Publications

Generating longer textual sequences when conditioned on the visual information is an interesting problem to explore. The challenge here proliferate over the standard vision conditioned sentence-level generation (e.g., image or video captioning) as it requires to produce a brief and coherent story describing the visual content. In this paper, we mask this Vision-to-Sequence as Graph-to-Sequence learning problem and approach it with the Transformer architecture. To be specific, we introduce Sparse Graph-to-Sequence Transformer (SGST) for encoding the graph and decoding a sequence. The encoder aims to directly encode graph-level semantics, while the decoder is used to generate longer sequences. Experiments conducted with the benchmark image paragraph dataset show that our proposed achieve 13.3% improvement on the CIDEr evaluation measure when comparing to the previous state-of-the-art approach.

• https://arxiv.org/abs/2007.06077

@inproceedings{mogadala2020sparse,
title = {Sparse Graph to Sequence Learning for Vision Conditioned Long Textual Sequence Generation},
author = {Aditya Mogadala and Marius Mosbach and Dietrich Klakow},
url = {https://arxiv.org/abs/2007.06077},
year = {2020},
date = {2020},
booktitle = {Bridge Between Perception and Reasoning: Graph Neural Networks & Beyond, Workshop at ICML},
abstract = {Generating longer textual sequences when conditioned on the visual information is an interesting problem to explore. The challenge here proliferate over the standard vision conditioned sentence-level generation (e.g., image or video captioning) as it requires to produce a brief and coherent story describing the visual content. In this paper, we mask this Vision-to-Sequence as Graph-to-Sequence learning problem and approach it with the Transformer architecture. To be specific, we introduce Sparse Graph-to-Sequence Transformer (SGST) for encoding the graph and decoding a sequence. The encoder aims to directly encode graph-level semantics, while the decoder is used to generate longer sequences. Experiments conducted with the benchmark image paragraph dataset show that our proposed achieve 13.3% improvement on the CIDEr evaluation measure when comparing to the previous state-of-the-art approach.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Transformer-based language models achieve high performance on various tasks, but we still lack understanding of the kind of linguistic knowledge they learn and rely on. We evaluate three models (BERT, RoBERTa, and ALBERT), testing their grammatical and semantic knowledge by sentence-level probing, diagnostic cases, and masked prediction tasks. We focus on relative clauses (in American English) as a complex phenomenon needing contextual information and antecedent identification to be resolved. Based on a naturalistic dataset, probing shows that all three models indeed capture linguistic knowledge about grammaticality, achieving high performance. Evaluation on diagnostic cases and masked prediction tasks considering fine-grained linguistic knowledge, however, shows pronounced model-specific weaknesses especially on semantic knowledge, strongly impacting models’ performance. Our results highlight the importance of (a) model comparison in evaluation task and (b) building up claims of model performance and the linguistic knowledge they capture beyond purely probing-based evaluations.

• https://aclanthology.org/2020.coling-main.67/

@inproceedings{Mosbach2020,
title = {A Closer Look at Linguistic Knowledge in Masked Language Models: The Case of Relative Clauses in American English},
author = {Marius Mosbach and Stefania Degaetano-Ortlieb and Marie-Pauline Krielke and Badr M. Abdullah and Dietrich Klakow},
url = {https://aclanthology.org/2020.coling-main.67/},
year = {2020},
date = {2020},
booktitle = {Proceedings of the 28th International Conference on Computational Linguistics},
pages = {771-787},
abstract = {Transformer-based language models achieve high performance on various tasks, but we still lack understanding of the kind of linguistic knowledge they learn and rely on. We evaluate three models (BERT, RoBERTa, and ALBERT), testing their grammatical and semantic knowledge by sentence-level probing, diagnostic cases, and masked prediction tasks. We focus on relative clauses (in American English) as a complex phenomenon needing contextual information and antecedent identification to be resolved. Based on a naturalistic dataset, probing shows that all three models indeed capture linguistic knowledge about grammaticality, achieving high performance. Evaluation on diagnostic cases and masked prediction tasks considering fine-grained linguistic knowledge, however, shows pronounced model-specific weaknesses especially on semantic knowledge, strongly impacting models’ performance. Our results highlight the importance of (a) model comparison in evaluation task and (b) building up claims of model performance and the linguistic knowledge they capture beyond purely probing-based evaluations.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

The lack of labeled training data has limited the development of natural language processing tools, such as named entity recognition, for many languages spoken in developing countries. Techniques such as distant and weak supervision can be used to create labeled data in a (semi-) automatic way.

Additionally, to alleviate some of the negative effects of the errors in automatic annotation, noise-handling methods can be integrated. Pretrained word embeddings are another key component of most neural named entity classifiers. With the advent of more complex contextual word embeddings, an interesting trade-off between model size and performance arises. While these techniques have been shown to work well in high-resource settings, we want to study how they perform in low-resource scenarios.

In this work, we perform named entity recognition for Hausa and Yorùbá, two languages that are widely spoken in several developing countries. We evaluate different embedding approaches and show that distant supervision can be successfully leveraged in a realistic low-resource scenario where it can more than double a classifier’s performance.

