![Arxiv:1711.00294V1 [Cs.CL] 1 Nov 2017 the Audience, and the Crosstalk Language Features Humor- Entertaining Dialogue](https://data.docslib.org/img/3a60ab92a6e30910dab9bd827208bcff-1.webp)
Towards Automatic Generation of Entertaining Dialogues in Chinese Crosstalks Shikang Du, Xiaojun Wan, Yajie Ye Institute of Computer Science and Technology, The MOE Key Laboratory of Computational Linguistics Peking University, Beijing 100871, China fdusk, wanxiaojun, [email protected] Abstract performance. In each turn, the leading role usually tells sto- ries and jokes, or does some sound imitation in his utterance, Crosstalk, also known by its Chinese name xiangsheng, is a and the supporting role points out the humorous point in the traditional Chinese comedic performing art featuring jokes leading role’s performance, or even adds fuel to the leading and funny dialogues, and one of China’s most popular cul- tural elements. It is typically in the form of a dialogue be- role’s performance, making it funnier. For example, tween two performers for the purpose of bringing laughter to A: Z国'+H原,五月初五{的,我们 the audience, with one person acting as the leading comedian 应该8远怀õH原。要/¡有H原, and the other as the supporting role. Though general dialogue 我们怎H能有这三)G期b? generation has been widely explored in previous studies, it The mid-autumn festival is in memory of Qu is unknown whether such entertaining dialogues can be au- Yuan. We should keep him in mind forever, tomatically generated or not. In this paper, we for the first because his death brings us this 3-day holiday. time investigate the possibility of automatic generation of en- 这*,代价'¹?。 tertaining dialogues in Chinese crosstalks. Given the utter- B: ance of the leading comedian in each dialogue, our task aims It costs him a lot (to have a holiday). to generate the replying utterance of the supporting role. We A: 我觉得应该再多>几)G。 propose a humor-enhanced translation model to address this I think it would be better with more holidays. task and human evaluation results demonstrate the efficacy of B: £得{多少ºJ。 our proposed model. The feasibility of automatic entertaining How many people would die then! dialogue generation is also verified. In this example, B acts as the supporting role. His last response unexpectedly links the number of holiday with the Introduction number of people died, which makes the whole dialogue more funny. But in many cases, the supporting one acts as Crosstalk, also known by its Chinese name ø a go-between, gives positive response (such as “S6/Of 声/xiangsheng, is a traditional Chinese comedic per- course” or “这7/That’s why”) or negative response (such forming art, and one of China’s most popular cultural as “J?/Ah?”), and sometimes repeats key points in the elements. It is typically in the form of a dialogue between leading role’s utterance, making the narration given by the two performers, but much less often can also be a mono- leading role go smoothly (e.g. A: }6S路崎岖,@x logue by a solo performer, or even less frequently, a group 还有蒙蒙月r/ Although the road is rough, the moonlight act by multiple performers. The crosstalk language, rich is bright. B:还能看Á¹/ We can still see things on the in puns and allusions, is delivered in a rapid, bantering road.) In brief, the crosstalk between two performers can be style. The purpose of Xiangsheng is to bring laughter to considered a special and challenging dialogue form - the arXiv:1711.00294v1 [cs.CL] 1 Nov 2017 the audience, and the crosstalk language features humor- entertaining dialogue. ous dialogues (Link 1979; Moser 1990; Terence 2013; Though general dialogue generation has been widely ex- Mackerras 2013). plored and achieved great success in previous studies (Li The language style of crosstalk is just like chatting or gos- et al. 2016; Sordoni et al. 2015; Ritter, Cherry, and Dolan sip, but is more funny and humorous, especially in crosstalks 2011), it is unknown whether such entertaining dialogues given by two performers. It would be an ideal resource for can be automatically generated or not. If computers can gen- studying humor in dialogue system. erate entertaining dialogues well, the AI ability of computer However, there are some special rules in crosstalks. For will be further validated. The function of generating en- the crosstalk between two performers, one person acts as the tertaining dialogues is also very useful in many interactive leading comedian (or 逗Ï/dougen in Chinese) and the other products, making them more appealing. In this study, we for as the supporting role (or gÏ/penggen). The two perform- the first time investigate the possibility of automatic gener- ers usually stand before an audience and deliver their lines ation of entertaining dialogues in Chinese crosstalks. Given in rapid fire by turn. They echo each other in the crosstalk the utterance of the leading comedian in each dialogue, our task aims to generate the replying words of the supporting role. ignoring the entertaining characteristic of crosstalk. In ma- We propose a humor-enhanced translation model to ad- chine translation, beam search is used in decoding process, dress this special and challenging task, and the model ex- which could generate multiple candidates with scores. Usu- plicitly leverages a sub-model to measure the humorous ally only the candidate with the highest score could be ac- characteristic of a dialogue. Human evaluation results on cepted. These scores reflects the similarity of the candidate a real Chinese crosstalk dataset demonstrate the efficacy of and reference. However, just like that some question may our proposed model, which can outperform several retrieval have many different answers, there might still be acceptable, based and generation based baselines. The feasibility of au- or even unexpected but wonderful candidates with lower tomatic entertaining dialogue generation is also verified. scores. It’s a pity to get these good response ignored just The contributions of this paper are summarized as fol- because they shares little similarity with the references in lows: a limited training dataset. To exploit them, and also to ad- 1) We are the first to investigate the new task of entertain- dress the crosstalk generation problem, we propose a humor- ing dialogue generation in Chinese crosstalks. enhanced machine translation model to generate response 2) We propose a humor-enhanced translation model to ad- utterance in crosstalk. Our proposed model leverages a sub- dress this challenging task by making use of a sub-model to model to explicitly model the degree of humor of a dialogue, measure the humorous characteristic of a dialogue. and integrate it with other sub-models, as illustrated in Fig- 3) Manual evaluation is performed to verify the efficacy ure 1. of our proposed model and the feasibility of automatic en- tertaining dialogue generation. CRG model In the rest of this paper, we will first describe the details of Translation Model (M1) our proposed model and then present and discuss the evalua- Input Output tion results. After that, we introduce the related work. Lastly, Language Model (M2) we conclude this paper. Humor Model (M3) Our Generation Method Given an utterance s of the leading role (i.e. dougen) in Chi- nese crosstalks, our task aims to generate the replying utter- Xiang- Weibo Anotated sheng ance r of the supporting role (i.e. penggen), which is called Corpus Xiang- Corpus crosstalk response generation (CRG). The generated utter- sheng ance needs to be fluent and related to the leading role’s ut- terance. Moreover, it is also expected that the generated ut- terance can make the dialogue more funny and entertaining. Figure 1: General architecture of our system As mentioned earlier, our task is a special form of dia- logue generation. In recent years, there are many methods proposed for dialogue generation based on a large set of Response Generation Model training data, including the deep learning methods (espe- We get pairs of aligned utterance and response from the dia- cially sequence-to-sequence models) (Li et al. 2016). How- logue fragments in Chinese crosstalks, which are considered ever, deep learning methods usually require a large train- monolingual parallel data. The two performers echo each ing set to achieve good performance in dialogue generation other in a crosstalk, and their roles keep consistent in the tasks, which is hard to obtain for our task. So, we choose a whole crosstalk, and the leading role and the supporting role more traditional but effective way based on machine transla- of each utterance can be easily identified. Then we segment tion to address the new task of crosstalk response generation. the utterances into words. Each pair consists of a sequence penggen dougen often gives comments on ’s utterance, of words s(fs ; s ; :::; s g) spoken by the leading role, and a penggen dougen 1 2 l sometimes even retells the ’s words but in a sequence of words ref replied by the supporting role, while more humorous way. We believe that the dougen’s response the response we generated is denoted as r(fr1; r2; :::; rlg). has some potential patterns according to the utterance given Given the leading role’s utterance s, we aim to generate penggen by , and treat response generation as a monolin- the best response utterance r by using our proposed gener- gual translation problem, in which the given input (utterance ation model. The proposed generation model has three sub- dougen given by ) is treated as the foreign language and the models(M1, M2, M3): translation model, language model humorous response as the source language. Machine trans- and humor model. We will introduce each sub-model and lation (MT) has already been successfully used in response then introduce the framework of model combination. generation (Ritter, Cherry, and Dolan 2011), in which input post was seen as a sequence of words, and word or phrase Translation Model (M1) The translation model translates based translation was made to generate another sequence of the given leading role’s utterance s into a sequence of words words as response.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages8 Page
-
File Size-