We propose a self-supervised embedding learning frameworkSelfLinKGto link concepts in heterogeneous knowledge graphs. Without any labeled data, SelfLinKG can achieve competitive performance against its supervised counterpart, and significantly outperforms state-of-the-art unsupervised methods by 26%-50%. The essential components of SelfLinKG are local attention-based encoding and momentum contrastive learning. The former aims to learn the graph representation using an attention network, while the latter is to learn a self-supervised model across knowledge graphs using contrastive learning. SelfLinKG has been deployed to build the the new version, called OAG_know of Open Academic Graph (OAG). All data and codes are publicly available.

Different knowledge graphs for the same domain are often uniquely housed on the Web. Effectively linking entities from different graphs is critical for building an open and comprehensive knowledge graph. However, linking entities across different sources has thus far faced various challenges, including the increasingly large-scale volume of the data, the heterogeneity of the graphs, and the ambiguity of real-world entities. To address them, we propose a unified framework LinKG. Specifically, we decouple the problem into different linking tasks based on the unique properties of each type of entity. To link word sequence based entities, we propose an LSTM-based method to capture word dependencies. To link entities of large scale, we utilize the hashing technique and convolutional neural networks for scalable and accurate linking. To link ambiguous entities, we propose heterogeneous graph attention networks to leverage heterogeneous structural information. Finally, to validate the design choices of different LinKG modules, we characterize the relationships between different tasks based on the single-domain and multi-domain transfer models. Extensive experiments demonstrate the effectiveness of LinKG with an overall F1-score of 95.15%, based on which we deploy and release the Open Academic Graph (OAG)—the largest publicly available heterogeneous academic graph to date.

Large-scale multi-label text classification (LMTC) aims to associate a document with its relevant labels from a large candidate set. Most existing LMTC approaches rely on massive human-annotated training data, which are often costly to obtain and suffer from a long-tailed label distribution (i.e., many labels occur only a few times in the training set). In this paper, we study LMTC under the zero-shot setting, which does not require any annotated documents with labels and only relies on label surface names and descriptions. To train a classifier that calculates the similarity score between a document and a label, we propose a novel metadata-induced contrastive learning (MICoL) method. Different from previous text-based contrastive learning techniques, MICoL exploits document metadata (e.g., authors, venues, and references of research papers), which are widely available on the Web, to derive similar document–document pairs. Experimental results on two large-scale datasets show that: (1) MICoL significantly outperforms strong zero-shot text classification and contrastive learning baselines; (2) MICoL is on par with the state-of-the-art supervised metadata-aware LMTC method trained on 10K–200K labeled documents; and (3) MICoL tends to predict more infrequent labels than supervised methods, thus alleviates the deteriorated performance on long-tailed labels.

Multi-label text classification refers to the problem of assigning each given document its most relevant labels from a label set. Commonly, the metadata of the given documents and the hierarchy of the labels are available in real-world applications. However, most existing studies focus on only modeling the text information, with a few attempts to utilize either metadata or hierarchy signals, but not both of them. In this paper, we bridge the gap by formalizing the problem of metadata-aware text classification in a large label hierarchy (e.g., with tens of thousands of labels). To address this problem, we present the MATCH1 solution—an end-to-end framework that leverages both metadata and hierarchy information. To incorporate metadata, we pre-train the embeddings of text and metadata in the same space and also leverage the fully-connected attentions to capture the interrelations between them. To leverage the label hierarchy, we propose different ways to regularize the parameters and output probability of each child label by its parents. Extensive experiments on two massive text datasets with large-scale label hierarchies demonstrate the effectiveness of MATCH over the state-of-the-art deep learning baselines.

We focus on a recently deployed system built for summarizing academic articles by concept tagging. The system has shown great coverage and high accuracy of concept identification which could be contributed by the knowledge acquired from millions of publications. Provided with the interpretable concepts and knowledge encoded in a pre-trained neural model, we investigate whether the tagged concepts can be applied to a broader class of applications. We propose transforming the tagged concepts into sparse vectors as representations of academic documents. The effectiveness of the representations is analyzed theoretically by a proposed framework. We also empirically show that the representations can have advantages on academic topic discovery and paper recommendation. On these applications, we reveal that the knowledge encoded in the tagging system can be effectively utilized and can help infer additional features from data with limited information.

Representation learning has offered a revolutionary learning paradigm for various AI domains. In this survey, we examine and review the problem of representation learning with the focus on heterogeneous networks, which consists of different types of vertices and relations. The goal of this problem is to automatically project objects, most commonly, vertices, in an input heterogeneous network into a latent embedding space such that both the structural and relational properties of the network can be encoded and preserved. The embeddings (representations) can be then used as the features to machine learning algorithms for addressing corresponding network tasks. To learn expressive embeddings, current research developments can fall into two major categories: shallow embedding learning and graph neural networks. After a thorough review of the existing literature, we identify several critical challenges that remain unaddressed and discuss future directions. Finally, we build the Heterogeneous Graph Benchmark to facilitate open research for this rapidly-developing topic.

Graph neural networks (GNNs) have been demonstrated to be powerful in modeling graph-structured data. However, training GNNs requires abundant task-specific labeled data, which is often arduously expensive to obtain. One effective way to reduce the labeling effort is to pre-train an expressive GNN model on unlabelled data with self-supervision and then transfer the learned model to downstream tasks with only a few labels. In this paper, we present the GPT-GNN framework to initialize GNNs by generative pre-training. GPT-GNN introduces a self-supervised attributed graph generation task to pre-train a GNN so that it can capture the structural and semantic properties of the graph. We factorize the likelihood of graph generation into two components: 1) attribute generation and 2) edge generation. By modeling both components, GPT-GNN captures the inherent dependency between node attributes and graph structure during the generative process. Comprehensive experiments on the billion-scale open academic graph and Amazon recommendation data demonstrate that GPT-GNN significantly outperforms state-of-the-art GNN models without pre-training by up to 9.1% across various downstream tasks?

