One of the persistent challenges in designing neural network models, particularly in the field of natural language processing, is managing the large number of learnable parameters, which often leads to increased computational costs and a greater risk of overfitting. In this paper, we propose a novel approach to address this issue by leveraging Rational Activation Functions, including both previously proposed variants and a new class of non-linear RAFs with learnable parameters. These functions are designed to enhance the representational power of neural networks while significantly reducing their architectural complexity. We demonstrate that RAFs can effectively emulate the behavior of deeper feed-forward neural networks, thereby enabling a reduction in the number of hidden layers required for a given task. This reduction translates into a lower total number of learnable parameters without compromising performance. Our method is evaluated on two fronts: first, through a series of function approximation experiments that highlight the expressiveness of RAFs; and second, on real-world NLP tasks involving text classification in noisy, spoken dialogue environments and an image classification task. The results confirm that networks incorporating RAFs maintain high accuracy while benefiting from increased efficiency and robustness. This work suggests that RAF-based architectures offer a promising direction for building lightweight, high-performance models, particularly in resource-constrained or real-time NLP applications.

Convolutional Neural Networks (CNN) have been used in Automatic Speech Recognition (ASR) to learn representations directly from the raw signal instead of hand-crafted acoustic features, providing a richer and lossless input signal. Recent researches propose to inject prior acoustic knowledge to the first convolutional layer by integrating the shape of the impulse responses in order to increase both the interpretability of the learnt acoustic model, and its performances. We propose to combine the complex Gabor filter with complex-valued deep neural networks to replace usual CNN weights kernels, to fully take advantage of its optimal time-frequency resolution and of the complex domain. The conducted experiments on the TIMIT phoneme recognition task shows that the proposed approach reaches top-of-the-line performances while remaining interpretable.

Modern end-to-end (E2E) Automatic Speech Recognition (ASR) systems rely on Deep Neural Networks (DNN) that are mostly trained on handcrafted and pre-computed acoustic features such as Mel-filter-banks or Mel-frequency cepstral coefficients. Nonetheless, and despite worse performances, E2E ASR models processing raw waveforms are an active research field due to the lossless nature of the input signal. In this paper, we propose the E2E-SincNet, a novel fully E2E ASR model that goes from the raw waveform to the text transcripts by merging two recent and powerful paradigms: SincNet and the joint CTC-attention training scheme. The conducted experiments on two different speech recognition tasks show that our approach outperforms previously investigated E2E systems relying either on the raw waveform or pre-computed acoustic features, with a reported top-of-the-line Word Error Rate (WER) of 4.7% on the Wall Street Journal (WSJ) dataset.

International audience

Machine learning (ML) and deep learning with deep neural networks (DNN), have drastically improved the performances of modern systems on numerous spoken language understanding (SLU) related tasks. Since most of current researches focus on new neural architectures to enhance the performances in realistic conditions, few recent works investigated the use of different algebras with neural networks (NN), to better represent the nature of the data being processed. To this extent, quaternion-valued neural networks (QNN) have shown better performances, and an important reduction of the number of neural parameters compared to traditional real-valued neural networks, when dealing with multidimensional signal. Nonetheless, the use of QNNs is strictly limited to quaternion input or output features. This article introduces a new unsupervised method based on a hybrid autoencoder (AE) called real-to-quaternion autoencoder (R2H), to extract a quaternion-valued input signal from any real-valued data, to be processed by QNNs. The experiments performed to identify the most related theme of a given telephone conversation from a customer care service (CCS), demonstrate that the R2H approach outperforms all the previously established models, either real- or quaternion-valued ones, in term of accuracy and with up to four times fewer neural parameters.

Convolutional neural networks (CNN) have recently achieved state-of-the-art results in various applications. In the case of image recognition, an ideal model has to learn independently of the training data, both local dependencies between the three components (R,G,B) of a pixel, and the global relations describing edges or shapes, making it efficient with small or heterogeneous datasets. Quaternion-valued convolutional neural networks (QCNN) solved this problematic by introducing multidimensional algebra to CNN. This paper proposes to explore the fundamental reason of the success of QCNN over CNN, by investigating the impact of the Hamilton product on a color image reconstruction task performed from a gray-scale only training. By learning independently both internal and external relations and with less parameters than real valued convolutional encoder-decoder (CAE), quaternion convolutional encoder-decoders (QCAE) perfectly reconstructed unseen color images while CAE produced worst and gray-scale versions.

