emnlp emnlp2010 emnlp2010-17 knowledge-graph by maker-knowledge-mining

17 emnlp-2010-An Efficient Algorithm for Unsupervised Word Segmentation with Branching Entropy and MDL

Source: pdf

Author: Valentin Zhikov ; Hiroya Takamura ; Manabu Okumura

Abstract: This paper proposes a fast and simple unsupervised word segmentation algorithm that utilizes the local predictability of adjacent character sequences, while searching for a leasteffort representation of the data. The model uses branching entropy as a means of constraining the hypothesis space, in order to efficiently obtain a solution that minimizes the length of a two-part MDL code. An evaluation with corpora in Japanese, Thai, English, and the ”CHILDES” corpus for research in language development reveals that the algorithm achieves an accuracy, comparable to that of the state-of-the-art methods in unsupervised word segmentation, in a significantly reduced . computational time.

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 The model uses branching entropy as a means of constraining the hypothesis space, in order to efficiently obtain a solution that minimizes the length of a two-part MDL code. [sent-9, score-0.483]

2 An evaluation with corpora in Japanese, Thai, English, and the ”CHILDES” corpus for research in language development reveals that the algorithm achieves an accuracy, comparable to that of the state-of-the-art methods in unsupervised word segmentation, in a significantly reduced . [sent-10, score-0.098]

3 1 Introduction As an inherent preprocessing step to nearly all NLP tasks for writing systems without orthographical marking of word boundaries, such as Japanese and Chinese, the importance of word segmentation has lead to the emergence of a micro-genre in NLP focused exclusively on this problem. [sent-12, score-0.18]

4 However, the development of the annotated training corpora necessary for their 832 t itech . [sent-15, score-0.117]

5 Recent advances in unsupervised word segmentation have been promoted by human cognition research, where it is involved in the modeling of the mechanisms that underlie language acquisition. [sent-22, score-0.286]

6 The minimum description length framework (MDL) (Rissanen, 1978) is an appealing means of formalizing such models, as it provides a robust foundation for learning and inference, based solely on compression. [sent-34, score-0.228]

7 The major problem in MDL-based word segmentation is the lack of standardized search algorithms for the exponential hypothesis space (Goldwater, 2006). [sent-35, score-0.271]

8 Yu (2000) proposed an EM optimization routine, which achieved a high accuracy, in spite of a lower compression than the gold standard segmentation. [sent-38, score-0.224]

9 As a solution to the aforementioned issue, the proposed method incorporates the local predictability of character sequences into the inference process. [sent-39, score-0.254]

10 The increase in uncertainty following a given word prefix is a well studied criterion for morpheme boundary prediction (Harris, 1955). [sent-43, score-0.123]

11 A good deal of research has been conducted on methods through which such local statistics can be applied to the word induction problem (e. [sent-44, score-0.11]

12 833 Most methods for unsupervised word segmentation based solely on local statistics presume a certain albeit minimum level of acquaintance with the target language. [sent-48, score-0.324]

13 The state-of-the-art in unsupervised word segmentation is represented by Bayesian models. [sent-51, score-0.232]

14 (2009) proposed extensions to this method, which included a nested character model and an optimized inference procedure. [sent-55, score-0.153]

15 Johnson and Goldwater (2009) have proposed a novel method based on adaptor grammars, whose accuracy surpasses the aforementioned methods by a large margin, when appropriate assumptions are – – made regarding the structural units of a language. [sent-56, score-0.117]

16 1 Word segmentation with MDL The proposed two-part code incorporates some extensions of models presented in related work, aimed at achieving a more precise estimation of the representation length. [sent-58, score-0.287]

17 , a lexicon of unique bwoodride types Mcod = {w1, . [sent-61, score-0.107]

18 The total description length amounts to the number of bits necessary for simultaneous transmission of the codebook and the source text. [sent-65, score-0.271]

19 Therefore, our objective is to minimize the combined description length of both terms: L(D, M) = L(M) + L(D|M). [sent-66, score-0.177]

20 The description length of the data given M is cal- culated using the Shannon-Fano code: L(D|M) = −Xj|M=1|#wjlog2P(wj), where #wj stands for the frequency of the word wj in the text. [sent-67, score-0.325]

21 Different strategies have been proposed in the literature for the calculation of the codebook cost. [sent-68, score-0.135]

22 A common technique in segmentation and morphology induction models is to calculate the product of the total length in characters of the lexicon and an estimate of the per-character entropy. [sent-69, score-0.49]

23 For instance, in Yu (2000) the average entropy per character is measured against the original corpus, but this model does not capture the effects of the word distributions on the observed character probabilities. [sent-72, score-0.34]

24 The de- scription length of the lexicon data D0 given M0 is then calculated as: L(D0|M0) = −X|C|#cilog2P(ci), iX= X1 where #ci denotes the frequency of a character ci in the lexicon of hypothesis M. [sent-75, score-0.626]

25 The term L(M0) is constant for any choice of hypothesis, as is represents the character set of a corpus. [sent-76, score-0.112]

26 The total description length under the proposed model is thus calculated as: L(M) + L(D|M) = L(M0) + L(D0|M0) + L(D|M) = X| XC| X| XM| −X#cilog2P(ci) −X#wjlog2P(wj) + O(1). [sent-77, score-0.265]

28 Depending on the choice of a universal code, the two approaches can overlap, as is the case with the twopart code discussed in this paper. [sent-85, score-0.113]

29 The MDL framework does not provide standard search algorithms for obtaining the hypotheses that minimize the description length. [sent-88, score-0.143]

30 In the rest of this section, we will describe an efficient technique suit- able for the word segmentation task. [sent-89, score-0.18]

31 2 Obtaining an initial hypothesis First, a rough initial hypothesis is built by an algorithm that combines the branching entropy and MDL criteria. [sent-91, score-0.442]

32 Given a set X, comprising all the characters found in a text, tah ese entropy ofbranching aet position ks foofuthned text is defined as: H(Xk |xk−1 , . [sent-92, score-0.167]

33 ,xk−n), xX∈X where xk represents the character found at position k, and n is the order of the Markov model over characters. [sent-101, score-0.454]

34 The above definition is extended to combine the entropy estimates in the left-to-right and right-toleft directions, as this factor has reportedly improved performance figures for models based on branching entropy (Jin and Tanaka-Ishii, 2006). [sent-106, score-0.514]

35 Suffix arrays are employed during the collection of frequency statistics. [sent-108, score-0.119]

36 For a character model of order n over a testing corpus of size t and a training corpus of size m, suffix arrays allow these to be acquired in O(tn log m) time. [sent-109, score-0.238]

37 The chunking technique we adopt is to insert a boundary when the branching entropy measured in sequences of length n exceeds a certain threshold value (H(X|xk−1:k−n) > β). [sent-113, score-0.604]

38 Within the described framework, the increase in context length n promotes precision and recall at first, but causes a performance degradation when the entropy estimates become unreliable due to the reduced frequencies of long strings. [sent-115, score-0.239]

39 Since the F-score curve obtained as decreasing values are assigned to the threshold is typically unimodal as in many applications of MDL, we employ a bisection search routine for the estimation of the threshold (Algorithm 1). [sent-117, score-0.37]

40 All positions of the dataset are sorted by their entropy values. [sent-118, score-0.218]

41 At each iteration, at most two new hypotheses are built, and their description lengths are calculated in time linear to the data size. [sent-119, score-0.141]

42 The computational complexity of the described routine is O(t log t), where t is the corpus length in characters. [sent-120, score-0.192]

43 The order of the Markov chain n used during the entropy calculation is the only input variable of the proposed model. [sent-121, score-0.157]