Graph representation learning has emerged as a powerful technique for addressing real-world problems. Various downstream graph learning tasks have benefited from its recent developments, such as node classification, similarity search, and graph classification. However, prior arts on graph representation learning focus on domain specific problems and train a dedicated model for each graph dataset, which is usually non-transferable to out-of-domain data. Inspired by the recent advances in pre-training from natural language processing and computer vision, we design Graph Contrastive Coding (GCC) --- a self-supervised graph neural network pre-training framework --- to capture the universal network topological properties across multiple networks. We design GCC's pre-training task as subgraph instance discrimination in and across networks and leverage contrastive learning to empower graph neural networks to learn the intrinsic and transferable structural representations. We conduct extensive experiments on three graph learning tasks and ten graph datasets. The results show that GCC pre-trained on a collection of diverse datasets can achieve competitive or better performance to its task-specific and trained-from-scratch counterparts. This suggests that the pre-training and fine-tuning paradigm presents great potential for graph representation learning.

Recent years have witnessed the emerging success of graph neural networks (GNNs) for modeling structured data. However, most GNNs are designed for homogeneous graphs, in which all nodes and edges belong to the same types, making it infeasible to represent heterogeneous structures. In this paper, we present the Heterogeneous Graph Transformer (HGT) architecture for modeling Web-scale heterogeneous graphs. To model heterogeneity, we design node- and edge-type dependent parameters to characterize the heterogeneous attention over each edge, empowering HGT to maintain dedicated representations for different types of nodes and edges. To handle Web-scale graph data, we design the heterogeneous mini-batch graph sampling algorithm—HGSampling—for efficient and scalable training. Extensive experiments on the Open Academic Graph of 179 million nodes and 2 billion edges show that the proposed HGT model consistently outperforms all the state-of-the-art GNN baselines by 9–21 on various downstream tasks. The dataset and source code of HGT are publicly available at https://github.com/acbull/pyHGT.

Taxonomies consist of machine-interpretable semantics and provide valuable knowledge for many web applications. For example, online retailers (e.g., Amazon and eBay) use taxonomies for product recommendation, and web search engines (e.g., Google and Bing) leverage taxonomies to enhance query understanding. Enormous efforts have been made on constructing taxonomies either manually or semi-automatically. However, with the fast-growing volume of web content, existing taxonomies will become outdated and fail to capture emerging knowledge. Therefore, in many applications, dynamic expansions of an existing taxonomy are in great demand. In this paper, we study how to expand an existing taxonomy by adding a set of new concepts. We propose a novel self-supervised framework, named TaxoExpan, which automatically generates a set of ⟨query concept, anchor concept⟩ pairs from the existing taxonomy as training data. Using such self-supervision data, TaxoExpan learns a model to predict whether a query concept is the direct hyponym of an anchor concept. We develop two innovative techniques in TaxoExpan: (1) a position-enhanced graph neural network that encodes the local structure of an anchor concept in the existing taxonomy, and (2) a noise-robust training objective that enables the learned model to be insensitive to the label noise in the self-supervision data. Extensive experiments on three large-scale datasets from different domains demonstrate both the effectiveness and the efficiency of TaxoExpan for taxonomy expansion.

Recent years have witnessed the rapid growth in the number of academics and practitioners who are interested in big scholarly data as well as closely-related areas. Quite a lot of papers reporting recent advancements in this area have been published in leading conferences and journals. Both non-commercial and commercial platforms and systems have been released in recent years, which provide innovative services built upon big scholarly data to the academic community. Examples include Microsoft Academic Graph, Google Scholar, DBLP, arXiv, CiteSeerX, Web of Knowledge, Udacity, Coursera, and edX. The workshop will contribute to the birth of a community having a shared interest around big scholarly data and exploring it using knowledge discovery, data science and analytics, network science, and other appropriate technologies.

Linking entities from different sources is a fundamental task in building open knowledge graphs. Despite much research conducted in related fields, the challenges of linkinglarge-scale heterogeneous entity graphs are far from resolved. Employing two billion-scale academic entity graphs (Microsoft Academic Graph and AMiner) as sources for our study, we propose a unified framework --- LinKG --- to address the problem of building a large-scale linked entity graph. LinKG is coupled with three linking modules, each of which addresses one category of entities. To link word-sequence-based entities (e.g., venues), we present a long short-term memory network-based method for capturing the dependencies. To link large-scale entities (e.g., papers), we leverage locality-sensitive hashing and convolutional neural networks for scalable and precise linking. To link entities with ambiguity (e.g., authors), we propose heterogeneous graph attention networks to model different types of entities. Our extensive experiments and systematical analysis demonstrate that LinKG can achieve linking accuracy with an F1-score of 0.9510, significantly outperforming the state-of-the-art. LinKG has been deployed to Microsoft Academic Search and AMiner to integrate the two large graphs. We have published the linked results---the Open Academic Graph (OAG)\footnote\urlhttps://www.openacademic.ai/oag/ , making it the largest publicly available heterogeneous academic graph to date.

We present the design and methodology for the large scale hybrid paper recommender system used by Microsoft Academic. The system provides recommendations for approximately 160 million English research papers and patents. Our approach handles incomplete citation information while also alleviating the cold-start problem that often affects other recommender systems. We use the Microsoft Academic Graph (MAG), titles, and available abstracts of research papers to build a recommendation list for all documents, thereby combining co-citation and content based approaches. Tuning system parameters also allows for blending and prioritization of each approach which, in turn, allows us to balance paper novelty versus authority in recommendation results. We evaluate the generated recommendations via a user study of 40 participants, with over 2400 recommendation pairs graded and discuss the quality of the results using [email protected] and nDCG scores. We see that there is a strong correlation between participant scores and the similarity rankings produced by our system but that additional focus needs to be put towards improving recommender precision, particularly for content based recommendations. The results of the user survey and associated analysis scripts are made available via GitHub and the recommendations produced by our system are available as part of the MAG on Azure to facilitate further research and light up novel research paper recommendation applications.