• 2003.08370     (0.23MB)
• https://arxiv.org/abs/2003.08370

@miscellaneous{Adelani2020,
title = {Distant Supervision and Noisy Label Learning for Low Resource Named Entity Recognition: A Study on Hausa and Yorùb{\'a}},
author = {David Adelani and Michael Hedderich and Dawei Zhu and Esther van Berg and Dietrich Klakow},
url = {https://arxiv.org/abs/2003.08370},
year = {2020},
date = {2020},
abstract = {The lack of labeled training data has limited the development of natural language processing tools, such as named entity recognition, for many languages spoken in developing countries. Techniques such as distant and weak supervision can be used to create labeled data in a (semi-) automatic way. Additionally, to alleviate some of the negative effects of the errors in automatic annotation, noise-handling methods can be integrated. Pretrained word embeddings are another key component of most neural named entity classifiers. With the advent of more complex contextual word embeddings, an interesting trade-off between model size and performance arises. While these techniques have been shown to work well in high-resource settings, we want to study how they perform in low-resource scenarios. In this work, we perform named entity recognition for Hausa and Yorùb{\'a}, two languages that are widely spoken in several developing countries. We evaluate different embedding approaches and show that distant supervision can be successfully leveraged in a realistic low-resource scenario where it can more than double a classifier's performance.},
pubstate = {published},
type = {miscellaneous}
}

Copy BibTeX to Clipboard

Multilingual transformer models like mBERT and XLM-RoBERTa have obtained great improvements for many NLP tasks on a variety of languages. However, recent works also showed that results from high-resource languages could not be easily transferred to realistic, low-resource scenarios. In this work, we study trends in performance for different amounts of available resources for the three African languages Hausa, isiXhosa and on both NER and topic classification. We show that in combination with transfer learning or distant supervision, these models can achieve with as little as 10 or 100 labeled sentences the same performance as baselines with much more supervised training data. However, we also find settings where this does not hold. Our discussions and additional experiments on assumptions such as time and hardware restrictions highlight challenges and opportunities in low-resource learning.

• https://www.aclweb.org/anthology/2020.emnlp-main.204
• https://doi.org/10.18653/v1/2020.emnlp-main.204

@inproceedings{hedderich-etal-2020-transfer,
title = {Transfer Learning and Distant Supervision for Multilingual Transformer Models: A Study on African Languages},
author = {Michael Hedderich and David Adelani and Dawei Zhu and Alabi Jesujoba and Markus Udia and Dietrich Klakow},
url = {https://www.aclweb.org/anthology/2020.emnlp-main.204},
doi = {https://doi.org/10.18653/v1/2020.emnlp-main.204},
year = {2020},
date = {2020},
booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
pages = {2580-2591},
publisher = {Association for Computational Linguistics},
abstract = {Multilingual transformer models like mBERT and XLM-RoBERTa have obtained great improvements for many NLP tasks on a variety of languages. However, recent works also showed that results from high-resource languages could not be easily transferred to realistic, low-resource scenarios. In this work, we study trends in performance for different amounts of available resources for the three African languages Hausa, isiXhosa and on both NER and topic classification. We show that in combination with transfer learning or distant supervision, these models can achieve with as little as 10 or 100 labeled sentences the same performance as baselines with much more supervised training data. However, we also find settings where this does not hold. Our discussions and additional experiments on assumptions such as time and hardware restrictions highlight challenges and opportunities in low-resource learning.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Fine-tuning pre-trained contextualized embedding models has become an integral part of the NLP pipeline. At the same time, probing has emerged as a way to investigate the linguistic knowledge captured by pre-trained models. Very little is, however, understood about how fine-tuning affects the representations of pre-trained models and thereby the linguistic knowledge they encode. This paper contributes towards closing this gap. We study three different pre-trained models: BERT, RoBERTa, and ALBERT, and investigate through sentence-level probing how fine-tuning affects their representations. We find that for some probing tasks fine-tuning leads to substantial changes in accuracy, possibly suggesting that fine-tuning introduces or even removes linguistic knowledge from a pre-trained model. These changes, however, vary greatly across different models, fine-tuning and probing tasks. Our analysis reveals that while fine-tuning indeed changes the representations of a pre-trained model and these changes are typically larger for higher layers, only in very few cases, fine-tuning has a positive effect on probing accuracy that is larger than just using the pre-trained model with a strong pooling method. Based on our findings, we argue that both positive and negative effects of fine-tuning on probing require a careful interpretation.

• https://www.aclweb.org/anthology/2020.findings-emnlp.227
• https://doi.org/10.18653/v1/2020.findings-emnlp.227

