Hongliang FU Qianqian LI Huawei TAO Chunhua ZHU Yue XIE Ruxue GUO
Speech emotion recognition (SER) is a key research technology to realize the third generation of artificial intelligence, which is widely used in human-computer interaction, emotion diagnosis, interpersonal communication and other fields. However, the aliasing of language and semantic information in speech tends to distort the alignment of emotion features, which affects the performance of cross-corpus SER system. This paper proposes a cross-corpus SER model based on causal emotion information representation (CEIR). The model uses the reconstruction loss of the deep autoencoder network and the source domain label information to realize the preliminary separation of causal features. Then, the causal correlation matrix is constructed, and the local maximum mean difference (LMMD) feature alignment technology is combined to make the causal features of different dimensions jointly distributed independent. Finally, the supervised fine-tuning of labeled data is used to achieve effective extraction of causal emotion information. The experimental results show that the average unweighted average recall (UAR) of the proposed algorithm is increased by 3.4% to 7.01% compared with the latest partial algorithms in the field.
Yang WANG Hongliang FU Huawei TAO Jing YANG Hongyi GE Yue XIE
This letter focuses on the cross-corpus speech emotion recognition (SER) task, in which the training and testing speech signals in cross-corpus SER belong to different speech corpora. Existing algorithms are incapable of effectively extracting common sentiment information between different corpora to facilitate knowledge transfer. To address this challenging problem, a novel convolutional auto-encoder and adversarial domain adaptation (CAEADA) framework for cross-corpus SER is proposed. The framework first constructs a one-dimensional convolutional auto-encoder (1D-CAE) for feature processing, which can explore the correlation among adjacent one-dimensional statistic features and the feature representation can be enhanced by the architecture based on encoder-decoder-style. Subsequently the adversarial domain adaptation (ADA) module alleviates the feature distributions discrepancy between the source and target domains by confusing domain discriminator, and specifically employs maximum mean discrepancy (MMD) to better accomplish feature transformation. To evaluate the proposed CAEADA, extensive experiments were conducted on EmoDB, eNTERFACE, and CASIA speech corpora, and the results show that the proposed method outperformed other approaches.
Qing-dao-er-ji REN Yuan LI Shi BAO Yong-chao LIU Xiu-hong CHEN
As the mainstream approach in the field of machine translation, neural machine translation (NMT) has achieved great improvements on many rich-source languages, but performance of NMT for low-resource languages ae not very good yet. This paper uses data enhancement technology to construct Mongolian-Chinese pseudo parallel corpus, so as to improve the translation ability of Mongolian-Chinese translation model. Experiments show that the above methods can improve the translation ability of the translation model. Finally, a translation model trained with large-scale pseudo parallel corpus and integrated with soft context data enhancement technology is obtained, and its BLEU value is 39.3.
Jiateng LIU Wenming ZHENG Yuan ZONG Cheng LU Chuangao TANG
In this letter, we propose a novel deep domain-adaptive convolutional neural network (DDACNN) model to handle the challenging cross-corpus speech emotion recognition (SER) problem. The framework of the DDACNN model consists of two components: a feature extraction model based on a deep convolutional neural network (DCNN) and a domain-adaptive (DA) layer added in the DCNN utilizing the maximum mean discrepancy (MMD) criterion. We use labeled spectrograms from source speech corpus combined with unlabeled spectrograms from target speech corpus as the input of two classic DCNNs to extract the emotional features of speech, and train the model with a special mixed loss combined with a cross-entrophy loss and an MMD loss. Compared to other classic cross-corpus SER methods, the major advantage of the DDACNN model is that it can extract robust speech features which are time-frequency related by spectrograms and narrow the discrepancies between feature distribution of source corpus and target corpus to get better cross-corpus performance. Through several cross-corpus SER experiments, our DDACNN achieved the state-of-the-art performance on three public emotion speech corpora and is proved to handle the cross-corpus SER problem efficiently.
Xiuzhen CHEN Xiaoyan ZHOU Cheng LU Yuan ZONG Wenming ZHENG Chuangao TANG
For cross-corpus speech emotion recognition (SER), how to obtain effective feature representation for the discrepancy elimination of feature distributions between source and target domains is a crucial issue. In this paper, we propose a Target-adapted Subspace Learning (TaSL) method for cross-corpus SER. The TaSL method trys to find a projection subspace, where the feature regress the label more accurately and the gap of feature distributions in target and source domains is bridged effectively. Then, in order to obtain more optimal projection matrix, ℓ1 norm and ℓ2,1 norm penalty terms are added to different regularization terms, respectively. Finally, we conduct extensive experiments on three public corpuses, EmoDB, eNTERFACE and AFEW 4.0. The experimental results show that our proposed method can achieve better performance compared with the state-of-the-art methods in the cross-corpus SER tasks.
Nurul LUBIS Dessi LESTARI Sakriani SAKTI Ayu PURWARIANTI Satoshi NAKAMURA
As interaction between human and computer continues to develop to the most natural form possible, it becomes increasingly urgent to incorporate emotion in the equation. This paper describes a step toward extending the research on emotion recognition to Indonesian. The field continues to develop, yet exploration of the subject in Indonesian is still lacking. In particular, this paper highlights two contributions: (1) the construction of the first emotional audio-visual database in Indonesian, and (2) the first multimodal emotion recognizer in Indonesian, built from the aforementioned corpus. In constructing the corpus, we aim at natural emotions that are corresponding to real life occurrences. However, the collection of emotional corpora is notably labor intensive and expensive. To diminish the cost, we collect the emotional data from television programs recordings, eliminating the need of an elaborate recording set up and experienced participants. In particular, we choose television talk shows due to its natural conversational content, yielding spontaneous emotion occurrences. To cover a broad range of emotions, we collected three episodes in different genres: politics, humanity, and entertainment. In this paper, we report points of analysis of the data and annotations. The acquisition of the emotion corpus serves as a foundation in further research on emotion. Subsequently, in the experiment, we employ the support vector machine (SVM) algorithm to model the emotions in the collected data. We perform multimodal emotion recognition utilizing the predictions of three modalities: acoustic, semantic, and visual. When compared to the unimodal result, in the multimodal feature combination, we attain identical accuracy for the arousal at 92.6%, and a significant improvement for the valence classification task at 93.8%. We hope to continue this work and move towards a finer-grain, more precise quantification of emotion.
Jing ZHANG Degen HUANG Kaiyu HUANG Zhuang LIU Fuji REN
Microblog data contains rich information of real-world events with great commercial values, so microblog-oriented natural language processing (NLP) tasks have grabbed considerable attention of researchers. However, the performance of microblog-oriented Chinese Word Segmentation (CWS) based on deep neural networks (DNNs) is still not satisfying. One critical reason is that the existing microblog-oriented training corpus is inadequate to train effective weight matrices for DNNs. In this paper, we propose a novel active learning method to extend the scale of the training corpus for DNNs. However, due to a large amount of partially overlapped sentences in the microblogs, it is difficult to select samples with high annotation values from raw microblogs during the active learning procedure. To select samples with higher annotation values, parameter λ is introduced to control the number of repeatedly selected samples. Meanwhile, various strategies are adopted to measure the overall annotation values of a sample during the active learning procedure. Experiments on the benchmark datasets of NLPCC 2015 show that our λ-active learning method outperforms the baseline system and the state-of-the-art method. Besides, the results also demonstrate that the performances of the DNNs trained on the extended corpus are significantly improved.
Reiko KUWA Tsuneo KATO Seiichi YAMAMOTO
This paper proposes a classification method of second-language-learner utterances for interactive computer-assisted language learning systems. This classification method uses three types of bilingual evaluation understudy (BLEU) scores as features for a classifier. The three BLEU scores are calculated in accordance with three subsets of a learner corpus divided according to the quality of utterances. For the purpose of overcoming the data-sparseness problem, this classification method uses the BLEU scores calculated using a mixture of word and part-of-speech (POS)-tag sequences converted from word sequences based on a POS-replacement rule according to which words are replaced with POS tags in n-grams. Experiments of classifying English utterances by Japanese demonstrated that the proposed classification method achieved classification accuracy of 78.2% which was 12.3 points higher than a baseline with one BLEU score.
One of the problems associated with voice conversion from a nonparallel corpus is how to find the best match or alignment between the source and the target vector sequences without linguistic information. In a previous study, alignment was achieved by minimizing the distance between the source vector and the transformed vector. This method, however, yielded a sequence of feature vectors that were not well matched with the underlying speaker model. In this letter, the vectors were selected from the candidates by maximizing the overall likelihood of the selected vectors with respect to the target model in the HMM context. Both objective and subjective evaluations were carried out using the CMU ARCTIC database to verify the effectiveness of the proposed method.
In this paper, we propose a method to find similar sentences based on language resources for building a parallel corpus between English and Korean from Wikipedia. We use a Wiki-dictionary consisted of document titles from the Wikipedia and bilingual example sentence pairs from Web dictionary instead of traditional machine readable dictionary. In this way, we perform similarity calculation between sentences using sequential matching of the language resources, and evaluate the extracted parallel sentences. In the experiments, the proposed parallel sentences extraction method finally shows 65.4% of F1-score.
Peng SONG Yun JIN Li ZHAO Minghai XIN
A major challenge for speech emotion recognition is that when the training and deployment conditions do not use the same speech corpus, the recognition rates will obviously drop. Transfer learning, which has successfully addressed the cross-domain classification or recognition problem, is presented for cross-corpus speech emotion recognition. First, by using the maximum mean discrepancy embedding (MMDE) optimization and dimension reduction algorithms, two close low-dimensional feature spaces are obtained for source and target speech corpora, respectively. Then, a classifier function is trained using the learned low-dimensional features in the labeled source corpus, and directly applied to the unlabeled target corpus for emotion label recognition. Experimental results demonstrate that the transfer learning method can significantly outperform the traditional automatic recognition technique for cross-corpus speech emotion recognition.
Nattapong TONGTEP Thanaruk THEERAMUNKONG
Automated or semi-automated annotation is a practical solution for large-scale corpus construction. However, the special characteristics of Thai language, such as lack of word-boundary and sentence-boundary markers, trigger several issues in automatic corpus annotation. This paper presents a multi-stage annotation framework, containing two stages of chunking and three stages of tagging. The two chunking stages are pattern matching-based named entity (NE) extraction and dictionary-based word segmentation while the three succeeding tagging stages are dictionary-, pattern- and statist09812490981249ical-based tagging. Applying heuristics of ambiguity priority, NE extraction is performed first on an original text using a set of patterns, in the order of pattern ambiguity. Next, the remaining text is segmented into words with a dictionary. The obtained chunks are then tagged with types of named entities or parts-of-speech (PoS) using dictionaries, patterns and statistics. Focusing on the reduction of human intervention in corpus construction, our experimental results show that the dictionary-based tagging process can assign unique tags to 64.92% of the words, with the remaining of 24.14% unknown words and 10.94% ambiguously tagged words. Later, the pattern-based tagging can reduce unknown words to only 13.34% while the statistical-based tagging can solve the ambiguously tagged words to only 3.01%.
This paper proposes a method for detecting determiner errors, which are highly frequent in learner English. To augment conventional methods, the proposed method exploits a strong tendency displayed by learners in determiner usage, i.e., mistakenly omitting determiners most of the time. Its basic idea is simple and applicable to almost any conventional method. This paper also proposes combining the method with countability prediction, which results in further improvement. Experiments show that the proposed method achieves an F-measure of 0.684 and significantly outperforms conventional methods.
Yoo Rhee OH Yong Guk KIM Mina KIM Hong Kook KIM Mi Suk LEE Hyun Joo BAE
In this paper, we propose a text corpus design method for a Korean stereo super-wideband speech database. Since a small-sized text corpus for speech coding is generally required for speech coding, the corpus should be designed to comply with the pronunciation behavior of natural conversation in order to ensure efficient speech quality tests. To this end, the proposed design method utilizes a similarity measure between the phoneme distribution occurring from natural conversation and that from the designed text corpus. In order to achieve this goal, we first collect and refine text data from textbooks and websites. Next, a corpus is designed from the refined text data based on the similarity measure to compare phoneme distributions. We then construct a Korean stereo super-wideband speech (K-SW) database using the designed text corpus, where the recording environment is set to meet the conditions defined by ITU-T. Finally, the subjective quality of the K-SW database is evaluated using an ITU-T super-wideband codec in order to demonstrate that the K-SW database is useful for developing and evaluating super-wideband codecs.
Tetsuo KOSAKA Yuui TAKEDA Takashi ITO Masaharu KATO Masaki KOHDA
In this paper, we propose a new speaker-class modeling and its adaptation method for the LVCSR system and evaluate the method on the Corpus of Spontaneous Japanese (CSJ). In this method, closer speakers are selected from training speakers and the acoustic models are trained by using their utterances for each evaluation speaker. One of the major issues of the speaker-class model is determining the selection range of speakers. In order to solve the problem, several models which have a variety of speaker range are prepared for each evaluation speaker in advance, and the most proper model is selected on a likelihood basis in the recognition step. In addition, we improved the recognition performance using unsupervised speaker adaptation with the speaker-class models. In the recognition experiments, a significant improvement could be obtained by using the proposed speaker adaptation based on speaker-class models compared with the conventional adaptation method.
Yoshihide KATO Shigeki MATSUBARA
This paper proposes a method of correcting annotation errors in a treebank. By using a synchronous grammar, the method transforms parse trees containing annotation errors into the ones whose errors are corrected. The synchronous grammar is automatically induced from the treebank. We report an experimental result of applying our method to the Penn Treebank. The result demonstrates that our method corrects syntactic annotation errors with high precision.
Do-Gil LEE Gumwon HONG Seok Kee LEE Hae-Chang RIM
The construction of annotated corpora requires considerable manual effort. This paper presents a pragmatic method to minimize human intervention for the construction of Korean part-of-speech (POS) tagged corpus. Instead of focusing on improving the performance of conventional automatic POS taggers, we devise a discriminative POS tagger which can selectively produce either a single analysis or multiple analyses based on the tagging reliability. The proposed approach uses two decision rules to judge the tagging reliability. Experimental results show that the proposed approach can effectively control the quality of corpus and the amount of manual annotation by the threshold value of the rule.
Ryohei SASANO Daisuke KAWAHARA Sadao KUROHASHI
This paper reports the effect of corpus size on case frame acquisition for predicate-argument structure analysis in Japanese. For this study, we collect a Japanese corpus consisting of up to 100 billion words, and construct case frames from corpora of six different sizes. Then, we apply these case frames to syntactic and case structure analysis, and zero anaphora resolution, in order to investigate the relationship between the corpus size for case frame acquisition and the performance of predicate-argument structure analysis. We obtained better analyses by using case frames constructed from larger corpora; the performance was not saturated even with a corpus size of 100 billion words.
Ryuichiro HIGASHINAKA Mikio NAKANO
This paper discusses the discourse understanding process in spoken dialogue systems. This process enables a system to understand user utterances from the context of a dialogue. Ambiguity in user utterances caused by multiple speech recognition hypotheses and parsing results sometimes makes it difficult for a system to decide on a single interpretation of a user intention. As a solution, the idea of retaining possible interpretations as multiple dialogue states and resolving the ambiguity using succeeding user utterances has been proposed. Although this approach has proven to improve discourse understanding accuracy, carefully created hand-crafted rules are necessary in order to accurately rank the dialogue states. This paper proposes automatically ranking multiple dialogue states using statistical information obtained from dialogue corpora. The experimental results in the train ticket reservation and weather information service domains show that the statistical information can significantly improve the ranking accuracy of dialogue states as well as the slot accuracy and the concept error rate of the top-ranked dialogue states.
Keiji YASUDA Hirofumi YAMAMOTO Eiichiro SUMITA
For statistical language model training, target domain matched corpora are required. However, training corpora sometimes include both target domain matched and unmatched sentences. In such a case, training set selection is effective for both reducing model size and improving model performance. In this paper, training set selection method for statistical language model training is described. The method provides two advantages for training a language model. One is its capacity to improve the language model performance, and the other is its capacity to reduce computational loads for the language model. The method has four steps. 1) Sentence clustering is applied to all available corpora. 2) Language models are trained on each cluster. 3) Perplexity on the development set is calculated using the language models. 4) For the final language model training, we use the clusters whose language models yield low perplexities. The experimental results indicate that the language model trained on the data selected by our method gives lower perplexity on an open test set than a language model trained on all available corpora.