Knowledge graph reasoning (KGR) is a key technology that infers missing facts in knowledge graphs (KGs). Given that real-world scenarios typically encounter unseen KGs with new entities and new relations, researchers have begun to explore fully inductive KGR methods. This setting presents greater challenges and has not been fully explored. Current methods primarily construct relation graphs based on the original KG to facilitate message passing between relations. These models have made significant progress in achieving fully inductive reasoning. However, as relation graphs focus solely on the co-occurrence patterns between relations, they often fail to capture reasoning patterns in KGs, which causes the model to struggle in effectively distinguishing between different relations and entities. This limitation severely restrict the reasoning capabilities of existing methods. In light of this, we propose the Adaptive Rule Pattern Transfer model (AdaRPT) for KGR. It aims to leverage logical rules for each relation in the KG to learn more comprehensive and transferable knowledge representations for entities and relations. For entities, we design a non-parameter message passing model that aggregates path information from the query entity to other entities. The path information for each entity is then matched with rules to obtain the transferable feature of each entity. And for relations, we extract both reasoning and co-occurrence patterns from KGs as transferable relation features. Finally, a path-based graph neural network (GNN) is employed on the transferable features of entities and relations to perform reasoning on KGs. Extensive experimental evaluations on 43 datasets for both inductive and transductive reasoning demonstrate the effectiveness and generalization capability of AdaRPT.

Dense passage retrieval has become the mainstream method in the first stage of open-domain question answering, which usually adopts a bi-encoder structure to learn the dense representation of questions and passages for semantic matching. One of the main challenges currently is to effectively utilize more informative hard negatives. Many efforts have been made to address this challenge by reducing the discrepancy between training and inference and training expensive cross-encoder for mining hard negatives. However, none of these approaches consider the theme information regarding the candidate passages. In this paper, we propose an approach for theme-enhanced hard negative sample mining, called THNSM. Firstly, we employ a topic sampling approach to mine hard negatives that are related to the topic. Secondly, we introduce a straightforward yet effective method for generating even more challenging hard negative samples by partially fusing their features. Empirical results on popular open-domain QA datasets demonstrate that THNSM significantly outperforms previous state-of-the-art models.

As the field of machine reading comprehension (MRC) continues to evolve, it is unlocking enormous potential for its practical application. However, the currently well-performing models predominantly rely on massive pre-trained language models with at least several hundred million or even over one hundred billion parameters. These complex models not only require immense computational power but also extensive storage, presenting challenges for resource-limited environments such as online education.Current research indicates that specific capabilities of larger models can be transferred to smaller models through knowledge distillation. However, prior to our work, there were no small models specifically designed for MRC task with complex reasoning abilities. In light of this, we present a novel multi-teacher multi-stage distillation approach, MTMS. It facilitates the easier deployment of reasoning-based MRC task on resource-constrained devices, thereby enabling effective applications. In this method, we design a multi-teacher distillation framework that includes both a logical teacher and a semantic teacher. This framework allows MTMS to simultaneously extract features from different perspectives of the text, mitigating the limitations inherent in single-teacher information representations. Furthermore, we introduce a multi-stage contrastive learning strategy. Through this strategy, the student model can progressively align with the teacher models, effectively bridging the gap between them. Extensive experimental outcomes on two inference-based datasets from real-world scenarios demonstrate that MTMS requires nearly 10 times fewer parameters compared with the teacher model size while achieving the competitive performance.