@inproceedings{mosbach-etal-2020-interplay-fine,
title = {On the Interplay Between Fine-tuning and Sentence-level Probing for Linguistic Knowledge in Pre-trained Transformers},
author = {Marius Mosbach and Anna Khokhlova and Michael Hedderich and Dietrich Klakow},
url = {https://www.aclweb.org/anthology/2020.findings-emnlp.227},
doi = {https://doi.org/10.18653/v1/2020.findings-emnlp.227},
year = {2020},
date = {2020},
booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2020},
pages = {2502-2516},
publisher = {Association for Computational Linguistics},
abstract = {Fine-tuning pre-trained contextualized embedding models has become an integral part of the NLP pipeline. At the same time, probing has emerged as a way to investigate the linguistic knowledge captured by pre-trained models. Very little is, however, understood about how fine-tuning affects the representations of pre-trained models and thereby the linguistic knowledge they encode. This paper contributes towards closing this gap. We study three different pre-trained models: BERT, RoBERTa, and ALBERT, and investigate through sentence-level probing how fine-tuning affects their representations. We find that for some probing tasks fine-tuning leads to substantial changes in accuracy, possibly suggesting that fine-tuning introduces or even removes linguistic knowledge from a pre-trained model. These changes, however, vary greatly across different models, fine-tuning and probing tasks. Our analysis reveals that while fine-tuning indeed changes the representations of a pre-trained model and these changes are typically larger for higher layers, only in very few cases, fine-tuning has a positive effect on probing accuracy that is larger than just using the pre-trained model with a strong pooling method. Based on our findings, we argue that both positive and negative effects of fine-tuning on probing require a careful interpretation.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Manual evaluation of individual results of natural language generation tasks is one of the bottlenecks. It is very time consuming and expensive if it is, for example, crowdsourced. In this work, we address this problem for the specific task of automatic image captioning. We automatically generate human-like judgements on grammatical correctness, image relevance and diversity of the captions obtained from a neural image caption generator. For this purpose, we use pool-based active learning with uncertainty sampling and represent the captions using fixed size vectors from Google’s Universal Sentence Encoder. In addition, we test common metrics, such as BLEU, ROUGE, METEOR, Levenshtein distance, and n-gram counts and report F1 score for the classifiers used under the active learning scheme for this task. To the best of our knowledge, our work is the first in this direction and promises to reduce time, cost, and human effort.

• https://link.springer.com/chapter/10.1007/978-3-030-31372-2_22

@inproceedings{Biswas2019,
title = {Automatic Judgement of Neural Network-Generated Image Captions},
author = {Rajarshi Biswas and Aditya Mogadala and Michael Barz and Daniel Sonntag and Dietrich Klakow},
url = {https://link.springer.com/chapter/10.1007/978-3-030-31372-2_22},
year = {2019},
date = {2019},
booktitle = {7th International Conference on Statistical Language and Speech Processing (SLSP2019)},
address = {Ljubljana, Slovenia},
abstract = {Manual evaluation of individual results of natural language generation tasks is one of the bottlenecks. It is very time consuming and expensive if it is, for example, crowdsourced. In this work, we address this problem for the specific task of automatic image captioning. We automatically generate human-like judgements on grammatical correctness, image relevance and diversity of the captions obtained from a neural image caption generator. For this purpose, we use pool-based active learning with uncertainty sampling and represent the captions using fixed size vectors from Google’s Universal Sentence Encoder. In addition, we test common metrics, such as BLEU, ROUGE, METEOR, Levenshtein distance, and n-gram counts and report F1 score for the classifiers used under the active learning scheme for this task. To the best of our knowledge, our work is the first in this direction and promises to reduce time, cost, and human effort.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

In low-resource settings, the performance of supervised labeling models can be improved with automatically annotated or distantly supervised data, which is cheap to create but often noisy. Previous works have shown that significant improvements can be reached by injecting information about the confusion between clean and noisy labels in this additional training data into the classifier training. However, for noise estimation, these approaches either do not take the input features (in our case word embeddings) into account, or they need to learn the noise modeling from scratch which can be difficult in a low-resource setting. We propose to cluster the training data using the input features and then compute different confusion matrices for each cluster. To the best of our knowledge, our approach is the first to leverage feature-dependent noise modeling with pre-initialized confusion matrices. We evaluate on low-resource named entity recognition settings in several languages, showing that our methods improve upon other confusion-matrix based methods by up to 9%.

• https://aclanthology.org/D19-1362/
• https://doi.org/10.18653/v1/D19-1362

@inproceedings{lange-etal-2019-feature,
title = {Feature-Dependent Confusion Matrices for Low-Resource NER Labeling with Noisy Labels},
author = {Lukas Lange and Michael Hedderich and Dietrich Klakow},
editor = {Kentaro Inui and Jing Jiang and Vincent Ng and Xiaojun Wan},
url = {https://aclanthology.org/D19-1362/},
doi = {https://doi.org/10.18653/v1/D19-1362},
year = {2019},
date = {2019},
booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
pages = {3552-3557},
publisher = {Association for Computational Linguistics},
address = {Hong Kong, China},
abstract = {In low-resource settings, the performance of supervised labeling models can be improved with automatically annotated or distantly supervised data, which is cheap to create but often noisy. Previous works have shown that significant improvements can be reached by injecting information about the confusion between clean and noisy labels in this additional training data into the classifier training. However, for noise estimation, these approaches either do not take the input features (in our case word embeddings) into account, or they need to learn the noise modeling from scratch which can be difficult in a low-resource setting. We propose to cluster the training data using the input features and then compute different confusion matrices for each cluster. To the best of our knowledge, our approach is the first to leverage feature-dependent noise modeling with pre-initialized confusion matrices. We evaluate on low-resource named entity recognition settings in several languages, showing that our methods improve upon other confusion-matrix based methods by up to 9%.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Languages may be differently distant from each other and their mutual intelligibility may be asymmetric. In this paper we introduce incom.py, a toolbox for calculating linguistic distances and asymmetries between related languages. incom.py allows linguist experts to quickly and easily perform statistical analyses and compare those with experimental results. We demonstrate the efficacy of incom.py in an incomprehension experiment on two Slavic languages: Bulgarian and Russian. Using incom.py we were able to validate three methods to measure linguistic distances and asymmetries: Levenshtein distance, word adaptation surprisal, and conditional entropy as predictors of success in a reading intercomprehension experiment.

