Sentence-level Language Research Articles

What Is Important to Measure in Sentence-Level Language Comprehension?

What Is Important to Measure in Sentence-Level Language Comprehension?

Transformer Grammars: Augmenting Transformer Language Models with Syntactic Inductive Biases at Scale

Abstract We introduce Transformer Grammars (TGs), a novel class of Transformer language models that combine (i) the expressive power, scalability, and strong performance of Transformers and (ii) recursive syntactic compositions, which here are implemented through a special attention mask and deterministic transformation of the linearized tree. We find that TGs outperform various strong baselines on sentence-level language modeling perplexity, as well as on multiple syntax-sensitive language modeling evaluation metrics. Additionally, we find that the recursive syntactic composition bottleneck which represents each sentence as a single vector harms perplexity on document-level language modeling, providing evidence that a different kind of memory mechanism—one that is independent of composed syntactic representations—plays an important role in current successful models of long text.

Speaking in the Brain: The Interaction between Words and Syntax in Sentence Production

This neuroimaging study investigated the neural infrastructure of sentence-level language production. We compared brain activation patterns, as measured with BOLD-fMRI, during production of sentences that differed in verb argument structures (intransitives, transitives, ditransitives) and the lexical status of the verb (known verbs or pseudoverbs). The experiment consisted of 30 mini-blocks of six sentences each. Each mini-block started with an example for the type of sentence to be produced in that block. On each trial in the mini-blocks, participants were first given the (pseudo-)verb followed by three geometric shapes to serve as verb arguments in the sentences. Production of sentences with known verbs yielded greater activation compared to sentences with pseudoverbs in the core language network of the left inferior frontal gyrus, the left posterior middle temporal gyrus, and a more posterior middle temporal region extending into the angular gyrus, analogous to effects observed in language comprehension. Increasing the number of verb arguments led to greater activation in an overlapping left posterior middle temporal gyrus/angular gyrus area, particularly for known verbs, as well as in the bilateral precuneus. Thus, producing sentences with more complex structures using existing verbs leads to increased activation in the language network, suggesting some reliance on memory retrieval of stored lexical-syntactic information during sentence production. This study thus provides evidence from sentence-level language production in line with functional models of the language network that have so far been mainly based on single-word production, comprehension, and language processing in aphasia.

Narrative Skills in Primary School Children with Autism in Relation to Language and Nonverbal Temporal Sequencing

Recent research has suggested that temporal sequencing of narrative events might be a domain-general ability that underlies oral narrative capacities. The current study investigated this issue in a group of children with known pragmatic and narrative difficulties, namely Autism Spectrum Disorder (ASD). We hypothesized (1) that children with ASD (n = 45) would retell narratives of poorer quality than both chronological age-matched (CAM) children and younger children matched on sentence-level language skills (LM), and (2) that nonverbal temporal sequencing skills would uniquely predict individual differences in oral narrative performance in children with ASD. The results show that children with ASD performed poorer on all measures of oral narrative quality compared with the CAM group, and on eight of ten measures compared with the LM group. Thus, our first hypothesis was confirmed, suggesting that narrative difficulties in ASD cannot be fully explained by impaired language. The second hypothesis was only partly confirmed: nonverbal temporal sequencing explained significant or marginally significant variance in some, but not all, aspects of oral narrative performance of children with ASD. These results are discussed from theoretical and clinical/educational perspectives, in relation to the heterogeneity of language skills in ASD and to domain-general features of narrative processing.

Seamless learning for oral presentations: designing for performance needs

This paper investigates English language learners’ oral presentation needs, alongside students’ and instructors’ perceptions towards mobile seamless language learning. The findings will be used to develop a mobile-based learning environment. Interviews with both instructors and students were used to help build a Likert questionnaire which was refined with input from experts. Results reveal that students have a need for oral presentation training in areas such as slide design, presentation-specific language, structure, and body language. Instructors and students are willing to work with peers on oral presentation projects despite lacking previous experience. Deeper exploration from interviews show students believe sentence level language issues are more important than the organization and design of a presentation. Furthermore, instructors expressed problems with incorporating technology and online collaborative learning into language courses suggesting they need more training with learning technology. Students also believed online collaboration would be convenient but difficult to manage, indicating a need for both online and face-to-face collaboration. These findings will help with the development of a mobile seamless language learning framework and app for oral presentations. Guidelines are suggested to help instructors to improve students’ oral presentation performance.

Age-sensitive associations of segmental and suprasegmental perception with sentence-level language skills in Mandarin-speaking children with cochlear implants

Background and aimIt remains unclear how recognition of segmental and suprasegmental phonemes contributes to sentence-level language processing skills in Mandarin-speaking children with cochlear implants (CIs). Our study examined the influence of implantation age on the recognition of consonants, lexical tones and sentences respectively, and more importantly, the contribution of phonological skills to sentence repetition accuracy in Mandarin-speaking children with CIs. MethodsThe participants were three groups of prelingually deaf children who received cochlear implants at various ages and their age-matched controls with normal hearing. Three tasks were administered to assess their consonant perception, lexical tone recognition and language skills in open-set sentence repetition. ResultsChildren with CIs lagged behind NH peers in all the three tests, and performances on segmental, suprasegmental and sentence-level processing were differentially modulated by implantation age. Furthermore, performances on recognition of consonants and lexical tones were significant predictors of sentence repetition accuracy in the children with CIs. ConclusionOverall, segmental and suprasegmental perception as well as sentence-level processing is impaired in Mandarin-speaking children with CIs compared with age-matched children with NH. In children with CIs recognition of segmental and suprasegmental phonemes at the lower level predicts sentence repetition accuracy at the higher level. More importantly, implantation age plays an important role in the development of phonological skills and higher-order language skills, suggesting that age-appropriate aural rehabilitation and speech intervention programs need to be developed in order to better help CI users who receive CIs at different ages.

Examining Sources and Mechanisms of Reading Comprehension Difficulties: Comparing English Learners and Non-English Learners within the Simple View of Reading.

In the present study, we compared the extent to which linguistic comprehension (vocabulary and listening comprehension) and word reading explain reading comprehension differentially for English learners (ELs) and non-ELs with reading difficulties, and we investigated whether different mechanisms of reading comprehension failure exist for each group. Using the simple view of reading as our framework, we tested a model in which vocabulary exerts a direct effect on reading comprehension and indirect effects through listening comprehension and word reading. Results from a multigroup structural equation model with a sample of 446 struggling fourth-grade readers (n = 229 for ELs; n = 211 for non-ELs) demonstrated both similarities and differences in the sources and mechanisms of reading comprehension difficulties for ELs and non-ELs with reading problems. Word reading was an important source of reading comprehension difficulty for both groups. For non-ELs, the effect of word reading was larger than the effects of linguistic comprehension (vocabulary and listening comprehension combined); however, for ELs, the effects of linguistic comprehension were greater than the effect of word reading. Vocabulary had indirect effects via both listening comprehension and word reading for ELs, but it demonstrated a direct effect on reading comprehension for non-ELs. These results suggest that developing a range of linguistic comprehension skills (e.g., word-level and sentence-level language skills) may be important for ELs with reading comprehension difficulties in the upper-elementary grades.

Morphological awareness and word-level reading in early and middle elementary school years

AbstractWe examined whether morphological awareness made a significant contribution to word-level reading across Grades 1 to 4. We test these relations specifically in a task measuring awareness of past-tense forms. A total of 375 children from Grades 1 to 4 completed tasks assessing past-tense morphological awareness along with real word and pseudoword reading. Children also completed control measures assessing phonological awareness, phonological short-term memory, sentence-level language skills, and nonverbal cognitive ability. After these controls, past-tense morphological awareness was a significant predictor of real word reading in Grades 1 and 2, but not in Grades 3 and 4. Further, following on all controls, past-tense morphological awareness was a consistent predictor of pseudoword reading across Grades 1 to 4. Morphological awareness, at least as measured with past-tense verbs, appears to have a role in word reading across the early to middle elementary school grades; for young readers, there are relations to reading of both known and novel words, and for older readers, relations are significant specifically in reading novel words. These findings are discussed within the context of theories of word reading development.

Language Model Score Regularization for Speech Recognition

Inspired by the fact that back-off and interpolated smoothing algorithms have significant effect on statistical language modeling, this paper proposes a sentence-level Language model (LM) score regularization algorithm to improve the fault-tolerance of LMs for recognition errors. The proposed algorithm is applicable to both count-based LMs and neural network LMs. Instead of predicting the occurrence of a sequence of words under a fixed order Markov assumption, we use a composite model consisting of different order models with either n-gram or skip-gram features to estimate the probability of the sequence of words. In order to simplify implementations, we derive a connection between bidirectional neural networks and the proposed algorithm. Experiments were carried out on the Switchboard corpus. Results on N-best lists re-scoring show that the proposed algorithm achieves consistent word error rate reduction when it is applied to count-based LMs, Feedforward neural network (FNN) LMs, and Recurrent neural network (RNN) LMs.

