acl acl2012 acl2012-55 knowledge-graph by maker-knowledge-mining

55 acl-2012-Community Answer Summarization for Multi-Sentence Question with Group L1 Regularization


Source: pdf

Author: Wen Chan ; Xiangdong Zhou ; Wei Wang ; Tat-Seng Chua

Abstract: We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i.e., the “best answer” of a complex multi-sentence question misses valuable information that is contained in other answers. In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. Various textual and non-textual QA features are explored. Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. Experimental results on a Yahoo! Answers dataset show that our proposed method significantly outperforms state-of-the-art methods on cQA summarization task.

Reference: text


Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 sg , , Abstract We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i. [sent-7, score-0.79]

2 In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. [sent-10, score-2.037]

3 Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. [sent-12, score-0.533]

4 To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. [sent-13, score-0.296]

5 1 Introduction Community Question and Answering services (cQAs) have become valuable resources for users to pose questions of their interests and share their knowledge by providing answers to questions. [sent-16, score-0.522]

6 Answers, a resolved question often gets more than one answers and a “best answer” will be chosen by the asker or voted by other community participants. [sent-21, score-0.67]

7 d I itn performs very well in simple factoid QA settings, where the answers to factoid questions often relate to a single named entity like a person, time or location. [sent-24, score-0.613]

8 That is, such question often comprises several sub questions in specific contexts and the asker wishes to get elaborated answers for as many aspects of the question as possible. [sent-26, score-1.189]

9 In which case, the single best answer that covers just one or few aspects may not be a good choice (Liu et al. [sent-27, score-0.606]

10 , 2008), the use of a single best answer often misses valuable human generated information contained in other answers. [sent-31, score-0.637]

11 (2008) reported that no more than 48% of the 400 best answers were indeed the unique best answers in 4 most popular Yahoo! [sent-33, score-0.63]

12 Gums that bleed could be a sign of a more serious issue like leukemia, an infection, gum disease, a blood disorder, or a vitamin deficiency. [sent-58, score-0.316]

13 your gum Table 1: An example of question with incomplete answer problem from Yahoo! [sent-61, score-1.01]

14 In fact, it is often the case, that a complex multi-sentence question could be answered from multiple aspects by different people focusing on different sub questions. [sent-66, score-0.378]

15 Therefore we address the incomplete answer problem by developing a novel summarization technique taking different sub questions and contexts into consideration. [sent-67, score-1.178]

16 Specifically we want to learn a concise summary from a set of corresponding answers as supplement or replacement to the “best answer”. [sent-68, score-0.375]

17 Various textual and non-textual question answering features are exploited in the work. [sent-71, score-0.328]

18 Second, we propose a group L1-regularization approach in the CRF model for automatic optimal feature learning to unleash the potential of the features and enhance the performance of answer summarization. [sent-72, score-0.794]

19 First we define a complex multi-sentence question as a question with the following properties: Definition: A complex multi-sentence question is one that contains multiple sub-questions. [sent-82, score-0.684]

20 In the cQAs scenario a question often consists of one or more main question sentences accompany by some context sentences described by askers. [sent-83, score-0.516]

21 We treat the original question and context as a whole single complex multi-sentence question and obtain the sub questions by question segmentation. [sent-84, score-1.039]

22 We study the issues of similarity threshold and the minimal number of stars empirically in the experimental section and show that they are useful in identifying questions with the incomplete answer problem. [sent-86, score-0.995]

23 2 Related Work There exist several attempts to alleviate the answer completeness problem in cQA. [sent-88, score-0.577]

24 One of them is to segment the multi-sentence question into a set of sub-questions along with their contexts, then sequentially retrieve the sub questions one by one, and return similar questions and their best answers (Wang et al. [sent-89, score-1.103]

25 On general problem of cQA answer summarization, Liu et al. [sent-92, score-0.606]

26 (2008) manually classified both questions and answers into different taxonomies and applied clustering algorithms for answer summarization. [sent-93, score-1.068]

27 Through exploiting metadata, Tomasoni and Huang(2010) introduced four characteristics (constraints) of summarized answer and combined them in an additional model as well as a multiplicative model. [sent-95, score-0.643]

28 However there is no previous work that explores the complex multi-sentence question segmentation and its contextual modeling for community answer summarization. [sent-100, score-1.02]

29 Some other works examined the evaluation of the quality of features for answers extracted from cQA services (Jeon et al. [sent-101, score-0.318]

30 (2010), a large number of features extracted for 584 predicting asker-rated quality of answers was evaluated by using a logistic regression model. [sent-105, score-0.318]

31 However, to the best of our knowledge, there is no work in evaluating the quality of features for community answer summarization. [sent-106, score-0.702]

32 In our work we model the feature learning and evaluation problem as a group L1 regularization problem (Schmidt , 2010) on different feature groups. [sent-107, score-0.301]

33 1 Conditional Random Fields We utilize the probabilistic graphical model to solve the answer summarization task, Figure 1gives some illustrations, in which the sites correspond to the sentences and the edges are utilized to model the interactions between sentences. [sent-109, score-0.886]

34 Specifically, let x be the sentence sequence to all answers within a question thread, and y be the corresponding label se- µ quence. [sent-110, score-0.569]

35 2 cQA Features and Contextual Modeling In this section, we give a detailed description of the different sentence-level cQA features and the contextual modeling between sentences used in our model for answer summarization. [sent-115, score-0.752]

36 Sentence-level Features Different from the conventional multi-document summarization in which only the textual features are utilized, we also explore a number of non-textual author related features (Shah et al. [sent-116, score-0.294]

37 Sentence Length: The length of the sentence in the answers with the stop words removed. [sent-119, score-0.374]

38 If a sentence is at the beginning or at the end of one answer, it might be a generation or viewpoint sentence and will be given higher weight in the summarization task. [sent-123, score-0.298]

39 Answer Length: The length of the answer to which the sentence belonged, again with the stop words removed. [sent-125, score-0.665]

40 Similarity to Question: Semantic similarity to the question and question context. [sent-136, score-0.506]

41 It imports the semantic information relevance to the question and question context. [sent-137, score-0.456]

42 Best Answer Star: The stars of the best answer received by the askers or voters. [sent-139, score-0.64]

43 Thumbs Up: The number of thumbs-ups the answer which contains the sentence receives. [sent-141, score-0.632]

44 Users are often used to support one answer by giving a thumbs up after reading some relevant or interesting information for their intentions. [sent-142, score-0.614]

45 Author Level: The level of stars the author who gives the answer sentence acquires. [sent-144, score-0.712]

46 Best Answer Rate: Rate of answers annotated as the best answer the author who gives the answer sentence receives. [sent-147, score-1.57]

47 Total Answer Number: The number of total answers by the author who gives the answer sentence. [sent-149, score-0.909]

48 Total Points: The total points that the author who gives the answer sentence receives. [sent-152, score-0.678]

49 Therefore, to explore the optimal combination of these features, we propose a group L1 regularization term in the general CRF model (Section 3. [sent-158, score-0.291]

50 This give rise to four contextual factors that we will explore for modeling the pairwise semantic interactions based on question segmentation. [sent-167, score-0.487]

51 Question sentence detection: every sentence in the original multi-sentence question is classified into question sentence and non-question (context) sentence. [sent-172, score-0.621]

52 We compute the semantic similarity(Simpson and Crowe, 2005) between sentences or sub ques- Figure 1: Four kinds of the contextual factors are considered for answer summarization in our general CRF based models. [sent-176, score-1.134]

53 One answer sentence may related to more than one sub questions to some extent. [sent-178, score-0.987]

54 Thus, we define the replied question Qri as the sub question with the maximal similarity to sentence xi: Qri = argmaxQj sim(xi, Qj). [sent-179, score-0.803]

55 It is intuitive that different summary sentences aim at answering different sub questions. [sent-180, score-0.302]

56 Therefore, we design the following two contextual factors based on the similarity of replied questions. [sent-181, score-0.341]

57 Dissimilar Replied Question Factor: Given two answer sentences xi , xj and their corresponding replied questions Qri, Qrj. [sent-182, score-1.026]

58 If the similarity2 of Qri and Qrj is below some threshold τlq, it means that xi and xj will present different viewpoints to answer different sub questions. [sent-183, score-0.881]

59 In this case, it is likely that xi and xj are both summary sentences; we ensure this by setting the contextual factor cf1 with a large value of exp ν, where ν is a positive real constant often assigned to value 1; otherwise we set cf1 to exp − ν for penalization. [sent-184, score-0.357]

60 586 swer sentences xi , xj and their corresponding replied questions Qri, Qrj. [sent-186, score-0.503]

61 If the similarity of Qri and Qrj is above some upper threshold τuq, this means that xi and xj are very similar and likely to provide similar viewpoint to answer similar questions. [sent-187, score-0.82]

62 Therefore, we propose the following two kinds of contextual factors for selecting the answer sentences in the CRF model. [sent-191, score-0.806]

63 Notice that this penalty term is indeed a L(1, 2) regularization because in every particular group we normalize the parameters in L2 norm while the weight of a whole group is summed in L1 form. [sent-198, score-0.412]

64 The first two terms of Equation 6 measure the difference between the empirical and the model expected values of feature j in group g, while the third term is the derivative of group L1 priors. [sent-220, score-0.305]

65 Although it is an approximation of the exact inference, we will see that it works well for our answer summarization task in the experiments. [sent-224, score-0.726]

66 1 Dataset To evaluate the performance of our CRF based answer summarization model, we conduct experiments on the Yahoo! [sent-226, score-0.726]

67 Our original dataset contains 1,300,559 questions and 2,770,896 answers in ten taxonomies from Yahoo! [sent-231, score-0.491]

68 After filtering the questions which have less than 5 answers and some trivial factoid questions using the features by (Tomasoni and Huang, 2010) , we reduce the dataset to 55,132 questions. [sent-233, score-0.789]

69 From this sub-set, we next select the questions with incomplete answers as defined in Section 2. [sent-234, score-0.588]

70 Specifically, we select the questions where the average similarity between the best answer and all sub questions is less than 0. [sent-236, score-1.216]

71 6 or when the star rating of the best answer is less than 4. [sent-237, score-0.671]

72 To evaluate the effectiveness of this method, we randomly choose 400 questions in the filtered dataset and invite 10 graduate candidate students (not in NLP research field) to verify whether a question suffers from the incomplete answer problem. [sent-239, score-1.107]

73 We consider the questions as the “incomplete answer questions” only when they are judged by both members in a group to be the case. [sent-241, score-0.9]

74 As a result, we find that 360 (90%) of these questions indeed suffer from the incomplete answer problem, which indicates that our automatic detection method is efficient. [sent-242, score-0.879]

75 This randomly selected 400 questions along with their 2559 answers are then further manually summarized for evaluation of automatically generated answer summaries by our model in experiments. [sent-243, score-1.134]

76 In our experiments, we also compare the precision, recall and F1 score in the ROUGE-1, ROUGE2 and ROUGE-L measures (Lin , 2004) for answer summarization performance. [sent-249, score-0.798]

77 For linear CRF system, we use all our textual and non-textual features as well as the local (exact previous and next) neighborhood contextual factors instead of the features of (Shen et al. [sent-254, score-0.329]

78 Table 2 shows that our general CRF model based on question segmentation with group L1 regularization out-performs the baselines significantly in all three measures (gCRF-QS-l1 is 13. [sent-261, score-0.591]

79 This is mainly because we have divided the question into several sub questions, and the system is able to select more novel sentences than just treating the original multi-sentence as a whole. [sent-273, score-0.408]

80 In addition, when we replace the default L2 regularization by the group L1 regularization for more efficient feature weight learning, we obtain a much better performance in 589 precision while not sacrificing the recall measurement statistically. [sent-274, score-0.47]

81 Therefore, the im- provements in these measures are more encouraging than those of the average classification accuracy for answer summarization. [sent-280, score-0.616]

82 From the viewpoint of ROUGE measures we observe that our question segmentation method can enhance the recall of the summaries significantly due to the more fine-grained modeling of sub questions. [sent-281, score-0.527]

83 We also find that the precision of the group L1 regularization is much better than that of the default L2 regularization while not hurting the recall significantly. [sent-282, score-0.405]

84 In general, the experimental results show that our proposed method is more effective than other baselines in answer summarization for addressing the incomplete answer problem in cQAs. [sent-283, score-1.428]

85 Figure 2: The accumulated weight of each site feature group in the group L1-regularization to our Yahoo! [sent-284, score-0.304]

86 To see how much the different textual and non-textual features contribute to community answer summarization, the accumulated weight of each group of sentence-level features5 is presented in Figure 2. [sent-289, score-0.858]

87 The main reasons that the feature 7 (Similarity to Question) has low contribution is that we have utilized the similarity to question in the contextual factors, and this similarity feature in the single site becomes redundant. [sent-291, score-0.548]

88 3 An Example of Summarized Answer To demonstrate the effectiveness of our proposed method, Table 4 shows the generated summary of the example question which is previously illustrated in Table 1 in the introduction section. [sent-297, score-0.317]

89 The best answer available in the system and the summarized answer generated by our model are compared in Table 4. [sent-298, score-1.249]

90 It is found that the summarized answer contains more valuable information about the original multi- sentence question, as it better answers the reason of blooding and offers some solution for it. [sent-299, score-1.046]

91 Storing and indexing this summarized answer in question archives should provide a better choice for anteeth swer reuse in question retrieval of cQAs. [sent-300, score-1.153]