• RANLP094     (0.63MB)
• https://aclanthology.org/R19-1094/
• https://doi.org/10.26615/978-954-452-056-4_094

@inproceedings{Mosbach2019,
title = {incom.py - A Toolbox for Calculating Linguistic Distances and Asymmetries between Related Languages},
author = {Marius Mosbach and Irina Stenger and Tania Avgustinova and Dietrich Klakow},
editor = {Galia Angelova and Ruslan Mitkov and Ivelina Nikolova and Irina Temnikova},
url = {https://aclanthology.org/R19-1094/},
doi = {https://doi.org/10.26615/978-954-452-056-4_094},
year = {2019},
date = {2019},
booktitle = {Proceedings of Recent Advances in Natural Language Processing, RANLP 2019, Varna, Bulgaria, 2-4 September 2019},
pages = {811-819},
address = {Varna, Bulgaria},
abstract = {Languages may be differently distant from each other and their mutual intelligibility may be asymmetric. In this paper we introduce incom.py, a toolbox for calculating linguistic distances and asymmetries between related languages. incom.py allows linguist experts to quickly and easily perform statistical analyses and compare those with experimental results. We demonstrate the efficacy of incom.py in an incomprehension experiment on two Slavic languages: Bulgarian and Russian. Using incom.py we were able to validate three methods to measure linguistic distances and asymmetries: Levenshtein distance, word adaptation surprisal, and conditional entropy as predictors of success in a reading intercomprehension experiment.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Recently, a weight-based attack on stochastic gradient descent inducing overfitting has been proposed. We show that the threat is broader: A task-independent permutation on the initial weights suffices to limit the achieved accuracy to for example 50% on the Fashion MNIST dataset from initially more than 90%. These findings are confirmed on MNIST and CIFAR. We formally confirm that the attack succeeds with high likelihood and does not depend on the data. Empirically, weight statistics and loss appear unsuspicious, making it hard to detect the attack if the user is not aware. Our paper is thus a call for action to acknowledge the importance of the initial weights in deep learning.

• https://arxiv.org/abs/1902.03020

@article{Grosse2019,
title = {On the security relevance of weights in deep learning},
author = {Kathrin Grosse and Thomas Trost and Marius Mosbach and Michael Backes and Dietrich Klakow},
url = {https://arxiv.org/abs/1902.03020},
year = {2019},
date = {2019},
journal = {CoRR},
abstract = {Recently, a weight-based attack on stochastic gradient descent inducing overfitting has been proposed. We show that the threat is broader: A task-independent permutation on the initial weights suffices to limit the achieved accuracy to for example 50% on the Fashion MNIST dataset from initially more than 90%. These findings are confirmed on MNIST and CIFAR. We formally confirm that the attack succeeds with high likelihood and does not depend on the data. Empirically, weight statistics and loss appear unsuspicious, making it hard to detect the attack if the user is not aware. Our paper is thus a call for action to acknowledge the importance of the initial weights in deep learning.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

For many real-word applications, the considered data is given as a time sequence that becomes available in an orderly fashion, where the order incorporates important information about the entities of interest. The work presented in this thesis deals with two such cases by introducing new sequential estimation solutions. More precisely, we introduce a: I. Sequential Bayesian estimation framework to solve the multiple speaker localization, detection and tracking problem. This framework is a complete pipeline that includes 1) new observation estimators, which extract a fixed number of potential locations per time frame; 2) new unsupervised Bayesian detectors, which classify these estimates into noise/speaker classes and 3) new Bayesian filters, which use the speaker class estimates to track multiple speakers.

This framework was developed to tackle the low overlap detection rate of multiple speakers and to reduce the number of constraints generally imposed in standard solutions. II. Sequential neural estimation framework for language modeling, which overcomes some of the shortcomings of standard approaches through merging of different models in a hybrid architecture. That is, we introduce two solutions that tightly merge particular models and then show how a generalization can be achieved through a new mixture model. In order to speed-up the training of large vocabulary language models, we introduce a new extension of the noise contrastive estimation approach to batch training.

• http://nbn-resolving.de/urn:nbn:de:bsz:291-scidok-ds-272280
• https://doi.org/http://dx.doi.org/10.22028/D291-27228

@phdthesis{Oualil2017b,
title = {Sequential estimation techniques and application to multiple speaker tracking and language modeling},
author = {Youssef Oualil},
url = {http://nbn-resolving.de/urn:nbn:de:bsz:291-scidok-ds-272280},
doi = {https://doi.org/http://dx.doi.org/10.22028/D291-27228},
year = {2017},
date = {2017},
school = {Saarland University},
address = {Saarbruecken, Germany},
abstract = {For many real-word applications, the considered data is given as a time sequence that becomes available in an orderly fashion, where the order incorporates important information about the entities of interest. The work presented in this thesis deals with two such cases by introducing new sequential estimation solutions. More precisely, we introduce a: I. Sequential Bayesian estimation framework to solve the multiple speaker localization, detection and tracking problem. This framework is a complete pipeline that includes 1) new observation estimators, which extract a fixed number of potential locations per time frame; 2) new unsupervised Bayesian detectors, which classify these estimates into noise/speaker classes and 3) new Bayesian filters, which use the speaker class estimates to track multiple speakers. This framework was developed to tackle the low overlap detection rate of multiple speakers and to reduce the number of constraints generally imposed in standard solutions. II. Sequential neural estimation framework for language modeling, which overcomes some of the shortcomings of standard approaches through merging of different models in a hybrid architecture. That is, we introduce two solutions that tightly merge particular models and then show how a generalization can be achieved through a new mixture model. In order to speed-up the training of large vocabulary language models, we introduce a new extension of the noise contrastive estimation approach to batch training.},
pubstate = {published},
type = {phdthesis}
}

