深澤 佑介

コンテキストはユーザの興味・嗜好に影響する重要な要因の1つである．本稿では，コンテキストの中でも特に家族や同僚など同行者に注目し，同行者依存のトピックモデルを提案する．第一に同行者クラスを考慮したモデル（CTM）を提案する．次にスイッチ変数を導入し同行者依存の単語を自動学習する仕組みを取り入れたモデルを提案する（sCTM）．さらに，同行者依存の単語をWeb全体から事前学習し，それをモデル内に反映するモデルを提案する（fCTM）．それぞれのモデルはCollapsed Gibbs Sampling（CGS）に基づき推論を行う．Webから同行者依存の投稿データを抽出し，提案モデル間の比較実験を実施した．文書の予測精度（Perplexity）の観点でCTM（ベースライン）とfCTM，sCTMを比較評価し提案手法の優位性を示した．また，質的評価としてfCTMの同行者のトピックに含まれる単語を確認し，妥当なモデル化が行われていることを確認した．Context is understood as an important factor that affects topics to be generated. We focus on companion of users (friends, wife, husband etc.) as one of the most important factors to determine the topic. Different from location and time context, context of companion does not appear explicitly with the documents but appears inside the document in the form of contextual words (e.g. friends). To discriminate contextual words, topical words and background words from documents, and obtain both precise and discriminative topics, we propose three kinds of context aware topic models. Firstly, we introduce context class (CTM) to extract context dominant topics from text. Secondly, we introduce switch variable (sCTM) to discriminate background words from contextual and topical words. Thirdly, we introduce fixed Dirichlet parameter learned from the web to sCTM (fCTM). We conduct experiments on data set extracted from the web, and they show that the proposed model (sCTM and fCTM) can capture interpretable and discriminative sets of topics than baseline CTM from the view point of perplexity and KL-Divergence.

Contexts such as time, location and companion play an important role to determine topics of documents. We focus on companion context (friends, wife, husband etc.) as one of the most important contexts to determine topics. We propose companion context dependent topic model by the extension of basic graphical model: LDA (Latent Dirichlet allocation). We propose three kinds of LDA extensions. Firstly, we incorporate context class to extract context dominant topics (CTM). Secondly, in addition to CTM, we incorporate predefined contextual words to make context dominant topics more discriminative (eCTM). Thirdly, in addition to eCTM, we incorporate switch variables (sCTM), which discriminate background words from contextual and topical words to associate context related words to the topics more precisely. We conduct experiments on two data sets, and they show that the proposed model (eCTM and sCTM) can predict the word distribution of test documents with high probability and generate discriminative sets of topics than baseline CTM from the viewpoint of perplexity.

we propose a method of predicting destinations by using Twitter posts with location information. The proposed method chooses base tweets, which is close to the current user's tweet, and then predict destination using the next set of tweets of base tweet. The base tweets are selected based on not only location closeness but also similarity of tweet content. We evaluate the proposed method by the error range of the distance between predicted destination and golden answer. We used three months of Twitter data with location information (almost 40 mil.) as the test set tweets. The experimental result demonstrates that the prediction accuracy of the proposed method is superior to the baseline, which only consider the location similarity.

We propose a topic model to better estimate activities from tweets. The whole estimation process consists of two phases: one is the cluster generation, and the other is the activity estimation. At the first phase, we obtain the expected tri-layer clusters with the components: a topic layer, an activity layer and a word layer. Then, at the second phase, we utilize the activity-specific word distribution derived from the training results to learn the activities of testing tweets. To prove the feasibility of this model, we evaluate the precision of activity estimation using 35 activities to extract 23,988 tweets for training and 350 for testing, respectively. The experimental results demonstrate that the reasonable topic-specific activity distribution contributes to the cluster generation, and the proposed model exhibits the superiority in activity estimation.