Recurrent neural networks (RNN) are at the core of modern automatic speech recognition (ASR) systems. In particular, long short-term memory (LSTM) recurrent neural networks have achieved state-of-the-art results in many speech recognition tasks, due to their efficient representation of long and short term dependencies in sequences of inter-dependent features. Nonetheless, internal dependencies within the element composing multidimensional features are weakly considered by traditional real-valued representations. We propose a novel quaternion long short-term memory (QL-STM) recurrent neural network that takes into account both the external relations between the features composing a sequence, and these internal latent structural dependencies with the quaternion algebra. QLSTMs are compared to LSTMs during a memory copy-task and a realistic application of speech recognition on the Wall Street Journal (WSJ) dataset. QLSTM reaches better performances during the two experiments with up to 2.8 times less learning parameters, leading to a more expressive representation of the information.

International audience

Deep neural networks (DNNs) and more precisely recurrent neural networks (RNNs) are at the core of modern automatic speech recognition systems, due to their efficiency to process input sequences. Recently, it has been shown that different input representations, based on multidimensional algebras, such as complex and quaternion numbers, are able to bring to neural networks a more natural, compressive and powerful representation of the input signal by outperforming common real-valued NNs. Indeed, quaternion-valued neural networks (QNNs) better learn both internal dependencies, such as the relation between the Mel-filter-bank value of a specific time frame and its time derivatives, and global dependencies, describing the relations that exist between time frames. Nonetheless, QNNs are limited to quaternion-valued input signals, and it is difficult to benefit from this powerful representation with real-valued input data. This paper proposes to tackle this weakness by introducing a real-to-quaternion encoder that allows QNNs to process any one dimensional input features, such as traditional Mel-filter-banks for automatic speech recognition.

Recently, the connectionist temporal classification (CTC) model coupled with recurrent (RNN) or convolutional neural networks (CNN), made it easier to train speech recognition systems in an end-to-end fashion. However in real-valued models, time frame components such as mel-filter-bank energies and the cepstral coefficients obtained from them, together with their first and second order derivatives, are processed as individual elements, while a natural alternative is to process such components as composed entities. We propose to group such elements in the form of quaternions and to process these quaternions using the established quaternion algebra. Quaternion numbers and quaternion neural networks have shown their efficiency to process multidimensional inputs as entities, to encode internal dependencies, and to solve many tasks with less learning parameters than real-valued models. This paper proposes to integrate multiple feature views in quaternion-valued convolutional neural network (QCNN), to be used for sequence-to-sequence mapping with the CTC model. Promising results are reported using simple QCNNs in phoneme recognition experiments with the TIMIT corpus. More precisely, QCNNs obtain a lower phoneme error rate (PER) with less learning parameters than a competing model based on real-valued CNNs.

Quaternion convolutional neural networks (QCNN) are powerful architectures to learn and model external dependencies that exist between neighbor features of an input vector, and internal latent dependencies within the feature. This paper proposes to evaluate the effectiveness of the QCNN on a realistic theme identification task of spoken telephone conversations between agents and customers from the call center of the Paris transportation system (RATP). We show that QCNNs are more suitable than real-valued CNN to process multidimensional data and to code internal dependencies. Indeed, real-valued CNNs deal with both internal and external relations at the same level since components of an entity are processed independently. Experimental evidence is provided that the proposed QCNN architecture always outperforms real-valued equivalent CNN models in the theme identification task of the DECODA corpus. It is also shown that QCNN accuracy results are the best achieved so far on this task, while reducing by a factor of 4 the number of model parameters.

Deep Neural Networks (DNN) received a great interest from researchers due to their capability to construct robust abstract representations of heterogeneous documents in a latent subspace. Nonetheless, mere real-valued deep neural networks require an appropriate adaptation, such as the convolution process, to capture latent relations between input features. Moreover, real-valued deep neural networks reveal little in way of document internal dependencies, by only considering words or topics contained in the document as an isolate basic element. Quaternion-valued multi-layer perceptrons (QMLP), and autoencoders (QAE) have been introduced to capture such latent dependencies, alongside to represent multidimensional data. Nonetheless, a three-layered neural network does not benefit from the high abstraction capability of DNNs. The paper proposes first to extend the hyper-complex algebra to deep neural networks (QDNN) and, then, introduces pre-trained deep quaternion neural networks (QDNN-AE) with dedicated quaternion encoder-decoders (QAE). The experiments conduced on a theme identification task of spoken dialogues from the DECODA data set show, inter alia, that the QDNN-AE reaches a promising gain of 2.2% compared to the standard real-valued DNN-AE.

International audience

Automatic transcription of spoken documents is affected by automatic transcription errors that are especially frequent when speech is acquired in severe noisy conditions. Automatic speech recognition errors induce errors in the linguistic features used for a variety of natural language processing tasks. Recently, denoisng autoencoders (DAE) and stacked autoencoders (SAE) have been proposed with interesting results for acoustic feature denoising tasks. This paper deals with the recovery of corrupted linguistic features in spoken documents. Solutions based on DAEs and SAEs are considered and evaluated in a spoken conversation analysis task. In order to improve conversation theme classification accuracy, the possibility of combining abstractions obtained from manual and automatic transcription features is considered. As a result, two original representations of highly imperfect spoken documents are introduced. They are based on bottleneck features of a supervised autoencoder that takes advantage of both noisy and clean transcriptions to improve the robustness of error prone representations. Experimental results on a spoken conversation theme identification task show substantial accuracy improvements obtained with the proposed recovery of corrupted features.

International audience

International audience

Most of the Natural and Spoken Language Processing tasks now employ Neural Networks (NN), allowing them to reach impressive performances. Embedding features allow the NLP systems to represent input vectors in a latent space and to improve the observed performances. In this context, Recurrent Neural Network (RNN) based architectures such as Long Short-Term Memory (LSTM) are well known for their capacity to encode sequential data into a non-sequential hidden vector representation, called sequence embedding. In this paper, we propose an LSTM-based multi-stream sequence embedding in order to encode parallel sequences by a single non-sequential latent representation vector. We then propose to map this embedding representation in a high-dimensional space using a Support Vector Machine (SVM) in order to classify the multi-stream sequences by finding out an optimal hyperplane. Multi-stream sequence embedding allowed the SVM classifier to more efficiently profit from information carried by both parallel streams and longer sequences. The system achieved the best performance, in a multi-stream sequence classification task, with a gain of 9 points in error rate compared to an SVM trained on the original input sequences.

Recently, machine learning methods have provided a broad spectrum of original and efficient algorithms based on Deep Neural Networks (DNN) to automatically predict an outcome with respect to a sequence of inputs. Recurrent hidden cells allow these DNN-based models to manage long-term dependencies such as Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM). Nevertheless, these RNNs process a single input stream in one (LSTM) or two (Bidirectional LSTM) directions. But most of the information available nowadays is from multistreams or multimedia documents, and require RNNs to process these information synchronously during the training. This paper presents an original LSTM-based architecture, named Parallel LSTM (PLSTM), that carries out multiple parallel synchronized input sequences in order to predict a common output. The proposed PLSTM method could be used for parallel sequence classification purposes. The PLSTM approach is evaluated on an automatic telecast genre sequences classification task and compared with different state-of-the-art architectures. Results show that the proposed PLSTM method outperforms the baseline n-gram models as well as the state-of-the-art LSTM approach.

