emnlp emnlp2010 emnlp2010-25 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Wei Lu ; Hwee Tou Ng
Abstract: This paper focuses on the task of inserting punctuation symbols into transcribed conversational speech texts, without relying on prosodic cues. We investigate limitations associated with previous methods, and propose a novel approach based on dynamic conditional random fields. Different from previous work, our proposed approach is designed to jointly perform both sentence boundary and sentence type prediction, and punctuation prediction on speech utterances. We performed evaluations on a transcribed conversational speech domain consisting of both English and Chinese texts. Empirical results show that our method outperforms an approach based on linear-chain conditional random fields and other previous approaches.
Reference: text
sentIndex sentText sentNum sentScore
1 s g Abstract This paper focuses on the task of inserting punctuation symbols into transcribed conversational speech texts, without relying on prosodic cues. [sent-4, score-1.018]
2 Different from previous work, our proposed approach is designed to jointly perform both sentence boundary and sentence type prediction, and punctuation prediction on speech utterances. [sent-6, score-0.961]
3 We performed evaluations on a transcribed conversational speech domain consisting of both English and Chinese texts. [sent-7, score-0.276]
4 Empirical results show that our method outperforms an approach based on linear-chain conditional random fields and other previous approaches. [sent-8, score-0.241]
5 , true case, sentence boundaries, punctuation symbols, etc) is not available. [sent-11, score-0.46]
6 We focus on the punctuation prediction task in this work. [sent-13, score-0.582]
7 Most previous punctuation prediction techniques, developed mostly by the speech processing community, exploit both lexical and prosodic cues. [sent-14, score-0.818]
8 However, in order to fully exploit prosodic features such as pitch and pause duration, it is necessary 177 to have access to the original raw speech waveforms. [sent-15, score-0.303]
9 In some scenarios where further natural language processing (NLP) tasks on the transcribed speech texts become the main concern, speech prosody information may not be readily available. [sent-16, score-0.491]
10 For example, in the recent evaluation campaign of the Inter- national Workshop on Spoken Language Translation (IWSLT) (Paul, 2009), only manually transcribed or automatically recognized speech texts are provided but the original raw speech waveforms are not available. [sent-17, score-0.594]
11 In this paper, we tackle the task of predicting punctuation symbols from a standard text processing perspective, where only the speech texts are available, without relying on additional prosodic features such as pitch and pause duration. [sent-18, score-1.002]
12 Specifically, we perform the punctuation prediction task on transcribed conversational speech texts, using the IWSLT corpus (Paul, 2009) as the evaluation data. [sent-19, score-0.858]
13 An example English utterance randomly selected from the IWSLT corpus, along with its punctuated and cased version, are shown below: you are quite welcome and by the way we may get other reservations so could you please call us as soon as you fix the date You are quite welcome . [sent-22, score-0.513]
14 2 Related Work Punctuation prediction has been extensively studied in the speech processing field. [sent-32, score-0.282]
15 Kim – and Woodland (2001) performed punctuation insertion during speech recognition. [sent-35, score-0.534]
16 Huang and Zweig (2002) presented a maximum entropy-based tagging approach to punctuation insertion in spontaneous English conversational speech, where both lexical and prosodic features were exploited. [sent-39, score-0.753]
17 (2005) focused on the sentence boundary detection task, by making use of conditional random fields (CRF) (Lafferty et al. [sent-41, score-0.465]
18 (2009) presented a purely n-gram based approach that jointly predicted punctuation and case information of English. [sent-49, score-0.442]
19 (1998) presented a “hidden event language model” that treated boundary detection and punctuation insertion as an interword hidden event detection task. [sent-51, score-1.028]
20 Recently, there are also several research efforts that try to optimize some downstream application after punctuation prediction, rather than the prediction task itself. [sent-56, score-0.619]
21 Examples of such downstream applications include punctuation prediction for part-ofspeech (POS) tagging and name tagging (Hillard et al. [sent-57, score-0.69]
22 3 Hidden Event Language Model Many previous research efforts consider the boundary detection and punctuation insertion task as a hidden event detection task. [sent-61, score-0.891]
23 Specifically, in this task word boundaries and punctuation symbols are encoded as interword events. [sent-65, score-0.657]
24 However, we argue that in many cases, modeling of longer range dependencies is required for punctuation insertion. [sent-72, score-0.432]