We propose a topic model capable of generating tri-layer clusters, each of which is composed of a topic layer, an activity layer and a word layer. The objective is to better predict activities involved in documents by considering general topics of the activities for clustering. The proposed model is a supervised topic model based on the Latent Dirichlet Allocation (LDA). As a follow-up study of word-pair generation LDA (wpLDA) model, the model introduces the topic-specific activity distribution as an external input, with an activity node inserted into the main generation thread. In addition, we refer to D. Ramage et al.'s one-to-one correspondence to directly learn word-activity tags. An experiment was conducted to prove the feasibility of this model. We chose ten top-listed activities from the wish clusters obtained by the previous wpLDA research, and used each as the key words to extract thirty tweets for training and five for testing, respectively, tagging the tweets with the corresponding activities. By applying the proposed model, we obtained the expected tri-layer clusters in the training phase. Then, in the testing phase, we utilized the activity-specific word distribution derived from the training results to learn the activities of the testing documents. The Stanford Classifier was put forward as the control group, and the activity prediction accuracy demonstrates that the proposed model exhibits the superiority in multi-activity prediction.

We demonstrate a generative model that incorporates word-pair connection into the smoothed LDA model to intuitively discover people's wish related activities. The widely used model, LDA topic model, generally generates clusters in the form of separate words. However, this form is not intuitive enough to express people's activities. Therefore, we consider the word-pairs led by verbs can better describe users' intentions and activities, and we prefer to present this collocation under topics as the clustering results. We mathematically present the relatedness between verbs and non-verb words through association rule, and build the physical connection of word-pairs and possible topics. By incorporating the connection lattice into the smoothed LDA, the word-pair LDA model is created. In the experiments, Twitter posts about "new year's resolutions" were chosen as the data source. The results show that the proposed model performs well on perplexity, and presents excellent intuitive character. © 2012 IEEE.

We used the twitter posts about New Year's resolutions as data source to capture users' long-term goals. New Year's resolutions are the commitments that people set for their personal goals, and generally, people plan to fulfill them for the whole following year. Therefore, we can think of such tweets as data source to explore people's possible long-term goals. The key words in each tweet were extracted for clustering. Considering the form of word-pairs led by verbs is a more intuitive and clearer way to express people's intentions than the one of separate words, we propose a generative model that incorporates word connections into the smoothed LDA to cluster the key words of long-term goals. The experiments demonstrate the proposed model is capable of clustering the word-pairs with better intuitive character, and clearly dividing people's long-term goals. © 2012 Springer-Verlag Berlin Heidelberg.

Activities, defined as actions someone can accomplish or things one can create, are located all around the web contained in various How-To and DIY sites. In this paper we are presenting our research on a system focused on retrieving information about activities related to a simple query which is given as input. An overview of the system's design which is based on semantic query expansion is given along with detailed explanation of the optimization of the system's parameters through the use of genetic algorithms. Finally experimental results gathered from both an offline and a limited online evaluation are presented in order to test the efficiency of the system. Copyright 2012 ACM.

In this paper, we investigated onomatopoeia usage pattern in food reviews by proposing LDA (Latent Dirichlet Allocation) based onomatopoeia usage pattern analysis model. We collected total 685 numbers of onomatopoeias which are distributed to 208 food categories from 3,581,808 food reviews of Japanese food review site Tabelog. From the experimental result, we found several patterns how the onomatopoeias are chosen. The onomatopoeia is chosen based on user's interest on the combination of {location of food, material of the food, cooking method} and {the texture of food, sound when eating, and looks of food people's status when eating the food}. In addition, we investigate how the precision of the clustering result changes depending on the N (number of onomatopoeia of each food categories). We found that the results of N=30 is better than one of N=100 as large number of onomatopoeia for each food categories like 100 is likely to include onomatopoeias that are irrelevant to food. © 2012 IEEE.

This paper proposes task-oriented content-based recommendation for cross service recommendation. The proposed method has two features: one is that task-based features are automatically mined from the web, second is that it estimates user's intention on task, which means what user wants to do and what problem user has by task-based profile representation. We evaluate improvement of recommendation accuracy by user evaluation, in which we collect ratings on variety of contents (i.e. mobile web content, TV programs, restaurants, sightseeing spots, and hotels) from 1,859 people and conduct cross validation. In an experiment, the combination of task-based profile representation and term-based representation yields a 17.7% improvement in MAE (Mean Absolute Error) compared to term-based profile representation only or domain (content category) based profile representation only. © 2012-IOS Press and the authors. All rights reserved.

