et al. (2013); Behaviour: Referential uncertainty, Local/global competition, Context-sensitive association learning |
Time-Varying Self-Organizing Maps are better at capturing co-variations than Hebbian learning | Does not benefit from prior knowledge in forming new associations |
| Mixed Models | |||||||
| Fontanari, Tikhanoff, Cangelosi, & Perlovsky (2009) | Symbolic | Neural Modelling Fields | Batch processing of input; NMF categorization mechanism | Noise/Clutter detection; Parametric models |
Data: Small artificial dataset Behaviour: Referential uncertainty |
Fuzziness in noise can be exploited to find the correct associations | The number of models is chosen a priori; No modelling of any empirical data |
| Kachergis & Yu (2018) | Symbolic | Mathematical | Extends Kachergis e.al. (2012) with uncertain responses during training | Uncertain response probability |
Data:
Kachergis & Yu (2018) Behaviour: Captures participant accuracy and uncertainty in learning trials |
Neither pure HT or extreme AL models can account for CSWL behaviours | Processes / representations that generate uncertain responses not specified |
| Smith, Smith, & Blythe (2011) | Symbolic | Probabilistic analysis | Comparison of different possible strategies in an associative model |
Data:
Smith et al. (2011) Behaviour: Learning under varying referential uncertainty and interleaving trials |
Continuum of possible strategies used, modulated task difficulty | Mathematical treatment is specific to the empirical work by the authors | |
| Stevens, Gleitman, Trueswell, & Yang (2017)) | Sub-Symbolic audio Symbolic visual | Probabilistic | Incremental; Combines selection, ME, reward based learning and associative learning; | Mutual exclusivity |
Data: CHILDES; Cartmill et al. (2013); Yu & Smith (2007); Trueswell et al. (2013); Koehne et al. (2013) Behaviour: CSWL under varying uncertainty |
Adults can retain multiple associations but always a single favoured hypothesis | Does not account for retaining multiple hypotheses for one word |
| Taniguchi, Taniguchi & Cangelosi (2017) | Sub-symbolic audiovisual | Embodied, Bayesian, Generative | Unsupervised machine learning based on a Bayesian generative model; Robotic implementation; Word learning for objects and actions | Mutual exclusivity; Taxonomic bias |
Data: Artificial experiment on a limited word-referent set Behaviour: learning under referential uncertainty; Learning of objects and actions |
Mutual exclusivity constraint is effective for lexical acquisition in CSL | Does not deal with major issues in CSWL |
| Yurovsky & Frank (2015) | Symbolic | Probabilistic | Incremental; shares intention/attention to create AL to HT spectrum | Intention distribution and memory decay |
Data:
Yurovsky & Frank (2015) Behaviour: CSWL under varying within and between trial uncertainty |
CSWL distributional but modulated by limited attention and memory | Even distribution of attention among non-hypothesized is arbitrary |