Semantic processing in children with cochlear implants: Evidence from event-related potentials

ABSTRACTIntroduction: Existing research has shown that children with significant hearing loss who use cochlear implants (CIs) perform worse than their hearing peers on behavioral measures of spoken language. The present study sought to examine how children with CIs process lexical–semantic incongruence, as indexed by electrophysiological evidence of the N400 effect.Method: Twelve children with CIs, aged between 6 and 9 years, participated in a spoken word–picture matching task while event-related potentials (ERPs) were recorded. To determine whether the N400 effect elicited in this group deviated from normal, independent samples t tests and analysis of variance (ANOVA) analyses were used to compare the results of children with CIs against those of a similarly aged typically hearing (TH) group (n = 30). Correlational analyses were also conducted within each group to gauge the degree to which the N400 effect related to behavioral measures of spoken language.Results: An N400 effect was elicited in both groups of CI and TH children. The amplitude and latency of the N400 effect did not differ significantly between groups. Despite the similarity in ERP responses, children with CIs scored significantly lower on behavioral measures of spoken word- and sentence-level comprehension. No significant correlations between ERP and behavioral measures were found, although there was a trending relationship between sentence-level spoken language comprehension and the TH group’s N400 effect mean amplitude (p = .060).Conclusions: The results suggest that, at a neural level, children with CIs can process lexical–semantic incongruence, and that other underlying processes not measured by the N400 effect contribute to this population’s spoken language difficulties.

Influence of social conversational features on language identification in highly multilingual online conversations

With the explosion of multilingual content on Web, particularly in social media platforms, identification of languages present in the text is becoming an important task for various applications. While automatic language identification (ALI) in social media text is considered to be a non-trivial task due to the presence of slang words, misspellings, creative spellings and special elements such as hashtags, user mentions etc., ALI in multilingual environment becomes even more challenging task. In a highly multilingual society, code-mixing without affecting the underlying language sense has become a natural phenomenon. In such a dynamic environment, conversational text alone often fails to identify the underlying languages present in the text. This paper proposes various methods of exploiting social conversational features for enhancing ALI performance. Although social conversational features for ALI have been explored previously using methods like probabilistic language modeling, these models often fail to address issues related to code-mixing, phonetic typing, out-of-vocabulary etc. which are prevalent in a highly multilingual environment. This paper differs in the way the social conversational features are used to propose text refinement strategies that are suitable for ALI in highly multilingual environment. The contributions in this paper therefore includes the following. First, this paper analyzes the characteristics of various social conversational features by exploiting language usage patterns. Second, various methods of text refinement suitable for language identification are proposed. Third, the effects of the proposed refinement methods are investigated using various sentence level language identification frameworks. From various experimental observations over three conversational datasets collected from Facebook, Youtube and Twitter social media platforms, it is evident that our proposed method of ALI using social conversational features outperforms the baseline counterparts.

Oscillatory Brain Dynamics during Sentence Reading: A Fixation-Related Spectral Perturbation Analysis.

The present study investigated oscillatory brain dynamics during self-paced sentence-level processing. Participants read fully correct sentences, sentences containing a semantic violation and “sentences” in which the order of the words was randomized. At the target word level, fixations on semantically unrelated words elicited a lower-beta band (13–18 Hz) desynchronization. At the sentence level, gamma power (31–55 Hz) increased linearly for syntactically correct sentences, but not when the order of the words was randomized. In the 300–900 ms time window after sentence onsets, theta power (4–7 Hz) was greater for syntactically correct sentences as compared to sentences where no syntactic structure was preserved (random words condition). We interpret our results as conforming with a recently formulated predictive-coding framework for oscillatory neural dynamics during sentence-level language comprehension. Additionally, we discuss how our results relate to previous findings with serial visual presentation vs. self-paced reading.

A Predictive Coding Perspective on Beta Oscillations during Sentence-Level Language Comprehension.

Oscillatory neural dynamics have been steadily receiving more attention as a robust and temporally precise signature of network activity related to language processing. We have recently proposed that oscillatory dynamics in the beta and gamma frequency ranges measured during sentence-level comprehension might be best explained from a predictive coding perspective. Under our proposal we related beta oscillations to both the maintenance/change of the neural network configuration responsible for the construction and representation of sentence-level meaning, and to top–down predictions about upcoming linguistic input based on that sentence-level meaning. Here we zoom in on these particular aspects of our proposal, and discuss both old and new supporting evidence. Finally, we present some preliminary magnetoencephalography data from an experiment comparing Dutch subject- and object-relative clauses that was specifically designed to test our predictive coding framework. Initial results support the first of the two suggested roles for beta oscillations in sentence-level language comprehension.

A suggested model of corrective feedback provision

Following Guénette (2007), Bitchener (2008), Bitchener & Knoch (2009), among others, I believe that previous studies on corrective feedback provision were flawed in terms of their “design, execution, and analysis” (Bitchener & Knoch, 2009: 204). As a commentary on previous research findings, the current paper aims to suggest a corrective feedback provision model on how future studies should be designed, so that comparisons can be safely made. The suggested model underlies three basic premises. These are: (1) combining error correction with error feedback; (2) targeting one linguistic structure at a time; and (3) providing error correction on all the functional uses of the targeted structure. This approach has made it imperative that corrective feedback be factored out into error correction vis-à-vis error feedback. Whereas error correction targets sentence-level language corrections for local and mechanical errors such as improving grammar, spelling, and vocabulary, error feedback targets global issues that affect meaning and organization. Additionally, I suggest drawing a line of demarcation between two types of focused feedback: providing focused feedback selectively versus providing focused feedback comprehensively. The suggested model then calls for adopting relatively an all-inclusive approach to feedback provision, a model that, I believe, might be helpful in theory-building, and thus in bridging the gap between the theory of corrective feedback provision and actual classroom practices in some FL contexts.

Frequency-based Segregation of Syntactic and Semantic Unification during Online Sentence Level Language Comprehension.

During sentence level language comprehension, semantic and syntactic unification are functionally distinct operations. Nevertheless, both recruit roughly the same brain areas (spatially overlapping networks in the left frontotemporal cortex) and happen at the same time (in the first few hundred milliseconds after word onset). We tested the hypothesis that semantic and syntactic unification are segregated by means of neuronal synchronization of the functionally relevant networks in different frequency ranges: gamma (40 Hz and up) for semantic unification and lower beta (10-20 Hz) for syntactic unification. EEG power changes were quantified as participants read either correct sentences, syntactically correct though meaningless sentences (syntactic prose), or sentences that did not contain any syntactic structure (random word lists). Other sentences contained either a semantic anomaly or a syntactic violation at a critical word in the sentence. Larger EEG gamma-band power was observed for semantically coherent than for semantically anomalous sentences. Similarly, beta-band power was larger for syntactically correct sentences than for incorrect ones. These results confirm the existence of a functional dissociation in EEG oscillatory dynamics during sentence level language comprehension that is compatible with the notion of a frequency-based segregation of syntactic and semantic unification.

Advantages of sentence-level fMRI language tasks in presurgical language mapping for temporal lobe epilepsy

fMRI language mapping has become increasingly utilized for determining language dominance before surgical intervention for temporal lobe epilepsy (TLE). This study aimed to examine the differences between two classes of fMRI word generation tasks used in our clinic: tasks using a single word cue, referred to as simple generative tasks (SGTs), and tasks also involving sentence-level processing, referred to as sentence-level language tasks (SLTs). Specifically, we aimed to investigate the extent and laterality of activation and frontal–temporal connectivity during these language tasks and their relationship to clinical language measures. Thirty-one patients with TLE (18 patients with left TLE and 13 patients with right TLE) performed four language tasks during an fMRI scan, two SGTs and two SLTs. We found significantly greater activity for SLTs over SGTs in bilateral inferior frontal and middle temporal gyri and the left temporal pole. Sentence-level language tasks also showed greater lateralization compared with SGTs. Finally, we found that while activation extent did not correlate with clinical language tests, the degree of left frontal–temporal connectivity was significantly correlated with naming and semantic fluency performance. These correlations also were more robust for SLTs than for SGTs. Taken together, these results provide a compelling argument for including some form of SLTs in fMRI language lateralization protocols for TLE as they allow for better characterization of language networks, particularly in the temporal lobes which are at risk in surgery.

Functional MRI of sentence-level language comprehension in schizophrenia: A coordinate-based analysis

Numerous authors have hypothesised that abnormal pathways for language play a key role in the pathophysiology of schizophrenia, a notion that is supported by structural imaging and post-mortem findings especially in patients with thought disorder and auditory verbal hallucinations. Recently, an increasing number of functional magnetic resonance imaging (fMRI) investigations addressed language comprehension schizophrenia.We present a systematic review of the fMRI-studies on sentence- and text-level language comprehension in schizophrenia.13 studies met the inclusion criteria. Additional studies specifically addressed language lateralization. Coordinates for differential contrasts for healthy controls>patients reported in these studies indicate that the left fronto-temporal language network is altered in schizophrenia. 33 out of the 51 reported coordinates are located in the left hemisphere. Overactivation in schizophrenia extends into premotor areas and is about equally divided among the left and right hemispheres. Several negative studies indicate heterogeneity within schizophrenia, which could possibly be related to the severity of thought disorder or auditory verbal hallucinations of patients. Activation changes related to thought disorder within schizophrenia (n=4 studies) include the inferior frontal and superior temporal gyri and are moderately lateralized to the left hemisphere.Although current fMRI literature is still insufficient to draw decisive conclusions, results point towards functionally altered pathways for language in schizophrenia. This notion is also plausible from the viewpoint of psychopathology especially since hallmark symptoms of the disease, thought disorder, auditory verbal hallucinations and alogia, are expressed in terms of language or represent abnormalities of language function.

Emergent latent symbol systems in recurrent neural networks

Fodor and Pylyshyn [(1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1–2), 3–71] famously argued that neural networks cannot behave systematically short of implementing a combinatorial symbol system. A recent response from Frank et al. [(2009). Connectionist semantic systematicity. Cognition, 110(3), 358–379] claimed to have trained a neural network to behave systematically without implementing a symbol system and without any in-built predisposition towards combinatorial representations. We believe systems like theirs may in fact implement a symbol system on a deeper and more interesting level: one where the symbols are latent – not visible at the level of network structure. In order to illustrate this possibility, we demonstrate our own recurrent neural network that learns to understand sentence-level language in terms of a scene. We demonstrate our model's learned understanding by testing it on novel sentences and scenes. By paring down our model into an architecturally minimal version, we demonstrate how it supports combinatorial computation over distributed representations by using the associative memory operations of Vector Symbolic Architectures. Knowledge of the model's memory scheme gives us tools to explain its errors and construct superior future models. We show how the model designs and manipulates a latent symbol system in which the combinatorial symbols are patterns of activation distributed across the layers of a neural network, instantiating a hybrid of classical symbolic and connectionist representations that combines advantages of both.

Beta oscillations relate to the N400m during language comprehension

The relationship between the evoked responses (ERPs/ERFs) and the event-related changes in EEG/MEG power that can be observed during sentence-level language comprehension is as yet unclear. This study addresses a possible relationship between MEG power changes and the N400m component of the event-related field. Whole-head MEG was recorded while subjects listened to spoken sentences with incongruent (IC) or congruent (C) sentence endings. A clear N400m was observed over the left hemisphere, and was larger for the IC sentences than for the C sentences. A time-frequency analysis of power revealed a decrease in alpha and beta power over the left hemisphere in roughly the same time range as the N400m for the IC relative to the C condition. A linear regression analysis revealed a positive linear relationship between N400m and beta power for the IC condition, not for the C condition. No such linear relation was found between N400m and alpha power for either condition. The sources of the beta decrease were estimated in the LIFG, a region known to be involved in semantic unification operations. One source of the N400m was estimated in the left superior temporal region, which has been related to lexical retrieval. We interpret our data within a framework in which beta oscillations are inversely related to the engagement of task-relevant brain networks. The source reconstructions of the beta power suppression and the N400m effect support the notion of a dynamic communication between the LIFG and the left superior temporal region during language comprehension.

Integration or Predictability? A Further Specification of the Functional Role of Gamma Oscillations in Language Comprehension

Gamma-band neuronal synchronization during sentence-level language comprehension has previously been linked with semantic unification. Here, we attempt to further narrow down the functional significance of gamma during language comprehension, by distinguishing between two aspects of semantic unification: successful integration of word meaning into the sentence context, and prediction of upcoming words. We computed event-related potentials (ERPs) and frequency band-specific electroencephalographic (EEG) power changes while participants read sentences that contained a critical word (CW) that was (1) both semantically congruent and predictable (high cloze, HC), (2) semantically congruent but unpredictable (low cloze, LC), or (3) semantically incongruent (and therefore also unpredictable; semantic violation, SV). The ERP analysis showed the expected parametric N400 modulation (HC < LC < SV). The time-frequency analysis showed qualitatively different results. In the gamma-frequency range, we observed a power increase in response to the CW in the HC condition, but not in the LC and the SV conditions. Additionally, in the theta frequency range we observed a power increase in the SV condition only. Our data provide evidence that gamma power increases are related to the predictability of an upcoming word based on the preceding sentence context, rather than to the integration of the incoming word’s semantics into the preceding context. Further, our theta band data are compatible with the notion that theta band synchronization in sentence comprehension might be related to the detection of an error in the language input.