44 However, the MDL objec- tive also enables unsupervised optimization against 835 Algorithm 1Generates an initial hypothesis. [sent-123, score-0.108]

45 5037 Table 1: Length in bits of the solutions proposed by Algorithm 1with respect to the character n-gram order. [sent-132, score-0.153]

46 The order that minimizes the description length of the data can be discovered in a few iterations of Algorithm 1 with increasing values of n, and it typically matches the optimal value of the parameter (Table 1). [sent-134, score-0.212]

47 Although an acceptable initial segmentation can be built using the described approach, it is possible to obtain higher accuracy with an extended model that takes into account the statistics of Markov chains from several orders during the entropy calculation. [sent-135, score-0.336]

48 3 Refining the initial hypothesis In the second phase of the proposed method, we will refine the initial hypothesis through the reorganization of local co-occurrences which produce redundant description length. [sent-138, score-0.408]

49 We opt for greedy optimization, as our primary interest is to further explore the impact that description length minimization has on accuracy. [sent-139, score-0.177]

50 Since a preliminary segmentation is available, it is convenient to proceed by inserting or removing boundaries in the text, thus splitting or merging the already discovered tokens. [sent-141, score-0.469]

51 The ranked positions involved in the previous step can be reused here, as this is a way to bias the search towards areas of the text where boundaries are more likely to occur. [sent-142, score-0.186]

52 Boundary insertion should start in regions where the branching entropy is high, and removal should first occur in regions where the entropy is close to zero. [sent-143, score-0.474]

53 A drawback of this approach is that it omits locations where the gains are not immediately obvious, as it cannot assess the cumulative gains arising from the merging or splitting of all occurrences of a certain pair (Algorithm 2). [sent-144, score-0.153]

54 It operates directly on the types found in the lexicon produced by Algorithm 2, and is capable of modifying a large number of occurrences of a given pair in a single step. [sent-146, score-0.107]

55 The lexicon types are sorted by their contribution to the total description length of the corpus. [sent-147, score-0.35]

56 For each word type, splitting or merging is attempted at every letter, beginning from the center. [sent-148, score-0.153]

57 The design of the merging routine makes it impossible to produce types longer than the ones already found in the lexicon, as an exhaustive search would be prohibitive. [sent-150, score-0.216]

58 The evaluation of each hypothetical change in the segmentation requires that the description length of the two-part code is recalculated. [sent-151, score-0.459]

59 In order to 836 Algorithm 2 Algorithm 2 Compresses local token co-occurrences. [sent-152, score-0.131]

60 The term L(D|M) can be rewritten as: L(D|M) = −Xj|M=1|#wjlog2#Nwj= −j|X=M1|#wjlog2#wj+ N log2N = T1+ T2, where #wj is the frequency of P|jM=1| wj in the segmented corpus, and N = #wj is the cumulative token count. [sent-157, score-0.238]

61 Their new values are computed for each hypothetical split or merge on the basis of the last values, and the expected description length is calculated as their sum. [sent-159, score-0.26]

62 Known boundaries, such as the sentence boundaries in the CHILDES corpus, are also taken into consideration. [sent-163, score-0.101]

63 We evaluated the performance of our learner in the cases when the few boundaries among the individual sentences are available to it (B), and when it starts from a blank state (N). [sent-167, score-0.137]

64 The BEST corpus for word segmentation and named entity recognition in Thai language combines text from a variety of sources in837 CorpusLanguage(SMizBe)(CKh)ars(TKo)kens(TKy)pes Table 2: Corpora used during the evaluation. [sent-169, score-0.18]

65 We report the obtained precision, recall and F- score values calculated using boundary, token and type counts. [sent-182, score-0.137]

66 Boundary, token and lexicon F-scores, denoted as B-F and T-F and L-F, are calculated as the ModelCorpus & SettingsB-PrecB-RecB-FT-PrecT-RecT-F(DbiLts)(Rbeift. [sent-184, score-0.244]

67 As a rule, boundary-based evaluation produces the highest scores among the three evaluation modes, as it only considers the correspondence between the proposed and the gold standard boundaries at the individual positions of the corpora. [sent-208, score-0.178]

68 Token-based evaluation is more strict it accepts a word as correct only if its beginning and end are identified accurately, and no additional boundaries lie in between. [sent-209, score-0.16]

69 – It provides another useful perspective for the error analysis, which in combination with token scores can give a better idea of the relationship between the accuracy of induction and item frequency. [sent-211, score-0.199]

70 The system was implemented in Java, however it handled the suffix arrays through an external C library called Sary. [sent-212, score-0.126]

71 The heuristic of Jin and Tanaka-Ishii takes advantage of the trend that branching entropy decreases as the observed character sequences become longer; sudden rises can thus be regarded as an indication of locations where a boundary is likely to exist. [sent-215, score-0.547]

72 net 838 entropy change throughout all n-gram orders, and combines the boundaries discovered in both directions in a separate step. [sent-218, score-0.288]

73 These properties of the method would lead to complications if we tried to employ it in the first phase of our method (i. [sent-219, score-0.148]

74 The proposed criterion with an automatically determined threshold value produced slightly worse results than that of Jin and Tanaka-Ishii at the CHILDES corpus. [sent-223, score-0.173]

75 On one hand, the correspondence between description length and F-score is not absolutely perfect, and this may pose an obstacle to the optimization process for relatively small language samples. [sent-226, score-0.233]

76 Another issue lies in the bisection search routine, which suggests approximations of the de- scription length minima. [sent-227, score-0.219]

77 The nmax parameter is set to the value which maximizes the compression during the initial phase, in order to make the results representative of the case in which no annotated development corpora are accessible to the algorithm. [sent-231, score-0.267]

78 It is evident that after the optimization carried out in the second phase, the description length is reduced to levels significantly lower than the ground truth. [sent-232, score-0.307]

79 We conducted experiments involving various initialization strategies: scattering boundaries at ran- dom throughout the text, starting from entirely unsegmented state, or considering each symbol of the text to be a separate token. [sent-236, score-0.228]

80 The results obtained with random initialization confirm the strong relationship between compression and segmentation accuracy, evident in the increase of token F-score between the random initialization and the termination of the algorithm, where description length is lower (Table 6). [sent-237, score-0.83]

81 They also reveal the importance of the branching entropy criterion to the generation of hypotheses that maximize the evaluation scores and compression, as well as the role it plays in the reduction of computational time. [sent-238, score-0.404]

82 839 The greedy algorithms fail to suggest any opti- mizations that improve the compression in the extreme cases when the boundaries/character ratio is either 0 or 1. [sent-246, score-0.127]

83 When no boundaries are given, splitting operations produce unique types with a low frequency that increase the cost of both parts of the MDL code, and are rejected. [sent-247, score-0.326]

84 Finally, we tried randomizing the search path for Algorithm 2 after an entropy-guided initialization, to observe a small deterioration in accuracy in the final segmentation (less than 1% on average). [sent-251, score-0.269]

85 Figure 1a illustrates the effect that training data size has on the accuracy of segmentation for the Kyoto corpus. [sent-252, score-0.22]

86 For the CHILDES corpus, which has a rather limited vocabulary, token F-score above 70% can be achieved for datasets as small as 5000 characters of training data, provided that reasonable values are set for the nmax parameter (we used the values presented in Table 4 throughout these experiments). [sent-254, score-0.271]

87 The initialization phase seems to have the highest contribution to the formation of the final segmentation, and the refinement phase is highly dependent on the output it produces. [sent-256, score-0.387]

88 For Japanese language with the setting for the nmax parameter that maximized compression, we observed an almost 4% increase in the token F-score produced at the end of the first phase with the Asahi corpus as training data. [sent-259, score-0.391]