We study the problem of large-scale network embedding, which aims to learn latent representations for network mining applications. Previous research shows that 1) popular network embedding benchmarks, such as DeepWalk, are in essence implicitly factorizing a matrix with a closed form, and 2) the explicit factorization of such matrix generates more powerful embeddings than existing methods. However, directly constructing and factorizing this matrix-which is dense-is prohibitively expensive in terms of both time and space, making it not scalable for large networks.

Social and information networking activities such as on Facebook, Twitter, WeChat, and Weibo have become an indispensable part of our everyday life, where we can easily access friends' behaviors and are in turn influenced by them. Consequently, an effective social influence prediction for each user is critical for a variety of applications such as online recommendation and advertising.

Since the invention of word2vec, the skip-gram model has significantly advanced the research of network embedding, such as the recent emergence of the DeepWalk, LINE, PTE, and node2vec approaches. In this work, we show that all of the aforementioned models with negative sampling can be unified into the matrix factorization framework with closed forms. Our analysis and proofs reveal that: (1) DeepWalk empirically produces a low-rank transformation of a network's normalized Laplacian matrix; (2) LINE, in theory, is a special case of DeepWalk when the size of vertices' context is set to one; (3) As an extension of LINE, PTE can be viewed as the joint factorization of multiple networks» Laplacians; (4) node2vec is factorizing a matrix related to the stationary distribution and transition probability tensor of a 2nd-order random walk. We further provide the theoretical connections between skip-gram based network embedding algorithms and the theory of graph Laplacian. Finally, we present the NetMF method as well as its approximation algorithm for computing network embedding. Our method offers significant improvements over DeepWalk and LINE for conventional network mining tasks. This work lays the theoretical foundation for skip-gram based network embedding methods, leading to a better understanding of latent network representation learning.

Progress in science has advanced the development of human society across history, with dramatic revolutions shaped by information theory, genetic cloning, and artificial intelligence, among the many scientific achievements produced in the 20th century. However, the way that science advances itself is much less well-understood. In this work, we study the evolution of scientific development over the past century by presenting an anatomy of 89 million digitalized papers published between 1900 and 2015. We find that science has benefited from the shift from individual work to collaborative effort, with over 90% of the world-leading innovations generated by collaborations in this century, nearly four times higher than they were in the 1900s. We discover that rather than the frequent myopic- and self-referencing that was common in the early 20th century, modern scientists instead tend to look for literature further back and farther around. Finally, we also observe the globalization of scientific development from 1900 to 2015, including 25-fold and 7-fold increases in international collaborations and citations, respectively, as well as a dramatic decline in the dominant accumulation of citations by the US, the UK, and Germany, from ~95% to ~50% over the same period. Our discoveries are meant to serve as a starter for exploring the visionary ways in which science has developed throughout the past century, generating insight into and an impact upon the current scientific innovations and funding policies.

This paper presents a novel research problem, Comparative Document Analysis (CDA), that is, joint discovery of commonalities and differences between two individual documents (or two sets of documents) in a large text corpus. Given any pair of documents from a (background) document collection, CDA aims to automatically identify sets of quality phrases to summarize the commonalities of both documents and highlight the distinctions of each with respect to the other informatively and concisely. Our solution uses a general graph-based framework to derive novel measures on phrase semantic commonality and pairwise distinction, where the background corpus is used for computing phrase-document semantic relevance. We use the measures to guide the selection of sets of phrases by solving two joint optimization problems. A scalable iterative algorithm is developed to integrate the maximization of phrase commonality or distinction measure with the learning of phrase-document semantic relevance. Experiments on large text corpora from two different domains---scientific papers and news---demonstrate the effectiveness and robustness of the proposed framework on comparing documents. Analysis on a 10GB+ text corpus demonstrates the scalability of our method, whose computation time grows linearly as the corpus size increases. Our case study on comparing news articles published at different dates shows the power of the proposed method on comparing sets of documents.

In this paper, we describe the WSDM Cup entity ranking challenge held in conjunction with the 2016 Web Search and Data Mining conference (WSDM 2016). Participants in the challenge were provided access to the Microsoft Academic Graph (MAG), a large heterogeneous graph of academic entities, and were invited to calculate the query-independent importance of each publication in the graph. Submissions for the challenge were open from August through November 2015, and a public leaderboard displayed teams? progress against a set of training judgements. Final evaluations were performed against a separate, withheld portion of the evaluation judgements. The top eight performing teams were then invited to submit papers to the WSDM Cup workshop, held at the WSDM 2016 conference.

Over the past few years, major web search engines have introduced knowledge bases to offer popular facts about people, places, and things on the entity pane next to regular search results. In addition to information about the entity searched by the user, the entity pane often provides a ranked list of related entities. To keep users engaged, it is important to develop a recommendation model that tailors the related entities to individual user interests. We propose a probabilistic Three-way Entity Model (TEM) that provides personalized recommendation of related entities using three data sources: knowledge base, search click log, and entity pane log. Specifically, TEM is capable of extracting hidden structures and capturing underlying correlations among users, main entities, and related entities. Moreover, the TEM model can also exploit the click signals derived from the entity pane log. We further provide an inference technique to learn the parameters in TEM, and propose a principled preference learning method specifically designed for ranking related entities. Extensive experiments with two real-world datasets show that TEM with our probabilistic framework significantly outperforms a state of the art baseline, confirming the effectiveness of TEM and our probabilistic framework in related entity recommendation.

No abstract available.