Automatically translating textual documents from one language to another inevitably results in translation errors. In addition to language specificities, this automatic translation appears more difficult in the context of spoken dialogues since, for example, the language register is far from “clean speech”. Speech analytics suffer from these translation errors. To tackle this difficulty, a solution consists in mapping translations into a space of hidden topics. In the classical topic-based representation obtained from a Latent Dirichlet Allocation (LDA), distribution of words into each topic is estimated automatically. Nonetheless, the targeted classes are ignored in the particular context of a classification task. In the DSTC5 main task, this targeted class information is crucial, the main objective being to track dialog states for sub-dialog segments. For this challenge, we propose to apply an original topic-based representation for each sub-dialogue based not only on the sub-dialogue content itself (words), but also on the dialogue state related to the sub-dialogue. This original representation is based on the Author-Topic (AT) model, previously successfully applied on a different classification task. Promising results confirmed the interest of such a method, the AT model reaching performance slightly better in terms of F-measure than baseline ones given by the task's organizers.

This paper proposes an original method which integrates contextual information of words into Word2vec neural networks that learn from words and their respective context windows. In the classical word embedding approach, context windows are represented as bag-of-words, i.e. every word in the context is treated equally. A log-linear weighting approach modeling the continuous context is proposed in our model to take into account the relative position of words in the surrounding context of the word. Quality improvements implied by this method are shown on the the Semantic-Syntactic Word Relationship test and on a real application framework implying a theme identification task of human dialogues. The promising gains of our adapted Word2vec model of 7 and 5 points for Skip-gram and CBOW approaches respectively demonstrate that the proposed models are a step forward for word and document representation.

Machine Learning (ML) techniques have allowed a great performance improvement of different challenging Spoken Language Understanding (SLU) tasks. Among these methods, Neural Networks (NN), or Multilayer Perceptron (MLP), recently received a great interest from researchers due to their representation capability of complex internal structures in a low dimensional subspace. However, MLPs employ document representations based on basic word level or topic-based features. Therefore, these basic representations reveal little in way of document statistical structure by only considering words or topics contained in the document as a “bag-of-words”, ignoring relations between them. We propose to remedy this weakness by extending the complex features based on Quaternion algebra presented in [1] to neural networks called QMLP. This original QMLP approach is based on hyper-complex algebra to take into consideration features dependencies in documents. New document features, based on the document structure itself, used as input of the QMLP, are also investigated in this paper, in comparison to those initially proposed in [1]. Experiments made on a SLU task from a real framework of human spoken dialogues showed that our QMLP approach associated with the proposed document features outperforms other approaches, with an accuracy gain of 2% with respect to the MLP based on real numbers and more than 3% with respect to the first Quaternion-based features proposed in [1]. We finally demonstrated that less iterations are needed by our QMLP architecture to be efficient and to reach promising accuracies.

This paper presents a method for speech analytics that integrates topic-space based representation into a feed-forward artificial neural network (FFANN), working as a document classifier. The proposed method consists in configuring the FFANN's topology and in initializing the weights according to a previously estimated topic-space. Setup based on thematic priors is expected to improve the efficiency of the FFANN's weight optimization process, while speeding-up the training process and improving the classification accuracy. This method is evaluated on a spoken dialogue categorization task which is composed of customer-agent dialogues from the call-centre of Paris Public Transportation Company. Results show the interest of the proposed setup method, with a gain of more than 4 points in terms of classification accuracy, compared to the baseline. Moreover, experiments highlight that performance is weakly dependent to FFANN's topology with the LDA-based configuration, in comparison to classical empirical setup.

In the context of noisy environments, Automatic Speech Recognition (ASR) systems usually produce poor transcription quality which also negatively impact performance of speech analyt-ics. Various methods have then been proposed to compensate the bad effect of ASR errors, mainly by projecting transcribed words in an abstract space. In this paper, we seek to identify themes from dialogues of telephone conversation services using latent topic-spaces estimated from a latent Dirichlet allocation (LDA). As an outcome, a document can be represented with a vector containing probabilities to be associated to each topic estimated with LDA. This vector should nonetheless be normalized to condition document representations. We propose to compare the original LDA vector representation (without normalization) with two normalization approaches, the Eigen Factor Radial (EFR) and the Feature Warping (FW) methods, already successfully applied in speaker recognition field, but never compared and evaluated in the context of a speech analytic task. Results show the interest of these normalization techniques for theme identification tasks using automatic transcriptions The EFR normalization approach allows a gain of 3.67 and 3.06 points respectively in comparison to the absence of normalization and to the FW normalization technique.