To realize the vision of intelligent agents on the web, agents need to be capable of understanding people's behavior. Such an understanding would enable them to better predict and support human activities on the web. If agents had access to knowledge about human goals, they could, for instance, recognize people's goals from their actions or reason about people's goals. In this work, we study to what extent it is feasible to automatically construct concept hierarchies of domain-specific human goals. This process consists of the following two steps: (1) extracting human goal instances from a search query log and (2) inferring hierarchical structures by applying clustering techniques. To compare resulting concept hierarchies, we manually construct a golden standard and calculate taxonomic overlaps. In our experiments, we achieve taxonomic overlaps of up to similar to 51% for the health domain and up to similar to 60% for individual health subdomains. In an illustration scenario, we provide a prototypical implementation to automatically complement goal concept hierarchies by means-ends relations, i.e. relating goals to actions which potentially contribute to their accomplishment. Our findings are particularly relevant for knowledge engineers interested in (i) acquiring knowledge about human goals as well as (ii) automating the process of constructing goal concept hierarchies.

We have been developing a task-based service navigation system that offers to the user services relevant to the task the user wants to perform. The system allows the user to concretize his/her request in the task-model developed by human-experts. In this study, to reduce the cost of collecting a wide variety of activities, we investigate the automatic modeling of users' real world activities from the web. To extract the widest possible variety of activities with high precision and recall, we investigate the appropriate number of contents and resources to extract. Our results show that we do not need to examine the entire web, which is too time consuming a limited number of search results (e.g. 900 from among 21,000,000 search results) from blog contents are needed. In addition, to estimate the hierarchical relationships present in the activity model with the lowest possible error rate, we propose a method that divides the representation of activities into a noun part and a verb part, and calculates the mutual information between them. The result shows almost 80% of the hierarchical relationships can be captured by the proposed method. © 2010 Copyright is held by the author/owner(s).

Mobile devices are becoming more and more difficult to use due to the sheer number of functions now supported. In this paper, we propose a menu customization system that ranks functions so as to make interesting functions including both frequently used and functions that are infrequently used but have the potential to satisfy the user, easy to access. Concretely, we define the features of the phone's functions by extracting keywords from the manufacturer's manual, and propose a method that uses the Ranking SVM (Support Vector Machine) to rank the functions based on user's operation history. We conduct a home-use test for one week to evaluate the efficiency of customization and the usability of menu customization. The results of this test show that the average rank at the last day was half that of the first day, and that the user could find, on average, 3.14 more kinds of new functions, ones that the user did not know about before the test, on a daily basis. This shows that the proposed customized menu supports the user by making it easier to access frequent items and to find new interesting functions. From interviews, almost 70 % of the users were satisfied with the ranking provided by menu customization as well as the usability of the resulting menus. In addition, interviews show that automatic cell phone menu customization is more appropriate for mobile phone beginners than expert users.

Mobile devices are becoming more and more difficult to use due to the sheer number of functions now supported. In this paper, we propose a menu customization system that ranks functions so as to make interesting functions, both frequently used functions and rarely used functions, easy to access. Concretely, we define the features of phone functions by extracting keywords from the manufacturer's manual, and propose the method that ranks the functions based on user operation history by using Ranking SVM (Support Vector Machine). We conduct a home-use test for one week to evaluate the efficiency of customization and the usability of menu customization. The results show that the average rank of used functions on the last day of the test is half of that of first day and almost 70 % of the users are satisfied with the ranking provided by menu customization and the usability of menus. In addition, interviews show that automatic mobile menu customization is more appropriate for mobile phone beginner rather than the master users.

We study the case of integrating situational reasoning into a mobile service recommendation system. Since mobile Internet services are rapidly proliferating, finding and using appropriate services require profound service descriptions. As a consequence, for average mobile users it is nowadays virtually impossible to find the most appropriate service among the many offered. To overcome these difficulties, task navigation systems have been proposed to guide users towards best-fitting services. Our goal is to improve the user experience of such task navigation systems making them context-aware (i.e. to optimize service navigation by taking the user's situation into account). We propose the integration of a situational reasoning engine that applies classification-based inference to qualitative context elements, gathered from multiple sources and represented using ontologies. The extended task navigator enables the delivery of situation-aware recommendations in a proactive way. Initial experiments with the extended system indicate a considerable improvement of the navigator's usability.

