Multi-behavior recommender systems, rapidly advancing across various domains, utilize plentiful auxiliary interactions on a variety of user behaviors to enhance recommendations for the target behavior, such as purchases. While previous methods have made strides in leveraging such interactions with advanced machine learning methods, they still face challenges in adequately using multi-faceted relationships among behaviors and handling uncertain auxiliary interactions that could potentially lead to purchases or not. In this paper, we propose MuLe (Multi-Grained Graph Learning), a novel graph-based model designed to address these limitations. We design a multi-grained graph learning strategy to capture diverse aspects of behaviors, ranging from unified to specific, and then to target-related behavior interactions. To handle uncertain interactions, we use graph attention, weighting the importance of those interactions related to the target behavior. Afterward, we use an attention mechanism to effectively aggregate diverse behavior embeddings obtained from the multi-grained graph encoders. Extensive experiments show that MuLe significantly outperforms the state-of-the-art methods, achieving improvements of up to 44.6% in HR@10 and 52.9% in NDCG@10, respectively. Our code and datasets are available at https://github.com/geonwooko/MULE.

The demand for a tool for summarizing emerging topics is increasing in modern life since the tool can deliver well-organized information to its users. Even though there are already a number of successful search systems, the system which automatically summarizes and organizes the content of emerging topics is still in its infancy. To fulfill such demand, we introduce an automated report generation system that generates a well-summarized human-readable report for emerging topics. In this report generation system, emerging topics are automatically discovered by a topic model and news articles are indexed by the discovered topics. Then, a topical summary and a timeline summary for each topic is generated by a topical multi-document summarizer and a timeline summarizer respectively. In order to enhance the apprehensibility of the users, the proposed report system provides two report modes. One is Today's Briefing which summarizes five discovered topics of every day, and the other is Full Report which shows a long-term view of each topic with a detailed topical summary and an important event timeline.

Two major problems of Korean part-of-speech (POS) tagging are that the word-spacing unit is not mapped one-to-one to a POS tag and that morphemes should be recovered during POS tagging. Therefore, this article proposes a novel two-step Korean POS tagger that solves the problems. This tagger first generates a sequence of lemmatized and recovered morphemes that can be mapped one-to-one to a POS tag using an encoder-decoder architecture derived from a POS-tagged corpus. Then, the POS tag of each morpheme in the generated sequence is finally determined by a standard sequence labeling method. Since the knowledge for segmenting and recovering morphemes is extracted automatically from a POS-tagged corpus by an encoder-decoder architecture, the POS tagger is constructed without a dictionary nor handcrafted linguistic rules. The experimental results on a standard dataset show that the proposed method outperforms existing POS taggers with its state-of-the-art performance.

Hyun-Je Song, Seong-Bae Park. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Dialog state tracking in spoken dialog system is the task that tracks the flow of a dialog and grasps what a user wants from the utterance precisely. Since the dialog success is related to catching the want of the user, dialog state tracking is a necessary component for spoken dialog systems. This paper proposes a neural dialog state tracker with the attention mechanism for focusing on valuable words and the hierarchical softmax for efficient training of the tracker. In addition, the proposed tracker combines a natural language understanding module and a dialog state module in an end-to-end style. As a result, the error propagation within a dialog system is minimized. To prove the effectiveness of the proposed model, we do experiments on dialog state tracking in the human-human task-oriented dialogs. Our experimental results show that the proposed method outperforms both the neural tracker without the attention mechanism and that without the hierarchical softmax.

The number of natural language queries submitted to search engines is increasing as search environments get diversified. However, legacy search engines are still optimized for short keyword queries. Thus, the use of natural language queries at legacy search engines degrades the retrieval performance of the engines. This paper proposes a novel method to translate a natural language query into a keyword query relevant to the natural language query for retrieving better search results without change of the engines. The proposed method formulates the translation as a generation task. That is, the method generates a keyword query from a natural language query by preserving the semantics of the natural language query. A recurrent neural network encoder-decoder architecture is adopted as a generator of keyword queries from natural language queries. In addition, an attention mechanism is also used to cope with long natural language queries.

