This study tackles unsupervised domain adaptation of reading comprehension (UDARC). Reading comprehension (RC) is a task to learn the capability for question answering with textual sources. State-of-the-art models on RC still do not have general linguistic intelligence; i.e., their accuracy worsens for out-domain datasets that are not used in the training. We hypothesize that this discrepancy is caused by a lack of the language modeling (LM) capability for the out-domain. The UDARC task allows models to use supervised RC training data in the source domain and only unlabeled passages in the target domain. To solve the UDARC problem, we provide two domain adaptation models. The first one learns the out-domain LM and in-domain RC task sequentially. The second one is the proposed model that uses a multi-task learning approach of LM and RC. The models can retain both the RC capability acquired from the supervised data in the source domain and the LM capability from the unlabeled data in the target domain. We evaluated the models on UDARC with five datasets in different domains. The models outperformed the model without domain adaptation. In particular, the proposed model yielded an improvement of 4.3/4.2 points in EM/F1 in an unseen biomedical domain.

Kosuke Nishida, Kyosuke Nishida, Masaaki Nagata, Atsushi Otsuka, Itsumi Saito, Hisako Asano, Junji Tomita. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

Kyosuke Nishida, Itsumi Saito, Kosuke Nishida, Kazutoshi Shinoda, Atsushi Otsuka, Hisako Asano, Junji Tomita. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

This study considers the task of machine reading at scale (MRS) wherein, given a question, a system first performs the information retrieval (IR) task of finding relevant passages in a knowledge source and then carries out the reading comprehension (RC) task of extracting an answer span from the passages. Previous MRS studies, in which the IR component was trained without considering answer spans, struggled to accurately find a small number of relevant passages from a large set of passages. In this paper, we propose a simple and effective approach that incorporates the IR and RC tasks by using supervised multi-task learning in order that the IR component can be trained by considering answer spans. Experimental results on the standard benchmark, answering SQuAD questions using the full Wikipedia as the knowledge source, showed that our model achieved state-of-the-art performance. Moreover, we thoroughly evaluated the individual contributions of our model components with our new Japanese dataset and SQuAD. The results showed significant improvements in the IR task and provided a new perspective on IR for RC: it is effective to teach which part of the passage answers the question rather than to give only a relevance score to the whole passage.

This study focuses on acquisition of commonsense knowledge.A previous study proposed a commonsense knowledge base completion (CKB completion) method that predicts a confidence score of triplet-style knowledge for improving the coverage of CKBs.To improve the accuracy of CKB completion and expand the size of CKBs, we formulate a new commonsense knowledge base generation task (CKB generation) and propose a joint learning method that incorporates both CKB completion and CKB generation.Experimental results show that the joint learning method improved completion accuracy and the generation model created reasonable knowledge.Our generation model could also be used to augment data and improve the accuracy of completion.

We explored the gaze behavior towards the end of utterances and dialogue act (DA), i.e., verbal-behavior information indicating the intension of an utterance, during turn-keeping/changing to estimate empathy skill levels in multiparty discussions. This is the first attempt to explore the relationship between such a combination. First, we collected data on Davis' Interpersonal Reactivity Index (which measures empathy skill level), utterances that include the DA categories of Provision, Self-disclosure, Empathy, Turn-yielding, and Others, and gaze behavior from participants in four-person discussions. The results of analysis indicate that the gaze behavior accompanying utterances that include these DA categories during turn-keeping/changing differs in accordance with people's empathy skill levels. The most noteworthy result was that speakers with low empathy skill levels tend to avoid making eye contact with the listener when the DA category is Self-disclosure during turn-keeping. However, they tend to maintain eye contact when the DA category is Empathy. A listener who has a high empathy skill level often looks away from the speaker during turn-changing when the DA category of a speaker's utterance is Provision or Empathy. There was also no difference in gaze behavior between empathy skill levels when the DA category of the speaker's utterance was turn-yielding. From these findings, we constructed and evaluated models for estimating empathy skill level using gaze behavior and DA information. The evaluation results indicate that using both gaze behavior and DA during turn-keeping/changing is effective for estimating an individual's empathy skill level in multi-party discussions.

We propose a model to automatically generate whole body motions accompanying utterances at appropriate times, similar to humans, by using various types of natural-language-analysis information obtained from spoken language. Specifically, we focus on the co-occurrence relationship between various types of natural-language-analysis information such as words included in the spoken language, parts of speech, a thesaurus, word positions, dialogue acts of the spoken language, and human motions. Our model automatically generates nods, head postures, facial expressions, hand gestures, and upper-body posture using such information. We first recorded a two-person dialogue and constructed a multimodal corpus including utterance and whole body motion information. Next, using the constructed corpus, we constructed our model for generating a motion for each phrase unit using machine learning and using words, parts of speech, a thesaurus, word positions, and speech acts of the entire spoken language as inputs. These types of natural-language-analysis information were useful for motion generation. The effectiveness of our model was verified through a subjective experiment using a virtual conversational agent. As a result, the agent's body motions and impressions regarding naturalness of motion, degree of coincidence between utterance and motion, humanness of the agent, and likability of the agent improved with our model.

A virtual agent in a dialogue system should express appropriate body motions according to utterances and effectively communicate with a user. We previously proposed a generation model of whole body motions such as head direction, nodding, facial expressions, hand gestures, and upper-body posture accompanying utterances at appropriate times similar to humans by using various types of natural-language-analysis information obtained from spoken language. As an attempt to promote this model, we constructed an API that can easily generate motions by using the generation model and constructed a demonstration system that can automatically control a virtual agent from only the spoken language. When inputting an arbitrary utterance language, synthesized sound and motion information are acquired from the speech synthesizer and motion-generation API, and the vocalization of the virtual agent and animated motion are generated. A dialog agent that is more attractive and can communicate smoothly by automatically generating natural motions is expected.

We propose a model to generate head nods accompanying an utterance from natural language. To the best of our knowledge, previous models generated simple nods from the final words at the end of an utterance, i.e., using bag of words. We focused on various text analyzed using various types of language information such as dialog act, part of speech, a large-scale Japanese thesaurus, and word position in a sentence. We also generated detailed parameters of speaker's nodding presence, frequency, and depth, which was the first attempt to do so. First, we compiled a Japanese corpus of 24 dialogues including utterance and nod information. Next, using the corpus, we constructed our generation model that estimates nodding presence, frequency, and depth, during a phrase by using such various types of language information as well as bag of words. The results indicate that our model outperformed simple automatic nod-generating models using only bag of words and chance level. The results also indicate that dialog act, part of speech, the large-scale Japanese thesaurus, and word position are useful for generating nods. We also evaluated, through subjective evaluation, if our nod-generation model is useful with conversational agents. The results show that the nodding generated with our model improves user impressions of naturalness, humanness, likability, and reliability toward a conversational agent.

Having consistent personalities is important for chatbots if we want them to be believable.Typically, many questionanswer pairs are prepared by hand for achieving consistent responses; however, the creation of such pairs is costly.In this study, our goal is to collect a large number of question-answer pairs for a particular character by using role playbased question-answering in which multiple users play the roles of certain characters and respond to questions by online users.Focusing on two famous characters, we conducted a large-scale experiment to collect question-answer pairs by using real users.We evaluated the effectiveness of role play-based questionanswering and found that, by using our proposed method, the collected pairs lead to good-quality chatbots that exhibit consistent personalities.

To provide a better discussion experience in current argumentative dialogue systems, it is necessary for the user to feel motivated to participate, even if the system already responds appropriately.In this paper, we propose a method that can smoothly introduce argumentative dialogue by inserting an initial discourse, consisting of question-answer pairs concerning personality.The system can induce interest of the users prior to agreement or disagreement during the main discourse.By disclosing their interests, the users will feel familiarity and motivation to further engage in the argumentative dialogue and understand the system's intent.To verify the effectiveness of a questionanswer dialogue inserted before the argument, a subjective experiment was conducted using a text chat interface.The results suggest that inserting the questionanswer dialogue enhances familiarity and naturalness.Notably, the results suggest that women more than men regard the dialogue as more natural and the argument as deepened, following an exchange concerning personality.

Social media texts, such as tweets from Twitter, contain many types of non-standard tokens, and the number of normalization approaches for handling such noisy text has been increasing. We present a method for automatically extracting pairs of a variant word and its normal form from unsegmented text on the basis of a pair-wise similarity approach. We incorporated the acquired variant-normalization pairs into Japanese morphological analysis. The experimental results show that our method can extract widely covered variants from large Twitter data and improve the recall of normalization without degrading the overall accuracy of Japanese morphological analysis.

Knowledge discovery from a large volumes of texts usually requires many complex analysis steps. The graph-based text representation model has been proposed to simplify the steps. The model represents texts in a formal manner, Subject Graphs, and provides text handling operations whose inputs and outputs are identical in form, i.e. a set of subject graphs, so they can be combined in any order. A subject graph uses node weight to represent the significance of each term, and link weight to represent that of each term-term association. This paper concentrates on the algorithms for making subject graphs and calculating the similarity between them. An evaluation shows that Subject Graphs can calculate the similarity between texts more precisely than term vectors, since they incorporate the significance of association between terms.