In the quest to give a formal compositional semantics to natural languages, semanticists have started turning their attention to phenomena that have been also considered as parts of pragmatics (e.g., discourse anaphora and presupposition projection). To account for these phenomena, the very kinds of meanings assigned to words and phrases are often revisited. To be more specific, in the prevalent paradigm of modeling natural language denotations using the simply-typed lambda calculus (higher-order logic) this means revisiting the types of denotations assigned to individual parts of speech. However, the lambda calculus also serves as a fundamental theory of computation, and in the study of computation, similar type shifts have been employed to give a meaning to side effects. Side effects in programming languages correspond to actions that go beyond the lexical scope of an expression (a thrown exception might propagate throughout a program, a variable modified at one point might later be read at an another) or even beyond the scope of the program itself (a program might interact with the outside world by e.g., printing documents, making sounds, operating robotic limbs...).
The IMPACT-es diachronic corpus of historical Spanish compiles over one hundred books --containing approximately 8 million words-- in addition to a complementary lexicon which links more than 10 thousand lemmas with attestations of the different variants found in the documents. This textual corpus and the accompanying lexicon have been released under an open license (Creative Commons by-nc-sa) in order to permit their intensive exploitation in linguistic research. Approximately 7% of the words in the corpus (a selection aimed at enhancing the coverage of the most frequent word forms) have been annotated with their lemma, part of speech, and modern equivalent. This paper describes the annotation criteria followed and the standards, based on the Text Encoding Initiative recommendations, used to the represent the texts in digital form. As an illustration of the possible synergies between diachronic textual resources and linguistic research, we describe the application of statistical machine translation techniques to infer probabilistic context-sensitive rules for the automatic modernisation of spelling. The automatic modernisation with this type of statistical methods leads to very low character error rates when the output is compared with the supervised modern version of the text.
We show that in language learning, contrary to received wisdom, keeping exceptional training instances in memory can be beneficial for generalization accuracy. We investigate this phenomenon empirically on a selection of benchmark natural language processing tasks: grapheme-to-phoneme conversion, part-of-speech tagging, prepositional-phrase attachment, and base noun phrase chunking. In a first series of experiments we combine memory-based learning with training set editing techniques, in which instances are edited based on their typicality and class prediction strength. Results show that editing exceptional instances (with low typicality or low class prediction strength) tends to harm generalization accuracy. In a second series of experiments we compare memory-based learning and decision-tree learning methods on the same selection of tasks, and find that decision-tree learning often performs worse than memory-based learning. Moreover, the decrease in performance can be linked to the degree of abstraction from exceptions (i.e., pruning or eagerness). We provide explanations for both results in terms of the properties of the natural language processing tasks and the learning algorithms.
Two classes of methods have been shown to be useful for resolving lexical ambiguity. The first relies on the presence of particular words within some distance of the ambiguous target word; the second uses the pattern of words and part-of-speech tags around the target word. These methods have complementary coverage: the former captures the lexical ``atmosphere'' (discourse topic, tense, etc.), while the latter captures local syntax. Yarowsky has exploited this complementarity by combining the two methods using decision lists. The idea is to pool the evidence provided by the component methods, and to then solve a target problem by applying the single strongest piece of evidence, whatever type it happens to be. This paper takes Yarowsky's work as a starting point, applying decision lists to the problem of context-sensitive spelling correction. Decision lists are found, by and large, to outperform either component method. However, it is found that further improvements can be obtained by taking into account not just the single strongest piece of evidence, but ALL the available evidence. A new hybrid method, based on Bayesian classifiers, is presented for doing this, and its performance improvements are demonstrated.
Model interpretability methods are often used to explain NLP model decisions on tasks such as text classification, where the output space is relatively small. However, when applied to language generation, where the output space often consists of tens of thousands of tokens, these methods are unable to provide informative explanations. Language models must consider various features to predict a token, such as its part of speech, number, tense, or semantics. Existing explanation methods conflate evidence for all these features into a single explanation, which is less interpretable for human understanding. To disentangle the different decisions in language modeling, we focus on explaining language models contrastively: we look for salient input tokens that explain why the model predicted one token instead of another. We demonstrate that contrastive explanations are quantifiably better than non-contrastive explanations in verifying major grammatical phenomena, and that they significantly improve contrastive model simulatability for human observers. We also identify groups of contrastive decisions where the model uses similar evidence, and we are able to characterize what input tokens models use during various language generation decisions.