Knowledge base completion aims to complete a knowledge base by filling up missing facts of the knowledge base. Neural knowledge base embeddings proposed to solve this task measure the plausibility of all candidate triples, and then select top-ranked triples by the plausibility as new facts for the knowledge base. The plausibility by neural embeddings allows true facts to be ranked at high positions, but not at top positions. This is because neural knowledge base embeddings are limited to using only the information within the knowledge base. Therefore, this paper proposes a re-ranking model for precise knowledge base completion. As a re-ranking model, a neural network which uses knowledge base schema and web statistic additionally is adopted. As a result, the proposed re-ranking model has an effect of using additional information for knowledge base completion. Thus, the candidate triples are first ranked by a neural knowledge base embedding, and then the result is re-ranked by the neural network. The experimental results show that the proposed re-ranking model improves the base neural embeddings up to 16% in Hits@1. This implies that the re-ranking model places true facts at top positions effectively.

Hee-Geun Yoon, Hyun-Je Song, Seong-Bae Park, Se-Young Park. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016.

Readability is used to provide users with highquality service in text recommendation or text visualization.With the increasing use of handheld devices, reading device is regarded as an important factor for readability.Therefore, this paper investigates the relationship between readability and reading devices such as a smart phone, a tablet, and paper.We suggest readability factors that are strongly related with the readability of a specific device by showing the correlations between various factors in each device and human-rated readability.Our experimental results show that each device has its own readability characteristics, and thus different weights should be imposed on readability factors according to the device type.In order to prove the usefulness of the results, we apply the device-dependent readability to news article recommendation.

User attribute is an essential factor for personalized recommendation and targeted advertising. Therefore, there have been a number of studies to identify user attributes automatically from SNS postings, since the postings reveal various attributes of writers. Many kinds of machine learning methods have been applied to automatic identification of user attributes as a candidate solution, but they suffer from two major problems. First, there are many postings in SNS that do not deliver any information about writers. Then, learning from SNS postings results in a biased model by these irrelevant postings. Second, the postings of a SNS user are somewhat related one another. However, most machine learning methods ignore this information, since they assume that data are independently and identically distributed. In order to solve these problems in user attribute identification, this paper proposes a novel method based on non-i.i.d. multi-instance learning. Since multi-instance learning treats all postings by a user as a bag and learns user attribute identification with such bags, not with postings, the first problem is solved. In addition, the proposed method assumes that the postings by a single user have a structure. By incorporating this assumption into the multi-instance learning, the second problem is solved. Our experimental results show that consideration of these two problems in automatic user attribute identification results in performance improvement.

With the increasing interest in location-based services, location comparison gains more and more attentions. One of the best ways to represent a location is to use topics that are generated near the location. In order to compare locations through such geographical topics, two conditions need to be met. One is that the topic set should be fixed but cover various aspects of all possible locations, and the other is that geographical topics often depend on each other. This paper proposes Probabilistic Explicit Semantic Analysis (PESA) that meets these conditions. PESA represents a location as a weighted topic vector where each topic is a Wikipedia concept. The number of Wikipedia concepts is fixed, but their enormous quantity allows PESA to be used to compare various locations. In addition, link information within Wikipedia articles is used to compute prior probabilities of topics considering their dependencies. That is, it enables PESA to model the topic dependency. PESA was evaluated using eighteen locations in three distinct geographical categories and compare it with LDA and ESA. The experimental results that PESA outperformed LDA and ESA highlighting its superiority in location comparison.

Text classification is a well-known problem for various applications. For last decades, it is beleived that a large corpus is one of the most important aspects for better classification. However, even though a great number of documents is available for training a classifier, it is practically impossible to achieve an ideal performance, since the distributions of labeled and unlabeled documents are often different. To overcome this problem, this paper describes a novel Naïve Bayes classifier for text classification under distribution difference between training and test data. The proposed method approximates test distribution by weighting labeled documents to cope with the distribution difference. Unlike other transfer learning which estimates the weights of labeled documents, the proposed method considerd both the documents and their estimated class labels. Therefore, the proposed method naturally combines the advantage of semi-supervised learning with those of transfer learning.

Information extraction from world wide web has been paid great attention to. Since a table is a well-organized and summarized knowledge expression for a domain, it is of great importance to extract information from the tables. However, many tables in web pages are used not to transfer information but to decorate the pages. Therefore, it is one of the most critical tasks in web table mining to discriminate the meaningful tables from the decorative ones. The main obstacle of this task comes from the difficulty of generating relevant features for the discrimination. This paper proposes a novel method to discriminate them using a composite kernel which combines a parse tree kernel and a linear kernel. Since a web table is represented as a parse tree by a HTML parser, the parse tree kernel can be naturally used in determining the similarity between trees, and the linear kernel with content features is used to make up for the weak points of the parse tree kernel. The support vector machines with the composite kernel distinguish with high accuracy the meaningful tables from the decorative ones. A series of experiments show that the proposed method achieves the state-of-the-art performance.