Web search is seeing a paradigm shift from keyword based search to an entity-centric organization of web data. To support web search with this deeper level of understanding, a web-scale entity linking system must have 3 key properties: First, its feature extraction must be robust to the diversity of web documents and their varied writing styles and content structures. Second, it must maintain high-precision linking for "tail" (unpopular) entities that is robust to the existence of confounding entities outside of the knowledge base and entity profiles with minimal information. Finally, the system must represent large-scale knowledge bases with a scalable and powerful feature representation. We have built and deployed a web-scale unsupervised entity linking system for a commercial search engine that addresses these requirements by combining new developments in sparse signal recovery to identify the most discriminative features from noisy, free-text web documents; explicit modeling of out-of-knowledge-base entities to improve precision at the tail; and the development of a new phrase-unigram language model to efficiently capture high-order dependencies in lexical features. Using a knowledge base of 100M unique people from a popular social networking site, we present experimental results in the challenging domain of people-linking at the tail, where most entities have limited web presence. Our experimental results show that this system substantially improves on the precision-recall tradeoff over baseline methods, achieving precision over 95% with recall over 60%.

Entity attribute values, such as "lord of the rings" for movie.title or "infant" for shoe.gender, are atomic components of entity expressions. Discovering alternative surface forms of attribute values is important for improving entity recognition and retrieval. In this work, we propose a novel compact clustering framework to jointly identify synonyms for a set of attribute values. The framework can integrate signals from multiple information sources into a similarity function between attribute values. And the weights of these signals are optimized in an unsupervised manner. Extensive experiments across multiple domains demonstrate the effectiveness of our clustering framework for mining entity attribute synonyms.

This paper addresses the problem of semantic entity resolution (SER), which aims to determine whether some or none of the entities in a knowledge base is mentioned in a given web document. The lexical features, e.g., words and phrases, which are critical to the resolution of the semantic entities are typically of a small amount compared to all lexical features in the web document, and therefore can be modeled as sparse signals. Two techniques leveraging the principles of sparse signal recovery are proposed to identify the sparse, salient lexical features: one technique, based on the Lasso algorithm with the l2-norm distance metric, attempts to recover all the salient lexical features at once; the other technique, namely Posterior Probability Pursuit (PPP), sequentially identifies salient features one after one using the negative log posterior probability as the distance metric. Using a knowledge base consisting of about 100 million entities, we show that the proposed techniques exploiting the sparsity nature underlying SER deliver substantial performance improvement over baseline methods without sparsity consideration, demonstrating the potentials of sparse signal techniques in entity-centric web information processing.

The semantic gap between low-level visual features and high-level semantics has been investigated for decades but still remains a big challenge in multimedia. When "search" became one of the most frequently used applications, "intent gap", the gap between query expressions and users' search intents, emerged. Researchers have been focusing on three approaches to bridge the semantic and intent gaps: 1) developing more representative features, 2) exploiting better learning approaches or statistical models to represent the semantics, and 3) collecting more training data with better quality. However, it remains a challenge to close the gaps. In this paper, we argue that the massive amount of click data from commercial search engines provides a data set that is unique in the bridging of the semantic and intent gap. Search engines generate millions of click data (a.k.a. image-query pairs), which provide almost "unlimited" yet strong connections between semantics and images, as well as connections between users' intents and queries. To study the intrinsic properties of click data and to investigate how to effectively leverage this huge amount of data to bridge semantic and intent gap is a promising direction to advance multimedia research. In the past, the primary obstacle is that there is no such dataset available to the public research community. This changes as Microsoft has released a new large-scale real-world image click data to public. This paper presents preliminary studies on the power of large-scale click data with a variety of experiments, such as building large-scale concept detectors, tag processing, search, definitive tag detection, intent analysis, etc., with the goal to inspire deeper researches based on this dataset.

In this paper, we present a log-based study on user search behavior comparisons on three different platforms: desktop, mobile and tablet. We use three-month search logs in 2012 from a commercial search engine for our study. Our objective is to better understand how and to what extent mobile and tablet searchers behave differently than desktop users. Our study spans a variety of aspects including query categorization, query length, search time distribution, search location distribution, user click patterns and so on. From our data set, we reveal that there are significant differences between user search patterns in these three platforms, and therefore use the same ranking system is not an optimal solution for all of them. Consequently, we propose a framework that leverages a set of domain-specific features, along with the training data from desktop search, to further improve the search relevance for mobile and tablet platforms. Experimental results demonstrate that by transferring knowledge from desktop search, search relevance on mobile and tablet can be greatly improved.

Web search components such as ranking and query suggestions analyze the user data provided in query and click logs. While this data is easy to collect and provides information about user behavior, it omits user interactions with the search engine that do not hit the server; these logs omit search data such as users' cursor movements. Just as clicks provide signals for relevance in search results, cursor hovering and scrolling can be additional implicit signals. In this work, we demonstrate a technique to extend models of the user's search result examination state to infer document relevance. We start by exploring recorded user interactions with the search results, both qualitatively and quantitatively. We find that cursor hovering and scrolling are signals telling us which search results were examined, and we use these interactions to reveal latent variables in searcher models to more accurately compute document attractiveness and satisfaction. Accuracy is evaluated by computing how well our model using these parameters can predict future clicks for a particular query. We are able to improve the click predictions compared to a basic searcher model for higher ranked search results using the additional log data.

Query segmentation is an important task toward understanding queries accurately, which is essential for improving search results. Existing segmentation models either use labeled data to predict the segmentation boundaries, for which the training data is expensive to collect, or employ unsupervised strategy based on a large text corpus, which might be inaccurate because of the lack of relevant information. In this paper, we propose a probabilistic model to exploit clickthrough data for query segmentation, where the model parameters are estimated via an efficient EM algorithm. We further study how to properly interpret the segmentation results and utilize them to improve retrieval accuracy. Specifically, we propose an integrated language model based on the standard bigram language model to exploit the probabilistic structure obtained through query segmentation. Experiment results on two datasets show that our segmentation model outperforms existing segmentation models. Furthermore, extensive experiments on a large retrieval dataset reveals that the results of query segmentation can be leveraged to improve retrieval relevance by using the proposed integrated language model.