Copy BibTeX to Clipboard

The performance of Neural Network (NN)-based language models is steadily improving due to the emergence of new architectures, which are able to learn different natural language characteristics. This paper presents a novel framework, which shows that a significant improvement can be achieved by combining different existing heterogeneous models in a single architecture. This is done through 1) a feature layer, which separately learns different NN-based models and 2) a mixture layer, which merges the resulting model features. In doing so, this architecture benefits from the learning capabilities of each model with no noticeable increase in the number of model parameters or the training time. Extensive experiments conducted on the Penn Treebank (PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a significant reduction of the perplexity when compared to state-of-the-art feedforward as well as recurrent neural network architectures.

• https://arxiv.org/abs/1708.06989

@inproceedings{Oualil2017b,
title = {A neural network approach for mixing language models},
author = {Youssef Oualil and Dietrich Klakow},
url = {https://arxiv.org/abs/1708.06989},
year = {2017},
date = {2017},
publisher = {ICASSP 2017},
abstract = {The performance of Neural Network (NN)-based language models is steadily improving due to the emergence of new architectures, which are able to learn different natural language characteristics. This paper presents a novel framework, which shows that a significant improvement can be achieved by combining different existing heterogeneous models in a single architecture. This is done through 1) a feature layer, which separately learns different NN-based models and 2) a mixture layer, which merges the resulting model features. In doing so, this architecture benefits from the learning capabilities of each model with no noticeable increase in the number of model parameters or the training time. Extensive experiments conducted on the Penn Treebank (PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a significant reduction of the perplexity when compared to state-of-the-art feedforward as well as recurrent neural network architectures.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Traditionally, short-range Language Models (LMs) like the conventional n-gram models have been used for language model adaptation. Recent work has improved performance for such tasks using adapted long-span models like Recurrent Neural Network LMs (RNNLMs). With the first pass performed using a large background n-gram LM, the adapted RNNLMs are mostly used to rescore lattices or N-best lists, as a second step in the decoding process. Ideally, these adapted RNNLMs should be applied for first-pass decoding. Thus, we introduce two ways of applying adapted long-short-term-memory (LSTM) based RNNLMs for first-pass decoding. Using available techniques to convert LSTMs to approximated versions for first-pass decoding, we compare approximated LSTMs adapted in a Fast Marginal Adaptation framework (FMA) and an approximated version of architecture-based-adaptation of LSTM. On a conversational speech recognition task, these differently approximated and adapted LSTMs combined with a trigram LM outperform other adapted and unadapted LMs. Here, the architecture-adapted LSTM combination obtains a 35.9 % word error rate (WER) and is outperformed by FMA-based LSTM combination obtaining the overall lowest WER of 34.4 %

• https://www.researchgate.net/publication/319185101_Approximated_and_Domain-Adapted_LSTM_Language_Models_for_First-Pass_Decoding_in_Speech_Recognition

@inproceedings{Singh2017,
title = {Approximated and domain-adapted LSTM language models for first-pass decoding in speech recognition},
author = {Mittul Singh and Youssef Oualil and Dietrich Klakow},
url = {https://www.researchgate.net/publication/319185101_Approximated_and_Domain-Adapted_LSTM_Language_Models_for_First-Pass_Decoding_in_Speech_Recognition},
year = {2017},
date = {2017},
publisher = {18th Annual Conference of the International Speech Communication Association (INTERSPEECH)},
address = {Stockholm, Sweden},
abstract = {Traditionally, short-range Language Models (LMs) like the conventional n-gram models have been used for language model adaptation. Recent work has improved performance for such tasks using adapted long-span models like Recurrent Neural Network LMs (RNNLMs). With the first pass performed using a large background n-gram LM, the adapted RNNLMs are mostly used to rescore lattices or N-best lists, as a second step in the decoding process. Ideally, these adapted RNNLMs should be applied for first-pass decoding. Thus, we introduce two ways of applying adapted long-short-term-memory (LSTM) based RNNLMs for first-pass decoding. Using available techniques to convert LSTMs to approximated versions for first-pass decoding, we compare approximated LSTMs adapted in a Fast Marginal Adaptation framework (FMA) and an approximated version of architecture-based-adaptation of LSTM. On a conversational speech recognition task, these differently approximated and adapted LSTMs combined with a trigram LM outperform other adapted and unadapted LMs. Here, the architecture-adapted LSTM combination obtains a 35.9 % word error rate (WER) and is outperformed by FMA-based LSTM combination obtaining the overall lowest WER of 34.4 %},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Word embeddings are high-dimensional vector representations of words and are thus difficult to interpret. In order to deal with this, we introduce an unsupervised parameter free method for creating a hierarchical graphical clustering of the full ensemble of word vectors and show that this structure is a geometrically meaningful representation of the original relations between the words. This newly obtained representation can be used for better understanding and thus improving the embedding algorithm and exhibits semantic meaning, so it can also be utilized in a variety of language processing tasks like categorization or measuring similarity.