Temporal knowledge graph reasoning aims to predict the missing links (facts) in the future timestamps. However, most existing methods have a common limitation: they focus on learning dynamic representations of temporal knowledge graphs and rarely consider static characteristics that remain unchanged over time. To address the above issues, we propose to learn the dynamic and static representations for temporal knowledge graph reasoning (DSTKG), which introduces two latent variables to capture the dynamic and static characteristics of entities in temporal knowledge graphs. First, we use a Bi-GRU-based inference network to learn the static latent representation of historical facts and a nonlinear discrete-time transition-based inference network to learn the dynamic latent representation. Then, we sample the latent variables multiple times using re-parameterization tricks to obtain high-quality embeddings and make predictions in the future timestamps. The empirical results on four benchmark datasets show that our model is more effective than state-of-the-art approaches. Compared with the strong baseline model DBKGE (RotatE), the proposed model achieves performance improvements of 2.69%,$1.59\%$, 1.18% and 1.22% on Yago11k, Wikidata12k, ICEWS14 and ICEWS05-15 respectively, regarding the evaluation metric MRR.

Complex knowledge graph question answering (KGQA) aims at answering natural language questions by entities retrieving from a knowledge graph (KG). Recently, the relation path-based models have shown the unique advantage for complex KGQA. However, these existing models ignore the dependency between different relation paths, which leads to aimless reasoning over the KG. To resolve this issue, we propose the question-directed reasoning with relation-aware graph attention network (QRGAT) that encodes the reasoning process as a reasoning graph. The relation-aware GAT can recognize neighbor entities along with the corresponding relations for each entity. With the relation-aware GAT stacked in multiple layers, it can collaboratively capture the dependency of different relation paths for each entity. The question-directed reasoning utilizes the information learned by the relation-aware GAT to solve the aimless reasoning on the KG by constructing a reasoning graph. Extensive experiments demonstrate that our QRGAT outperforms the baseline models on both popular datasets WebQuestionsSP and ComplexWebQuestions. Compared with the strong GNN-based baseline NSM$_{+h}$, our QRGAT achieves the performance improvements of 2.3% on WebQuestionsSP and 3.6% on ComplexWebQuestions by the metric Hits@1.

The task of machine reading comprehension (MRC) is to enable machine to read and understand a piece of text and then answer the corresponding question correctly. This task requires machine to not only be able to perform semantic understanding but also possess logical reasoning capabilities. Just like human reading, it involves thinking about the text from two interacting perspectives of semantics and logic. However, previous methods based on reading comprehension either consider only the logical structure of the text or only the semantic structure of the text and cannot simultaneously balance semantic understanding and logical reasoning. This single form of reasoning cannot make the machine fully understand the meaning of the text. Additionally, the issue of sparsity in composition presents a significant challenge for models that rely on graph-based reasoning. To this end, a cross-graph knowledge propagation network (CGKPN) with adaptive connection is presented to address the above issues. The model first performs self-view node embedding on the constructed logical graph and semantic graph to update the representations of the graphs. Specifically, a relevance matrix between nodes is introduced to adaptively adjust node connections in response to the challenge posed by sparse graph. Subsequently, CGKPN conducts cross-graph knowledge propagation on nodes that are identical in both graphs, effectively resolving conflicts arising from identical nodes in different views, and enabling the model to better integrate the logical and semantic relationships of the text through efficient interaction. Experiments on the two MRC datasets ReClor and LogiQA indicate the superior performance of our proposed model CGKPN compared to other existing baselines.

Knowledge Graph Completion (KGC) has garnered massive research interest recently, and most existing methods are designed following a transductive setting where all entities are observed during training. Despite the great progress on the transductive KGC, these methods struggle to conduct reasoning on emerging KGs involving unseen entities. Thus, inductive KGC, which aims to deduce missing links among unseen entities, has become a new trend. Many existing studies transform inductive KGC as a graph classification problem by extracting enclosing subgraphs surrounding each candidate triple. Unfortunately, they still face certain challenges, such as the expensive time consumption caused by the repeat extraction of enclosing subgraphs, and the deficiency of entity-independent feature learning. To address these issues, we propose a global-local anchor representation (GLAR) learning method for inductive KGC. Unlike previous methods that utilize enclosing subgraphs, we extract a shared opening subgraph for all candidates and perform reasoning on it, enabling the model to perform reasoning more efficiently. Moreover, we design some transferable global and local anchors to learn rich entity-independent features for emerging entities. Finally, a global-local graph reasoning model is applied on the opening subgraph to rank all candidates. Extensive experiments show that our GLAR outperforms most existing state-of-the-art methods.

Temporal knowledge graph embedding (TKGE) aims to learn the embedding of entities and relations in a temporal knowledge graph (TKG). Although the previous graph neural networks (GNN) based models have achieved promising results, they cannot directly capture the interactions of multi-facts at different timestamps. To address the above limitation, we propose a time-aware relational graph attention model (TARGAT), which takes the multi-facts at different timestamps as a unified graph. First, we develop a relational generator to dynamically generate a series of time-aware relational message transformation matrices, which jointly models the relations and the timestamp information into a unified way. Then, we apply the generated message transformation matrices to project the neighborhood features into different time-aware spaces and aggregate these neighborhood features to explicitly capture the interactions of multi-facts. Finally, a temporal transformer classifier is applied to learn the representation of the query quadruples and predict the missing entities. The experimental results show that our TARGAT model beats the GNN-based models by a large margin and achieves new state-of-the-art results on four popular benchmark datasets.

Question generation over knowledge bases is an important research topic. How to deal with rare and low-frequency words in traditional generation models is a key challenge for question generation. Although the copy mechanism provides significant performance improvements, the original copy mechanism weakens the focus on aspect generation in the overall representations. In this article, we present a novel method to improve question generation with a question type enhanced representation and attention-based copy mechanism. The proposed method exploits the advantages of the generate mode in the copy mechanism and replaces objects in the factual triples with question types, which attempts to improve the output quality in the generate mode and effectively generate questions with proper interrogative words. We evaluate the proposed method on two standard benchmark datasets. The experimental results demonstrate that our proposed method can produce higher-quality questions than these of the Encoder-Decoder-based and CopyNet-based methods.

This article considers the task of text style transfer: transforming a specific style of sentence into another while preserving its style-independent content. A dominate approach to text style transfer is to learn a good content factor of text, define a fixed vector for every style and recombine them to generate text in the required style. In fact, there are a large number of different words to convey the same style from different aspects. Thus, using a fixed vector to represent one style is very inefficient, which causes the weak representation power of the style vector and limits text diversity of the same style. To address this problem, we propose a novel neural generative model called Adversarial Separation Network (ASN), which can learn the content and style vector jointly and the learnt vectors have strong representation power and good interpretabilities. In our method, adversarial learning is implemented to enhance our model’s capability of disentangling the two factors. To evaluate our method, we conduct experiments on two benchmark datasets. Experimental results show our method can perform style transfer better than strong comparison systems. We also demonstrate the strong interpretability of the learnt latent vectors.

Designing algorithms to solve math word problems (MWPs) is an important research topic in natural language processing and smart education domains. The task of solving MWPs involves transforming math problem texts into math equations. Although recent Graph2Tree-based models, which adopt homogeneous graph encoders to learn quantity representations, have obtained very promising results in generating math equations, they do not consider the heterogeneous issue and the long-distance dependencies of heterogeneous nodes. In this paper, we propose a novel hierarchical heterogeneous graph encoding called HGEN for MWPs. Specifically, HGEN first introduces a heterogeneous graph consisting of a node-level attention layer and a type-aware attention layer to learn the heterogeneous node embedding. HGEN then captures the long-distance dependent information by propagating the multi-hop nodes in a hierarchical manner. We conduct extensive experiments on two popular MWP datasets. Our empirical results show that HGEN significantly outperforms the state-of-the-art Graph2Tree-based models in the literature.

