Automatic music transcription (AMT) is commonly defined as the process of converting an acoustic musical signal into some form of musical notation, and can be split into two separate phases: (1) multi-pitch detection, the conversion of an audio signal into a time-frequency representation similar to a MIDI file; and (2) converting from this time-frequency representation into a musical score. A substantial amount of AMT research in recent years has concentrated on multi-pitch detection, and yet, in the case of the transcription of polyphonic music, there has been little progress. There are many potential reasons for this slow progress, but this thesis concentrates on the (lack of) use of music language models during the transcription process. In particular, a music language model would impart to a transcription system the background knowledge of music theory upon which a human transcriber relies. In the related field of automatic speech recognition, it has been shown that the use of a language model drawn from the field of natural language processing (NLP) is an essential component of a system for transcribing spoken word into text, and there is no reason to believe that music should be any different. This thesis will show that a music language model inspired by NLP techniques can be used successfully for transcription. In fact, this thesis will create the blueprint for such a music language model. We begin with a brief overview of existing multi-pitch detection systems, in particular noting four key properties which any music language model should have to be useful for integration into a joint system for AMT: it should (1) be probabilistic, (2) not use any data a priori, (3) be able to run on live performance data, and (4) be incremental. We then investigate voice separation, creating a model which achieves state-of-the-art performance on the task, and show that, used as a simple music language model, it improves multi-pitch detection performance significantly. This is followed by an investigation of metrical detection and alignment, where we introduce a grammar crafted for the task which, combined with a beat-tracking model, achieves state-of-the-art results on metrical alignment. This system's success adds more evidence to the long-existing hypothesis that music and language consist of extremely similar structures. We end by investigating the joint analysis of music, in particular showing that a combination of our two models running jointly outperforms each running independently. We also introduce a new joint, automatic, quantitative metric for the complete transcription of an audio recording into an annotated musical score, something which the field currently lacks.

A statistical language model attempts to characterise the patterns present in a natural language as a probability distribution defined over word sequences. Typically, they are trained using word co-occurrence statistics from a large sample of text. In some language modelling applications, such as automatic speech recognition (ASR), the availability of acoustic data provides an additional source of knowledge. This contains, amongst other things, the melodic and rhythmic aspects of speech referred to as prosody. Although prosody has been found to be an important factor in human speech recognition, its use in ASR has been limited. The goal of this research is to investigate how prosodic information can be employed to improve the language modelling component of a continuous speech recognition system. Because prosodic features are largely suprasegmental, operating over units larger than the phonetic segment, the language model is an appropriate place to incorporate such information. The prosodic features and standard language model features are combined under the maximum entropy framework, which provides an elegant solution to modelling information obtained from multiple, differing knowledge sources. We derive features for the model based on perceptually transcribed Tones and Break Indices (ToBI) labels, and analyse their contribution to the word recognition task. While ToBI has a solid foundation in linguistic theory, the need for human transcribers conflicts with the statistical model's requirement for a large quantity of training data. We therefore also examine the applicability of features which can be automatically extracted from the speech signal. We develop representations of an utterance's prosodic context using fundamental frequency, energy and duration features, which can be directly incorporated into the model without the need for manual labelling. Dimensionality reduction techniques are also explored with the aim of reducing the computational costs associated with training a maximum entropy model. Experiments on a prosodically transcribed corpus show that small but statistically significant reductions to perplexity and word error rates can be obtained by using both manually transcribed and automatically extracted features.

With the increasing availability of large datasets machine learning techniques are becoming an increasingly attractive alternative to expert-designed approaches to solving complex problems in domains where data is abundant. In this thesis we introduce several models for large sparse discrete datasets. Our approach, which is based on probabilistic models that use distributed representations to alleviate the effects of data sparsity, is applied to statistical language modelling and collaborative filtering. We introduce three probabilistic language models that represent words using learned real-valued vectors. Two of the models are based on the Restricted Boltzmann Machine (RBM) architecture while the third one is a simple deterministic model. We show that the deterministic model outperforms the widely used n-gram models and learns sensible word representations. To reduce the time complexity of training and making predictions with the deterministic model, we introduce a hierarchical version of the model, that can be exponentially faster. The speedup is achieved by structuring the vocabulary as a tree over words and taking advantage of this structure. We propose a simple feature-based algorithm for automatic construction of trees over words from data and show that the resulting models can outperform non-hierarchical neural models as well as the best n-gram models. We then turn our attention to collaborative filtering and show how RBM models can be used to model the distribution of sparse high-dimensional user rating vectors efficiently, presenting inference and learning algorithms that scale linearly in the number of observed ratings. We also introduce the Probabilistic Matrix Factorization model which is based on the probabilistic formulation of the low-rank matrix approximation problem for partially observed matrices. The two models are then extended to allow conditioning on the identities of the rated items whether or not the actual rating values are known. Our results on the Netflix Prize dataset show that both RBM and PMF models outperform online SVD models.

With the increasing availability of large datasets machine learning techniques are becoming an increasingly attractive alternative to expert-designed approaches to solving complex problems in domains where data is abundant. In this thesis we introduce several models for large sparse discrete datasets. Our approach, which is based on probabilistic models that use distributed representations to alleviate the effects of data sparsity, is applied to statistical language modelling and collaborative filtering. We introduce three probabilistic language models that represent words using learned real-valued vectors. Two of the models are based on the Restricted Boltzmann Machine (RBM) architecture while the third one is a simple deterministic model. We show that the deterministic model outperforms the widely used n-gram models and learns sensible word representations. To reduce the time complexity of training and making predictions with the deterministic model, we introduce a hierarchical version of the model, that can be exponentially faster. The speedup is achieved by structuring the vocabulary as a tree over words and taking advantage of this structure. We propose a simple feature-based algorithm for automatic construction of trees over words from data and show that the resulting models can outperform non-hierarchical neural models as well as the best n-gram models. We then turn our attention to collaborative filtering and show how RBM models can be used to model the distribution of sparse high-dimensional user rating vectors efficiently, presenting inference and learning algorithms that scale linearly in the number of observed ratings. We also introduce the Probabilistic Matrix Factorization model which is based on the probabilistic formulation of the low-rank matrix approximation problem for partially observed matrices. The two models are then extended to allow conditioning on the identities of the rated items whether or not the actual rating values are known. Our results on the Netflix Prize dataset show that both RBM and PMF models outperform online SVD models.

The move towards larger vocabulary Automatic Speech Recognition (ASR) systems places greater demands on language models. In a large vocabulary system, acoustic confusion is greater, thus there is more reliance placed on the language model for disambiguation. In addition to this, ASR systems are increasingly being deployed in situations where the speaker is not conscious of their interaction with the system, such as in recorded meetings and surveillance scenarios. This results in more natural speech, which contains many false starts and disfluencies. In this thesis we investigate a novel approach to the modelling of speech corrections. We propose a syntactic model of speech corrections, and seek to determine if this model can improve on the performance of standard language modelling approaches when applied to conversational speech. We investigate a number of related variations to our basic approach and compare these approaches against the class-based N-gram. We also investigate the modelling of styles of speech. Specifically, we investigate whether the incorporation of prior knowledge about sentence types can improve the performance of language models. We propose a sentence mixture model based on word-class N-grams, in which the sentence mixture models and the word-class membership probabilities are jointly trained. We compare this approach with word-based sentence mixture models.

Le travail présenté dans cette thèse explore les méthodes pratiques utilisées pour faciliter l'entraînement et améliorer les performances des modèles de langues munis de très grands vocabulaires. La principale limite à l'utilisation des modèles de langue neuronaux est leur coût computationnel: il dépend de la taille du vocabulaire avec laquelle il grandit linéairement. La façon la plus aisée de réduire le temps de calcul de ces modèles reste de limiter la taille du vocabulaire, ce qui est loin d'être satisfaisant pour de nombreuses tâches. La plupart des méthodes existantes pour l'entraînement de ces modèles à grand vocabulaire évitent le calcul de la fonction de partition, qui est utilisée pour forcer la distribution de sortie du modèle à être normalisée en une distribution de probabilités. Ici, nous nous concentrons sur les méthodes à base d'échantillonnage, dont le sampling par importance et l'estimation contrastive bruitée. Ces méthodes permettent de calculer facilement une approximation de cette fonction de partition. L'examen des mécanismes de l'estimation contrastive bruitée nous permet de proposer des solutions qui vont considérablement faciliter l'entraînement, ce que nous montrons expérimentalement. Ensuite, nous utilisons la généralisation d'un ensemble d'objectifs basés sur l'échantillonnage comme divergences de Bregman pour expérimenter avec de nouvelles fonctions objectif. Enfin, nous exploitons les informations données par les unités sous-mots pour enrichir les représentations en sortie du modèle. Nous expérimentons avec différentes architectures, sur le Tchèque, et montrons que les représentations basées sur les caractères permettent l'amélioration des résultats, d'autant plus lorsque l'on réduit conjointement l'utilisation des représentations de mots. / This work investigates practical methods to ease training and improve performances of neural language models with large vocabularies. The main limitation of neural language models is their expensive computational cost: it depends on the size of the vocabulary, with which it grows linearly. Despite several training tricks, the most straightforward way to limit computation time is to limit the vocabulary size, which is not a satisfactory solution for numerous tasks. Most of the existing methods used to train large-vocabulary language models revolve around avoiding the computation of the partition function, ensuring that output scores are normalized into a probability distribution. Here, we focus on sampling-based approaches, including importance sampling and noise contrastive estimation. These methods allow an approximate computation of the partition function. After examining the mechanism of self-normalization in noise-contrastive estimation, we first propose to improve its efficiency with solutions that are adapted to the inner workings of the method and experimentally show that they considerably ease training. Our second contribution is to expand on a generalization of several sampling based objectives as Bregman divergences, in order to experiment with new objectives. We use Beta divergences to derive a set of objectives from which noise contrastive estimation is a particular case. Finally, we aim at improving performances on full vocabulary language models, by augmenting output words representation with subwords. We experiment on a Czech dataset and show that using character-based representations besides word embeddings for output representations gives better results. We also show that reducing the size of the output look-up table improves results even more.

The current surge of interest in search and comparison tasks in natural language processing has brought with it a focus on vector space approaches and vector space dimensionality reduction techniques. Presenting data as points in hyperspace provides opportunities to use a variety of welldeveloped tools pertinent to this representation. Dimensionality reduction allows data to be compressed and generalised. Eigen decomposition and related algorithms are one category of approaches to dimensionality reduction, providing a principled way to reduce data dimensionality that has time and again shown itself capable of enabling access to powerful generalisations in the data. Issues with the approach, however, include computational complexity and limitations on the size of dataset that can reasonably be processed in this way. Large datasets are a persistent feature of natural language processing tasks. This thesis focuses on two main questions. Firstly, in what ways can eigen decomposition and related techniques be extended to larger datasets? Secondly, this having been achieved, of what value is the resulting approach to information retrieval and to statistical language modelling at the ngram level? The applicability of eigen decomposition is shown to be extendable through the use of an extant algorithm; the Generalized Hebbian Algorithm (GHA), and the novel extension of this algorithm to paired data; the Asymmetric Generalized Hebbian Algorithm (AGHA). Several original extensions to the these algorithms are also presented, improving their applicability in various domains. The applicability of GHA to Latent Semantic Analysisstyle tasks is investigated. Finally, AGHA is used to investigate the value of singular value decomposition, an eigen decomposition variant, to ngram language modelling. A sizeable perplexity reduction is demonstrated.

Thesis (M.Sc.(Computer Science )) -- University of Limpopo, 2020 / Language modelling (LM) work for under-resourced languages that does not consider most linguistic information inherent in a language produces language models that in adequately represent the language, thereby leading to under-development of natural language processing tools and systems such as speech recognition systems. This study investigated the influence that the orthography (i.e., writing system) of a lan guage has on the quality and/or robustness of the language models created for the text of that language. The unique conjunctive and disjunctive writing systems of isiN debele (Ndebele) and Sepedi (Pedi) were studied. The text data from the LWAZI and NCHLT speech corpora were used to develop lan guage models. The LM techniques that were implemented included: word-based n gram LM, LM smoothing, LM linear interpolation, and higher-order n-gram LM. The toolkits used for development were: HTK LM, SRILM, and CMU-Cam SLM toolkits. From the findings of the study – found on text preparation, data pooling and sizing, higher n-gram models, and interpolation of models – it is concluded that the orthogra phy of the selected languages does have effect on the quality of the language models created for their text. The following recommendations are made as part of LM devel opment for the concerned languages. 1) Special preparation and normalisation of the text data before LM development – paying attention to within sentence text markers and annotation tags that may incorrectly form part of sentences, word sequences, and n-gram contexts. 2) Enable interpolation during training. 3) Develop pentagram and hexagram language models for Pedi texts, and trigrams and quadrigrams for Ndebele texts. 4) Investigate efficient smoothing method for the different languages, especially for different text sizes and different text domains / National Research Foundation (NRF) Telkom University of Limpopo

