(PDF) When Your Users Are Not Serious - Using Web-based Associations Affect and Humor for Generating Appropriate Utterances for

114 人工知能学会論文誌 25 巻 1 号 SP-L（2010 年） 特集論文 When Your Users Are Not Serious Using Web-based Associations, Affect and Humor for Generating Appropriate Utterances for Inappropriate Input Rafal Rzepka Hokkaido University

, http://sig.media...jp/˜kabura/ Shinsuke Higuchi (affiliation as previous author) shin

, http://sig.media...jp/˜shin h/ Michal Ptaszynski (affiliation as previous author)

, http://sig.media...jp/˜ptaszynski/ Pawel Dybala (affiliation as previous author)

, http://sig.media...jp/˜paweldybala/ Kenji Araki (affiliation as previous author)

, http://sig.media...jp/˜araki/ keywords: non-task oriented dialogue, web-mining, affect analysis, pun generation Summary In this paper we propose a method for generating simple but semantically correct replies to user inputs which are not related to a given task of a task-oriented information kiosk or any other natural language interface placed in a public place. We describe our method for retrieving meaningful associations from the Web and adding modality based on chatlog data. After showing the results of the evaluation experiments, we introduce an implementation of an affect analysis algorithm and pun generator to increase users’ satisfaction level. 1. Introduction be?”, ”I’m in love” or ”you are ugly” appearing in infor- mation kiosks logs should not simply be answered by ”I Together with rapid technological development, more don’t understand”, ”please repeat” or ”I have no info on and more natural language processing systems are appear- this topic”. ing in streets and buildings, where anyone passing by is A system developer could add a chatbot which would able to ask an automatic guide or an information kiosk react after discovering whether keywords belong or not to about route or place details. Considerable progress has a task. However, such programs have obvious problems. been achieved in many subfields concerning task-oriented Two well-known examples of non-task-oriented dialogue dialogue systems [Liu 03, Reitter 06]. However, there is systems are ELIZA [Weizenbaum 66] and A.L.I.C.E∗1 . no research on solutions to the common problem of reply- Both systems and their countless imitators use many rules ing to inputs which are not related to a task as described coded by hand. ELIZA is able to generate a response to in the literature [Gustafson 00, Kopp 05]. We believe any input, but these responses are only information re- that the main reason for this is that an unrestricted do- quests which do not provide any new information to the main is disproportionately difficult compared to the pos- user. In the case of A.L.I.C.E, the knowledge resource is sible problems such input could cause. It is very hard limited to the existing database. Creating such databases to predict the contents and topics of user utterances, and is costly and a programmer must learn the AIML mark-up therefore it is almost impossible to prepare conversational language to build it. Although there have been attempts at scenarios. Furthermore, scenarios usually need specific updating AIML databases automatically, [De Pietro 05], goals to be useful. However, in our opinion, by combin- the scale was rather limited. ing task-oriented dialogue systems with non-task-oriented As mentioned above, these examples and many other ones, we are able to create more human-like architectures ”chatbots” need hand-crafted rules, and are thus often ig- which should be more trustful, and give better impressions nored by computer scientists and rarely become a research of a company or organization which sets the automatic ∗1 Wallace, R. The Anatomy of A.L.I.C.E. informer. Utterances such as ”what’s the weather gonna http://www.alicebot.org/anatomy.html. When Your Users Are Not Serious 115 topic. However, they have proved to be useful for e-learning utterances. We use Google∗2 search engine snippets to ex- [De Pietro 05] and language acquisition [Araki 06] sup- tract word associations in real time without using earlier port. prepared resources, such as off-line databases. Building a system using automatic methods, like we do, seems to be the most realistic strategy for inputs of unre- 2 ·1 Extracting Word Associations from the Web stricted domains. Considering the large cost of developing In