92 Gums that bleed could be a sign of a more serious issue like leukemia, an infection, gum disease, a blood disorder, or a vitamin deficiency. [sent-310, score-0.316]

93 Table 4: Summarized answer by our general CRF based model for the question in Table 1. [sent-312, score-0.834]

94 6 Conclusions We proposed a general CRF based community swer summarization method an- to deal with the in- complete answer problem for deep understanding of complex multi-sentence questions. [sent-313, score-0.873]

95 Our main contributions are that we proposed a systematic way for modeling semantic contextual interactions between the answer sentences based on question segmentation and we explored both the textual and nontextual answer features learned via a group L1 regularization. [sent-314, score-1.808]

96 We showed that our method is able to achieve significant improvements in performance of answer summarization compared to other baselines and previous methods on Yahoo! [sent-315, score-0.754]

97 We planed to extend our proposed model with more advanced feature learning as well as enriching our summarized answer with more available Web resources. [sent-317, score-0.679]

98 Retrieving answers from frequently asked questions pages on the web. [sent-373, score-0.491]

99 Understanding and summarizing answers in community-based question answering services. [sent-405, score-0.547]

100 Statistical machine translation for query expansion in answer retrieval. [sent-414, score-0.577]


similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('answer', 0.577), ('answers', 0.286), ('question', 0.228), ('questions', 0.205), ('cqa', 0.154), ('sub', 0.15), ('summarization', 0.149), ('yahoo', 0.142), ('teeth', 0.138), ('crf', 0.137), ('group', 0.118), ('contextual', 0.113), ('regularization', 0.111), ('bleed', 0.108), ('gum', 0.108), ('gums', 0.108), ('incomplete', 0.097), ('asker', 0.092), ('replied', 0.092), ('uppercase', 0.092), ('summary', 0.089), ('factors', 0.086), ('qri', 0.08), ('cqas', 0.077), ('lcrf', 0.077), ('xi', 0.069), ('sigir', 0.068), ('disease', 0.067), ('summarized', 0.066), ('star', 0.065), ('community', 0.064), ('eexxpp', 0.061), ('factoid', 0.061), ('gcrf', 0.061), ('shah', 0.061), ('interactions', 0.06), ('sentence', 0.055), ('blood', 0.054), ('swer', 0.054), ('xj', 0.053), ('similarity', 0.05), ('salt', 0.049), ('gingivitis', 0.046), ('lq', 0.046), ('periodontal', 0.046), ('tomasoni', 0.046), ('vitamin', 0.046), ('author', 0.046), ('novelty', 0.046), ('rouge', 0.046), ('yt', 0.045), ('denotes', 0.042), ('rate', 0.042), ('viewpoint', 0.039), ('measures', 0.039), ('segmentation', 0.038), ('morning', 0.037), ('thumbs', 0.037), ('uq', 0.037), ('equation', 0.036), ('feature', 0.036), ('textual', 0.035), ('sites', 0.035), ('utilized', 0.035), ('stars', 0.034), ('lr', 0.034), ('factor', 0.033), ('conditional', 0.033), ('answering', 0.033), ('recall', 0.033), ('stop', 0.033), ('term', 0.033), ('accumulated', 0.032), ('norm', 0.032), ('features', 0.032), ('precision', 0.032), ('threshold', 0.032), ('valuable', 0.031), ('ls', 0.031), ('blooding', 0.031), ('brush', 0.031), ('caked', 0.031), ('crowe', 0.031), ('dentist', 0.031), ('disorder', 0.031), ('gyongyi', 0.031), ('jeon', 0.031), ('mouth', 0.031), ('qrj', 0.031), ('simpson', 0.031), ('takechi', 0.031), ('unleash', 0.031), ('warm', 0.031), ('local', 0.031), ('sentences', 0.03), ('measurement', 0.029), ('thread', 0.029), ('general', 0.029), ('best', 0.029), ('baselines', 0.028)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.99999917 55 acl-2012-Community Answer Summarization for Multi-Sentence Question with Group L1 Regularization

Author: Wen Chan ; Xiangdong Zhou ; Wei Wang ; Tat-Seng Chua

Abstract: We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i.e., the “best answer” of a complex multi-sentence question misses valuable information that is contained in other answers. In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. Various textual and non-textual QA features are explored. Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. Experimental results on a Yahoo! Answers dataset show that our proposed method significantly outperforms state-of-the-art methods on cQA summarization task.