Mapping text documents in an LDA-based topic-space is a classical way to extract high-level representation of text documents. Unfortunately, LDA is highly sensitive to hyper-parameters related to the number of classes, or word and topic distribution, and there is no systematic way to pre-estimate optimal configurations. Moreover, various hyper-parameter configurations offer complementary views on the document. In this paper, we propose a method based on a two-step process that, first, expands the representation space by using a set of topic spaces and, second, compacts the representation space by removing poorly relevant dimensions. These two steps are based respectively on multi-view LDA-based representation spaces and factor-analysis models. This model provides a view-independent representation of documents while extracting complementary information from a massive multi-view representation. Experiments are conducted on the DECODA conversation corpus and the Reuters-21578 textual dataset. Results show the efficiency of the proposed multiview compact representation paradigm. The proposed categorization system reaches an accuracy of 86.5% with automatic transcriptions of conversations from DECODA corpus and a Macro-F1 of 80% during a classification task of the well-known Reuters-21578 corpus, with a significant gain compared to the baseline (best single topic space configuration), as well as methods and document representations previously studied.

Mohamed Morchid, Mohamed Bouallegue, Richard Dufour, Georges Linarès, Driss Matrouf, Renato de Mori. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014.

In this paper, we study the impact of dialogue representations and classification methods in the task of theme identification of telephone conversation services having highly imperfect automatic transcriptions. Two dialogue representations are firstly compared: the classical Term Frequency-Inverse Document Frequency with Gini purity criteria (TF-IDF-Gini) method and the Latent Dirichlet Allocation (LDA) approach. We then propose to study an original classification method that takes advantage of the LDA topic space representation, highlighted as the best dialogue representation. To do so, two assumptions about topic representation led us to choose a Gaussian process (GP) based method. This approach is compared with a Support Vector Machine (SVM) classification method. Results show that the GP approach is a better solution to deal with the multiple theme complexity of a dialogue, no matter the conditions studied (manual or automatic transcriptions). We finally discuss the impact of the topic space reduction on the classification accuracy.

International audience

International audience

The performance of Automatic Speech Recognition (ASR) systems drops dramatically when used in noisy environments. Speech analytics suffer from this poor quality of automatic transcriptions. In this paper, we seek to identify themes from dialogues of telephone conversation services using multiple topic-spaces estimated with a Latent Dirichlet Allocation (LDA) approach. This technique consists in estimating several topic models that offer different views of the document. Unfortunately, such a multi-model approach also introduces additional vari-abilities due to the model diversity. We propose to extract the useful information from the full model-set by using an i-vector based approach, previously developed in the context of speaker recognition. Experiments are conducted on the DECODA corpus , that contains records from the call center of the Paris Transportation Company. Results show the effectiveness of the proposed representation paradigm, our identification system reaching an accuracy of 84.7%, with a gain of 3.3 points compared to the baseline.

International audience

Performance of Automatic Speech Recognition (ASR) systems drops dramatically when transcribing conversations recorded in noisy conditions. Speech analytics suffer from this poor automatic transcription quality. To tackle this difficulty, a solution consists in mapping transcriptions into a space of hidden topics. This abstract representation allows to substantiate the drawbacks of the ASR process. The well-known and commonly used one is the topic-based representation from a Latent Dirichlet Allocation (LDA). Several studies demonstrate the effectiveness and reliability of this high-level representation. During the LDA learning process, distribution of words into each topic is estimated automatically. Nonetheless, in the context of a classification task, no consideration is made for the targeted classes. Thus, if the targeted application is to find out the main theme related to a dialogue, this information should be taken into consideration. In this paper, we propose to compare a classical topic-based representation of a dialogue, with a new one based not only on the dialogue content itself (words), but also on the theme related to the dialogue. This original representation is based on the author-topic (AT) model. The effectiveness of the proposed representation is evaluated on a classification task from automatic dialogue transcriptions between an agent and a customer of the Paris Transportation Company. Experiments confirmed that this author-topic model approach outperforms by far the classical topic representation, with a substantial gain of more than 7% in terms of correctly labeled conversations.

International audience