In this paper, an algorithm that acquires the intermediate goals between the initial and goal states is proposed for an agent executing multiple tasks. We demonstrate the algorithm in the problem of rearranging multiple objects. The result shows that the moving distance to transfer the entire objects to their goal configuration is 1/15 of that without using intermediate goals. We experiment using a real robot to confirm that the intermediate goal can be adapted to a real environment. Our experimental results showed that an agent could, adapt the intermediate goals, which were acquired in the simulation, to the experimental environment.

We have been developing a task-based service navigation system that offers to the user for his selected services relevant to the task the user wants to perform. We observed that the tasks likely to be performed in a given situation depend on the user's role such as businessman or father. To further our research, we constructed a role-ontology and utilized it to improve the usability of task-based service navigation. We have enhanced a basic task-model by associating tasks with role-concepts defined in the new role-ontology. We can generate a task-list that is precisely tuned to the user's current role. In addition, we can generate a personalized task-list from the task-model based on the user's task selection history. Because services are associated with tasks, our approach makes it much easier to navigate a user to the most appropriate services. In this paper, we describe the construction of our role-ontology and the task-based service navigation system based on the role-ontology.

Mobile access to the Internet is increasing drastically, and this is raising the importance of information retrieval via mobile units. We have developed a task-oriented service navigation system[6] that allows a user to find the mobile contents desired from the viewpoint of the task that the user wants to do. However, the user is still faced with the problem of having to select the most appropriate task from among the vast number of task candidates this is difficult due to the fact that mobile devices have several limitations such as small displays and poor input methods. This paper tackles this issue by proposing a framework for retrieving only those tasks that suit the abstraction level of the user's intention. If the user has settled on a specific object, the abstraction level is concrete, and tasks related to the handling of the specific object are selected if not, tasks related to general objects are selected. Finally, we introduce two task retrieval applications that realize the proposed framework. By using this framework, we can reduce the number of retrieved tasks irrelevant to the user simulations show that roughly 30% fewer tasks are displayed to the user as retrieval results. © Springer-Verlag Berlin Heidelberg 2005.

This paper offers a proposal for realizing the exploration and rearrangement of multiple unknown objects that lay scattered in working environments. The objective of the exploration task is to find all the objects in the environments. On the other hand, the objective of the rearrangement task is to carry all the objects to their goal position. Many applications are possible if the exploration and rearrangement tasks are combined. Some of them are cleaning, mine detecting and housework. An algorithm that integrates two tasks is presented with respect to the effectiveness of the path length and computational cost. In addition, an exploration algorithm is proposed that can work well in an environment that has many objects. In order to verify the algorithm, experiments are conducted with an actual robot. In the experimentals, an environmental recognition method is developed by attaching a mark to the objects. The robot recognizes the objects by finding the mark. It then obtains information from the mark. The mark is also used to modify the odometry error of the robot by computing its configuration relative to a mark attached to a wall. The success rate of this experiment was almost 80% in 20 trials.

In this paper we present a method for cooperative exploration by multiple mobile robots in a restricted working area. By generating and sharing the minimal-length path for each mobile robot in which mobile robots go through several observation points to explore the working area. The observation points should be arranged at a fixed distance from at least one corresponding observation point. Therefore, mobile robots can obtain the complete information of the working area by only moving along all of the observation points. To reduce the redundancy of exploration task, the number of observation points and the length of observation paths for each mobile robot are to be minimized. To ensure the advantages from using multiple mobile robots, exploration path must be equally assigned to each mobile robot. Simulations had proved the efficiency of cooperating mobile robots using our proposed method. The feasible of our method is real world with actual mobile robots was demonstrated in cooperative exploration experiment.

In this paper, exploration path algorithm is proposed for mobile robots to make a map of working environment. The exploration task is defined as a problem of generating minimal-cost path, in which robots go through several observation points and observe a working environment. Both number of the observation points and path length should be minimized. The proposed algorithm has two characteristics: efficiency in exploration and adaptability to dynamic environmental changes. Our method can be realized with the combination of (a) distribution of observation points by a reaction-diffusion equation on a graph, and (b) generation of a Hamiltonian circle that connects all observation points. The observation points dynamically change their arrangements in accordance with the recognized environmental situation. The calculation cost for exploration path generation is shown to be in order of N1.5, where N is the number of the observation points. Our method can be extended into cooperative exploration path planning method. Our method homogenized the arrangement of the observation points, then only a basic partition method can equally part exploration task for each robot. The effectiveness of our method is shown by both simulation and real robot experiments.