25 This hurts the punctuation prediction performance for our task since we are particularly interested in conversational speech texts where question sentences appear frequently. [sent-74, score-0.961]
26 Examples include relocating or duplicating punctuation symbols to different positions of a sentence such that they appear closer to the indicative words (e. [sent-76, score-0.798]
27 One such technique was introduced by the organizers of the IWSLT evaluation campaign, who suggested duplicating the ending punctuation symbol to the beginning of each sentence before training the language model1 . [sent-79, score-0.704]
28 Furthermore, a direct application of such a method may fail in the event of multiple sentences per utterance without clearly annotated sentence boundaries within an utterance. [sent-82, score-0.367]
29 Another drawback associated with such an approach is that the method encodes strong dependency assumptions between the punctuation symbol to be inserted and its surrounding words. [sent-83, score-0.58]
30 Thus, it lacks the robustness to handle cases where noisy or out-of-vocabulary (OOV) words frequently appear, such as in texts automatically recognized by ASR systems. [sent-84, score-0.262]
31 In this paper, we devise techniques based on conditional random fields to tackle the difficulties due to long range dependencies. [sent-85, score-0.241]
32 4 Linear-Chain Conditional Random Fields One natural approach to relax the strong dependency assumptions encoded by the hidden event language model is to adopt an undirected graphical model, where arbitrary overlapping features can be exploited. [sent-86, score-0.239]
33 ) Table 1: The set of all possible tags for linear-chain CRF We can model the punctuation prediction task as the process of assigning a tag to each word, where the set of possible tags is given in Table 1. [sent-106, score-0.724]
34 That is, we assume each word can be associated with an event, which tells us which punctuation sym- bol (possibly NONE) should be inserted after the word. [sent-107, score-0.444]
35 The training data consists of a set of utterances where punctuation symbols are encoded as tags that are assigned to the individual words. [sent-108, score-0.708]
36 The tag NONE means no punctuation symbol is inserted after the current word. [sent-109, score-0.537]
37 Any other tag refers to inserting the corresponding punctuation symbol. [sent-110, score-0.435]
38 QMARK Figure 2: An example tagging of a training sentence for the linear-chain CRF predicted and the punctuated text can then be constructed from such an output. [sent-125, score-0.284]
39 , 2007), we factorize a feature of conditional random fields as a product of a binary function on assignment of the set of cliques at the current time step (in this case an edge), and a feature function solely defined on the observation sequence. [sent-128, score-0.294]
40 Special start and end symbols are used beyond the utterance boundaries. [sent-131, score-0.316]
41 Such a linear-chain CRF model is capable ofmodeling dependencies between words and punctuation symbols with arbitrary overlapping features, thus avoiding the strong dependency assumptions in the hidden event language model. [sent-133, score-0.8]
42 However, the linearchain CRF model still exhibits several problems for the punctuation task. [sent-134, score-0.429]
43 In particular, the dependency between the punctuation symbols and the indicative words cannot be captured adequately, if they appear too far away from each other. [sent-135, score-0.644]
44 For example, in the sample utterance shown in Figure 2, the long range dependency between the ending question mark and the indicative words would you which appear very far away cannot be directly captured. [sent-136, score-0.362]
45 Hence, it would be more reasonable to hypothesize that the punctuation symbols are annotated at the sentence level, rather than relying on a limited window of surrounding words. [sent-138, score-0.668]
46 A model that can 180 jointly perform sentence segmentation and sentence type prediction, together with word level punctuation prediction would be more beneficial for our task. [sent-139, score-0.798]
47 However, it is not clear how such a model can benefit our task, which requires wordlevel labeling in addition to sentence boundary detection and sentence type prediction. [sent-145, score-0.341]
48 As we have discussed above, since we would like to jointly model both the word-level labeling task and the sentence-level annotation task (sentence boundary detection and sentence type prediction), introducing an additional layer of tags to perform both tasks together would be desirable. [sent-148, score-0.51]
49 The F-CRF as a specific case of dynamic condi- tional random fields was originally motivated from dynamic Bayesian networks, where an identical structure repeats over different time steps. [sent-152, score-0.26]
50 Formally, dynamic conditional random fields define the conditional probability of a sequence of label vectors y given the observation x as: pλ(y|x) =Z(1x)exp XtXXkλkfk(x,y(c,t),t)! [sent-155, score-0.402]