Bidirectional Encoder Representations from Transformers (BERT) is a recently proposed language representation model, designed to pre-train deep bidirectional representations, with the goal of extracting context-sensitive features from an input text [1]. One of the challenging problems in the field of Natural Language Processing is Conversational Machine Comprehension (CMC). Given a context passage, a conversational question and the conversational history, the system should predict the answer span of the question in the context passage. The main challenge in this task is how to effectively encode the conversational history into the prediction of the next answer. In this thesis work, we investigate the use of the BERT language model for the CMC task. We propose a new architecture, named BERT-CMC, using the BERT model as a base. This architecture includes a new module for encoding the conversational history, inspired by the Transformer-XL model [2]. This module serves the role of memory throughout the conversation. The proposed model is trained and evaluated on the Conversational Question Answering dataset (CoQA) [3]. Our hypothesis is that the BERT-CMC model will effectively learn the underlying context of the conversation, leading to better performance than the baseline model proposed for CoQA. Our results of evaluating the BERT-CMC on the CoQA dataset show that the model performs poorly (44.7% F1 score), comparing to the CoQA baseline model (66.2% F1 score). In the light of model explainability, we also perform a qualitative analysis of the model behavior in questions with various linguistic phenomena eg coreference, pragmatic reasoning. Additionally, we motivate the critical design choices made, by performing an ablation study of the effect of these choices on the model performance. The results suggest that fine tuning the BERT layers boost the model performance. Moreover, it is shown that increasing the number of extra layers on top of BERT leads to bigger capacity of the conversational memory. / Bidirectional Encoder Representations from Transformers (BERT) är en nyligen föreslagen språkrepresentationsmodell, utformad för att förträna djupa dubbelriktade representationer, med målet att extrahera kontextkänsliga särdrag från en inmatningstext [1]. Ett utmanande problem inom området naturligtspråkbehandling är konversationsförståelse (förkortat CMC). Givet en bakgrundstext, en fråga och konversationshistoriken ska systemet förutsäga vilken del av bakgrundstexten som utgör svaret på frågan. Den viktigaste utmaningen i denna uppgift är hur man effektivt kan kodifiera konversationshistoriken i förutsägelsen av nästa svar. I detta examensarbete undersöker vi användningen av BERT-språkmodellen för CMC-uppgiften. Vi föreslår en ny arkitektur med namnet BERT-CMC med BERT-modellen som bas. Denna arkitektur innehåller en ny modul för kodning av konversationshistoriken, inspirerad av Transformer-XL-modellen [2]. Den här modulen tjänar minnets roll under hela konversationen. Den föreslagna modellen tränas och utvärderas på en datamängd för samtalsfrågesvar (CoQA) [3]. Vår hypotes är att BERT-CMC-modellen effektivt kommer att lära sig det underliggande sammanhanget för konversationen, vilket leder till bättre resultat än basmodellen som har föreslagits för CoQA. Våra resultat av utvärdering av BERT-CMC på CoQA-datasetet visar att modellen fungerar dåligt (44.7% F1 resultat), jämfört med CoQAbasmodellen (66.2% F1 resultat). För att bättre kunna förklara modellen utför vi också en kvalitativ analys av modellbeteendet i frågor med olika språkliga fenomen, t.ex. koreferens, pragmatiska resonemang. Dessutom motiverar vi de kritiska designvalen som gjorts genom att utföra en ablationsstudie av effekten av dessa val på modellens prestanda. Resultaten tyder på att finjustering av BERT-lager ökar modellens prestanda. Dessutom visas att ökning av antalet extra lager ovanpå BERT leder till större konversationsminne.

This thesis investigates how the sub-structure of words can be accounted for in probabilistic models of language. Such models play an important role in natural language processing tasks such as translation or speech recognition, but often rely on the simplistic assumption that words are opaque symbols. This assumption does not fit morphologically complex language well, where words can have rich internal structure and sub-word elements are shared across distinct word forms. Our approach is to encode basic notions of morphology into the assumptions of three different types of language models, with the intention that leveraging shared sub-word structure can improve model performance and help overcome data sparsity that arises from morphological processes. In the context of n-gram language modelling, we formulate a new Bayesian model that relies on the decomposition of compound words to attain better smoothing, and we develop a new distributed language model that learns vector representations of morphemes and leverages them to link together morphologically related words. In both cases, we show that accounting for word sub-structure improves the models' intrinsic performance and provides benefits when applied to other tasks, including machine translation. We then shift the focus beyond the modelling of word sequences and consider models that automatically learn what the sub-word elements of a given language are, given an unannotated list of words. We formulate a novel model that can learn discontiguous morphemes in addition to the more conventional contiguous morphemes that most previous models are limited to. This approach is demonstrated on Semitic languages, and we find that modelling discontiguous sub-word structures leads to improvements in the task of segmenting words into their contiguous morphemes.