89 We attributed this Figure 1: a) corpus size / accuracy relationship (Kyoto); b) accuracy levels by phase; c) accuracy levels by phase with various corpora for frequency statistics (Kyoto); d) accuracy levels by phase with different corpora for frequency statistics (BEST). [sent-262, score-0.64]

90 Through the increase of compression in the refinement phase of the algorithm, accuracy is improved by around 3%, and the scores approach those of the explicit probabilistic models of Goldwater et al. [sent-268, score-0.315]

91 The proposed learner surpasses the other unsupervised word induction models in terms of processing speed. [sent-271, score-0.198]

92 Furthermore, different segmentation standards exist for Japanese, and therefore the ”ground truth” provided by the Kyoto corpus cannot be considered an ideal measure of accuracy. [sent-274, score-0.18]

93 New instantiations of the branching entropy and MDL criteria have been proposed and evaluated against corpora in different languages. [sent-276, score-0.445]

94 The MDL-based optimization eliminates the discretion in the choice of the context length and threshold parameters, common in segmentation models based on local statistics. [sent-277, score-0.446]

95 At the same time, the branching entropy criterion enables a constrained search through the hypothesis space, allowing the proposed method to demonstrate a very high ModelCorpusT-PrecT-RecT-FL-PrecL-RecL-FTime EHN P nDY t-P(M 2 3) DLC CH HI L LD DE ES S0 0. [sent-278, score-0.536]

96 Possible improvements of the proposed method include modeling the dependencies among neighboring tokens, which would allow the evaluation of the context to be reflected in the cost function. [sent-296, score-0.125]

97 Mechanisms for stochastic optimization imple- mented in the place of the greedy algorithms could provide an additional flexibility of search for such more complex models. [sent-297, score-0.105]

98 As the proposed approach provides significant performance improvements, it could be utilized in the development of more sophisticated novel word induction schemes, e. [sent-298, score-0.11]

99 Improving nonparameteric Bayesian inference: experiments on unsupervised word segmentation with adaptor grammars. [sent-340, score-0.232]

100 Bayesian unsupervised word segmentation with nested Pitman-Yor language modeling. [sent-375, score-0.232]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('xk', 0.342), ('mdl', 0.306), ('branching', 0.242), ('dl', 0.182), ('segmentation', 0.18), ('phase', 0.148), ('childes', 0.141), ('kyoto', 0.141), ('compression', 0.127), ('mochihashi', 0.118), ('nextthreshold', 0.118), ('entropy', 0.116), ('goldwater', 0.115), ('character', 0.112), ('routine', 0.109), ('jin', 0.107), ('lexicon', 0.107), ('wj', 0.102), ('boundaries', 0.101), ('splitting', 0.095), ('description', 0.094), ('asahi', 0.094), ('codebook', 0.094), ('nmax', 0.094), ('initialization', 0.091), ('token', 0.09), ('threshold', 0.086), ('japanese', 0.085), ('cost', 0.084), ('length', 0.083), ('boundary', 0.077), ('evident', 0.074), ('thresholds', 0.074), ('arrays', 0.073), ('itech', 0.071), ('lefttoken', 0.071), ('lefttype', 0.071), ('longtype', 0.071), ('righttoken', 0.071), ('righttype', 0.071), ('thai', 0.071), ('induction', 0.069), ('tokyo', 0.067), ('sorted', 0.066), ('code', 0.066), ('xp', 0.061), ('predictability', 0.06), ('end', 0.059), ('merging', 0.058), ('optimization', 0.056), ('cognition', 0.054), ('xx', 0.053), ('suffix', 0.053), ('unsupervised', 0.052), ('minimum', 0.051), ('characters', 0.051), ('search', 0.049), ('calculated', 0.047), ('hutchens', 0.047), ('infants', 0.047), ('kempe', 0.047), ('kurohashi', 0.047), ('longtoken', 0.047), ('manber', 0.047), ('powers', 0.047), ('rissanen', 0.047), ('saffran', 0.047), ('scription', 0.047), ('tmims', 0.047), ('twopart', 0.047), ('zhikov', 0.047), ('zipf', 0.047), ('frequency', 0.046), ('criterion', 0.046), ('corpora', 0.046), ('bayesian', 0.045), ('jp', 0.044), ('yu', 0.044), ('hypothesis', 0.042), ('acquisition', 0.042), ('local', 0.041), ('proposed', 0.041), ('bisection', 0.04), ('harris', 0.04), ('kit', 0.04), ('tsuboi', 0.04), ('estimates', 0.04), ('accuracy', 0.04), ('sharon', 0.037), ('blank', 0.036), ('ky', 0.036), ('brent', 0.036), ('hypothetical', 0.036), ('surpasses', 0.036), ('update', 0.036), ('positions', 0.036), ('throughout', 0.036), ('discovered', 0.035), ('ci', 0.035), ('articles', 0.035)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.9999994 17 emnlp-2010-An Efficient Algorithm for Unsupervised Word Segmentation with Branching Entropy and MDL

Author: Valentin Zhikov ; Hiroya Takamura ; Manabu Okumura

2 0.12972561 43 emnlp-2010-Enhancing Domain Portability of Chinese Segmentation Model Using Chi-Square Statistics and Bootstrapping

Author: Baobao Chang ; Dongxu Han

Abstract: Almost all Chinese language processing tasks involve word segmentation of the language input as their first steps, thus robust and reliable segmentation techniques are always required to make sure those tasks wellperformed. In recent years, machine learning and sequence labeling models such as Conditional Random Fields (CRFs) are often used in segmenting Chinese texts. Compared with traditional lexicon-driven models, machine learned models achieve higher F-measure scores. But machine learned models heavily depend on training materials. Although they can effectively process texts from the same domain as the training texts, they perform relatively poorly when texts from new domains are to be processed. In this paper, we propose to use χ2 statistics when training an SVM-HMM based segmentation model to im- prove its ability to recall OOV words and then use bootstrapping strategies to maintain its ability to recall IV words. Experiments show the approach proposed in this paper enhances the domain portability of the Chinese word segmentation model and prevents drastic decline in performance when processing texts across domains.

3 0.11462383 2 emnlp-2010-A Fast Decoder for Joint Word Segmentation and POS-Tagging Using a Single Discriminative Model