2 0.31717315 177 acl-2012-Sentence Dependency Tagging in Online Question Answering Forums

Author: Zhonghua Qu ; Yang Liu

Abstract: Online forums are becoming a popular resource in the state of the art question answering (QA) systems. Because of its nature as an online community, it contains more updated knowledge than other places. However, going through tedious and redundant posts to look for answers could be very time consuming. Most prior work focused on extracting only question answering sentences from user conversations. In this paper, we introduce the task of sentence dependency tagging. Finding dependency structure can not only help find answer quickly but also allow users to trace back how the answer is concluded through user conversations. We use linear-chain conditional random fields (CRF) for sentence type tagging, and a 2D CRF to label the dependency relation between sentences. Our experimental results show that our proposed approach performs well for sentence dependency tagging. This dependency information can benefit other tasks such as thread ranking and answer summarization in online forums.

3 0.17006302 144 acl-2012-Modeling Review Comments

Author: Arjun Mukherjee ; Bing Liu

Abstract: Writing comments about news articles, blogs, or reviews have become a popular activity in social media. In this paper, we analyze reader comments about reviews. Analyzing review comments is important because reviews only tell the experiences and evaluations of reviewers about the reviewed products or services. Comments, on the other hand, are readers’ evaluations of reviews, their questions and concerns. Clearly, the information in comments is valuable for both future readers and brands. This paper proposes two latent variable models to simultaneously model and extract these key pieces of information. The results also enable classification of comments accurately. Experiments using Amazon review comments demonstrate the effectiveness of the proposed models.

4 0.13369702 56 acl-2012-Computational Approaches to Sentence Completion

Author: Geoffrey Zweig ; John C. Platt ; Christopher Meek ; Christopher J.C. Burges ; Ainur Yessenalina ; Qiang Liu

Abstract: This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. We find that by fusing local and global information, we can exceed 50% on this task (chance baseline is 20%), and we suggest some avenues for further research.

5 0.10709044 35 acl-2012-Automatically Mining Question Reformulation Patterns from Search Log Data

Author: Xiaobing Xue ; Yu Tao ; Daxin Jiang ; Hang Li

Abstract: Natural language questions have become popular in web search. However, various questions can be formulated to convey the same information need, which poses a great challenge to search systems. In this paper, we automatically mined 5w1h question reformulation patterns from large scale search log data. The question reformulations generated from these patterns are further incorporated into the retrieval model. Experiments show that using question reformulation patterns can significantly improve the search performance of natural language questions.

6 0.078927495 101 acl-2012-Fully Abstractive Approach to Guided Summarization

7 0.075998798 134 acl-2012-Learning to Find Translations and Transliterations on the Web

8 0.075671144 52 acl-2012-Combining Coherence Models and Machine Translation Evaluation Metrics for Summarization Evaluation

9 0.068576403 65 acl-2012-Crowdsourcing Inference-Rule Evaluation

10 0.065799095 33 acl-2012-Automatic Event Extraction with Structured Preference Modeling

11 0.06007478 94 acl-2012-Fast Online Training with Frequency-Adaptive Learning Rates for Chinese Word Segmentation and New Word Detection

12 0.055124704 123 acl-2012-Joint Feature Selection in Distributed Stochastic Learning for Large-Scale Discriminative Training in SMT

13 0.050822493 182 acl-2012-Spice it up? Mining Refinements to Online Instructions from User Generated Content

14 0.049311485 3 acl-2012-A Class-Based Agreement Model for Generating Accurately Inflected Translations

15 0.048725754 100 acl-2012-Fine Granular Aspect Analysis using Latent Structural Models

16 0.047949482 117 acl-2012-Improving Word Representations via Global Context and Multiple Word Prototypes

17 0.047591235 40 acl-2012-Big Data versus the Crowd: Looking for Relationships in All the Right Places

18 0.047260452 9 acl-2012-A Cost Sensitive Part-of-Speech Tagging: Differentiating Serious Errors from Minor Errors

19 0.046760492 141 acl-2012-Maximum Expected BLEU Training of Phrase and Lexicon Translation Models