• https://aclanthology.org/W17-2404
• https://doi.org/10.18653/v1/W17-2404"

@inproceedings{TroKla2017,
title = {Parameter Free Hierarchical Graph-Based Clustering for Analyzing Continuous Word Embeddings},
author = {Dietrich Klakow and Thomas Trost},
url = {https://aclanthology.org/W17-2404},
doi = {https://doi.org/10.18653/v1/W17-2404"},
year = {2017},
date = {2017},
booktitle = {Proceedings of TextGraphs-11: Graph-based Methods for Natural Language Processing (Workshop at ACL 2017)},
pages = {30-38},
publisher = {Association for Computational Linguistics},
address = {Vancouver, Canada},
abstract = {Word embeddings are high-dimensional vector representations of words and are thus difficult to interpret. In order to deal with this, we introduce an unsupervised parameter free method for creating a hierarchical graphical clustering of the full ensemble of word vectors and show that this structure is a geometrically meaningful representation of the original relations between the words. This newly obtained representation can be used for better understanding and thus improving the embedding algorithm and exhibits semantic meaning, so it can also be utilized in a variety of language processing tasks like categorization or measuring similarity.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Training large vocabulary Neural Network Language Models (NNLMs) is a difficult task due to the explicit requirement of the output layer normalization, which typically involves the evaluation of the full softmax function over the complete vocabulary. This paper proposes a Batch Noise Contrastive Estimation (B-NCE) approach to alleviate this problem. This is achieved by reducing the vocabulary, at each time step, to the target words in the batch and then replacing the softmax by the noise contrastive estimation approach, where these words play the role of targets and noise samples at the same time. In doing so, the proposed approach can be fully formulated and implemented using optimal dense matrix operations. Applying B-NCE to train different NNLMs on the Large Text Compression Benchmark (LTCB) and the One Billion Word Benchmark (OBWB) shows a significant reduction of the training time with no noticeable degradation of the models performance. This paper also presents a new baseline comparative study of different standard NNLMs on the large OBWB on a single Titan-X GPU.

• https://arxiv.org/abs/1708.05997

@inproceedings{Oualil2017,
title = {A batch noise contrastive estimation approach for training large vocabulary language models},
author = {Youssef Oualil and Dietrich Klakow},
url = {https://arxiv.org/abs/1708.05997},
year = {2017},
date = {2017},
publisher = {18th Annual Conference of the International Speech Communication Association (INTERSPEECH)},
abstract = {Training large vocabulary Neural Network Language Models (NNLMs) is a difficult task due to the explicit requirement of the output layer normalization, which typically involves the evaluation of the full softmax function over the complete vocabulary. This paper proposes a Batch Noise Contrastive Estimation (B-NCE) approach to alleviate this problem. This is achieved by reducing the vocabulary, at each time step, to the target words in the batch and then replacing the softmax by the noise contrastive estimation approach, where these words play the role of targets and noise samples at the same time. In doing so, the proposed approach can be fully formulated and implemented using optimal dense matrix operations. Applying B-NCE to train different NNLMs on the Large Text Compression Benchmark (LTCB) and the One Billion Word Benchmark (OBWB) shows a significant reduction of the training time with no noticeable degradation of the models performance. This paper also presents a new baseline comparative study of different standard NNLMs on the large OBWB on a single Titan-X GPU.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Training good word embeddings requires large amounts of data. Out-of-vocabulary words will still be encountered at test-time, leaving these words without embeddings. To overcome this lack of embeddings for rare words, existing methods leverage morphological features to generate embeddings. While the existing methods use computationally-intensive rule-based (Soricut and Och, 2015) or tool-based (Botha and Blunsom, 2014) morphological analysis to generate embeddings, our system applies a computationally-simpler sub-word search on words that have existing embeddings.

Embeddings of the sub-word search results are then combined using string similarity functions to generate rare word embeddings. We augmented pre-trained word embeddings with these novel embeddings and evaluated on a rare word similarity task, obtaining up to 3 times improvement in correlation over the original set of embeddings. Applying our technique to embeddings trained on larger datasets led to on-par performance with the existing state-of-the-art for this task. Additionally, while analysing augmented embeddings in a log-bilinear language model, we observed up to 50% reduction in rare word perplexity in comparison to other more complex language models.