51 xyzn n Tables2:ntTlweahyorcedstQopENDfrXMoEOaBNplAoE Rsp,GeoCKd,sEODtiXbEaMlgIeANstR,AaQgKPNsEBpRrIoOGpD,QseNdIfo,reach Figure 3: A two-layer factorial CRF layer We build two layers of labels for this task, as listed in Table 2. [sent-165, score-0.318]
52 The word layer tags are responsible for inserting a punctuation symbol (including NONE) after each word, while the sentence layer 181 tags are used for annotating sentence boundaries and identifying the sentence type (declarative, question, or exclamatory). [sent-166, score-1.137]
53 The sentence layer tags are designed for three types of sentences. [sent-168, score-0.253]
54 When learning the sentence layer tags together with the word layer tags, the F-CRF model is capable of leveraging useful clues learned from the sentence layer about sentence type (e. [sent-172, score-0.617]
55 ), which can be used to guide the prediction of the punctuation symbol at each word, hence improving the performance at the word layer. [sent-180, score-0.675]
56 For example, consider jointly labeling the utterance shown in Figure 4. [sent-181, score-0.25]
57 Intuitively, when evidences show that the utterance consists of two sentences a declarative sentence followed by a question sentence, the model tends to annotate the second half of the utterance with the sequence QNBEG QNIN . [sent-182, score-0.554]
58 This in turn helps to predict the word level tag at the end of the utterance as QMARK, given the dependencies between the two layers existing at each time step. [sent-186, score-0.268]
59 In practice, during the learning process, the two layers of tags are jointly learned, thus providing evidences that influence each other’s tagging process. [sent-187, score-0.252]
60 – 6 Experiments We perform experiments on part of the corpus of the IWSLT09 evaluation campaign (Paul, 2009), where both Chinese and English conversational speech Sentence: no , please do not . [sent-191, score-0.313]
61 QNIN Figure 4: An example tagging of a training sentence for the factorial CRF texts are used. [sent-211, score-0.383]
62 We randomly split each of the two datasets into two portions, where 90% of the utterances are used for training the punctuation prediction models, and the remaining 10% for evaluating the prediction performance. [sent-215, score-0.886]
63 33 Table 3: Statistics of the BTEC and CT datasets For the methods based on the hidden event language model, we design extensive experiments due 2http://www. [sent-236, score-0.241]
64 Specifically, these experiments can be divided into two categories: with or without duplicating the ending punctuation symbol to the start of a sentence before training. [sent-241, score-0.704]
65 This setting can be used to assess the impact of the proximity between the punctuation symbol and the indicative words for the prediction task. [sent-242, score-0.76]
66 The single pass approach performs prediction in one single step, where all the punctuation symbols are predicted sequentially from left to right. [sent-244, score-0.747]
67 In the cascaded approach, we format the training sentences by replacing all sentence-ending punctuation symbols with special sentence boundary symbols first. [sent-245, score-0.895]
68 A model for sentence boundary prediction is learned based on such training data. [sent-246, score-0.359]
69 This step is then followed by predicting the actual punctuation symbols. [sent-247, score-0.394]
70 To assess the performance of the punctuation prediction task, we compute precision (prec. [sent-251, score-0.582]
71 1 Performance on Correctly Recognized Texts The performance of punctuation prediction on both Chinese (CN) and English (EN) texts in the correctly recognized output of the BTEC and CT datasets are presented in Table 4 and Table 5 respectively. [sent-258, score-0.882]
72 Specifically, for English, duplicating the ending punctuation symbol to the start of a sentence before training is shown to be very helpful in improving the overall prediction performance. [sent-290, score-0.892]
73 An English question sentence usually starts with indicative words such as do you or where that distinguish it from a declarative sentence. [sent-293, score-0.304]
74 Thus, duplicating the ending punctuation symbol to the start of a sentence so that it is near these indicative words helps to improve the prediction accuracy. [sent-294, score-0.977]
75 These pose difficulties for the simple hidden event language model, which only encodes simple dependencies over surrounding words by means of n-gram language modeling. [sent-322, score-0.284]
76 By introducing an additional layer of tags for performing sentence segmentation and sentence type prediction, the F-CRF model further boosts the performance over the L-CRF model. [sent-324, score-0.355]
77 2 Performance on Automatically Recognized Texts So far we only evaluated punctuation prediction performance on transcribed texts consisting of correctly recognized words. [sent-349, score-0.944]