the first step, the system analyzes user utterances us- a program that can talk about any topic, it is appealing to ing the morphological analyzer MeCab∗3 in order to spot turn to the huge, cheap textual resource that is the Internet. query keywords for extracting word associations lists. We At this very moment, millions of people [Kumar 03] are define nouns, verbs, adjectives, and unknown words as updating their blogs and writing articles on every possible query keywords. The reason we chose these word classes topic. These are available on the Web which we can ac- is that these word classes can be treated as important and, cess at any time, and in a faster and faster manner as search to some extent, describe the context. We define a noun engines grow more and more efficient. Thus, the Web is as the longest set of nouns in a compound noun. For ex- well suited to extracting word associations from user utter- ample, the compound noun shizen gengo shori∗9 (natural ances from conversations with a topic-free dialogue sys- language processing) is treated by MeCab as three words: tem. We describe a system making use of this, details of (shizen - natural), (gengo - language) and (shori - process- which were presented at [Higuchi 09] and demonstrated ing). Our system, however, treats it as one noun. at [Rzepka 09]. It automatically extracts word associa- In the next step, the system uses these keywords as query tion lists using all keywords in a given utterance without words for the Google search engine. The system extracts choosing a specific one (which most other systems that the nouns from the search results and sorts them in fre- ignore the context do) then generates a reply using one as- quency order. This process is based on the idea that words sociation from the strongest associations found. Modality which co-occur frequently with the input words are of high is then added to the reply, resulting in the system’s output. relevance to them. The number of extracted snippets is Our system is built upon the idea that human utterances 500. This value was set experimentally, taking the pro- consist of a proposition and a modality [Nitta 89]. In this cessing time and output quality into account. The top ten paper we present an algorithm for extracting word asso- words of a list are treated as word associations, see Table 1 ciations from the Web and a method for adding modality for an example of a noun group. to statements. We evaluate both the word associations and Table 1 Example of word associations extracted for a user utterance the use of modality. We also suggest some future possible Sapporo wa samui. (Sapporo (city) is cold.) extensions of the system and show the results of a small experiment with adding humor to the system. Association frequency ranking: 1 yuki (snow) 52 The system described in this paper works for Japanese 2 fuyu (winter) 50 and uses text as input and output. Although the final goal 3 kion (temperature) 16 of our research is to help develop a freely talking car nav- 4 jiki (season) 12 igation system which, by using chatting abilities, can help 5 Tokyo (Tokyo) 12 to prevent drowsiness while driving, at this stage of de- 6 tenki (weather) 11 velopment we are concentrating on proposition generation 7 chiiki (area) 10 and modality processing. Therefore, at present we work 8 heya (room) 10 only with text. We plan to combine this project with re- search on in-car voice recognition and generation. To the best of the authors’ knowledge, their system is 2 ·2 Evaluation the only one that does not use preprocessed data, and gen- We asked evaluators to use our system and to evalu- erates new phrases without citing existing ones previously ate the correctness of word lists generated by the system. created by humans. First, an evaluator freely inputs an utterance, for which the system retrieves ten association words. Next, he or 2. Extracting Word Associations she rated these words using a scale of one to three with ∗2 Google, http://www.google.co.jp/ ∗3 MeCab: Yet Another Part-of-Speech and Morphological Ana- In this section, we present a method for automatic ex- lyzer, http://mecab.sourceforge.jp/ traction of word associations based on keywords from user ∗9 All Japanese transcriptions will be written in italics. 116 人工知能学会論文誌 25 巻 1 号 SP-L（2010 年） 3 meaning ”perfectly correct”, 2 -”partially correct” and 1 - ”incorrect”. In this research we consider words that receive 2 or 3 as usable. Three evaluators repeated the ex- periment ten times, so the final amount of evaluated words was 300. Table 2 shows the top 10 words, sorted by the frequency of appearance. Table 3 shows the top 5 words. What constitutes a correct word association was left to each evaluator to decide subjectively, as in a casual con- versation setting, associations are hard to define strictly. Table 2 Top 10 word associations score evaluators (A, B, C) total 3 40, 52, 57 149 2 37, 17, 27 81 1 23, 31, 16 70 usability[%] 77, 69, 84 77 Fig. 1 System flow 3 ·1 Extraction of Keywords from User Utterances Table 3 Top 5 word associations The system applies morphological analysis to the user’s score evaluators (A, B, C) total 3 20, 29, 36 85 utterances in the same way as described in Section 2 ·1 2 17, 9, 10 36 and extracts keywords based on part of speech. These 1 13, 12, 4 29 keywords create association groups by using methods also usability[%] 74, 76, 92 81 introduced in the same section. As shown in Table 2, approximately 77% of the word 3 ·2 Generation of Proposition Using Word Associations associations were judged as usable, but there were individ- Using the word associations, the system generates the ual differences between the evaluators. This shows that the proposition of a sentence to be used as a reply to the user definition of word associations is different for each partic- input. A proposition is an expression representing an ob- ipant. Table 3 shows that approximately 80% of the word jective, declarative statement, which does not contain any associations were judged as usable. It is thus highly likely form of affect the modality usually conveys. The proposi- that the top words from the frequency lists are correct as- tion is generated by applying word associations to a propo- sociations. The results show that automatic extracting of sition template like [(noun) (topic indicating particle wa) word associations using a Web search engine is feasible. (adjective)]. We prepared 8 proposition templates manu- The main reason for extracting word associations from the ally (see Table 4). The templates were chosen subjectively Web is that, due to this method, the system can handle after examining statistics from IRC∗4 chat logs, checking new information, proper names, technical terms, etc. As their flexibility to be grammatically correct after combin- the system uses snippets from the search engine, the word ing them with different parts of speech. In order to ensure association extraction takes no more than few seconds. greater diversity of utterances, the proposition templates are applied in a predetermined order exactly the same as the one shown in Table 4. However, since the generated 3. General Description of the System proposition is not always a natural statement, the system uses exact matching searches of the whole phrase in a The system generates responses in the following pattern search engine to check the naturalness of each proposi- (Figure 1 shows the system flow): tion. If the frequency of occurrence of the proposition is • Extraction of keywords from user utterance low, it is defined as unnatural and deleted. This processing • Extraction of word associations from the Web is based on the idea that the phrases existing on the Web • Generation of sentence proposition using word asso- in large numbers are most probably correct grammatically ciations ∗4 Internet Relay Chat Protocol, http://www.irchelp.org/ • Addition of modality to the sentence proposition irchelp/rfc/rfc.html When Your Users Are Not Serious 117 and semantically. If an unnatural, low frequency proposi- order. The words appearing at the top of the list were tion is generated, the system repeats the proposition gener- correct, but even the ones appearing only once were still ation using the same preposition template but with differ- deemed as usable. For example, the question expression ent, randomly chosen top associations. In this experiment ”janakatta deshita-kke?” is a correct expression, but ap- the system used propositions for which the hit number ex- peared only once in the 100,000 utterances. Hence, we ceeded 1,000 hits using Google. confirmed that chat logs include various modality expres- sions, and only a few of them are incorrect. Therefore the Table 4 Proposition templates system randomly chooses from the whole set of correct (noun) (wa) (adjective) modalities and sets them with the flexible proposition pat- (noun) (ga) (adjective) terns we picked up beforehand. See Table 5 and Table 6 (noun) (ga) (verb) for examples. (noun) (wa) (verb) Table 5 Examples of informative expression modality (so-re) (wa) (verb) (noun) informative expression frequency (adjective) maa - kedo 21 (verb) (Well , it can be said - but -) maa - dana 16 (Well , it can be said -) 3 ·3 Adding Modality to the Propositions maa - desu-ga 16 Finally, the system adds modality to the generated propo- (Well , it appears that -) sition. By modality we mean a set of grammatical and soko-de - desu-yo 15 pragmatic rules to express subjective judgments and atti- (Here , it is said that -) tudes. It is realized through adverbs at the end of a sen- maa - da-ga 14 tence [Nitta 89]. In our system, a pair of sentence-head (Well , it can be said - but -) and sentence-end auxiliary verbs are defined as ”modal- maa - desu-yo 12 ity”. (Well , it is that -) § 1 Extracting Modality There is no standard definition of what constitutes modal- Table 6 Examples of question modality ity in Japanese. In this paper, modality of casual conver- sation is classified into questions and informative expres- question frequency sions. Questions are expressions that request information desuka? 232 from the user. Informative expressions are expressions (Is it that - ?) that transmit information to the user. Patterns for these kana? 90 modalities are extracted automatically from IRC chat logs (Maybe - ?) (100,000 utterances) in advance. Modality patterns are ex- da-kke? 87 tracted in the following ways: (Is it right that - ?) • a pair of a grammatical particle and an auxiliary verb masu-ka? 69 placed at the end of a sentence (Is it that - ?) • sentences with question marks are defined as ques- nano? 68 tions (Is it that - ?) • adverbs, emotive words, and connectives at the begin- toka? 55 ning of a sentence are defined as informative expres- ( - , isn’t it ?) sions • candidate patterns thus obtained are sorted by fre- § 2 Adding Modality quency The system adds the modality from Section 3 ·3 ·1 to 685 patterns were obtained for informative expressions. the proposition from Section 3 ·2 to generate the system 550 of these informative expression patterns were consid- output. This process is based on the idea that human ut- ered by the authors as correct (80%). For questions, 396 terances consist of proposition and modality. A modal- patterns were obtained, and 292 patterns (73%) were eval- ity pattern is selected randomly. For example, if the sys- uated as correct. We sorted these candidates in frequency tem generates the proposition fuyu wa samui (Winter is 118 人工知能学会論文誌 25 巻 1 号 SP-L（2010 年） Table 8 Examples of dialogues with system β cold.) and selects the modality iyaa ... desu-yo (Ooh ... isn’t it?), the generated output will be iyaa, fuyu-wa samui user: Nanika sukina tabemono aru? desu-yo (Winter is cold, you know). However, there is a (What food do you like ?) possibility that the system generates unnatural output like system: Maa, tabemono-wa oishii desu. fuyu-wa samui dayo-ne, depending on the pair of propo- (Well, food tastes good.) sition and modality. Regarding this problem, the system user: Saikin-wa osake-mo uses the Google search engine to filter out unnatural out- sukini nattekitanda. put. The system performs a phrase search on the end of (Recently, I begin to like alcohol too.) the sentence. If the number of search hits are higher than system: Demo, sake-wa yowai-no-yo-ne. the threshold, the output is judged as correct. If the num- (But, I can’t drink much.) ber of a search hits is lower than the threshold, the output user: Yowai-n-da... Sugu yochau-no? is judged as incorrect and discarded, and a new reply is (Really...? generated. Here, we set the threshold to 100 hits. Do you get drunk quickly?) 4. Experimental Results system: Maa, sore-wa sugoi-wa-ne. (Whow, that’s amazing.) We used system α, generating only the proposition, and user: Boku-wa wari-to osake-wa system β, generating both proposition and modality. 5 par- tsuyoi hou-nan-da. ticipants used each system for conversations of 10 turns (I have a rather high and evaluated the conversations on a 5-point scale. Eval- tolerance for alcohol.) uation criteria were ”willingness to continue the conver- system: Iya, sore-wa tsuyoi-n-de. sation” (A), ”grammatical naturalness of dialogues” (B), (Well, it is strong, I can say.) ”semantical naturalness of dialogues” (C), ”vocabulary rich- ness” (D), ”knowledge richness” (E), and ”humanity of user: Kyou-wa sukoshi hadazamui-yo-ne. the system” (F)∗5 . Table 7 shows average scores for the Natsu-nanoni. evaluations of each system. System β that uses modality (It is a little cold today. scored much higher than system α. Table 8 shows exam- Even though it is summer.) ples of actual dialogue. In the evaluation, the participants system: Asa-wa hadazamui-toka ? expressed the opinion that an utterance like (xx ha yy) is (Isn’t it chilly in morning?) unnatural and using a modality like (maa), (moo) is very user: Asa-kara hadazamukatta-ne. natural. Thus we can say that modality expressions make (It was chilly since morning.) the utterances of the system more natural. system: Aa kyou-wa hayai-na. (Ah, it’s early. ) 5. System Expandability Examples The simplicity of our system, real-time processing capa- 5 ·1 Testing Affect Analysis bilities and promising results showing that users do not get Ptaszynski et al. [Ptaszynski 08] have developed a method bored quickly, encouraged us to perform trials with other for affect analysis of Japanese text called ML-Ask. Their ongoing projects, and to experiment with the system as a method is based on cross-referencing lexical emotive ele- platform for adding various modules and algorithms in or- ments with emotive expressions appearing in text. In the der to make the utterances more natural. By using our sys- process of analysis, first a general emotive context is deter- tem, it is possible to perform tests determining whether a mined and then the specific types of emotional states con- new idea will support or improve Human-Computer inter- veyed in an utterance are extracted. Their method achieved action or not. Here we will briefly describe two such trials human level performance in determining the emotiveness - one on guessing emotive values of utterances, and one of utterances, and 85% of human level performance in ex- on improving the system’s overall evaluation by adding a tracting the specific types of emotion was achieved by pun generator. adding a Web mining technique. They used our base- ∗5 The evaluation sheets were created by setting following polari- line and humor-equipped systems to prove that their af- ties: A - I do not want to continue (1) / I wish to continue (5); B,C - unnatural (1) / natural (5); D,E - not rich (1) / rich (5); F - not fect analysis could replace human evaluators. Ability to human-like (1) / human-like. recognize user emotions is a very important indication of When Your Users Are Not Serious 119 Table 7 Evaluation Results System α (proposition) system β (proposition + modality) Evaluation criteria A B C D E F A B C D E F Evaluator a 1 3 2 2 4 2 4 4 3 4 3 5 Evaluator b 1 3 1 2 1 1 4 4 4 5 4 3 Evaluator c 1 2 1 2 1 1 1 2 1 2 1 1 Evaluator d 1 3 1 3 1 2 4 3 1 3 3 4 Evaluator e 1 4 1 1 2 1 3 2 2 4 5 4 Average 1.0 3.0 1.2 2.0 1.8 1.4 3.2 3.0 2.2 3.6 3.2 3.4 intelligence, and is becoming a crucial part of our system, that the system was human-like? G) Do you think the sys- needed for modules such as the one introduced in the next tem tried to make the dialogue more funny and interest- subsection. ing? and H) Did you find the system’s utterances interest- ing and funny?∗6 Answers were given on a 5-point scale 5 ·2 Improving the System Using Humor and the results are shown in Table 9. In this trial, an experiment showing that humor can im- A third-person evaluation experiment was also performed. prove a non-task oriented conversational system’s overall Again, the humor-equipped system scored higher than the performance was conducted. non-humor one. The question asked in this evaluation § 1 Implementing PUNDA system was: ”Which dialogue do you find most interesting and By using a simplified version of Dybala’s PUNDA sys- funny?”. Evaluators could choose between 3 options: Dia- tem [Dybala 08], a pun-generator was added to our base- logue 1 (Baseline system’s first 3 turns), Dialogue 2 (Humor- line system. The PUNDA algorithm consists of two parts: equipped system’s first 3 turns, with system’s third re- a Candidate Selection Algorithm and a Sentence Integra- sponse replaced by pun generator’s output) and Dialogue tion Engine. The former generates a candidate for a pun 3 (the first 3 turns of the baseline system with joking abil- by analyzing an input utterance and selecting words or ity). Among 25 evaluators, only 5 (20%) responded that phrases that could be transformed into a pun by one of four Dialogue 1 was most interesting and funny. 10 chose Di- generation patterns: homophony, initial mora addition, in- alogue 2 and the other 10 chose Dialogue 3 (40% respec- ternal mora addition or final mora addition. The latter part tively). This means that both of the humor equipped dia- generates a sentence including the candidate extracted in logues received evaluations double that of non-humor di- the previous step. To make the system’s response more alogue. relavant to the user’s input, each sentence which includes a joke starts with the pattern [base phrase] to ieba (”Speak- 5 ·3 Timing Problem - Combining Affect and Humor ing of [base phrase]”). The remaining part of the sentence In the experiment described above, the system tells jokes is extracted from the Web, where the candidate is used as a (puns) at every third turn of dialogue. In future, timing query word and the list of sentences including this word is problems could be solved by replacing this rule with a tim- retrieved. Then the shortest sentence with an exclamation ing algorithm based on emotiveness analysis of users’ ut- mark is selected, as most jokes convey some emotions. terances. To perform the analysis, Ptaszynski’s idea [Ptaszyn- When the candidate list is empty, the system selects one ski 08] mentioned in the ”Testing Affect Analysis” subsec- random pun from a pun database. tion could be useful as it detects user’s emotional states § 2 Experiment results from the textual layer of speech, by which it can discover After using one of the systems (baseline or humor-equipped), if an utterance is positive, negative or neutral. During con- evaluators were asked to evaluate both systems’ perfor- versation with the humor-equipped talking system, each mances by answering the following questions: A) Do you user’s utterance would be analyzed with ML-Ask, then want to continue the dialogue? B) Was the system’s utter- based on the analysis results, the system would decide ances grammatically natural? C) Was the system’s utter- whether it is appropriate to tell a pun. The decision about ances semantically natural? D) Was the system’s vocab- ∗6 Polarity answers were the same as in section 4 for A-F; G - did ulary rich? E) Did you get an impression that the system not try at all (1) / tried hard (5); H - not funny at all (1) / very funny possesses any knowledge? F) Did you get an impression (5). 120 人工知能学会論文誌 25 巻 1 号 SP-L（2010 年） Table 9 Results of humor experiments Evaluation Criteria A B C D E F G H Baseline System 3.0 2.2 2.4 2.4 2.0 2.8 2.2 2.8 With pun generator 3.2 3.0 2.8 2.8 2.2 3.0 3.4 3.6 appropriateness of pun-telling would be made based on 6. Conclusion and Future Work conclusions drawn from broad humor literature [Martin 96, Newman 96, Cann 99, Danzer 90, Labott 87, Dien- In this research, we investigated if word associations ex- stbier 96, Vilaythong 03, Takayanagi 07, Robinson 91, tracted automatically from the Web are reasonable (i.e., Frankl 60, Hennman 00, Bonanno 97]: semantically on topic) and if they can be successfully used a) If the user’s emotive state is negative (stress, depres- in non-task-oriented dialogue systems. We also imple- sion, anxiety etc.), a pun can be told to help him/her deal mented an extraction module which is able to automat- with it. For example, if the user says: ”You know, I’m ically generate in real-time responses to user utterances, feeling kind of down today...”, the system, after detecting by generating a proposition and adding modality retrieved negative emotion (sadness), could tell a joke to make the from IRC chat logs. We conducted evaluation experiments user’s mood better. on the overall influence of the modality usage and it im- b) If the user’s state is neutral, a pun can be told to induce proved the system. Therefore, we showed that it is possi- a good mood. ble to construct a dialogue system that automatically gen- These rules, however, should be limited to situations in erates understandable on-topic utterances without the need which there would be no risk of inducing negative instead to create vast amounts of