Author: Yue Zhang ; Stephen Clark

Abstract: We show that the standard beam-search algorithm can be used as an efficient decoder for the global linear model of Zhang and Clark (2008) for joint word segmentation and POS-tagging, achieving a significant speed improvement. Such decoding is enabled by: (1) separating full word features from partial word features so that feature templates can be instantiated incrementally, according to whether the current character is separated or appended; (2) deciding the POS-tag of a potential word when its first character is processed. Early-update is used with perceptron training so that the linear model gives a high score to a correct partial candidate as well as a full output. Effective scoring of partial structures allows the decoder to give high accuracy with a small beam-size of 16. In our 10-fold crossvalidation experiments with the Chinese Tree- . bank, our system performed over 10 times as fast as Zhang and Clark (2008) with little accuracy loss. The accuracy of our system on the standard CTB 5 test was competitive with the best in the literature. 1 Introduction and Motivation Several approaches have been proposed to solve word segmentation and POS-tagging jointly, including the reranking approach (Shi and Wang, 2007; Jiang et al., 2008b), the hybrid approach (Nakagawa and Uchimoto, 2007; Jiang et al., 2008a), and the single-model approach (Ng and Low, 2004; Zhang and Clark, 2008; Kruengkrai et al., 2009). These methods led to accuracy improvements over the traditional, pipelined segmentation and POS-tagging . . . 843 clark} @ cl cam ac uk baseline by avoiding segmentation error propagation and making use of part-of-speech information to improve segmentation. The single-model approach to joint segmentation and POS-tagging offers consistent training of all in- formation, concerning words, characters and partsof-speech. However, exact inference with dynamic programming can be infeasible if features are defined over a large enough range of the output, such as over a two-word history. In our previous work (Zhang and Clark, 2008), which we refer to as Z&C08; from now on, we used an approximate decoding algorithm that keeps track of a set of partially built structures for each character, which can be seen as a dynamic programming chart which is greatly reduced by pruning. In this paper we follow the line of single-model research, in particular the global linear model of Z&C08.; We show that effective decoding can be achieved with standard beam-search, which gives significant speed improvements compared to the decoding algorithm of Z&C08;, and achieves accuracies that are competitive with the state-of-the-art. Our research is also in line with recent research on improving the speed of NLP systems with little or no accuracy loss (Charniak et al., 2006; Roark and Hollingshead, 2008). Our speed improvement is achieved by the use of a single-beam decoder. Given an input sentence, candidate outputs are built incrementally, one character at a time. When each character is processed, it is combined with existing candidates in all possible ways to generate new candidates, and an agenda is used to keep the N-best candidate outputs from ProceMedITin,g Ms oasfs thaceh 2u0se1t0ts C,o UnSfAer,e n9c-e1 on O Ectmobpeir ic 2a0l1 M0.e ?tc ho2d0s10 in A Nsastoucira tlio Lnan fogru Cagoem Ppruotcaetisosninagl, L pinag eusis 8t4ic3s–852, the begining of the sentence to the current character. Compared to the multiple-beam search algorithm of Z&C08;, the use of a single beam can lead to an order of magnitude faster decoding speed. 1.1 The processing of partial words An important problem that we solve in this paper is the handling of partial words with a single beam decoder for the global model. As we pointed out in Z&C08;, it is very difficult to score partial words properly when they are compared with full words, although such comparison is necessary for incremental decoding with a single-beam. To allow comparisons with full words, partial words can either be treated as full words, or handled differently. We showed in Z&C08; that a naive single-beam decoder which treats partial words in the same way as full words failed to give a competitive accu- racy. An important reason for the low accuracy is over-segmentation during beam-search. Consider the three characters “ 自来水 (tap water)”. The first two characters do not make sense when put together as a single word. Rather, when treated as two singlecharacter words, they can make sense in a sentence such as “请 (please) 自 (self) 来 (come) 取 (take)”. Therefore, when using single-beam search to process “ 自来水 (tap water)”, the two-character word candidate “ 自来” is likely to have been thrown off the agenda before the third character “水” is considered, leading to an unrecoverable segmentation error. This problem is even more severe for a joint segmentor and POS-tagger than for a pure word segmentor, since the POS-tags and POS-tag bigram of “ 自 and “来” further supports them being separated when ”来” is considered. The multiple-beam search decoder we proposed in Z&C08; can be seen as a means to ensure that the three characters “ 自来水” always have a chance to be considered as a single word. It explores candidate segmentations from the beginning of the sentence until each character, and avoids the problem of processing partial words by considering only full words. However, since it ex- ” plores a larger part of the search space than a singlebeam decoder, its time complexity is correspondingly higher. In this paper, we treat partial words differently from full words, so that in the previous example, 844 the decoder can take the first two characters in “ 自来水 (tap water)” as a partial word, and keep it in the beam before the third character is processed. One challenge is the representation of POS-tags for partial words. The POS of a partial word is undefined without the corresponding full word information. Though a partial word can make sense with a particular POS-tag when it is treated as a complete word, this POS-tag is not necessarily the POS of the full word which contains the partial word. Take the three-character sequence “下雨天” as an example. The first character “下” represents a singlecharacter word “below”, for which the POS can be LC or VV. The first two characters “下雨” represent a two-character word “rain”, for which the POS can be VV. Moreover, all three characters when put together make the word “rainy day”, for which the POS is NN. As discussed above, assigning POS tags to partial words as if they were full words leads to low accuracy. An obvious solution to the above problem is not to assign a POS to a partial word until it becomes a full word. However, lack of POS information for partial words makes them less competitive compared to full words in the beam, since the scores of full words are futher supported by POS and POS ngram information. Therefore, not assigning POS to partial words potentially leads to over segmentation. In our experiments, this method did not give comparable accuracies to our Z&C08; system. In this paper, we take a different approach, and assign a POS-tag to a partial word when its first character is separated from the final character of the previous word. When more characters are appended to a partial word, the POS is not changed. The idea is to use the POS of a partial word as the predicted POS of the full word it will become. Possible predictions are made with the first character of the word, and the likely ones will be kept in the beam for the next processing steps. For example, with the three characters “下雨天”, we try to keep two partial words (besides full words) in the beam when the first word “下” is processed, with the POS being VV and NN, respectively. The first POS predicts the two-character word “下雨” ， and the second the three-character word “下雨天”. Now when the second character is processed, we still need to maintain the possible POS NN in the agenda, which predicts the three-character word “下雨天”. As a main contribution of this paper, we show that the mechanism ofpredicting the POS at the first character gives competitive accuracy. This mechanism can be justified theoretically. Unlike alphabetical languages, each Chinese character represents some specific meanings. Given a character, it is natural for a human speaker to know immediately what types of words it can start. The allows the knowledge of possible POS-tags of words that a character can start, using information about the character from the training data. Moreover, the POS of the previous words to the current word are also useful in deciding possible POS for the word.1 The mechanism of first-character decision of POS also boosts the efficiency, since the enumeration of POS is unecessary when a character is appended to the end of an existing word. As a result, the complexity of each processing step is reduce by half compared to a method without POS prediction. Finally, an intuitive way to represent the status of a partial word is using a flag explicitly, which means an early decision of the segmentation of the next incoming character. We take a simpler alternative approach, and treat every word as a partial word until the next incoming character is separated from the last character of this word. Before a word is confirmed as a full word, we only apply to it features that represent its current partial status, such as character bigrams, its starting character and its part-ofspeech, etc. Full word features, including the first and last characters of a word, are applied immediately after a word is confirmed as complete. An important component for our proposed system is the training process, which needs to ensure that the model scores a partial word with predicted POS properly. We use the averaged perceptron (Collins, 2002) for training, together with the “early update” mechanism of Collins and Roark (2004). Rather than updating the parameters after decoding is com- plete, the modified algorithm updates parameters at any processing step if the correct partial candidate falls out of the beam. In our experiments using the Chinese Treebank 1The next incoming characters are also a useful source of information for predicting the POS. However, our system achieved competitive accuracy with Z&C08; without such character lookahead features. 845 data, our system ran an order of magnitude faster than our Z&C08; system with little loss of accuracy. The accuracy of our system was competitive with other recent models. 