20 0.045724094 43 acl-2012-Building Trainable Taggers in a Web-based, UIMA-Supported NLP Workbench


similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, -0.167), (1, 0.066), (2, 0.001), (3, 0.011), (4, 0.013), (5, 0.096), (6, -0.0), (7, -0.004), (8, -0.047), (9, 0.026), (10, -0.054), (11, 0.059), (12, -0.067), (13, 0.035), (14, -0.014), (15, 0.091), (16, 0.16), (17, -0.075), (18, 0.081), (19, -0.083), (20, 0.27), (21, -0.005), (22, 0.141), (23, -0.041), (24, -0.117), (25, -0.148), (26, -0.267), (27, -0.184), (28, -0.041), (29, 0.071), (30, -0.085), (31, -0.159), (32, 0.073), (33, 0.056), (34, 0.045), (35, 0.079), (36, 0.171), (37, -0.001), (38, 0.135), (39, -0.155), (40, -0.074), (41, -0.143), (42, -0.052), (43, 0.119), (44, -0.044), (45, -0.169), (46, 0.077), (47, 0.001), (48, -0.002), (49, 0.025)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.98153657 55 acl-2012-Community Answer Summarization for Multi-Sentence Question with Group L1 Regularization

Author: Wen Chan ; Xiangdong Zhou ; Wei Wang ; Tat-Seng Chua

Abstract: We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i.e., the “best answer” of a complex multi-sentence question misses valuable information that is contained in other answers. In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. Various textual and non-textual QA features are explored. Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. Experimental results on a Yahoo! Answers dataset show that our proposed method significantly outperforms state-of-the-art methods on cQA summarization task.

2 0.81135392 177 acl-2012-Sentence Dependency Tagging in Online Question Answering Forums

Author: Zhonghua Qu ; Yang Liu

Abstract: Online forums are becoming a popular resource in the state of the art question answering (QA) systems. Because of its nature as an online community, it contains more updated knowledge than other places. However, going through tedious and redundant posts to look for answers could be very time consuming. Most prior work focused on extracting only question answering sentences from user conversations. In this paper, we introduce the task of sentence dependency tagging. Finding dependency structure can not only help find answer quickly but also allow users to trace back how the answer is concluded through user conversations. We use linear-chain conditional random fields (CRF) for sentence type tagging, and a 2D CRF to label the dependency relation between sentences. Our experimental results show that our proposed approach performs well for sentence dependency tagging. This dependency information can benefit other tasks such as thread ranking and answer summarization in online forums.

3 0.58710164 35 acl-2012-Automatically Mining Question Reformulation Patterns from Search Log Data

Author: Xiaobing Xue ; Yu Tao ; Daxin Jiang ; Hang Li

Abstract: Natural language questions have become popular in web search. However, various questions can be formulated to convey the same information need, which poses a great challenge to search systems. In this paper, we automatically mined 5w1h question reformulation patterns from large scale search log data. The question reformulations generated from these patterns are further incorporated into the retrieval model. Experiments show that using question reformulation patterns can significantly improve the search performance of natural language questions.

4 0.55143857 56 acl-2012-Computational Approaches to Sentence Completion

Author: Geoffrey Zweig ; John C. Platt ; Christopher Meek ; Christopher J.C. Burges ; Ainur Yessenalina ; Qiang Liu

Abstract: This paper studies the problem of sentencelevel semantic coherence by answering SATstyle sentence completion questions. These questions test the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena. We tackle the problem with two approaches: methods that use local lexical information, such as the n-grams of a classical language model; and methods that evaluate global coherence, such as latent semantic analysis. We evaluate these methods on a suite of practice SAT questions, and on a recently released sentence completion task based on data taken from five Conan Doyle novels. We find that by fusing local and global information, we can exceed 50% on this task (chance baseline is 20%), and we suggest some avenues for further research.

5 0.53922403 144 acl-2012-Modeling Review Comments

Author: Arjun Mukherjee ; Bing Liu

Abstract: Writing comments about news articles, blogs, or reviews have become a popular activity in social media. In this paper, we analyze reader comments about reviews. Analyzing review comments is important because reviews only tell the experiences and evaluations of reviewers about the reviewed products or services. Comments, on the other hand, are readers’ evaluations of reviews, their questions and concerns. Clearly, the information in comments is valuable for both future readers and brands. This paper proposes two latent variable models to simultaneously model and extract these key pieces of information. The results also enable classification of comments accurately. Experiments using Amazon review comments demonstrate the effectiveness of the proposed models.

6 0.33590558 101 acl-2012-Fully Abstractive Approach to Guided Summarization

7 0.3072387 182 acl-2012-Spice it up? Mining Refinements to Online Instructions from User Generated Content

8 0.30047274 94 acl-2012-Fast Online Training with Frequency-Adaptive Learning Rates for Chinese Word Segmentation and New Word Detection

9 0.2929832 43 acl-2012-Building Trainable Taggers in a Web-based, UIMA-Supported NLP Workbench

10 0.28404757 178 acl-2012-Sentence Simplification by Monolingual Machine Translation

11 0.27114582 112 acl-2012-Humor as Circuits in Semantic Networks

12 0.2679919 23 acl-2012-A Two-step Approach to Sentence Compression of Spoken Utterances

13 0.26634777 145 acl-2012-Modeling Sentences in the Latent Space

14 0.25858727 52 acl-2012-Combining Coherence Models and Machine Translation Evaluation Metrics for Summarization Evaluation

15 0.25689331 117 acl-2012-Improving Word Representations via Global Context and Multiple Word Prototypes

16 0.25231418 65 acl-2012-Crowdsourcing Inference-Rule Evaluation

17 0.2509037 131 acl-2012-Learning Translation Consensus with Structured Label Propagation

18 0.2501604 100 acl-2012-Fine Granular Aspect Analysis using Latent Structural Models

19 0.24521035 33 acl-2012-Automatic Event Extraction with Structured Preference Modeling

20 0.23175231 57 acl-2012-Concept-to-text Generation via Discriminative Reranking


similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(0, 0.248), (25, 0.015), (26, 0.033), (28, 0.053), (30, 0.023), (37, 0.023), (39, 0.046), (74, 0.037), (82, 0.026), (84, 0.019), (85, 0.029), (90, 0.224), (92, 0.053), (94, 0.03), (99, 0.055)]

similar papers list:

simIndex simValue paperId paperTitle

1 0.88993293 151 acl-2012-Multilingual Subjectivity and Sentiment Analysis

Author: Rada Mihalcea ; Carmen Banea ; Janyce Wiebe

Abstract: Subjectivity and sentiment analysis focuses on the automatic identification of private states, such as opinions, emotions, sentiments, evaluations, beliefs, and speculations in natural language. While subjectivity classification labels text as either subjective or objective, sentiment classification adds an additional level of granularity, by further classifying subjective text as either positive, negative or neutral. While much of the research work in this area has been applied to English, research on other languages is growing, including Japanese, Chinese, German, Spanish, Romanian. While most of the researchers in the field are familiar with the methods applied on English, few of them have closely looked at the original research carried out in other languages. For example, in languages such as Chinese, researchers have been looking at the ability of characters to carry sentiment information (Ku et al., 2005; Xiang, 2011). In Romanian, due to markers of politeness and additional verbal modes embedded in the language, experiments have hinted that subjectivity detection may be easier to achieve (Banea et al., 2008). These additional sources ofinformation may not be available across all languages, yet, various articles have pointed out that by investigating a synergistic approach for detecting subjectivity and sentiment in multiple languages at the same time, improvements can be achieved not only in other languages, but in English as well. The development and interest in these methods is also highly motivated by the fact that only 27% of Internet users speak English (www.internetworldstats.com/stats.htm, 4 . unt . edu wiebe @ c s . pitt . edu Oct 11, 2011), and that number diminishes further every year, as more people across the globe gain Internet access. The aim of this tutorial is to familiarize the attendees with the subjectivity and sentiment research carried out on languages other than English in order to enable and promote crossfertilization. Specifically, we will review work along three main directions. First, we will present methods where the resources and tools have been specifically developed for a given target language. In this category, we will also briefly overview the main methods that have been proposed for English, but which can be easily ported to other languages. Second, we will describe cross-lingual approaches, including several methods that have been proposed to leverage on the resources and tools available in English by using cross-lingual projections. Finally, third, we will show how the expression of opinions and polarity pervades language boundaries, and thus methods that holistically explore multiple languages at the same time can be effectively considered. References C. Banea, R. Mihalcea, and J. Wiebe. 2008. A Bootstrapping method for building subjectivity lexicons for languages with scarce resources. In Proceedings of LREC 2008, Marrakech, Morocco. L. W. Ku, T. H. Wu, L. Y. Lee, and H. H. Chen. 2005. Construction of an Evaluation Corpus for Opinion Extraction. In Proceedings of NTCIR-5, Tokyo, Japan. L. Xiang. 2011. Ideogram Based Chinese Sentiment Word Orientation Computation. Computing Research Repository, page 4, October. Jeju, Republic of Korea,T 8ut Jourliya 2l0 A1b2s.tr ?ac c2t0s1 o2f A ACssLo 2c0ia1t2io,n p faogre C 4o,mputational Linguistics

same-paper 2 0.83074409 55 acl-2012-Community Answer Summarization for Multi-Sentence Question with Group L1 Regularization

Author: Wen Chan ; Xiangdong Zhou ; Wei Wang ; Tat-Seng Chua

Abstract: We present a novel answer summarization method for community Question Answering services (cQAs) to address the problem of “incomplete answer”, i.e., the “best answer” of a complex multi-sentence question misses valuable information that is contained in other answers. In order to automatically generate a novel and non-redundant community answer summary, we segment the complex original multi-sentence question into several sub questions and then propose a general Conditional Random Field (CRF) based answer summary method with group L1 regularization. Various textual and non-textual QA features are explored. Specifically, we explore four different types of contextual factors, namely, the information novelty and non-redundancy modeling for local and non-local sentence interactions under question segmentation. To further unleash the potential of the abundant cQA features, we introduce the group L1 regularization for feature learning. Experimental results on a Yahoo! Answers dataset show that our proposed method significantly outperforms state-of-the-art methods on cQA summarization task.

3 0.71171492 3 acl-2012-A Class-Based Agreement Model for Generating Accurately Inflected Translations

Author: Spence Green ; John DeNero

Abstract: When automatically translating from a weakly inflected source language like English to a target language with richer grammatical features such as gender and dual number, the output commonly contains morpho-syntactic agreement errors. To address this issue, we present a target-side, class-based agreement model. Agreement is promoted by scoring a sequence of fine-grained morpho-syntactic classes that are predicted during decoding for each translation hypothesis. For English-to-Arabic translation, our model yields a +1.04 BLEU average improvement over a state-of-the-art baseline. The model does not require bitext or phrase table annotations and can be easily implemented as a feature in many phrase-based decoders. 1

4 0.71119177 150 acl-2012-Multilingual Named Entity Recognition using Parallel Data and Metadata from Wikipedia

Author: Sungchul Kim ; Kristina Toutanova ; Hwanjo Yu

Abstract: In this paper we propose a method to automatically label multi-lingual data with named entity tags. We build on prior work utilizing Wikipedia metadata and show how to effectively combine the weak annotations stemming from Wikipedia metadata with information obtained through English-foreign language parallel Wikipedia sentences. The combination is achieved using a novel semi-CRF model for foreign sentence tagging in the context of a parallel English sentence. The model outperforms both standard annotation projection methods and methods based solely on Wikipedia metadata.

5 0.7109887 45 acl-2012-Capturing Paradigmatic and Syntagmatic Lexical Relations: Towards Accurate Chinese Part-of-Speech Tagging

Author: Weiwei Sun ; Hans Uszkoreit

Abstract: From the perspective of structural linguistics, we explore paradigmatic and syntagmatic lexical relations for Chinese POS tagging, an important and challenging task for Chinese language processing. Paradigmatic lexical relations are explicitly captured by word clustering on large-scale unlabeled data and are used to design new features to enhance a discriminative tagger. Syntagmatic lexical relations are implicitly captured by constituent parsing and are utilized via system combination. Experiments on the Penn Chinese Treebank demonstrate the importance of both paradigmatic and syntagmatic relations. Our linguistically motivated approaches yield a relative error reduction of 18% in total over a stateof-the-art baseline.

6 0.7099359 172 acl-2012-Selective Sharing for Multilingual Dependency Parsing

7 0.70912051 131 acl-2012-Learning Translation Consensus with Structured Label Propagation

8 0.70911497 9 acl-2012-A Cost Sensitive Part-of-Speech Tagging: Differentiating Serious Errors from Minor Errors

9 0.70793003 127 acl-2012-Large-Scale Syntactic Language Modeling with Treelets

10 0.70741892 123 acl-2012-Joint Feature Selection in Distributed Stochastic Learning for Large-Scale Discriminative Training in SMT

11 0.70645195 140 acl-2012-Machine Translation without Words through Substring Alignment

12 0.70317298 217 acl-2012-Word Sense Disambiguation Improves Information Retrieval

13 0.70217383 119 acl-2012-Incremental Joint Approach to Word Segmentation, POS Tagging, and Dependency Parsing in Chinese

14 0.70217079 116 acl-2012-Improve SMT Quality with Automatically Extracted Paraphrase Rules

15 0.70206195 168 acl-2012-Reducing Approximation and Estimation Errors for Chinese Lexical Processing with Heterogeneous Annotations

16 0.70163465 156 acl-2012-Online Plagiarized Detection Through Exploiting Lexical, Syntax, and Semantic Information

17 0.70080322 137 acl-2012-Lemmatisation as a Tagging Task

18 0.70068139 203 acl-2012-Translation Model Adaptation for Statistical Machine Translation with Monolingual Topic Information

19 0.7006036 213 acl-2012-Utilizing Dependency Language Models for Graph-based Dependency Parsing Models

20 0.70005697 25 acl-2012-An Exploration of Forest-to-String Translation: Does Translation Help or Hurt Parsing?