rules and data beforehand. We of positive reaction. also confirmed that our system’s performance can be im- The flow of such a combined algorithm is shown in Fig- proved by joke generation and affect analysis and we in- ure 2. troduced an idea regarding how these two topics could be combined to achieve an even more natural human-computer interface. There is still a lot of work to be done. It is necessary for a non-task-oriented dialogue system to obtain not only word associations, but also different kinds of knowledge - of the user’s preferences or of dialogue itself - for ex- ample, conversational strategies. At this moment, the sys- tem generates utterances by applying word associations to the proposition templates and adding modality. We also need to consider semantics, speech acts and context more deeply to create a more advanced system. Finally, the sys- tem needs to recognize not only keywords, but also the user’s modality. We assume that the affect recognition mentioned above will help us to achieve this goal in near future and this will be our next step. Acknowledgments This work was partially supported by the Research Grant from the Nissan Science Foundation. We are especially grateful to Mr. Takashi Sunda from Mobility Laboratory at Nissan Research Center for his valuable comments. ♦ References ♦ Fig. 2 Three systems combined into one chatbot. [Araki 06] K. Araki, and M. Kuroda (2006) Generality of a Spoken Dialogue System Using SeGA-IL for Different Languages, Proceed- ings of the IASTED International Conference COMPUTER INTEL- LIGENCE, pp. 70-75, San Francisco, CA, U.S.A. , 2006. When Your Users Are Not Serious 121 [Bonanno 97] Bonanno, G. A. and Keltner, D.: Facial expressions of No. 2, pp.51-57 emotion and the course of conjugal bereavement, Journal of Abnor- [Vilaythong 03] Vilaythong, A. P., R. C. Arnau, D. H. Rosen, and mal Psychology, 106(1), pp.126-137, 1997. Mascard, N. Humor and hope: Can humor increase hope? In: Hu- [Cann 99] Cann, A., Holt, K. and Calhoun, L. G.: The roles of humor mor: International Journal of Humor Research 16(1), pp. 79.89, and sense of humor in response to stressors, Humor: International 2003. Journal of Humor Research,12(2), pp. 177-193, 1999. [Weizenbaum 66] J. Weizenbaum, ELIZA - computer program for the [Danzer 90] Danzer, A., Dale, J. A. and Klions, H. L.: Effect of expo- study of natural language communication between man and machine, sure to humorous stimuli on induced depression, Psychological Re- Communication (ACM), vol.9, no.1 ,pp. 36-45,1966. ports, 66(3, Pt 1), pp. 1027-1036, 1990. [De Pietro 05] M. De Pietro, M. De Rose, and G. Frontera(2005) Au- 〔担当委員：山本 健一〕 tomatic Update of AIML Knowledge Base in E-Learning Environ- ment, In Proceedings of Computers and Advanced Technology in Received May 1, 2009. Education. , Oranjestad, Aruba, August 29-31, 2005. [Dienstbier 96] Dienstbier, R. A., The impact of humor on energy, tension, task choices, and attributions: Exploring hypotheses from toughness theory, Motivation and Emotion, 19(4), pp. 255-267, 1995. [Dybala 08] P. Dybala, M. Ptaszynski, R. Rzepka and K. Araki (2008) Author’s Profile Proceedings of LIBM’08 First International Workshop on Laughter in Interaction and Body Movement, pp. 46-51, 2008 Rzepka, Rafal (Member) [Frankl 60] Frankl, V. E.: Man’s search for meaning, New York Washington Press, 1960. Born in Szczecin, Poland in 1974. Received M.A. from [Gustafson 00] J. Gustafson and L. Bell (2000). Speech technology Adam Mickiewicz University in Poznan, Poland in 1999 and Ph.D. from Hokkaido University, Japan in 2004. Cur- on trial: Experiences from the August system. In Natural Language rently he is an assistant professor at the Graduate School of Engineering, 1 (1), pp. 1-15. Information Science and Technology, Hokkaido University. [Hennman 00] Hennman, L. D.: Humor as a coping mechanism: Research interests: NLP, Web-mining, Commonsense Re- Lessons from POWs, Humor: International Journal of Humor Re- trieval, Affect Analysis and Machine Ethics. Member of: search, 14(1), pp. 83-04, 2000. AAAI, ACL, IPSJ, IEICE, JCSS and NLP. [Higuchi 09] Higuchi, S., Rzepka, R. and Araki, K. 2009 A Casual Conversation System Using Modality and Word Associations Re- Higuchi, Shinsuke trieved from the Web, Proceedings of the 2008 Conference on Em- Born in Sapporo, Japan in 1983. He is a second year Mas- pirical Methods in Natural Language Processing, pp. 382-390, Hon- ter Course student at the Graduate School of Information olulu, USA, October 2008 Science and Technology, Hokkaido University, Japan. His [Kopp 05] Kopp, S., Gesellensetter, L., Kramer, N., & Wachsmuth, research interests include Natural Language Processing, Di- alogue Processing, and Web-mining. I. (2005). A Conversational Agent as Museum Guide - Design and Evaluation of a Real-World Application. In Panayiotopoulos et al., (Eds.), Intelligent Virtual Agents, LNAI 3661, pp. 329-343. [Kumar 03] Ravi Kumar, Jasmine Novak, Prabhakar Raghavan, An- Ptaszynski, Michal (Member) drew Tomkins :On the Bursty Evolution of Blogspace, The Twelfth Born in Wroclaw, Poland in 1981. He received his M.A. International World Wide Web Conference (2003). from Adam Mickiewicz University in Poznan, Poland in [Labott 87] Labott, S. M. and Martin, R. B.: The stressmoderating 2006. He was a research student at Otaru University of effects of weeping and humor, Journal of Human Stress, 13(4), pp. Commerce, and since 2007 he has been studying towards 159-164, 1987. his Ph.D. degree at the Graduate School of Information Sci- [Liu 03] B. Liu, L. Du, S. Yu. The method of building expectation ence and Technology, Hokkaido University, Japan. His re- search interests include Natural Language Processing, Dia- model in task-oriented dialogue systems and its realization algo- logue Processing, Affect Analysis, Sentiment Analysis, HCI rithms. Natural Language Processing and Knowledge Engineering, and Information Retrieval. He is a member of SOFT and NLP. 2003. Proceedings. 2003 International Conference on, pp. 174- 179 [Martin 96] Martin, R. A. and Lefcourt, H. M.: Humor and life stress. Dybala, Pawel Antidote to adversity. New York, Springer Verlag, 1996 Born in Ostrow Wielkopolski, Poland in 1981. He received [Newman 96] Newman, M. G. and Stone, A. A.: Does humor mod- his M.A. from Jagiellonian University in Krakow, Poland erate the effects of experimentally induced stress? Annals of Behav- in 2006. He was a research student at Hokkaido Univer- ioral Medicine, 18(2), pp. 101-109, 1996. sity, and since 2007 he has been studying towards his Ph.D. [Nitta 89] Nitta. Y, and T. Masuoka, ”Japanese modality (Nihongo no degree at the Graduate School of Information Science and modariti)” Kuroshio,1989. Technology, Hokkaido University, Japan. His research in- [Ptaszynski 08] Ptaszynski, M. DYBALA, P. Rzepka, R. and Araki, terests include Natural Language Processing, Dialogue Pro- cessing, Humor Processing, HCI and Information Retrieval. K. Double Standpoint Evaluation Method for Affect Analysis Sys- tems The 22nd Annual Conference of Japanese Society for Artificial Intelligence (JSAI 2008), CD-ROM Proceedings 2P2-07, 2008 Araki, Kenji (Member) [Reitter 06] D. Reitter, J. D. Moore and F. Keller. Priming of syn- Born in 1959 in Otaru, Japan. Received B.E., M.E. and tactic rules in task-oriented dialogue and spontaneous conversation. Ph.D. degrees in electronics engineeringfrom Hokkaido Uni- In Proc. 28th Annual Conference of the Cognitive Science Society versity, Sapporo, Japan in 1982, 1985 and 1988, respec- tively. In April 1988, he joined Hokkai Gakuen Univer- (CogSci), Vancouver, Canada, 2006. sity, Sapporo, Japan. He was a professor of Hokkai Gakuen [Robinson 91] Robinson, V. M.: Humor and the Health Professions. University. He joined Hokkaido University in 1998 as an Thorofare, NJ, Slack, Inc., 1991 associate professor of the Division of Electronics and In- [Rzepka 09] Rzepka, R., Dybala, P., Shi, W., Higuchi, S., Ptaszynski, formation Engineering. He was a professor of the Division M. and Araki, K. 2009 Serious Processing for Frivolous Purpose - A of Electronics and Information Engineering of Hokkaido University from 2002. Chatbot Using Web-mining Supported Affect Analysis and Pun Gen- Currently, he is a professor of the Division of Media and Network Technologies of Hokkaido University. Research interests: NLP, Spoken Dialogue Processing, eration, Proceedings of IUI’09 - International Conference on Intelli- Machine Translation and Language Acquisition. Member of: AAAI, IEEE, IPSJ, gent User Interfaces, pp 487-488, Sanibel Island, FL, USA. (ACM IEICE and JCSS. Press). [Takayanagi 07] Takayanagi, K.: The Laughter Therapy, Japanese Journal of Complementary and Alternative Medicine, Vol. 4, 2007, View publication stats