In this paper, we provide an overview of the MSR-Bing Web Scale Speller Challenge of 2011. We describe the motivation and outline the algorithmic and engineering challenges posed by this activity. The design and the evaluation methods are also reviewed, and the online resources that will remain publicly available to the community are also described. The Challenge will culminate in a workshop after the time of the writing where the top prize winners will publish their approaches. The main findings and the lessons learned will be summarized and shared in the Industry Track presentation accompanying this paper.

This paper presents Partially Observable Markov model with Duration (POMD), a statistical method that addresses the challenge of understanding sophisticated user behaviors from the search log in which some user actions, such as reading and skipping search results, cannot be observed and recorded. POMD utilizes not only the positional but also the temporal information of the clicks in the log. In this work, we treat the user engagements with a search engine as a Markov process, and model the unobservable engagements as hidden states. POMD differs from the traditional hidden Markov model (HMM) in that not all the hidden state transitions emit observable events, and that the duration of staying in each state is explicitly factored into the core statistical model. To address the training and decoding issues emerged as the results of the variations, we propose an iterative two-stage training algorithm and a greedy segmental decoding algorithm respectively. We validate the proposed algorithm with two sets of experiments. First, we show that the search behavioral patterns inferred by POMD match well with those reported in the eye tracking experiments. Secondly, through a series of A/B comparison experiments, we demonstrate that POMD can distinguish the ranking qualities of different search engine configurations much better than the patterns inferred by the model proposed in the previous work. Both of the experimental results suggest that POMD can provide a statistical and quantitative way to understand the sophisticated search behaviors by simply mining the search logs.

This paper uses the URL word breaking task as an example to elaborate what we identify as crucial in designing statistical natural language processing (NLP) algorithms for Web scale applications: (1) rudimentary multilingual capabilities to cope with the global nature of the Web, (2) multi-style modeling to handle diverse language styles seen in the Web contents, (3) fast adaptation to keep pace with the dynamic changes of the Web, (4) minimal heuristic assumptions for generalizability and robustness, and (5) possibilities of efficient implementations and minimal manual efforts for processing massive amount of data at a reasonable cost. We first show that the state-of-the-art word breaking techniques can be unified and generalized under the Bayesian minimum risk (BMR) framework that, using a Web scale N-gram, can meet the first three requirements. We discuss how the existing techniques can be viewed as introducing additional assumptions to the basic BMR framework, and describe a generic yet efficient implementation called word synchronous beam search. Testing the framework and its implementation on a series of large scale experiments reveals the following. First, the language style used to build the model plays a critical role in the word breaking task, and the most suitable for the URL word breaking task appears to be that of the document title where the best performance is obtained. Models created from other language styles, such as from document body, anchor text, and even queries, exhibit varying degrees of mismatch. Although all styles benefit from increasing modeling power which, in our experiments, corresponds to the use of a higher order N-gram, the gain is most recognizable for the title model. The heuristics proposed by the prior arts do contribute to the word breaking performance for mismatched or less powerful models, but are less effective and, in many cases, lead to poorer performance than the matched model with minimal assumptions. For the matched model based on document titles, an accuracy rate of 97.18% can already be achieved using simple trigram without any heuristics.

Analyzing clickthrough log data is important for improving search performance as well as understanding user behaviors. In this paper, we propose a novel collaborative ranking model to tackle two difficulties in analyzing clickthrough log. First, previous studies have shown that users tend to click top-ranked results even they are less relevant. Therefore, we use pairwise ranking relation to avoid the position bias in clicks. Second, since click data are extremely sparse with respect to each query or user, we construct a collaboration model to eliminate the sparseness problem. We also find that the proposed model and previous popular used click-based models address different aspects of clickthrough log data. We further propose a hybrid model that can achieve significant improvement compared to the baselines on a large-scale real world dataset.

Web documents are typically associated with many text streams, including the body, the title and the URL that are determined by the authors, and the anchor text or search queries used by others to refer to the documents. Through a systematic large scale analysis on their cross entropy, we show that these text streams appear to be composed in different language styles, and hence warrant respective language models to properly describe their properties. We propose a language modeling approach to Web document retrieval in which each document is characterized by a mixture model with components corresponding to the various text streams associated with the document. Immediate issues for such a mixture model arise as all the text streams are not always present for the documents, and they do not share the same lexicon, making it challenging to properly combine the statistics from the mixture components. To address these issues, we introduce an 'open-vocabulary' smoothing technique so that all the component language models have the same cardinality and their scores can simply be linearly combined. To ensure that the approach can cope with Web scale applications, the model training algorithm is designed to require no labeled data and can be fully automated with few heuristics and no empirical parameter tunings. The evaluation on Web document ranking tasks shows that the component language models indeed have varying degrees of capabilities as predicted by the cross-entropy analysis, and the combined mixture model outperforms the state-of-the-art BM25F based system.

This article describes an application of the partially observable Markov (POM) model to the analysis of a large scale commercial web search log. Mathematically, POM is a variant of the hidden Markov model in which all the hidden state transitions do not necessarily emit observable events. This property of POM is used to model, as the hidden process, a common search behavior that users would read and skip search results, leaving no observable user actions to record in the search logs. The Markov nature of the model further lends support to cope with the facts that a single observed sequence can be probabilistically associated with many hidden sequences that have variable lengths, and the search results can be read in various temporal orders that are not necessarily reflected in the observed sequence of user actions. To tackle the implementation challenges accompanying the flexibility and analytic powers of POM, we introduce segmental Viterbi algorithm based on segmental decoding and Viterbi training to train the POM model parameters and apply them to uncover hidden processes from the search logs. To validate the model, the latent variables modeling the browsing patterns on the search result page are compared with the experimental data of the eye tracking stu-dies. The close agreements suggest that the search logs do contain rich information of user behaviors in browsing the search result page even though they are not directly observable, and that using POM to understand these sophisticated search behaviors is a promising approach.