Language models for historical states of language are becoming increasingly important to allow the optimal digitisation and analysis of old textual sources. Because these historical states are at the same time more complex to process and more scarce in the corpora available, specific efforts are necessary to train natural language processing (NLP) tools adapted to the data. In this paper, we present our efforts to develop NLP tools for Early Modern French (historical French from the 16$^\text{th}$ to the 18$^\text{th}$ centuries). We present the $\text{FreEM}_{\text{max}}$ corpus of Early Modern French and D'AlemBERT, a RoBERTa-based language model trained on $\text{FreEM}_{\text{max}}$. We evaluate the usefulness of D'AlemBERT by fine-tuning it on a part-of-speech tagging task, outperforming previous work on the test set. Importantly, we find evidence for the transfer learning capacity of the language model, since its performance on lesser-resourced time periods appears to have been boosted by the more resourced ones. We release D'AlemBERT and the open-sourced subpart of the $\text{FreEM}_{\text{max}}$ corpus.
Audio segmentation and sound event detection are crucial topics in machine listening that aim to detect acoustic classes and their respective boundaries. It is useful for audio-content analysis, speech recognition, audio-indexing, and music information retrieval. In recent years, most research articles adopt segmentation-by-classification. This technique divides audio into small frames and individually performs classification on these frames. In this paper, we present a novel approach called You Only Hear Once (YOHO), which is inspired by the YOLO algorithm popularly adopted in Computer Vision. We convert the detection of acoustic boundaries into a regression problem instead of frame-based classification. This is done by having separate output neurons to detect the presence of an audio class and predict its start and end points. YOHO obtained a higher F-measure and lower error rate than the state-of-the-art Convolutional Recurrent Neural Network on multiple datasets. As YOHO is purely a convolutional neural network and has no recurrent layers, it is faster during inference. In addition, as this approach is more end-to-end and predicts acoustic boundaries directly, it is significantly quicker during post-processing and smoothing.
The research direction of identifying acoustic bio-markers of respiratory diseases has received renewed interest following the onset of COVID-19 pandemic. In this paper, we design an approach to COVID-19 diagnostic using crowd-sourced multi-modal data. The data resource, consisting of acoustic signals like cough, breathing, and speech signals, along with the data of symptoms, are recorded using a web-application over a period of ten months. We investigate the use of statistical descriptors of simple time-frequency features for acoustic signals and binary features for the presence of symptoms. Unlike previous works, we primarily focus on the application of simple linear classifiers like logistic regression and support vector machines for acoustic data while decision tree models are employed on the symptoms data. We show that a multi-modal integration of acoustics and symptoms classifiers achieves an area-under-curve (AUC) of 92.40, a significant improvement over any individual modality. Several ablation experiments are also provided which highlight the acoustic and symptom dimensions that are important for the task of COVID-19 diagnostics.
Recently, Mixture of Experts (MoE) based Transformer has shown promising results in many domains. This is largely due to the following advantages of this architecture: firstly, MoE based Transformer can increase model capacity without computational cost increasing both at training and inference time. Besides, MoE based Transformer is a dynamic network which can adapt to the varying complexity of input instances in realworld applications. In this work, we explore the MoE based model for speech recognition, named SpeechMoE. To further control the sparsity of router activation and improve the diversity of gate values, we propose a sparsity L1 loss and a mean importance loss respectively. In addition, a new router architecture is used in SpeechMoE which can simultaneously utilize the information from a shared embedding network and the hierarchical representation of different MoE layers. Experimental results show that SpeechMoE can achieve lower character error rate (CER) with comparable computation cost than traditional static networks, providing 7.0%-23.0% relative CER improvements on four evaluation datasets.
Add to Chrome
Add to Firefox