Cross-lingual entity alignment has attracted considerable attention in recent years. Past studies using conventional approaches to match entities share the common problem of missing important structural information beyond entities in the modeling process. This allows graph neural network models to step in. Most existing graph neural network approaches model individual knowledge graphs (KGs) separately with a small amount of pre-aligned entities served as anchors to connect different KG embedding spaces. However, this characteristic can cause several major problems, including performance restraint due to the insufficiency of available seed alignments and ignorance of pre-aligned links that are useful in contextual information in-between nodes. In this article, we propose DuGa-DIT, a dual gated graph attention network with dynamic iterative training, to address these problems in a unified model. The DuGa-DIT model captures neighborhood and cross-KG alignment features by using intra-KG attention and cross-KG attention layers. With the dynamic iterative process, we can dynamically update the cross-KG attention score matrices, which enables our model to capture more cross-KG information. We conduct extensive experiments on two benchmark datasets and a case study in cross-lingual personalized search. Our experimental results demonstrate that DuGa-DIT outperforms state-of-the-art methods.

The graph attention network (GAT) [1] has started to become a mainstream neural network architecture since 2018, yielding remarkable performance gains in various natural language processing (NLP) tasks. Although GAT has reached the state-of-the-art (SOTA) performance as a recent success in relation prediction in knowledge graph, the current model is still limited by the following two aspects: (1) the existing model only considers the neighbors from the inbound-direction of the given entity, but ignores the rich neighborhood information from outbound-directions; (2) the existing model only uses the k-th hop output to learn the multi-hop embeddings, which leads to the loss of a large amount of early-stage embedding information (e.g., one-hop) at the graph attention step. In this study, we propose a novel bidirectional graph attention network (BiGAT) to learn the hierarchical neighbor propagation. In our proposed BiGAT, an inbound-directional GAT and an outbound-directional GAT are introduced to capture sufficient neighborhood information before propagating the bidirectional neighborhood information to learn the multi-hop feature embeddings in a hierarchical manner. Experiments conducted on the four publicly available datasets show that BiGAT achieves the competitive results in comparison to other SOTA methods.

The goal of Knowledge graph embedding (KGE) is to learn how to represent the low dimensional vectors for entities and relations based on the observed triples. The conventional shallow models are limited to their expressiveness. ConvE (Dettmers et al., 2018) takes advantage of CNN and improves the expressive power with parameter efficient operators by increasing the interactions between head and relation embeddings. However, there is no structural information in the embedding space of ConvE, and the performance is still limited by the number of interactions. The recent KBGAT (Nathani et al., 2019) provides another way to learn embeddings by adaptively utilizing structural information. In this paper, we take the benefits of ConvE and KBGAT together and propose a Relation-aware Inception network with joint local-global structural information for knowledge graph Embedding (ReInceptionE). Specifically, we first explore the Inception network to learn query embedding, which aims to further increase the interactions between head and relation embeddings. Then, we propose to use a relation-aware attention mechanism to enrich the query embedding with the local neighborhood and global entity information. Experimental results on both WN18RR and FB15k-237 datasets demonstrate that ReInceptionE achieves competitive performance compared with state-of-the-art methods.

Cross-lingual entity alignment, which aims to match equivalent entities in KGs with different languages, has attracted considerable focus in recent years.Recently, many graph neural network (GNN) based methods are proposed for entity alignment and obtain promising results.However, existing GNN-based methods consider the two KGs independently and learn embeddings for different KGs separately, which ignore the useful pre-aligned links between two KGs.In this paper, we propose a novel Contextual Alignment Enhanced Cross Graph Attention Network (CAECGAT) for the task of cross-lingual entity alignment, which is able to jointly learn the embeddings in different KGs by propagating cross-KG information through pre-aligned seed alignments.We conduct extensive experiments on three benchmark cross-lingual entity alignment datasets.The experimental results demonstrate that our proposed method obtains remarkable performance gains compared to state-of-the-art methods.

An important aspect of developing dialogue agents involves endowing a conversation system with emotion perception and interaction. Most existing emotion dialogue models lack the adaptability and extensibility of different scenes because of their limitation to require a specified emotion category or their reliance on a fixed emotional dictionary. To overcome these limitations, we propose a neural conversation generation with auxiliary emotional supervised model (nCG-ESM) comprising a sequence-to-sequence (Seq2Seq) generation model and an emotional classifier used as an auxiliary model. The emotional classifier was trained to predict the emotion distributions of the dialogues, which were then used as emotion supervised signals to guide the generation model to generate diverse emotional responses. The proposed nCG-ESM is flexible enough to generate responses with emotional diversity, including specified or unspecified emotions, which can be adapted and extended to different scenarios. We conducted extensive experiments on the popular dataset of Weibo post--response pairs. Experimental results showed that the proposed model was capable of producing more diverse, appropriate, and emotionally rich responses, yielding substantial gains in diversity scores and human evaluations.

Software developers can search, share and learn development experience, solutions, bug fixes and open source projects in software information sites such as StackOverflow and Freecode. Many software information sites rely on tags to classify their contents, i.e. software objects, in order to improve the performance and accuracy of various operations on the sites. The quality of tags thus has a significant impact on the usefulness of these sites. High quality tags are expected to be concise and can describe the most important features of the software objects. Unfortunately tagging is inherently an uncoordinated process. The choice of tags made by individual software developers is dependent not only on a developer's understanding of the software object but also on the developer's English skills and preferences. As a result, the number of different tags grows rapidly along with continuous addition of software objects. With thousands of different tags, many of which introduce noise, software objects become poorly classified. Such phenomenon affects negatively the speed and accuracy of developers' queries. In this paper, we propose a tool called TagMulRec to automatically recommend tags and classify software objects in evolving large-scale software information sites. Given a new software object, TagMulRec locates the software objects that are semantically similar to the new one and exploit their tags. We have evaluated TagMulRec on four software information sites, StackOverflow, AskUbuntu, AskDifferent and Freecode. According to our empirical study, TagMulRec is not only accurate but also scalable that can handle a large-scale software information site with millions of software objects and thousands of tags.

Community question answering (cQA) has become an important issue due to the popularity of cQA archives on the Web. This paper focuses on addressing the lexical gap problem in question retrieval. Question retrieval in cQA archives aims to find the existing questions that are semantically equivalent or relevant to the queried questions. However, the lexical gap problem brings a new challenge for question retrieval in cQA. In this paper, we propose to model and learn distributed word representations with metadata of category information within cQA pages for question retrieval using two novel category powered models. One is a basic category powered model called MB-NET and the other one is an enhanced category powered model called ME-NET which can better learn the distributed word representations and alleviate the lexical gap problem. To deal with the variable size of word representation vectors, we employ the framework of fisher kernel to transform them into the fixed-length vectors. Experimental results on large-scale English and Chinese cQA data sets show that our proposed approaches can significantly outperform state-of-the-art retrieval models for question retrieval in cQA. Moreover, we further conduct our approaches on large-scale automatic evaluation experiments. The evaluation results show that promising and significant performance improvements can be achieved.

Sentiment classification aims to automatically predict sentiment polarity (e.g., positive or negative) of user generated sentiment data (e.g., reviews, blogs). In real applications, these user-generated sentiment data can span so many different domains that it is difficult to label the training data for all of them. Therefore, we study the problem of sentiment classification adaptation task in this article. That is, a system is trained to label reviews from one source domain but is meant to be used on the target domain. One of the biggest challenges for sentiment classification adaptation task is how to deal with the problem when two data distributions between the source domain and target domain are significantly different from one another. However, our observation is that there might exist some domain shared knowledge among certain input dimensions of different domains. In this article, we present a novel method for modeling and mining the domain shared knowledge from different sentiment review domains via a joint non-negative matrix factorization–based framework. In this proposed framework, we attempt to learn the domain shared knowledge and the domain-specific information from different sentiment review domains with several various regularization constraints. The advantage of the proposed method can promote the correspondence under the topic space between the source domain and the target domain, which can significantly reduce the data distribution gap across two domains. We conduct extensive experiments on two real-world balanced data sets from Amazon product reviews for sentence-level and document-level binary sentiment classification. Experimental results show that our proposed approach significantly outperforms several strong baselines and achieves an accuracy that is competitive with the most well-known methods for sentiment classification adaptation.

Sentiment classification aims to automatically predict sentiment polarity (e.g., positive or negative) of user generated sentiment data (e.g., reviews, blogs).Due to the mismatch among different domains, a sentiment classifier trained in one domain may not work well when directly applied to other domains.Thus, domain adaptation for sentiment classification algorithms are highly desirable to reduce the domain discrepancy and manual labeling costs.To address the above challenge, we propose a novel domain adaptation method, called Bi-Transferring Deep Neural Networks (BTDNNs).The proposed BTDNNs attempts to transfer the source domain examples to the target domain, and also transfer the target domain examples to the source domain.The linear transformation of BTDNNs ensures the feasibility of transferring between domains, and the distribution consistency between the transferred domain and the desirable domain is constrained with a linear data reconstruction manner.As a result, the transferred source domain is supervised and follows similar distribution as the target domain.Therefore, any supervised method can be used on the transferred source domain to train a classifier for sentiment classification in a target domain.We conduct experiments on a benchmark composed of reviews of 4 types of Amazon products.Experimental results show that our proposed approach significantly outperforms the several baseline methods, and achieves an accuracy which is competitive with the state-of-the-art method for domain adaptation.

Community question answering (CQA) has become an increasingly popular research topic. In this paper, we focus on the problem of question retrieval. Question retrieval in CQA can automatically find the most relevant and recent questions that have been solved by other users. However, the word ambiguity and word mismatch problems bring about new challenges for question retrieval in CQA. State-of-the-art approaches address these issues by implicitly expanding the queried questions with additional words or phrases using monolingual translation models. While useful, the effectiveness of these models is highly dependent on the availability of quality parallel monolingual corpora (e.g., question-answer pairs) in the absence of which they are troubled by noise issues. In this work, we propose an alternative way to address the word ambiguity and word mismatch problems by taking advantage of potentially rich semantic information drawn from other languages. Our proposed method employs statistical machine translation to improve question retrieval and enriches the question representation with the translated words from other languages via non-negative matrix factorization. Experiments conducted on real CQA data sets show that our proposed approach is promising.

We present Q2P, a system that discovers query templates from search engines via their query autocompletion services. Q2P is distinct from the existing works in that it does not rely on query logs of search engines that are typically not readily available. Q2P is also unique in that it uses a trie to economically store queries sampled from a search engine and employs a beam-search strategy that focuses the expansion of the trie on its most promising nodes. Furthermore, Q2P leverages the trie-based storage of query sample to discover query templates using only two passes over the trie. Q2P is a key part of our ongoing project Deep2Q on a template-driven data integration on the Deep Web, where the templates learned by Q2P are used to guide the integration process in Deep2Q. Experimental results on four major search engines indicate that (1) Q2P sends only a moderate number of queries (ranging from 597 to 1,135) to the engines, while obtaining a significant number of completions per query (ranging from 4.2 to 8.5 on the average); (2) a significant number of templates (ranging from 8 to 32 when the minimum support for frequent templates is set to 1%) may be discovered from the samples.

Community question answering (cQA) has become an important issue due to the popularity of cQA archives on the web. This paper is concerned with the problem of question retrieval. Question retrieval in cQA archives aims to find the existing questions that are semantically equivalent or relevant to the queried questions. However, the lexical gap problem brings about new challenge for question retrieval in cQA. In this paper, we propose to learn continuous word embeddings with metadata of category information within cQA pages for question retrieval. To deal with the variable size of word embedding vectors, we employ the framework of fisher kernel to aggregated them into the fixedlength vectors. Experimental results on large-scale real world cQA data set show that our approach can significantly outperform state-of-the-art translation models and topic-based models for question re-

Community question answering (cQA) has become an important service due to the popularity of cQA archives on the web. This paper is concerned with the problem of question search. Question search in cQA aims to find the historical questions that are semantically equivalent or similar to the queried questions. In this paper, we propose a faster and better retrieval model for question search by leveraging user chosen category. After introducing the question category, we can filter certain amount of irrelevant historical questions under a wide range of leaf categories. Experimental results conducted on real cQA data set demonstrate that the proposed techniques are more effective and efficient than a variety of baseline methods.

This work investigates selecting concise labels for the newly-arising topics in community question answer. Previous methods of generating labels do not take the information of the existing category hierarchy into consideration. The main motivation of our paper is to utilize this information into the label generation process. We propose a general framework to address this problem. Firstly, we map the questions into Wikipedia concept sets, which are more meaningful than terms. Secondly, important concepts are identified to represent the main focus of the newly-arising topics. Thirdly, candidate labels are extracted from Wikipedia category graph. Finally, candidate labels are filtered and reranked by combination of structure information of existing category hierarchy and Wikipedia category graph. The experiments show that in our test collections, about 80% "correct" labels appear in the top ten labels recommended by our system.

In this paper, we address the problem of expert finding in community question answering (CQA). Most of the existing approaches attempt to find experts in CQA by means of link analysis techniques. However, these traditional techniques only consider the link structure while ignore the topical similarity among users (askers and answerers) and user expertise and user reputation. In this study, we propose a topic-sensitive probabilistic model, which is an extension of PageRank algorithm to find experts in CQA. Compared to the traditional link analysis techniques, our proposed method is more effective because it finds the experts by taking into account both the link structure and the topical similarity among users. We conduct experiments on real world data set from Yahoo! Answers. Experimental results show that our proposed method significantly outperforms the traditional link analysis techniques and achieves the state-of-the-art performance for expert finding in CQA.

Community question answering (cQA) has become a popular service for users to ask and answer questions. In recent years, the efficiency of cQA service is hindered by a sharp increase of questions in the community. This paper is concerned with the problem of question routing. Question routing in cQA aims to route new questions to the eligible answerers who can give high quality answers. However, the traditional methods suffer from the following two problems: (1) word mismatch between the new questions and the users' answering history; (2) high variance in perceived answer quality. To solve the above two problems, this paper proposes a novel joint learning method by taking both word mismatch and answer quality into a unified framework for question routing. We conduct experiments on large-scale real world data set from Yahoo! Answers. Experimental results show that our proposed method significantly outperforms the traditional query likelihood language model (QLLM) as well as state-of-the-art cluster-based language model (CBLM) and category-sensitive query likelihood language model (TCSLM).

With the flourishing of community-based question answering (cQA) services like Yahoo! Answers, more and more web users seek their information need from these sites. Understanding user's information need expressed through their search questions is crucial to information providers. Question classification in cQA is studied for this purpose. However, there are two main difficulties in applying traditional methods (question classification in TREC QA and text classification) to cQA: (1) Traditional methods confine themselves to classify a text or question into two or a few predefined categories. While in cQA, the number of categories is much larger, such as Yahoo! Answers, there contains 1,263 categories. Our empirical results show that with the increasing of the number of categories to moderate size, the performance of the classification accuracy dramatically decreases. (2) Unlike the normal texts, questions in cQA are very short, which cannot provide sufficient word co-occurrence or shared information for a good similarity measure due to the data sparseness. In this paper, we propose a two-stage approach for question classification in cQA that can tackle the difficulties of the traditional methods. In the first stage, we preform a search process to prune the large-scale categories to focus our classification effort on a small subset. In the second stage, we enrich questions by leveraging Wikipedia semantic knowledge to tackle the data sparseness. As a result, the classification model is trained on the enriched small subset. We demonstrate the performance of our proposed method on Yahoo! Answers with 1,263 categories. The experimental results show that our proposed method significantly outperforms the baseline method (with error reductions of 23.21%).