Finding Structure in One Child's Linguistic Experience

Neural network models have recently made striking progress in natural language processing, but they are typically trained on orders of magnitude more language input than children receive. What can these neural networks, which are primarily distributional learners, learn from a naturalistic subset of a single child's experience? We examine this question using a recent longitudinal dataset collected from a single child, consisting of egocentric visual data paired with text transcripts. We train both language-only and vision-and-language neural networks and analyze the linguistic knowledge they acquire. In parallel with findings from Jeffrey Elman's seminal work, the neural networks form emergent clusters of words corresponding to syntactic (nouns, transitive and intransitive verbs) and semantic categories (e.g., animals and clothing), based solely on one child's linguistic input. The networks also acquire sensitivity to acceptability contrasts from linguistic phenomena, such as determiner-noun agreement and argument structure. We find that incorporating visual information produces an incremental gain in predicting words in context, especially for syntactic categories that are comparatively more easily grounded, such as nouns and verbs, but the underlying linguistic representations are not fundamentally altered. Our findings demonstrate which kinds of linguistic knowledge are learnable from a snapshot of a single child's real developmental experience.

The emergence of word-internal repetition through iterated learning: Explaining the mismatch between learning biases and language design

The idea that natural language is shaped by biases in learning plays a key role in our understanding of how human language is structured, but its corollary that there should be a correspondence between typological generalisations and ease of acquisition is not always supported. For example, natural languages tend to avoid close repetitions of consonants within a word, but developmental evidence suggests that, if anything, words containing sound repetitions are more, not less, likely to be acquired than those without. In this study, we use word-internal repetition as a test case to provide a cultural evolutionary explanation of when and how learning biases impact on language design. Two artificial language experiments showed that adult speakers possess a bias for both consonant and vowel repetitions when learning novel words, but the effects of this bias were observable in language transmission only when there was a relatively high learning pressure on the lexicon. Based on these results, we argue that whether the design of a language reflects biases in learning depends on the relative strength of pressures from learnability and communication efficiency exerted on the linguistic system during cultural transmission.

Quantifying the generalization error in deep learning in terms of data distribution and neural network smoothness

The accuracy of deep learning, i.e., deep neural networks, can be characterized by dividing the total error into three main types: approximation error, optimization error, and generalization error. Whereas there are some satisfactory answers to the problems of approximation and optimization, much less is known about the theory of generalization. Most existing theoretical works for generalization fail to explain the performance of neural networks in practice. To derive a meaningful bound, we study the generalization error of neural networks for classification problems in terms of data distribution and neural network smoothness. We introduce the cover complexity (CC) to measure the difficulty of learning a data set and the inverse of the modulus of continuity to quantify neural network smoothness. A quantitative bound for expected accuracy/error is derived by considering both the CC and neural network smoothness. Although most of the analysis is general and not specific to neural networks, we validate our theoretical assumptions and results numerically for neural networks by several data sets of images. The numerical results confirm that the expected error of trained networks scaled with the square root of the number of classes has a linear relationship with respect to the CC. We also observe a clear consistency between test loss and neural network smoothness during the training process. In addition, we demonstrate empirically that the neural network smoothness decreases when the network size increases whereas the smoothness is insensitive to training dataset size.

Cross-linguistic frequency and the learnability of semantics: Artificial language learning studies of evidentiality

It is often assumed that cross-linguistically more prevalent distinctions are easier to learn (Typological Prevalence Hypothesis; TPH). Prior work supports this idea in phonology, morphology and syntax but has not addressed semantics. Using Artificial Language Learning experiments with adults, we test predictions made by the TPH about the relative learnability of semantic distinctions in the domain of evidentiality, i.e., the linguistic encoding of information source. As the TPH predicted, when exposed to miniature evidential morphological systems, adult speakers of English whose language does not encode evidentiality grammatically learned the typologically most prevalent system (marking indirect, reportative information) better compared to less-attested systems (Experiments 1-2). Similar patterns were observed when non-linguistic symbols were used to encode evidential distinctions (Experiment 3). Our data support the conjecture that some semantic distinctions are marked preferentially and acquired more easily compared to others in both language and other symbolic systems.

Are ergonomically designed road signs more easily learned?

Traffic signs are used to control and regulate traffic and create a safe environment for road users. Therefore, it is of great importance to ensure that road signs are easily understood and easily learned, especially by drivers. The current study, conducted in Israel, focused on the ability to learn unfamiliar road signs and on the effect of the compliance of those signs with principles of physical and conceptual compatibility (ergonomic principles for symbol design). A preliminary experiment tested the extent to which each of 30 signs, not used in Israel, complied with the compatibility principle. Based on this evaluation, the signs were divided into two distinctive groups - "ergonomic" signs and "non-ergonomic" signs. In the second part of the study, we tested the differences between the two groups of signs in terms of comprehension and learnability, conducting recognition-recall tests among 33 teenagers without a driving license. Results showed a significant design main effect, with a higher comprehension level associated with the ergonomic signs compared to the non-ergonomic signs. A significant learnability main effect was also found, with comprehension levels in the recall test significantly higher than in the recognition test for both groups of signs. In addition, the probability for improvement in comprehension based on learning the meanings of the signs was higher among the ergonomic group of signs. These findings demonstrate the importance of ergonomic design in shaping the understandability and learnability of road signs. The implications of the current results can be most relevant to tourists, as well as novice drivers.

Evaluating the usability of speech recognition to create clinical documentation using a commercial electronic health record

OBJECTIVE

To conduct a usability study exploring the value of using speech recognition (SR) for clinical documentation tasks within an electronic health record (EHR) system.

METHODS

Thirty-five emergency department clinicians completed a system usability scale (SUS) questionnaire. The study was undertaken after participants undertook randomly allocated clinical documentation tasks using keyboard and mouse (KBM) or SR. SUS scores were analyzed and the results with KBM were compared to SR results.

RESULTS

Significant difference in SUS scores between EHR system use with and without SR were observed (KBM 67, SR 61; P = 0.045; CI, 0.1 to 12.0). Nineteen of 35 participants scored higher for EHR with KBM, 11 higher for EHR with SR and 5 gave the same score for both. Factor analysis showed no significant difference in scores for the sub-element of usability (EHR with KBM 65, EHR with SR 62; P = 0.255; CI, -2.6 to 9.5). Scores for the sub-element of learnability were significantly different (KBM 72, SR 55; P < 0.001; CI, 9.8 to 23.5). A significant correlation was found between the perceived usability of the two system configurations (EHR with KBM or SR) and the efficiency of documentation (time to document) (P = 0.002; CI, 10.5 to -0.1) but not with safety (number of errors) (P = 0.90; CI, -2.3 to 2.6).

DISCUSSION

SR was associated with significantly reduced overall usability scores, even though it is often positioned as ease of use technology. SR was perceived to impose larger costs in terms of learnability via training and support requirements for EHR based documentation when compared to using KBM. Lower usability scores were significantly associated with longer documentation times.

CONCLUSION

The usability of EHR systems with any input modality is an area that requires continued development. The addition of an SR component to an EHR system may cause a significant reduction in terms of perceived usability by clinicians.