It has been widely observed that search queries are composed in a very different style from that of the body or the title of a document. Many techniques explicitly accounting for this language style discrepancy have shown promising results for information retrieval, yet a large scale analysis on the extent of the language differences has been lacking. In this paper, we present an extensive study on this issue by examining the language model properties of search queries and the three text streams associated with each web document: the body, the title, and the anchor text. Our information theoretical analysis shows that queries seem to be composed in a way most similar to how authors summarize documents in anchor texts or titles, offering a quantitative explanation to the observations in past work.

Although higher order language models (LMs) have shown benefit of capturing word dependencies for Information retrieval(IR), the tuning of the increased number of free parameters remains a formidable engineering challenge. Consequently,in many real world retrieval systems, applying higher order LMs is an exception rather than the rule. In this study, we address the parameter tuning problem using a framework based on a linear ranking model in which different component models are incorporated as features. Using unigram and bigram LMs with 2 stage smoothing as examples, we show that our method leads to a bigram LM that outperforms significantly its unigram counterpart and the well-tuned BM25 model.

In this paper, we propose a semi-supervised learning approach for classifying program (bot) generated web search traffic from that of genuine human users. The work is motivated by the challenge that the enormous amount of search data pose to traditional approaches that rely on fully annotated training samples. We propose a semi-supervised framework that addresses the problem in multiple fronts. First, we use the CAPTCHA technique and simple heuristics to extract from the data logs a large set of training samples with initial labels, though directly using these training data is problematic because the data thus sampled are biased. To tackle this problem, we further develop a semi-supervised learning algorithm to take advantage of the unlabeled data to improve the classification performance. These two proposed algorithms can be seamlessly combined and very cost efficient to scale the training process. In our experiment, the proposed approach showed significant (i.e. 2:1) improvement compared to the traditional supervised approach.

In this paper, we describe CALM, a method for building statistical language models for the Web. CALM addresses several unique challenges dealing with the Web contents. First, CALM does not rely on the whole corpus to be available to build the language model. Instead, we design CALM to progressively adapt itself as Web chunks are made available by the crawler. Second, given the dynamic and dramatic changes in the Web contents, CALM is designed to quickly enrich its lexicon and N-grams as new vocabulary and phrases are discovered. To reduce the amount of heuristics and human interventions typically needed for model adaptation, we derive an information theoretical formula for CALM to facilitate the optimal adaptation in the maximum a posteriori (MAP) sense. Testing against a collection of Web chunks where new vocabulary and phrases are dominant, we show CALM can achieve comparable and satisfactory model measured in perplexity. We also show CALM is robust against over training and the initial condition, suggesting that any assumptions made in obtaining the initial model can gradually see their impacts diminished as CALM runs its full course and adapt to more data.

In this article, we report our efforts in mining the information encoded as clickthrough data in the server logs to evaluate and monitor the relevance ranking quality of a commercial web search engine. We describe a metric called pSkip that aims to quantify the ranking quality by estimating the probability of users encountering non relevant results that cost them the efforts to read and skip. A search engine with a lower pSkip is regarded as having a better ranking quality. A key design goal of pSkip is to integrate the findings from two sets of user studies that utilize eye-tracking devices to track users' browsing patterns on the search result pages, and that use specially instrumented browsers to actively solicit users' explicit judgments on their search activities. We present the derivation of the maximum likelihood estimation of pSkip and demonstrate its efficacy in describing the user study data. The mathematical properties of pSkip are further analyzed and compared with several objective metrics as well as the cumulated gain method that uses subjective judgments. Experimental data show that pSkip can measure aspects of the search quality that these existing metrics are not designed or fail to address, such as identifying the real search intents expressed in the ambiguous queries. Although effective and superior in many ways, we also report a series of experiments that show pSkip may be influenced by system issues that are not directly related to relevance ranking, suggesting that measurements complementary to pSkip are still needed in order to form a holistic and accurate characterization of the ranking quality.

Field speech data pose great challenges to statistical modeling because the speech signal is often intermixed with extraneous sounds and other environmental noise either that are too difficult to compensate dynamically or for which it is too expensive to collect sufficient data for proper offline training. We propose a detection based method in which the speech recognizer can sharply tune to only the "meaningful" speech and gracefully ignore the "unwanted" audio segments. The method is designed to be integrated with the frame synchronous search for a single pass processing. In contrast to the conventional keyword spotting techniques, this integration allows the use of the language model for better predicting the detection targets during the search. To study its efficacy, we apply the framework to a spontaneous speech understanding application where cohesive phrases congruent to the domain semantics and application context are used as the salient feature for selective hearing. Experimental results on the effectiveness of the system in dealing with out of domain phrases and other spontaneous speech effects are encouraging.

Computer Supported Telecommunication Applications (CSTA) is a widely adopted ECMA/ISO standard suite for global and enterprise communications. As it becomes evident that spoken language systems are playing an ever important role in telecommunications, CSTA has been revised to include a specification on spoken language interfaces so as to promote the interoperability and adoption of spoken language technologies. The new specification shares the same design philosophy as the Speech Application Language Tags, or SALT, that provides simple yet rich functionality for spoken language processing. The tight integration of speech into the telecommunication infrastructure not only presents many technical benefits but also ushers in an era that can potentially make spoken language technologies even more prevalent in our daily lives. This paper describes the design considerations for the specification, and shows how this new standard can be used in conjunction with other standards to create powerful speech applications.

SALT is an industrial standard that enables speech input/output for Web applications. Although the core design is to make simple tasks easy, SALT gives the designers ample fine-grained controls to create advanced user interface. The paper exploits a speech input mode in which SALT would dynamically report partial semantic parses while audio capturing is still ongoing. The semantic parses can be evaluated and the outcome reported immediately back to the user. The potential impact for the dialog systems is that tasks conventionally performed in a system turn can now be carried out in the midst of a user turn, thereby presenting a significant departure from the conventional turn-taking. To assess the efficacy of such highly interactive interface, more user studies are undoubtedly needed. This paper demonstrates how SALT can be employed to facilitate such studies.

Speech recognition of names in Personal Information Management (PIM) systems is an important yet difficult task. The difficulty arises from various sources: the large number of possible names that users may speak, different ways a person may be referred to, ambiguity when only first names are used, and mismatched pronunciations. In this paper we present our recent work on name recognition with User Modeling (UM), i.e., automatic modeling of user’s behavior patterns. We show that UM and our learning algorithm lead to significant improvement in the perplexity, Out Of Vocabulary rate, recognition speed, and accuracy of the top recognized candidate. The use of an exponential window reduces the perplexity by more than 30%. 1.

This paper describes the Speech Application Language Tags, or SALT, an emerging spoken language interface standard for multimodal or speech-only applications. A key premise in SALT design is speech-enabled user interface shares a lot of the design principles and computational requirements with the graphical user interface (GUI). As a result, it is logical to introduce into speech the object-oriented, event-driven model that is known to be flexible and powerful enough in meeting the requirements for realizing sophisticated GUIs. It is hopeful that reusing this rich infrastructure can enable dialog designers to focus more on the core user interface design issues than on the computer and software engineering details. The paper centers the discussion on the Web-based distributed computing environment and elaborates how SALT can be used to implement multimodal dialog systems. How advanced dialog effects (e.g., cross-modality reference resolution, implicit confirmation, multimedia synchronization) can be realized in SALT is also discussed.

This paper describes the main components of MiPad (multimodal interactive PAD) and especially its distributed speech processing aspects. MiPad is a wireless mobile PDA prototype that enables users to accomplish many common tasks using a multimodal spoken language interface and wireless-data technologies. It fully integrates continuous speech recognition and spoken language understanding, and provides a novel solution for data entry in PDAs or smart phones, often done by pecking with tiny styluses or typing on minuscule keyboards. Our user study indicates that the throughput of MiPad is significantly superior to that of the existing pen-based PDA interface. Acoustic modeling and noise robustness in distributed speech recognition are key components in MiPad's design and implementation. In a typical scenario, the user speaks to the device at a distance so that he or she can see the screen. The built-in microphone thus picks up a lot of background noise, which requires MiPad be noise robust. For complex tasks, such as dictating e-mails, resource limitations demand the use of a client-server (peer-to-peer) architecture, where the PDA performs primitive feature extraction, feature quantization, and error protection, while the transmitted features to the server are subject to further speech feature enhancement, speech decoding and understanding before a dialog is carried out and actions rendered. Noise robustness can be achieved at the client, at the server or both. Various speech processing aspects of this type of distributed computation as related to MiPad's potential deployment are presented. Previous user interface study results are also described. Finally, we point out future research directions as related to several key MiPad functionalities.

Dr. Who is a Microsoft research project aiming at creating a speech-centric multimodal interaction framework, which serves as the foundation for the NET natural user interface. MiPad is the application prototype that demonstrates compelling user advantages for wireless personal digital assistant (PDA) devices, MiPad fully integrates continuous speech recognition (CSR) and spoken language understanding (SLU) to enable users to accomplish many common tasks using a multimodal interface and wireless technologies. It tries to solve the problem of pecking with tiny styluses or typing on minuscule keyboards in today's PDAs. Unlike a cellular phone, MiPad avoids speech-only interaction. It incorporates a built-in microphone that activates whenever a field is selected. As a user taps the screen or uses a built in roller to navigate, the tapping action narrows the number of possible instructions for spoken word understanding. MiPad currently runs on a Windows CE Pocket PC with a Windows 2000 machine where speech recognition is performed. The Dr Who CSR engine uses a unified CFG and n-gram language model. The Dr Who SLU engine is based on a robust chart parser and a plan-based dialog manager. The paper discusses MiPad's design, implementation work in progress, and preliminary user study in comparison to the existing pen-based PDA interface.

In this paper, we propose an analytically tractable framework that integrates the frame and segment based acoustic modeling techniques. We combine the two approaches by jointly modeling their respective hidden Markov processes. Since the joint process is based on the same mathematical framework, conventional search and training techniques, such as Viterbi and EM algorithms, can be directly applied. It also allows the score from either model to contribute to the training and decoding of the other, reaching a jointly optimal decision. We conducted two series of experiments to verify our hypotheses. In the phone-pair classification experiments, our segment models show a 24% error reduction over state-of-the-art HMM-based system. The superior quality of segment models contributes to an 8.2% reduction in word error rates for the unified system on the WSJ dictation task.

MiPad is one of the application prototypes in a project codenamed Dr Who. As a wireless Personal Digital Assistant (PDA), MiPad fully integrates continuous speech recognition (CSR) and spoken language understanding (SLU) to enable users to accomplish many common tasks using a multimodal interface and wireless technologies. It tries to solve the problem of pecking with tiny styluses or typing on minuscule keyboards in today’s PDAs or smart phones. It also avoids the problem of being a cellular telephone that depends on speech-only interaction. MiPad incorporates a built-in microphone that activates whenever a field is selected. As a user taps the screen or uses a built-in roller to navigate, the tapping action narrows the number of possible instructions for spoken language processing. MiPad currently runs on a Windows CE Pocket PC with a Windows 2000 Server where speech recognition is performed. The Dr Who CSR engine has a 64k word vocabulary with a unified context-free grammar and n-gram language model. The Dr Who SLU engine is based on a robust chart parser and a plan-based dialog manager. This paper discusses MiPad’s design, implementation work in progress, and preliminary user study in comparison to the existing pen-based PDA interface. 1.

In this paper, we describe an implementation of a plan-based multimodal dialog system using the extensible markup language (XML). The dialog manager receives semantic objects representing the user’s utterance at the end of each user’s turn. We define a semantic markup language (SML), based on XML, to describe these semantic objects. Following the principles of XML Schema, we define the schema of SML in another XML called semantic definition language (SDL). In addition to supporting many discourse and dialog features, SDL is also designed to represent the domain knowledge via the application schema and the hierarchy of the semantic objects. We show that, with a thoughtful design in SDL, SML can be expressive enough that the behavior of a dialog planner can be fully specified in the extensible stylesheet language (XSL), a standardized language with a logical programming model that is most popular for implementing intelligent systems.

In this paper, we present a dialog system that extends the planbased approach with two features. Instead of Boolean inference, we include into the system the probabilistic measures from the front-end speech and language processes. As a result, rules can be activated and facts gathered based on statistical confidence measures. We also introduce the notion of entity types to classify the rules and facts. The entity types, derived from the schema of the knowledge base, assist the semantic evaluation process by indicating which rules and facts are interoperable. The semantic evaluation and dialog planning can therefore be better insulated among tasks, and be encapsulated into reusable components. To assess the feasibility and discover the areas for improvements for this framework, we launched the project DR. WHO, in which we strive to develop a speech interface for a personal information management application. We describe the current progress in the discourse area in this paper.

This paper reports our progress in building a mixed initiative, goal-oriented dialogue system for human machine interactions. The dialogue model embraces the spirits of the so-called plan-based approach in that the dialogue flow is not statically authored but dynamically generated by the system as a natural outcome of the semantic evaluation process. With multimodal applications in mind, the dialogue system is designed in an event driven architecture that is commonly seen in the core of a graphical user interface (GUI) environment. In the same manner that GUI events are handled by graphical objects, the proposed dialogue model assigns dialogue events to semantic objects that encapsulate the knowledge for handling events under various discourse contexts. Currently, we have found that four types of events, namely, dialogue object instantiation, semantic evaluation, dialogue repair, and discourse binding, are sufficient for a wide range of applications. 1.

In this article, a mathematical framework that jointly optimizes the parameters of classifier and feature extractor is presented. In this approach, feature extraction is formulated as a process of projecting the signals onto a smaller subspace in which the statistical properties of the signal can be efficiently modeled. An algorithm, called statistical matching pursuit (SMP), is proposed to learn from the training data the optimal projection dimensions and the extent of signal reduction. The algorithm is designed to achieve unconditional convergence and can be seamlessly incorporated into the expectation-maximization (EM) algorithm employed to train the classifier. Finally, we report some experimental results on speech recognition and elaborate the potential of the proposed method.

Modern speech processing research may be categorized into three broad areas: statistical, physiological, and perceptual. Statistical research investigates the nature of the variability of the speech waveform from a signal processing viewpoint. This approach relates to the processing of speech in order to obtain measurements of speech characteristics which demonstrate manageable variabilities across a wide range of the talker population, in the presence of noise or competing speakers as well as the interaction of speech with the channel through which it is transmitted, and under the inherent interaction of the information content of speech itself (i.e., the contextual factor). Physiological research aims at constructing accurate models of the articulatory and auditory process, helping to limit the signal space for speech processing. In the perceptual realm, work focuses on understanding the psychoacoustic and possibly the psycholinguistic aspects of the speech communication process that the human so conveniently conducts. By studying this working analysis/recognition system, insights may be garnered that will lead to improved methods of speech processing. Conversely by studying the limitations of this system, particularly how it reduces the information rate of the received signal through, for example, masking and adaptation improvements may be made in the efficiency of speech coding schemes without impacting the quality of the reconstructed speech. Thus comprehension of speech production and perception impacts methods of speech processing, and vice-versa. This paper enunciates such a position, focusing on how modern time-frequency signal analysis methods could help expedite needed advances in these areas.

We utilize speech information to improve the quality of audio-visual communications such as videotelephony and videoconferencing. In particular, the marriage of speech analysis and image processing can solve problems related to lip synchronization. We present a technique called speech-assisted frame-rate conversion. Demonstration sequences are presented. Other applications, including speech-assisted video coding, are outlined.>

A model of spectral shape analysis in the central auditory system is developed based on neurophysiological mappings in the primary auditory cortex and on results from psychoacoustical experiments in human subjects. The model suggests that the auditory system analyzes an input spectral pattern along three independent dimensions: a logarithmic frequency axis, a local symmetry axis, and a local spectral bandwidth axis. It is shown that this representation is equivalent to performing an affine wavelet transform of the spectral pattern and preserving both the magnitude (a measure of the scale or local bandwidth of the spectrum) and phase (a measure of the local symmetry of the spectrum). Such an analysis is in the spirit of the cepstral analysis commonly used in speech recognition systems, the major difference being that the double Fourier-like transformation that the auditory system employs is carried out in a local fashion. Examples of such a representation for various speech and synthetic signals are discussed, together with its potential significance and applications for speech and audio processing.>

A common sequence of operations in the early stages of most sensory systems is a multiscale transform followed by a compressive nonlinearity. The authors explore the contribution of these operations to the formation of robust and perceptually significant representation in the early auditory system. It is shown that auditory representation of the acoustic spectrum is effectively a self-normalized spectral analysis, i.e., the auditory system computes a spectrum divided by a smoothed version of itself. Such a self-normalization induces significant effects such as spectral shape enhancement and robustness against scaling and noise corruption. Examples using synthesized signals and a natural speech vowel are presented to illustrate these results. Furthermore, the characteristics of auditory representation are discussed in the context of several psychoacoustical findings, together with the possible benefits of this model for various engineering applications.>

An analytically tractable framework is presented to describe mechanical and neural processing in the early stages of the auditory system. Algorithms are developed to assess the integrity of the acoustic spectrum at all processing stages. The algorithms employ wavelet representations, multiresolution processing, and the method of convex projections to construct a close replica of the input stimulus. Reconstructions using natural speech sounds demonstrate minimal loss of information along the auditory pathway. Close inspection of the final auditory patterns reveals spectral enhancements and noise suppression that have close perceptual correlates. The functional significance of the various auditory processing stages is discussed in light of the model, together with their potential applications in automatic speech recognition and low bit-rate data compression.>