• http://aclweb.org/anthology/C16-1194

@inproceedings{singh-EtAl:2016:COLING1,
title = {Sub-Word Similarity based Search for Embeddings: Inducing Rare-Word Embeddings for Word Similarity Tasks and Language Modelling},
author = {Mittul Singh and Clayton Greenberg and Youssef Oualil and Dietrich Klakow},
url = {http://aclweb.org/anthology/C16-1194},
year = {2016},
date = {2016-12-01},
booktitle = {Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers},
publisher = {The COLING 2016 Organizing Committee},
address = {Osaka, Japan},
abstract = {Training good word embeddings requires large amounts of data. Out-of-vocabulary words will still be encountered at test-time, leaving these words without embeddings. To overcome this lack of embeddings for rare words, existing methods leverage morphological features to generate embeddings. While the existing methods use computationally-intensive rule-based (Soricut and Och, 2015) or tool-based (Botha and Blunsom, 2014) morphological analysis to generate embeddings, our system applies a computationally-simpler sub-word search on words that have existing embeddings. Embeddings of the sub-word search results are then combined using string similarity functions to generate rare word embeddings. We augmented pre-trained word embeddings with these novel embeddings and evaluated on a rare word similarity task, obtaining up to 3 times improvement in correlation over the original set of embeddings. Applying our technique to embeddings trained on larger datasets led to on-par performance with the existing state-of-the-art for this task. Additionally, while analysing augmented embeddings in a log-bilinear language model, we observed up to 50% reduction in rare word perplexity in comparison to other more complex language models.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Significant correlations between words can be observed over long distances, but contemporary language models like N-grams, Skip grams, and recurrent neural network language models (RNNLMs) require a large number of parameters to capture these dependencies, if the models can do so at all. In this paper, we propose the Custom Decay Language Model (CDLM), which captures long range correlations while maintaining sub-linear increase in parameters with vocabulary size. This model has a robust and stable training procedure (unlike RNNLMs), a more powerful modeling scheme than the Skip models, and a customizable representation. In perplexity experiments, CDLMs outperform the Skip models using fewer number of parameters. A CDLM also nominally outperformed a similar-sized RNNLM, meaning that it learned as much as the RNNLM but without recurrence.

• http://dx.doi.org/10.1007/978-3-319-45510-5_39
• https://doi.org/10.1007/978-3-319-45510-5_39

@inbook{Singh2016,
title = {The Custom Decay Language Model for Long Range Dependencies},
author = {Mittul Singh and Clayton Greenberg and Dietrich Klakow},
url = {http://dx.doi.org/10.1007/978-3-319-45510-5_39},
doi = {https://doi.org/10.1007/978-3-319-45510-5_39},
year = {2016},
date = {2016},
booktitle = {Text, Speech, and Dialogue: 19th International Conference, TSD 2016, Brno , Czech Republic, September 12-16, 2016, Proceedings},
isbn = {978-3-319-45510-5},
pages = {343-351},
publisher = {Springer International Publishing},
address = {Cham},
abstract = {Significant correlations between words can be observed over long distances, but contemporary language models like N-grams, Skip grams, and recurrent neural network language models (RNNLMs) require a large number of parameters to capture these dependencies, if the models can do so at all. In this paper, we propose the Custom Decay Language Model (CDLM), which captures long range correlations while maintaining sub-linear increase in parameters with vocabulary size. This model has a robust and stable training procedure (unlike RNNLMs), a more powerful modeling scheme than the Skip models, and a customizable representation. In perplexity experiments, CDLMs outperform the Skip models using fewer number of parameters. A CDLM also nominally outperformed a similar-sized RNNLM, meaning that it learned as much as the RNNLM but without recurrence.},
pubstate = {published},
type = {inbook}
}

Copy BibTeX to Clipboard

Feedforward Neural Network (FNN)-based language models estimate the probability of the next word based on the history of the last N words, whereas Recurrent Neural Networks (RNN) perform the same task based only on the last word and some context information that cycles in the network. This paper presents a novel approach, which bridges the gap between these two categories of networks. In particular, we propose an architecture which takes advantage of the explicit, sequential enumeration of the word history in FNN structure while enhancing each word representation at the projection layer through recurrent context information that evolves in the network. The context integration is performed using an additional word-dependent weight matrix that is also learned during the training. Extensive experiments conducted on the Penn Treebank (PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a significant reduction of the perplexity when compared to state-of-the-art feedforward as well as recurrent neural network architectures.

• https://arxiv.org/abs/1703.08068

@article{oualil2016sequential,
title = {Sequential recurrent neural networks for language modeling},
author = {Youssef Oualil and Clayton Greenberg and Mittul Singh and Dietrich Klakow andYoussef Oualil and Singh Mittul},
url = {https://arxiv.org/abs/1703.08068},
year = {2016},
date = {2016},
journal = {Interspeech 2016},
pages = {3509-3513},
abstract = {Feedforward Neural Network (FNN)-based language models estimate the probability of the next word based on the history of the last N words, whereas Recurrent Neural Networks (RNN) perform the same task based only on the last word and some context information that cycles in the network. This paper presents a novel approach, which bridges the gap between these two categories of networks. In particular, we propose an architecture which takes advantage of the explicit, sequential enumeration of the word history in FNN structure while enhancing each word representation at the projection layer through recurrent context information that evolves in the network. The context integration is performed using an additional word-dependent weight matrix that is also learned during the training. Extensive experiments conducted on the Penn Treebank (PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a significant reduction of the perplexity when compared to state-of-the-art feedforward as well as recurrent neural network architectures.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

In this paper, we discuss the human thematic fit judgement correlation task in the context of real-valued vector space word representations. Thematic fit is the extent to which an argument fulfils the selectional preference of a verb given a role: for example, how well “cake” fulfils the patient role of “cut”. In recent work, systems have been evaluated on this task by finding the correlations of their output judgements with human-collected judgement data. This task is a representationindependent way of evaluating models that can be applied whenever a system score can be generated, and it is applicable wherever predicate-argument relations are significant to performance in end-user tasks. Significant progress has been made on this cognitive modeling task, leaving considerable space for future, more comprehensive types of evaluation.