2 Model and Feature Templates We use a linear model to score both partial and full candidate outputs. Given an input x, the score of a candidate output y is computed as: Score(y) = Φ(y) · where Φ(y) is the global feature vector extracted from y, and is the parameter vector of the model. Figure 1 shows the feature templates for the model, where templates 1 14 contain only segmentation information and templates 15 29 contain w~ , w~ – – both segmentation and POS information. Each template is instantiated according to the current character in the decoding process. Row “For” shows the conditions for template instantiation, where “s” indicates that the corresponding template is instantiated when the current character starts a new word, and “a” indicates that the corresponding template is instantiated when the current character does not start a new word. In the row for feature templates, w, t and c are used to represent a word, a POS-tag and a character, respectively. The subscripts are based on the current character, where w−1 represents the first word to the left of the current character, and p−2 represents the POS-tag on the second word to the left of the current character, and so on. As an example, feature template 1is instantiated when the current character starts a new word, and the resulting feature value is the word to the left of this character. start(w), end(w) and len(w) represent the first character, the last character and the length of word w, respectively. The length of a word is normalized to 16 if it is larger than 16. cat(c) represents the POS category of character c, which is the set of POS-tags seen on character c, as we used in Z&C08.; Given a partial or complete candidate y, its global feature vector Φ(y) is computed by instantiating all applicable feature templates from Table 1 for each character in y, according to whether or not the character is separated from the previous character. The feature templates are mostly taken from, or inspired by, the feature templates of Z&C08.; Templates 1, 2, 3, 4, 5, 8, 10, 12, 13, 14, 15, 19, 20, Feature templateFor 24, 27 and 29 concern complete word information, and they are used in the model to differentiate correct and incorrect output structures in the same way as our Z&C08; model. Templates 6, 7, 9, 16, 17, 18, 21, 22, 23, 25, 26 and 28 concern partial word information, whose role in the model is to indicate the likelihood that the partial word including the current character will become a correct full word. They act as guidance for the action to take for the cur846 function DECODE(sent, agenda): CLEAR(agenda) ADDITEM(agenda, “”) for index in [0..LEN(sent)]: for cand in agenda: new ← APPEND(cand, sent[index]) ADDITEM(agenda, new) for pos in TAGSET(): new ← SEP(cand, sent[index], pos) ADDITEM(agenda, new) agenda ← N-BEST(agenda) retaugrenn BEST(agenda) Figure 1: The incremental beam-search decoder. rent character according to the context, and are the crucial reason for the effectiveness of the algorithm with a small beam-size. 2.1 Decoding The decoding algorithm builds an output candidate incrementally, one character at a time. Each character can either be attached to the current word or separated as the start a new word. When the current character starts a new word, a POS-tag is assigned to the new word. An agenda is used by the decoder to keep the N-best candidates during the incremental process. Before decoding starts, the agenda is initialized with an empty sentence. When a character is processed, existing candidates are removed from the agenda and extended with the current character in all possible ways, and the N-best newly generated candidates are put back onto the agenda. After all input characters have been processed, the highest-scored candidate from the agenda is taken as the output. Pseudo code for the decoder is shown in Figure 1. CLEAR removes all items from the agenda, ADDITEM adds a new item onto the agenda, N-BEST returns the N highest-scored items from the agenda, and BEST returns the highest-scored item from the agenda. LEN returns the number of characters in a sentence, and sent[i] returns the ith character from the sentence. APPEND appends a character to the last word in a candidate, and SEP joins a character as the start of a new word in a candidate, assigning a POS-tag to the new word. Both our decoding algorithm and the decoding algorithm of Z&C08; run in linear time. However, in order to generate possible candidates for each character, Z&C08; uses an extra loop to search for possible words that end with the current character. A restriction to the maximum word length is applied to limit the number of iterations in this loop, without which the algorithm would have quadratic time complexity. In contrast, our decoder does not search backword for the possible starting character of any word. Segmentation ambiguities are resolved by binary choices between the actions append or separate for each character, and no POS enumeration is required when the character is appended. This improves the speed by a significant factor. 2.2 Training The learning algorithm is based on the generalized perceptron (Collins, 2002), but parameter adjustments can be performed at any character during the decoding process, using the “early update” mechanism of Collins and Roark (2004). The parameter vector of the model is initialized as all zeros before training, and used to decode training examples. Each training example is turned into the raw input format, and processed in the same way as decoding. After each character is processed, partial candidates in the agenda are compared to the corresponding gold-standard output for the same characters. If none of the candidates in the agenda are correct, the decoding is stopped and the parameter vector is updated by adding the global feature vector of the gold-standard partial output and subtracting the global feature vector of the highest-scored partial candidate in the agenda. The training process then moves on to the next example. However, if any item in the agenda is the same as the corresponding gold-standard, the decoding process moves to the next character, without any change to the parameter values. After all characters are processed, the decoder prediction is compared with the training example. If the prediction is correct, the parameter vector is not changed; otherwise it is updated by adding the global feature vector of the training example and subtracting the global feature vector of the decoder prediction, just as the perceptron algorithm does. The same training examples can be used to train the model for multiple iterations. We use 847 the averaged parameter vector (Collins, 2002) as the final model. Pseudocode for the training algorithm is shown in Figure 2. It is based on the decoding algorithm in Figure 1, and the main differences are: (1) the training algorithm takes the gold-standard output and the parameter vector as two additional arguments; (2) the training algorithm does not return a prediction, but modifies the parameter vector when necessary; (3) lines 11to 20 are additional lines of code for parameter updates. Without lines 11 to 16, the training algorithm is exactly the same as the generalized perceptron algorithm. These lines are added to ensure that the agenda contains highly probable candidates during the whole beam-search process, and they are crucial to the high accuracy of the system. As stated earlier, the decoder relies on proper scoring of partial words to maintain a set of high quality candidates in the agenda. Updating the value of the parameter vector for partial outputs can be seen as a means to ensure correct scoring of partial candidates at any character. 2.3 Pruning We follow Z&C08; and use several pruning methods, most of which serve to to improve the accuracy by removing irrelevant candidates from the beam. First, the system records the maximum number of characters that a word with a particular POS-tag can have. For example, from the Chinese Treebank that we used for our experiments, most POS are associated with only with one- or two-character words. The only POS-tags that are seen with words over ten characters long are NN (noun), NR (proper noun) and CD (numbers). The maximum word length information is initialized as all ones, and updated according to each training example before it is processed. Second, a tag dictionary is used to record POStags associated with each word. During decoding, frequent words and words with “closed set” tags2 are only allowed POS-tags according to the tag dictionary, while other words are allowed every POS-tag to make candidate outputs. Whether a word is a frequent word is decided by the number of times it has been seen in the training process. Denoting the num2“Closed set” tags are the set of POS-tags which are only associated with a fixed set of words, according to the Penn Chinese Treebank specifications (Xia, 2000). function TRAIN(sent, agenda, gold-standard, w~ ): 01: CLEAR(agenda) 02: ADDITEM(agenda, “”) 03: for index in [0..LEN(sent)]: 04: 05: 06: 07: 08: 09: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: for cand in agenda: new ← APPEND(cand, sent[index]) ADDITEM(agenda, new) for pos in TAGSET(): new ← SEP(cand, sent[index], pos) ADDITEM(agenda, new) agenda ← N-BEST(agenda) faogre cnadnad ← ←in agenda: if cand = gold-standard[0:index] : CONTINUE w~ ← w~ + Φ(gold-standard[0:index]) ww~~ ← ww~ ~ - Φ(BEST(agenda)) wr~et ←urn w~ if BEST(agenda) gold-standard: w~ ← a ~wg + Φ(gold-standard) ww~~ ← ww~ ~ - Φ(BEST(agenda)) wr~et ←urn w~ return = Figure 2: The incremental learning function. ber of times the most frequent word has been seen with M, a word is a frequent word if it has been seen more than M/5000 5 times. The threshold value is taken from Z&C08;, and we did not adjust it during development. Word frequencies are initialized as zeros and updated according to each training example before it is processed; the tag dictionary is initialized as empty and updated according to each training example before it is processed. Third, we make an additional record of the initial characters for words with “closed set” tags. During decoding, when the current character is added as the start of a new word, “closed set” tags are only assigned to the word if it is consistent with the record. This type of pruning is used in addition to the tag + dictionary to prune invalid partial words, while the tag dictionary is used to prune complete words. The record for initial character and POS is initially empty, and udpated according to each training example before it is processed. Finally, at any decoding step, we group partial 848 candidates that are generated by separating the current character as the start of a new word by the signature p0p−1w−1, and keep only the best among those having the same p0p−1w−1. The signature p0p−1w−1 is decided by the feature templates we use: it can be shown that if two candidates cand1 and cand2 generated at the same step have the same signature, and the score of cand1 is higher than the score of cand2, then at any future step, the highest scored candidate generated from cand1 will always have a higher score than the highest scored candidate generated from cand2. From the above pruning methods, only the third was not used by Z&C08.; It can be seen as an extra mechanism to help keep likely partial words in the agenda and improve the accuracy, but which does not give our system a speed advantage over Z&C08.; 3 Experiments We used the Chinese Treebank (CTB) data to perform one set of development tests and two sets of fi- Training iteration Figure 3: The influence of beam-sizes, and the convergence of the perceptron. nal tests. The CTB 4 was split into two parts, with the CTB 3 being used for a 10-fold cross validation test to compare speed and accuracies with Z&C08;, and the rest being used for development. The CTB 5 was used to perform the additional set of experiments to compare accuracies with other recent work. We use the standard F-measure to evaluate output accuracies. For word segmentation, precision is defined as the number of correctly segmented words divided by the total number of words in the output, and recall is defined as the number of correctly segmented words divided by the total number of words in the gold-standard output. For joint segmentation and POS-tagging, precision is defined as the number of correctly segmented and POS-tagged words divided by the total number of words from the output, and recall is defined as the correctly segmented and POS-tagged words divided by the total number of words in the gold-standard output. All our experiments were performed on a Linux platform, and a single 2.66GHz Intel Core 2 CPU. 3.1 Development tests Our development data consists of 150K words in 4798 sentences. 80% of the data were randomly chosen as the development training data, while the rest were used as the development test data. Our development tests were mainly used to decide the size ofthe beam, the number oftraining iterations, the ef- fect of partial features in beam-search decoding, and the effect of incremental learning (i.e. early update). 849 Figure 3 shows the accuracy curves for joint segmentation and POS-tagging by the number of training iterations, using different beam sizes. With the size of the beam increasing from 1to 32, the accuracies generally increase, while the amount of increase becomes small when the size of the beam becomes 16. After the 10th iteration, a beam size of 32 does not always give better accuracies than a beam size of 16. We therefore chose 16 as the size of the beam for our system. The testing times for each beam size between 1 and 32 are 7.16s, 11.90s, 18.42s, 27.82s, 46.77s and 89.21s, respectively. The corresponding speeds in the number of sentences per second are 111.45, 67.06, 43.32, 28.68, 17.06 and 8.95, respectively. Figure 3 also shows that the accuracy increases with an increased number of training iterations, but the amount of increase becomes small after the 25th iteration. We chose 29 as the number of iterations to train our system. The effect of incremental training: We compare the accuracies by incremental training using early update and normal perceptron training. In the normal perceptron training case, lines 11to 16 are taken out of the training algorithm in Figure 2. The algorithm reached the best performance at the 22nd iteration, with the segmentation F-score being 90.58% and joint F-score being 83.38%. In the incremental training case, the algorithm reached the best accuracy at the 30th training iteration, obtaining a segmentation F-score of 91.14% and a joint F-score of 84.06%. 3.2 Final tests using CTB 3 CTB 3 consists of 150K words in 10364 sentences. We follow Z&C08; and split it into 10 equal-sized parts. In each test, one part is taken as the test data and the other nine are combined together as the training data. We compare the speed and accuracy with the joint segmentor and tagger of Z&C08;, which is publicly available as the ZPar system, version 0.23. The results are shown in Table 2, where each row shows one cross validation test. The column head- ings “sf”, “jf”, “time” and “speed” refer to segmentation F-measure, joint F-measure, testing time (in 3http://www.sourceforge.net/projects/zpar #sZf&C08jftimespeed; tshfis papjefrtimespeed seconds) and testing speed (in the number of sentences per second), respectively. Our system gave a joint segmentation and POStagging F-score of 91.37%, which is only 0.04% lower than that of ZPar 0.2. The speed of our system was over 10 times as fast as ZPar 0.2. 3.3 Final tests using CTB 5 We follow Kruengkrai et al. (2009) and split the CTB 5 into training, development testing and testing sets, as shown in Table 3. We ignored the development test data since our system had been developed in previous experiments. Kruengkrai et al. (2009) made use of character type knowledge for spaces, numerals, symbols, alphabets, Chinese and other characters. In the previous experiments, our system did not use any knowledge beyond the training data. To make the comparison fairer, we included knowledge of English letters and Arabic numbers in this experiment. During both training and decoding, English letters and Arabic numbers are segmented using simple rules, treating consecutive English letters or Arabic numbers as a single word. The results are shown in Table 4, where row “N07” refers to the model of Nakagawa and Uchimoto (2007), rows “J08a” and “b” refer to the models of Jiang et al. (2008a) and Jiang et al. (2008b), and row “K09” refers to the models of Kruengkrai et al. (2009). Columns “sf” and “jf” refer to segmentation and joint accuracies, respectively. Our system 850 SectionsSentencesWords T Daerbsvltien3:gTrain14230i–7n021 g–71,3d90–21e035v 1elopm1e385n40t,8and5tes a648t,903o2n,18C92TB5. TJoKNab0ul8re79abs4(y:rtAesomcl-indurea)vycom9 pa7 r.i87s34o59n w3 i.t64h2710recntsudio sfjf CTB 5. gave comparable accuracies to these recent works, obtaining the best (same as the error-driven version of K09) joint F-score. 4 Related Work The effectiveness of our beam-search decoder showed that the joint segmentation and tagging problem may be less complex than previously perceived (Zhang and Clark, 2008; Jiang et al., 2008a). At the very least, the single model approach with a simple decoder achieved competitive accuracies to what has been achieved so far by the reranking (Shi and Wang, 2007; Jiang et al., 2008b) models and an ensemble model using machine-translation techniques (Jiang et al., 2008a). This may shed new light on joint segmentation and POS-tagging methods. Kruengkrai et al. (2009) and Zhang and Clark (2008) are the most similar to our system among related work. Both systems use a discriminatively trained linear model to score candidate outputs. The work of Kruengkrai et al. (2009) is based on Nakagawa and Uchimoto (2007), which separates the processing of known words and unknown words, and uses a set of segmentation tags to represent the segmentation of characters. In contrast, our model is conceptually simpler, and does not differentiate known words and unknown words. Moreover, our model is based on our previous work, in line with Zhang and Clark (2007), which does not treat word segmentation as character sequence labeling. Our learning and decoding algorithms are also different from Kruengkrai et al. (2009). While Kruengkrai et al. (2009) perform dynamic programming and MIRA learning, we use beam-search to perform incremental decoding, and the early-update version of the perceptron algorithm to train the model. Dynamic programming is exact inference, for which the time complexity is decided by the locality of feature templates. In contrast, beam-search is approximate and can run in linear time. The parameter updating for our algorithm is conceptually and computationally simpler than MIRA, though its performance can be slightly lower. However, the earlyupdate mechanism we use is consistent with our incremental approach, and improves the learning of the beam-search process. 5 Conclusion We showed that a simple beam-search decoding algorithm can be effectively applied to the decoding problem for a global linear model for joint word segmentation and POS-tagging. By guiding search with partial word information and performing learning for partial candidates, our system achieved sig- nificantly faster speed with little accuracy loss compared to the system of Z&C08.; The source code of our joint segmentor and POStagger can be found at: www.sourceforge.net/projects/zpar, version 0.4. 851 Acknowledgements We thank Canasai Kruengkrai for discussion on efficiency issues, and the anonymous reviewers for their suggestions. Yue Zhang and Stephen Clark are supported by the European Union Seventh Framework Programme (FP7-ICT-2009-4) under grant agreement no. 247762. References Eugene Charniak, Mark Johnson, Micha Elsner, Joseph Austerweil, David Ellis, Isaac Haxton, Catherine Hill, R. Shrivaths, Jeremy Moore, Michael Pozar, and Theresa Vu. 2006. Multilevel coarse-to-fine PCFG parsing. In Proceedings of HLT/NAACL, pages 168– 175, New York City, USA, June. Association for Computational Linguistics. Michael Collins and Brian Roark. 2004. Incremental parsing with the perceptron algorithm. In Proceedings of ACL, pages 111–1 18, Barcelona, Spain, July. Michael Collins. 2002. Discriminative training methods for hidden Markov models: Theory and experiments with perceptron algorithms. In Proceedings of EMNLP, pages 1–8, Philadelphia, USA, July. Wenbin Jiang, Liang Huang, Qun Liu, and Yajuan L u¨. 2008a. A cascaded linear model for joint Chinese word segmentation and part-of-speech tagging. In Proceedings of ACL/HLT, pages 897–904, Columbus, Ohio, June. Wenbin Jiang, Haitao Mi, and Qun Liu. 2008b. Word lattice reranking for Chinese word segmentation and part-of-speech tagging. In Proceedings of COLING, pages 385–392, Manchester, UK, August. Canasai Kruengkrai, Kiyotaka Uchimoto, Jun’ichi Kazama, Yiou Wang, Kentaro Torisawa, and Hitoshi Isahara. 2009. An error-driven word-character hybrid model for joint Chinese word segmentation and POS tagging. In Proceedings of ACL/AFNLP, pages 5 13– 521, Suntec, Singapore, August. Tetsuji Nakagawa and Kiyotaka Uchimoto. 2007. A hybrid approach to word segmentation and POS tagging. In Proceedings of ACL Demo and Poster Session, Prague, Czech Republic, June. Hwee Tou Ng and Jin Kiat Low. 2004. Chinese part-ofspeech tagging: One-at-a-time or all-at-once? word- based or character-based? In Proceedings of EMNLP, Barcelona, Spain. Brian Roark and Kristy Hollingshead. 2008. Classifying chart cells for quadratic complexity context-free inference. In Proceedings of COLING, pages 745– 752, Manchester, UK, August. Coling 2008 Organizing Committee. Yanxin Shi and Mengqiu Wang. 2007. A dual-layer CRF based joint decoding method for cascade segmentation and labelling tasks. In Proceedings of IJCAI, Hyderabad, India. Fei Xia, 2000. The part-of-speech tagging guidelines for the Chinese Treebank (3.0). Yue Zhang and Stephen Clark. 2007. Chinese segmentation with a word-based perceptron algorithm. In Proceedings of ACL, pages 840–847, Prague, Czech Republic, June. Yue Zhang and Stephen Clark. 2008. Joint word segmentation and POS tagging using a single perceptron. In Proceedings of ACL/HLT, pages 888–896, Columbus, Ohio, June. 852

4 0.11160584 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks

Author: Xian Qian ; Qi Zhang ; Yaqian Zhou ; Xuanjing Huang ; Lide Wu

Abstract: Many sequence labeling tasks in NLP require solving a cascade of segmentation and tagging subtasks, such as Chinese POS tagging, named entity recognition, and so on. Traditional pipeline approaches usually suffer from error propagation. Joint training/decoding in the cross-product state space could cause too many parameters and high inference complexity. In this paper, we present a novel method which integrates graph structures of two subtasks into one using virtual nodes, and performs joint training and decoding in the factorized state space. Experimental evaluations on CoNLL 2000 shallow parsing data set and Fourth SIGHAN Bakeoff CTB POS tagging data set demonstrate the superiority of our method over cross-product, pipeline and candidate reranking approaches.

5 0.081743076 97 emnlp-2010-Simple Type-Level Unsupervised POS Tagging

Author: Yoong Keok Lee ; Aria Haghighi ; Regina Barzilay

Abstract: Part-of-speech (POS) tag distributions are known to exhibit sparsity a word is likely to take a single predominant tag in a corpus. Recent research has demonstrated that incorporating this sparsity constraint improves tagging accuracy. However, in existing systems, this expansion come with a steep increase in model complexity. This paper proposes a simple and effective tagging method that directly models tag sparsity and other distributional properties of valid POS tag assignments. In addition, this formulation results in a dramatic reduction in the number of model parameters thereby, enabling unusually rapid training. Our experiments consistently demonstrate that this model architecture yields substantial performance gains over more complex tagging — counterparts. On several languages, we report performance exceeding that of more complex state-of-the art systems.1

6 0.080561355 111 emnlp-2010-Two Decades of Unsupervised POS Induction: How Far Have We Come?

7 0.078673773 71 emnlp-2010-Latent-Descriptor Clustering for Unsupervised POS Induction

8 0.077248126 114 emnlp-2010-Unsupervised Parse Selection for HPSG

9 0.076972216 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction

10 0.073380336 41 emnlp-2010-Efficient Graph-Based Semi-Supervised Learning of Structured Tagging Models

11 0.072073303 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification

12 0.067965835 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment

13 0.06732516 25 emnlp-2010-Better Punctuation Prediction with Dynamic Conditional Random Fields

14 0.066486321 108 emnlp-2010-Training Continuous Space Language Models: Some Practical Issues

15 0.064770713 39 emnlp-2010-EMNLP 044

16 0.064554736 64 emnlp-2010-Incorporating Content Structure into Text Analysis Applications

17 0.062194742 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice

18 0.062056713 10 emnlp-2010-A Probabilistic Morphological Analyzer for Syriac

19 0.061298087 84 emnlp-2010-NLP on Spoken Documents Without ASR

20 0.05924527 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, 0.236), (1, 0.102), (2, 0.031), (3, -0.048), (4, -0.182), (5, 0.041), (6, -0.046), (7, -0.035), (8, -0.098), (9, -0.014), (10, 0.001), (11, -0.01), (12, 0.041), (13, -0.096), (14, 0.024), (15, -0.009), (16, 0.096), (17, -0.023), (18, -0.08), (19, -0.066), (20, 0.101), (21, -0.096), (22, -0.119), (23, 0.161), (24, 0.007), (25, 0.098), (26, 0.007), (27, -0.149), (28, 0.053), (29, -0.077), (30, 0.085), (31, -0.073), (32, 0.028), (33, 0.106), (34, -0.185), (35, 0.159), (36, 0.001), (37, -0.009), (38, 0.049), (39, -0.189), (40, -0.129), (41, -0.098), (42, -0.044), (43, 0.033), (44, 0.16), (45, 0.005), (46, 0.032), (47, 0.047), (48, -0.021), (49, -0.031)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.95828158 17 emnlp-2010-An Efficient Algorithm for Unsupervised Word Segmentation with Branching Entropy and MDL

Author: Valentin Zhikov ; Hiroya Takamura ; Manabu Okumura

2 0.64105111 43 emnlp-2010-Enhancing Domain Portability of Chinese Segmentation Model Using Chi-Square Statistics and Bootstrapping

Author: Baobao Chang ; Dongxu Han

3 0.58544868 2 emnlp-2010-A Fast Decoder for Joint Word Segmentation and POS-Tagging Using a Single Discriminative Model

Author: Yue Zhang ; Stephen Clark

4 0.46073902 108 emnlp-2010-Training Continuous Space Language Models: Some Practical Issues

Author: Hai Son Le ; Alexandre Allauzen ; Guillaume Wisniewski ; Francois Yvon

Abstract: Using multi-layer neural networks to estimate the probabilities of word sequences is a promising research area in statistical language modeling, with applications in speech recognition and statistical machine translation. However, training such models for large vocabulary tasks is computationally challenging which does not scale easily to the huge corpora that are nowadays available. In this work, we study the performance and behavior of two neural statistical language models so as to highlight some important caveats of the classical training algorithms. The induced word embeddings for extreme cases are also analysed, thus providing insight into the convergence issues. A new initialization scheme and new training techniques are then introduced. These methods are shown to greatly reduce the training time and to significantly improve performance, both in terms ofperplexity and on a large-scale translation task.

5 0.42951411 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks

Author: Xian Qian ; Qi Zhang ; Yaqian Zhou ; Xuanjing Huang ; Lide Wu

6 0.3726652 71 emnlp-2010-Latent-Descriptor Clustering for Unsupervised POS Induction

7 0.37062988 111 emnlp-2010-Two Decades of Unsupervised POS Induction: How Far Have We Come?

8 0.36147597 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification

9 0.35767469 114 emnlp-2010-Unsupervised Parse Selection for HPSG

10 0.35760257 10 emnlp-2010-A Probabilistic Morphological Analyzer for Syriac

11 0.3562797 25 emnlp-2010-Better Punctuation Prediction with Dynamic Conditional Random Fields

12 0.35313129 39 emnlp-2010-EMNLP 044

13 0.34529233 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction

14 0.32871684 113 emnlp-2010-Unsupervised Induction of Tree Substitution Grammars for Dependency Parsing

15 0.3138037 45 emnlp-2010-Evaluating Models of Latent Document Semantics in the Presence of OCR Errors

16 0.3052938 81 emnlp-2010-Modeling Perspective Using Adaptor Grammars

17 0.30163431 83 emnlp-2010-Multi-Level Structured Models for Document-Level Sentiment Classification

18 0.29977268 97 emnlp-2010-Simple Type-Level Unsupervised POS Tagging

19 0.29125872 122 emnlp-2010-WikiWars: A New Corpus for Research on Temporal Expressions

20 0.29068398 117 emnlp-2010-Using Unknown Word Techniques to Learn Known Words

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(3, 0.015), (10, 0.013), (12, 0.025), (29, 0.063), (30, 0.013), (32, 0.02), (52, 0.024), (56, 0.054), (66, 0.085), (72, 0.581), (76, 0.017)]

similar papers list:

simIndex simValue paperId paperTitle

1 0.94157422 122 emnlp-2010-WikiWars: A New Corpus for Research on Temporal Expressions

Author: Pawel Mazur ; Robert Dale

Abstract: The reliable extraction of knowledge from text requires an appropriate treatment of the time at which reported events take place. Unfortunately, there are very few annotated data sets that support the development of techniques for event time-stamping and tracking the progression of time through a narrative. In this paper, we present a new corpus of temporally-rich documents sourced from English Wikipedia, which we have annotated with TIMEX2 tags. The corpus contains around 120000 tokens, and 2600 TIMEX2 expressions, thus comparing favourably in size to other existing corpora used in these areas. We describe the prepa- ration of the corpus, and compare the profile of the data with other existing temporally annotated corpora. We also report the results obtained when we use DANTE, our temporal expression tagger, to process this corpus, and point to where further work is required. The corpus is publicly available for research purposes.

same-paper 2 0.93065804 17 emnlp-2010-An Efficient Algorithm for Unsupervised Word Segmentation with Branching Entropy and MDL

Author: Valentin Zhikov ; Hiroya Takamura ; Manabu Okumura

3 0.88323396 117 emnlp-2010-Using Unknown Word Techniques to Learn Known Words

Author: Kostadin Cholakov ; Gertjan van Noord

Abstract: Unknown words are a hindrance to the performance of hand-crafted computational grammars of natural language. However, words with incomplete and incorrect lexical entries pose an even bigger problem because they can be the cause of a parsing failure despite being listed in the lexicon of the grammar. Such lexical entries are hard to detect and even harder to correct. We employ an error miner to pinpoint words with problematic lexical entries. An automated lexical acquisition technique is then used to learn new entries for those words which allows the grammar to parse previously uncovered sentences successfully. We test our method on a large-scale grammar of Dutch and a set of sentences for which this grammar fails to produce a parse. The application of the method enables the grammar to cover 83.76% of those sentences with an accuracy of 86.15%.

4 0.54150182 11 emnlp-2010-A Semi-Supervised Approach to Improve Classification of Infrequent Discourse Relations Using Feature Vector Extension

Author: Hugo Hernault ; Danushka Bollegala ; Mitsuru Ishizuka

Abstract: Several recent discourse parsers have employed fully-supervised machine learning approaches. These methods require human annotators to beforehand create an extensive training corpus, which is a time-consuming and costly process. On the other hand, unlabeled data is abundant and cheap to collect. In this paper, we propose a novel semi-supervised method for discourse relation classification based on the analysis of cooccurring features in unlabeled data, which is then taken into account for extending the feature vectors given to a classifier. Our experimental results on the RST Discourse Treebank corpus and Penn Discourse Treebank indicate that the proposed method brings a significant improvement in classification accuracy and macro-average F-score when small training datasets are used. For instance, with training sets of c.a. 1000 labeled instances, the proposed method brings improvements in accuracy and macro-average F-score up to 50% compared to a baseline classifier. We believe that the proposed method is a first step towards detecting low-occurrence relations, which is useful for domains with a lack of annotated data.

5 0.48531586 73 emnlp-2010-Learning Recurrent Event Queries for Web Search

Author: Ruiqiang Zhang ; Yuki Konda ; Anlei Dong ; Pranam Kolari ; Yi Chang ; Zhaohui Zheng

Abstract: Recurrent event queries (REQ) constitute a special class of search queries occurring at regular, predictable time intervals. The freshness of documents ranked for such queries is generally of critical importance. REQ forms a significant volume, as much as 6% of query traffic received by search engines. In this work, we develop an improved REQ classifier that could provide significant improvements in addressing this problem. We analyze REQ queries, and develop novel features from multiple sources, and evaluate them using machine learning techniques. From historical query logs, we develop features utilizing query frequency, click information, and user intent dynamics within a search session. We also develop temporal features by time series analysis from query frequency. Other generated features include word matching with recurrent event seed words and time sensitivity of search result set. We use Naive Bayes, SVM and decision tree based logistic regres- sion model to train REQ classifier. The results on test data show that our models outperformed baseline approach significantly. Experiments on a commercial Web search engine also show significant gains in overall relevance, and thus overall user experience.

6 0.48474866 32 emnlp-2010-Context Comparison of Bursty Events in Web Search and Online Media

7 0.44451159 53 emnlp-2010-Fusing Eye Gaze with Speech Recognition Hypotheses to Resolve Exophoric References in Situated Dialogue

8 0.42779991 2 emnlp-2010-A Fast Decoder for Joint Word Segmentation and POS-Tagging Using a Single Discriminative Model

9 0.41880572 20 emnlp-2010-Automatic Detection and Classification of Social Events

10 0.4165594 45 emnlp-2010-Evaluating Models of Latent Document Semantics in the Presence of OCR Errors

11 0.41429257 24 emnlp-2010-Automatically Producing Plot Unit Representations for Narrative Text

12 0.41044044 49 emnlp-2010-Extracting Opinion Targets in a Single and Cross-Domain Setting with Conditional Random Fields

13 0.40651837 123 emnlp-2010-Word-Based Dialect Identification with Georeferenced Rules

14 0.40260431 92 emnlp-2010-Predicting the Semantic Compositionality of Prefix Verbs

15 0.39969793 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks

16 0.39917871 43 emnlp-2010-Enhancing Domain Portability of Chinese Segmentation Model Using Chi-Square Statistics and Bootstrapping

17 0.39726305 51 emnlp-2010-Function-Based Question Classification for General QA

18 0.39184919 80 emnlp-2010-Modeling Organization in Student Essays

19 0.39114875 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation

20 0.38875607 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice