We propose a novel method for selecting coherent and diverse responses for a given dialogue context. The proposed method re-ranks response candidates generated from conversational models by using event causality relations between events in a dialogue history and response candidates (e.g., ``be stressed out'' precedes ``relieve stress''). We use distributed event representation based on the Role Factored Tensor Model for a robust matching of event causality relations due to limited event causality knowledge of the system. Experimental results showed that the proposed method improved coherency and dialogue continuity of system responses.
The most common way for humans to communicate is by speech. But perhaps a language system cannot know what it is communicating without a connection to the real world by image perception. In fact, humans perceive these multiple sources of information together to build a general concept. However, constructing a machine that can alleviate these modalities together in a supervised learning fashion is difficult, because a parallel dataset is required among speech, image, and text modalities altogether that is often unavailable. A machine speech chain based on sequence-to-sequence deep learning was previously proposed to achieve semi-supervised learning that enabled automatic speech recognition (ASR) and text-to-speech synthesis (TTS) to teach each other when they receive unpaired data. In this research, we take a further step by expanding the speech chain into a multimodal chain and design a closely knit chain architecture that connects ASR, TTS, image captioning (IC), and image retrieval (IR) models into a single framework. ASR, TTS, IC, and IR components can be trained in a semi-supervised fashion by assisting each other given incomplete datasets and leveraging cross-modal data augmentation within the chain.
We describe our submitted system for the ZeroSpeech Challenge 2019. The current challenge theme addresses the difficulty of constructing a speech synthesizer without any text or phonetic labels and requires a system that can (1) discover subword units in an unsupervised way, and (2) synthesize the speech with a target speaker's voice. Moreover, the system should also balance the discrimination score ABX, the bit-rate compression rate, and the naturalness and the intelligibility of the constructed voice. To tackle these problems and achieve the best trade-off, we utilize a vector quantized variational autoencoder (VQ-VAE) and a multi-scale codebook-to-spectrogram (Code2Spec) inverter trained by mean square error and adversarial loss. The VQ-VAE extracts the speech to a latent space, forces itself to map it into the nearest codebook and produces compressed representation. Next, the inverter generates a magnitude spectrogram to the target voice, given the codebook vectors from VQ-VAE. In our experiments, we also investigated several other clustering algorithms, including K-Means and GMM, and compared them with the VQ-VAE result on ABX scores and bit rates. Our proposed approach significantly improved the intelligibility (in CER), the MOS, and discrimination ABX scores compared to the official ZeroSpeech 2019 baseline or even the topline.
In a human-machine dialog scenario, deciding the appropriate time for the machine to take the turn is an open research problem. In contrast, humans engaged in conversations are able to timely decide when to interrupt the speaker for competitive or non-competitive reasons. In state-of-the-art turn-by-turn dialog systems the decision on the next dialog action is taken at the end of the utterance. In this paper, we propose a token-by-token prediction of the dialog state from incremental transcriptions of the user utterance. To identify the point of maximal understanding in an ongoing utterance, we a) implement an incremental Dialog State Tracker which is updated on a token basis (iDST) b) re-label the Dialog State Tracking Challenge 2 (DSTC2) dataset and c) adapt it to the incremental turn-taking experimental scenario. The re-labeling consists of assigning a binary value to each token in the user utterance that allows to identify the appropriate point for taking the turn. Finally, we implement an incremental Turn Taking Decider (iTTD) that is trained on these new labels for the turn-taking decision. We show that the proposed model can achieve a better performance compared to a deterministic handcrafted turn-taking algorithm.