• https://www.researchgate.net/publication/306094219_Thematic_fit_evaluation_an_aspect_of_selectional_preferences

@article{Sayeed2016,
title = {Thematic fit evaluation: an aspect of selectional preferences},
author = {Asad Sayeed and Clayton Greenberg and Vera Demberg},
url = {https://www.researchgate.net/publication/306094219_Thematic_fit_evaluation_an_aspect_of_selectional_preferences},
year = {2016},
date = {2016},
journal = {Proceedings of the 1st Workshop on Evaluating Vector Space Representations for NLP},
pages = {99-105},
abstract = {In this paper, we discuss the human thematic fit judgement correlation task in the context of real-valued vector space word representations. Thematic fit is the extent to which an argument fulfils the selectional preference of a verb given a role: for example, how well “cake” fulfils the patient role of “cut”. In recent work, systems have been evaluated on this task by finding the correlations of their output judgements with human-collected judgement data. This task is a representationindependent way of evaluating models that can be applied whenever a system score can be generated, and it is applicable wherever predicate-argument relations are significant to performance in end-user tasks. Significant progress has been made on this cognitive modeling task, leaving considerable space for future, more comprehensive types of evaluation.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

The goal of language modeling techniques is to capture the statistical and structural properties of natural languages from training corpora. This task typically involves the learning of short range dependencies, which generally model the syntactic properties of a language and/or long range dependencies, which are semantic in nature. We propose in this paper a new multi-span architecture, which separately models the short and long context information while it dynamically merges them to perform the language modeling task. This is done through a novel recurrent Long-Short Range Context (LSRC) network, which explicitly models the local (short) and global (long) context using two separate hidden states that evolve in time. This new architecture is an adaptation of the Long-Short Term Memory network (LSTM) to take into account the linguistic properties. Extensive experiments conducted on the Penn Treebank (PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a significant reduction of the perplexity when compared to state-of-the-art language modeling techniques.

• https://aclanthology.org/D16-1154/

@inproceedings{Oualil2016,
title = {Long-short range context neural networks for language models},
author = {Youssef Oualil and Mittul Singh and Clayton Greenberg and Dietrich Klakow},
url = {https://aclanthology.org/D16-1154/},
year = {2016},
date = {2016},
publisher = {EMLP 2016},
abstract = {The goal of language modeling techniques is to capture the statistical and structural properties of natural languages from training corpora. This task typically involves the learning of short range dependencies, which generally model the syntactic properties of a language and/or long range dependencies, which are semantic in nature. We propose in this paper a new multi-span architecture, which separately models the short and long context information while it dynamically merges them to perform the language modeling task. This is done through a novel recurrent Long-Short Range Context (LSRC) network, which explicitly models the local (short) and global (long) context using two separate hidden states that evolve in time. This new architecture is an adaptation of the Long-Short Term Memory network (LSTM) to take into account the linguistic properties. Extensive experiments conducted on the Penn Treebank (PTB) and the Large Text Compression Benchmark (LTCB) corpus showed a significant reduction of the perplexity when compared to state-of-the-art language modeling techniques.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Secure user identification is important for the increasing number of eyewear computers but limited input capabilities pose significant usability challenges for established knowledge-based schemes, such as passwords or PINs. We present SkullConduct, a biometric system that uses bone conduction of sound through the user’s skull as well as a microphone readily integrated into many of these devices, such as Google Glass. At the core of SkullConduct is a method to analyze the characteristic frequency response created by the user’s skull using a combination of Mel Frequency Cepstral Coefficient (MFCC) features as well as a computationally light-weight 1NN classifier. We report on a controlled experiment with 10 participants that shows that this frequency response is person-specific and stable — even when taking off and putting on the device multiple times — and thus serves as a robust biometric. We show that our method can identify users with 97.0% accuracy and authenticate them with an equal error rate of 6.9%, thereby bringing biometric user identification to eyewear computers equipped with bone conduction technology.

• http://doi.acm.org/10.1145/2858036.2858152
• https://doi.org/10.1145/2858036.2858152

@inproceedings{Schneegass:2016:SBU:2858036.2858152,
title = {SkullConduct: Biometric User Identification on Eyewear Computers Using Bone Conduction Through the Skull},
author = {Stefan Schneegass and Youssef Oualil and Andreas Bulling},
url = {http://doi.acm.org/10.1145/2858036.2858152},
doi = {https://doi.org/10.1145/2858036.2858152},
year = {2016},
date = {2016},
booktitle = {Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems},
isbn = {978-1-4503-3362-7},
pages = {1379-1384},
publisher = {ACM},
address = {New York, NY, USA},
abstract = {Secure user identification is important for the increasing number of eyewear computers but limited input capabilities pose significant usability challenges for established knowledge-based schemes, such as passwords or PINs. We present SkullConduct, a biometric system that uses bone conduction of sound through the user's skull as well as a microphone readily integrated into many of these devices, such as Google Glass. At the core of SkullConduct is a method to analyze the characteristic frequency response created by the user's skull using a combination of Mel Frequency Cepstral Coefficient (MFCC) features as well as a computationally light-weight 1NN classifier. We report on a controlled experiment with 10 participants that shows that this frequency response is person-specific and stable -- even when taking off and putting on the device multiple times -- and thus serves as a robust biometric. We show that our method can identify users with 97.0% accuracy and authenticate them with an equal error rate of 6.9%, thereby bringing biometric user identification to eyewear computers equipped with bone conduction technology.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard