Computer Speech and Language最新文献

This paper proposed a Parallel Speech Corpus of Northern-central Thai (PaSCoNT). The purpose of this research is not only to understand the different linguistic characteristics between Northern and Central Thai, but also to utilize this corpus for automatic speech recognition. The corpus is composed of speech data from dialogues of daily life among northern Thai people. We designed 2,000 Northern Thai sentences covering all phonemes, in collaboration with linguists specialized in the Northern Thai dialect. The samples in this study are 200 Northern Thai dialect speakers who had been living in Chiang Mai province for more than 18 years. The speech was recorded in both open and closed environments. In the speech recording, each speaker must read 100 pairs of Northern-Central Thai sentences to ensure that the speech data comes from the same speaker. In total, 100 h of speech were recorded: 50 h of Northern Thai and 50 h of Central Thai. Overall, PaSCoNT consists of 907,832 words and 6,279 vocabulary items. Statistical analysis of the PaSCoNT corpus revealed that 49.64 % of words in the lexicon belongs to the Northern Thai dialect, 50.36 % from the Central Thai dialect, and 1,621 vocabulary items appeared in both Northern and Central Thai. Statistical analysis is used to examine the difference in speech tempo, i.e. time per phoneme (TTP), syllable per minute (SPM), between Northern and Central Thai. The results revealed that there were statistically significant differences speech tempo between Central and Northern Thai. The TTP speaking and articulation rate of Central Thai is lower than Northern Thai whereas SPM speaking and articulation rate of Central Thai is higher than Northern Thai. The results also showed that the ASR model training using Northern Thai speech corpus provides the lower WER% when testing using Northern Thai testing speech data and provides the higher WER% when testing using Central Thai Testing speech data and vice versa. However, the ASR model training using the PaSCoNT speech corpus provides the lower WER% for both Northern Thai and Central Thai testing speech data.

Automatic identification of hateful and abusive content is vital in combating the spread of harmful online content and its damaging effects. Most existing works evaluate models by examining the generalization error on train–test splits on hate speech datasets. These datasets often differ in their definitions and labeling criteria, leading to poor generalization performance when predicting across new domains and datasets. This work proposes a new Multi-task Learning (MTL) pipeline that trains simultaneously across multiple hate speech datasets to construct a more encompassing classification model. Using a dataset-level leave-one-out evaluation (designating a dataset for testing and jointly training on all others), we trial the MTL detection on new, previously unseen datasets. Our results consistently outperform a large sample of existing work. We show strong results when examining the generalization error in train–test splits and substantial improvements when predicting on previously unseen datasets. Furthermore, we assemble a novel dataset, dubbed PubFigs, focusing on the problematic speech of American Public Political Figures. We crowdsource-label using Amazon MTurk more than 20,000 tweets and machine-label problematic speech in all the 305,235 tweets in PubFigs. We find that the abusive and hate tweeting mainly originates from right-leaning figures and relates to six topics, including Islam, women, ethnicity, and immigrants. We show that MTL builds embeddings that can simultaneously separate abusive from hate speech, and identify its topics.

In text classification task, models have shown remarkable accuracy across various datasets. However, confusion often arises when certain categories within the dataset are too similar, causing misclassification of certain samples. This paper proposes an improved method for this problem, through the creation of a three-layer text graph for the corpus, which is used to calculate the Category Correlation Matrix (CCM). Additionally, this paper introduces category-adaptive contrastive learning for text embedding from the encoder, enhancing the model’s ability to distinguish between samples in confusable categories that are easily confused. Soft labels are generated using this matrix to guide the classifier, preventing the model from becoming overconfident with one-hot vectors. The efficacy of this approach was demonstrated through experimental evaluations on three text encoders and six different datasets.

Knowledge graph embedding (KGE) aims to embed entities and relations in knowledge graphs (KGs) into a continuous, low-dimensional vector space. It has been shown as an effective tool for integrating knowledge graphs to improve various intelligent applications, such as question answering and information extraction. However, previous KGE models ignore the hidden natural order of knowledge learning on learning the embeddings of entities and relations, leaving room for improvement in their performance. Inspired by the easy-to-hard pattern used in human knowledge learning, this paper proposes a Curriculum learning-based KGE (C-KGE) model, which learns the embeddings of entities and relations from “basic knowledge” to “domain knowledge”. Specifically, a seed set representing the basic knowledge and several knowledge subsets are identified from KG. Then, entity overlap is employed to score the learning difficulty of each subset. Finally, C-KGE trains the entities and relations in each subset according to the learning difficulty score of each subset. C-KGE leverages trained embeddings of the seed set as prior knowledge and learns knowledge subsets iteratively to transfer knowledge between the seed set and subsets, smoothing the learning process of knowledge facts. Experimental results on real-world datasets demonstrate that the proposed model achieves improved embedding performances as well as reducing training time. Our codes and data will be released later.

Long text generation is a hot topic in natural language processing. To address the problem of insufficient semantic representation and incoherent text generation in existing long text models, the Seq2Seq dynamic planning network progressive text generation model (DPPG-BART) is proposed. In the data pre-processing stage, the lexical division sorting algorithm is used. To obtain hierarchical sequences of keywords with clear information content, word weight values are calculated and ranked by TF-IDF of word embedding. To enhance the input representation, the dynamic planning progressive generation network is constructed. Positional features and word embedding vector features are integrated at the input side of the model. At the same time, to enrich the semantic information and expand the content of the text, the relevant concept words are generated by the concept expansion module. The scoring network and feedback mechanism are used to adjust the concept expansion module. Experimental results show that the DPPG-BART model is optimized over GPT2-S, GPT2-L, BART and ProGen-2 model approaches in terms of metric values of MSJ, B-BLEU and FBD on long text datasets from two different domains, CNN and Writing Prompts.

The objective of the relation classification task is to extract relations between entities. Recent studies have found that R-BERT (Wu and He, 2019) based on pre-trained BERT (Devlin et al., 2019) acquires extremely good results in the relation classification task. However, this method does not take into account the semantic differences between different kinds of entities and global semantic information either. In this paper, we set two different fully connected layers to take into account the semantic difference between subject and object entities. Besides, we build a new module named Concat Module to fully fuse the semantic information among the subject entity vector, object entity vector, and the whole sample sentence representation vector. In addition, we apply the average pooling to acquire a better entity representation of each entity and add the activation operation with a new fully connected layer after our Concat Module. Modifying R-BERT, we propose a new model named BERT with Global Semantic Information (GSR-BERT) for relation classification tasks. We use our approach on two datasets: the SemEval-2010 Task 8 dataset and the Chinese character relationship classification dataset. Our approach achieves a significant improvement over the two datasets. It means that our approach enjoys transferability across different datasets. Furthermore, we prove that these policies we used in our approach also enjoy applicability to named entity recognition task.

Supervised models for speech enhancement are trained using artificially generated mixtures of clean speech and noise signals. However, the synthetic training conditions may not accurately reflect real-world conditions encountered during testing. This discrepancy can result in poor performance when the test domain significantly differs from the synthetic training domain. To tackle this issue, the UDASE task of the 7th CHiME challenge aimed to leverage real-world noisy speech recordings from the test domain for unsupervised domain adaptation of speech enhancement models. Specifically, this test domain corresponds to the CHiME-5 dataset, characterized by real multi-speaker and conversational speech recordings made in noisy and reverberant domestic environments, for which ground-truth clean speech signals are not available. In this paper, we present the objective and subjective evaluations of the systems that were submitted to the CHiME-7 UDASE task, and we provide an analysis of the results. This analysis reveals a limited correlation between subjective ratings and several supervised nonintrusive performance metrics recently proposed for speech enhancement. Conversely, the results suggest that more traditional intrusive objective metrics can be used for in-domain performance evaluation using the reverberant LibriCHiME-5 dataset developed for the challenge. The subjective evaluation indicates that all systems successfully reduced the background noise, but always at the expense of increased distortion. Out of the four speech enhancement methods evaluated subjectively, only one demonstrated an improvement in overall quality compared to the unprocessed noisy speech, highlighting the difficulty of the task. The tools and audio material created for the CHiME-7 UDASE task are shared with the community.

Speech intelligibility (SI) prediction models are a valuable tool for the development of speech processing algorithms for hearing aids or consumer electronics. For the use in realistic environments it is desirable that the SI model is non-intrusive (does not require separate input of original and degraded speech, transcripts or a-priori knowledge about the signals) and does a binaural processing of the audio signals. Most of the existing SI models do not fulfill all of these criteria. In this study, we propose an SI model based on phone probabilities obtained from a deep neural net. The model comprises a binaural enhancement stage for prediction of the speech recognition threshold (SRT) in realistic acoustic scenes. In the first part of the study, SRT predictions in different spatial configurations are compared to the results from normal-hearing listeners. On average, our approach produces lower errors and higher correlations compared to three intrusive baseline models. In the second part, we explore if measures relevant in spatial hearing, i.e., the intelligibility level difference (ILD) and the binaural ILD (BILD), can be predicted with our modeling approach. We also investigate if a language mismatch between training and testing the model plays a role when predicting ILD and BILD. This point is especially important for low-resource languages, where not thousands of hours of language material are available for training. Binaural benefits are predicted by our model with an error of 1.5 dB. This is slightly higher than the error with a competitive baseline MBSTOI (1.1 dB), but does not require separate input of original and degraded speech. We also find that good binaural predictions can be obtained with models that are not specifically trained with the target language.