Argumentation-based dialogue systems, which can handle and exchange arguments through dialogue, have been widely researched. It is required that these systems have sufficient supporting information to argue their claims rationally; however, the systems often do not have enough of such information in realistic situations. One way to fill in the gap is acquiring such missing information from dialogue partners (information-seeking dialogue). Existing information-seeking dialogue systems are based on handcrafted dialogue strategies that exhaustively examine missing information. However, the proposed strategies are not specialized in collecting information for constructing rational arguments. Moreover, the number of system's inquiry candidates grows in accordance with the size of the argument set that the system deal with. In this paper, we formalize the process of information-seeking dialogue as Markov decision processes (MDPs) and apply deep reinforcement learning (DRL) for automatically optimizing a dialogue strategy. By utilizing DRL, our dialogue strategy can successfully minimize objective functions, the number of turns it takes for our system to collect necessary information in a dialogue. We conducted dialogue experiments using two datasets from different domains of argumentative dialogue. Experimental results show that the proposed formalization based on MDP works well, and the policy optimized by DRL outperformed existing heuristic dialogue strategies.
Although generation-based dialogue systems have been widely researched, the response generations by most existing systems have very low diversities. The most likely reason for this problem is Maximum Likelihood Estimation (MLE) with Softmax Cross-Entropy (SCE) loss. MLE trains models to generate the most frequent responses from enormous generation candidates, although in actual dialogues there are various responses based on the context. In this paper, we propose a new objective function called Inverse Token Frequency (ITF) loss, which individually scales smaller loss for frequent token classes and larger loss for rare token classes. This function encourages the model to generate rare tokens rather than frequent tokens. It does not complicate the model and its training is stable because we only replace the objective function. On the OpenSubtitles dialogue dataset, our loss model establishes a state-of-the-art DIST-1 of 7.56, which is the unigram diversity score, while maintaining a good BLEU-1 score. On a Japanese Twitter replies dataset, our loss model achieves a DIST-1 score comparable to the ground truth.
The speech chain mechanism integrates automatic speech recognition (ASR) and text-to-speech synthesis (TTS) modules into a single cycle during training. In our previous work, we applied a speech chain mechanism as a semi-supervised learning. It provides the ability for ASR and TTS to assist each other when they receive unpaired data and let them infer the missing pair and optimize the model with reconstruction loss. If we only have speech without transcription, ASR generates the most likely transcription from the speech data, and then TTS uses the generated transcription to reconstruct the original speech features. However, in previous papers, we just limited our back-propagation to the closest module, which is the TTS part. One reason is that back-propagating the error through the ASR is challenging due to the output of the ASR are discrete tokens, creating non-differentiability between the TTS and ASR. In this paper, we address this problem and describe how to thoroughly train a speech chain end-to-end for reconstruction loss using a straight-through estimator (ST). Experimental results revealed that, with sampling from ST-Gumbel-Softmax, we were able to update ASR parameters and improve the ASR performances by 11\% relative CER reduction compared to the baseline.
Multi-source translation systems translate from multiple languages to a single target language. By using information from these multiple sources, these systems achieve large gains in accuracy. To train these systems, it is necessary to have corpora with parallel text in multiple sources and the target language. However, these corpora are rarely complete in practice due to the difficulty of providing human translations in all of the relevant languages. In this paper, we propose a data augmentation approach to fill such incomplete parts using multi-source neural machine translation (NMT). In our experiments, results varied over different language combinations but significant gains were observed when using a source language similar to the target language.
In neural machine translation (NMT), the computational cost at the output layer increases with the size of the target-side vocabulary. Using a limited-size vocabulary instead may cause a significant decrease in translation quality. This trade-off is derived from a softmax-based loss function that handles in-dictionary words independently, in which word similarity is not considered. In this paper, we propose a novel NMT loss function that includes word similarity in forms of distances in a word embedding space. The proposed loss function encourages an NMT decoder to generate words close to their references in the embedding space; this helps the decoder to choose similar acceptable words when the actual best candidates are not included in the vocabulary due to its size limitation. In experiments using ASPEC Japanese-to-English and IWSLT17 English-to-French data sets, the proposed method showed improvements against a standard NMT baseline in both datasets; especially with IWSLT17 En-Fr, it achieved up to +1.72 in BLEU and +1.99 in METEOR. When the target-side vocabulary was very limited to 1,000 words, the proposed method demonstrated a substantial gain, +1.72 in METEOR with ASPEC Ja-En.