78 Unlike the correctly recognized texts described in Section 6. [sent-353, score-0.262]
79 In the dataset released by the IWSLT organizers, the correct punctuation symbols are not annotated in the ASR outputs. [sent-355, score-0.606]
80 To conduct our experimental evaluation, we manually annotated the correct punctuation symbols on the ASR outputs. [sent-356, score-0.559]
81 2 requires substantial manual efforts to annotate the correct punctuation symbols. [sent-364, score-0.431]
82 In this section, we instead adopt an indirect approach to automatically evaluate the performance of punctuation prediction on ASR output texts by feeding the punctuated ASR texts to a state-of-the-art machine translation system, and evaluate the resulting translation performance. [sent-365, score-1.214]
83 For tuning the parameters of Moses, we use the official IWSLT05 evaluation set where the correct punctuation symbols are present. [sent-399, score-0.59]
84 Evaluations are performed on the ASR outputs of the IWSLT08 BTEC evaluation dataset, with punctuation symbols inserted by each punctuation prediction method. [sent-400, score-1.191]
85 The best translation performances for both translation directions are achieved by applying F-CRF as the punctuation prediction model to the ASR texts. [sent-407, score-0.77]
86 In addition, we also assess the translation performance when the manually annotated punctuation symbols as mentioned in Section 6. [sent-411, score-0.653]
87 16 (English to Chinese) respectively, which show that our punctuation prediction method gives competitive performance for spoken language trans185 lation. [sent-415, score-0.582]
88 It is important to note that in this work, we only focus on optimizing the punctuation prediction performance in the form of F1-measure, without regard to the subsequent NLP tasks. [sent-416, score-0.582]
89 How to perform punctuation prediction so as to optimize translation performance is an important research topic that is beyond the scope of this paper and needs further investigation in future work. [sent-417, score-0.676]
90 7 Conclusion In this paper, we have proposed a novel approach for predicting punctuation symbols for transcribed conversational speech texts. [sent-418, score-0.835]
91 Our proposed approach is built on top of a dynamic conditional random fields framework, which jointly performs punctuation prediction together with sentence boundary and sentence type prediction on speech utterances. [sent-419, score-1.439]
92 Experimental results have shown that our proposed approach outperforms the widely used approach based on the hidden event language model, and also outperforms a method based on linear-chain conditional random fields. [sent-421, score-0.333]
93 Our proposed approach has been shown to be general, working well on both Chinese and English, and on both correctly recognized and automatically recognized texts. [sent-422, score-0.244]
94 Our proposed approach also results in better translation accuracy when the punctuated automatically recognized texts are used in subsequent translation. [sent-423, score-0.52]
95 Impact of automatic comma prediction on POS/name tagging of speech. [sent-487, score-0.242]
96 The use of prosody in a combined system for punctuation generation and speech recognition. [sent-503, score-0.551]
97 Using conditional random fields for sentence boundary detection in speech. [sent-562, score-0.465]
98 Automatic sentence segmentation and punctuation prediction for spoken language translation. [sent-570, score-0.684]
99 Dynamic conditional random fields for jointly labeling multiple sequences. [sent-578, score-0.34]
100 Automatic detection of sentence boundaries and disfluencies based on recognized words. [sent-635, score-0.278]
wordName wordTfidf (topN-words)
[('punctuation', 0.394), ('btec', 0.245), ('iwslt', 0.233), ('crf', 0.22), ('prediction', 0.188), ('qnin', 0.184), ('asr', 0.18), ('symbols', 0.165), ('punctuated', 0.164), ('utterance', 0.151), ('prosodic', 0.142), ('texts', 0.14), ('dein', 0.123), ('factorial', 0.123), ('ct', 0.122), ('recognized', 0.122), ('layer', 0.116), ('event', 0.113), ('fields', 0.111), ('sutton', 0.11), ('boundary', 0.105), ('transcribed', 0.1), ('translation', 0.094), ('speech', 0.094), ('please', 0.093), ('symbol', 0.093), ('hidden', 0.09), ('declarative', 0.09), ('duplicating', 0.088), ('indicative', 0.085), ('conversational', 0.082), ('qnbeg', 0.082), ('conditional', 0.079), ('layers', 0.079), ('chinese', 0.078), ('utterances', 0.078), ('tags', 0.071), ('sentence', 0.066), ('question', 0.063), ('prosody', 0.063), ('ending', 0.063), ('debeg', 0.061), ('exclamatory', 0.061), ('interword', 0.061), ('lcrf', 0.061), ('tagging', 0.054), ('detection', 0.053), ('cliques', 0.053), ('hillard', 0.053), ('stolcke', 0.051), ('labeling', 0.051), ('random', 0.051), ('inserted', 0.05), ('dynamic', 0.049), ('jointly', 0.048), ('icslp', 0.047), ('dataset', 0.047), ('insertion', 0.046), ('talk', 0.044), ('campaign', 0.044), ('moses', 0.043), ('surrounding', 0.043), ('beeferman', 0.041), ('christensen', 0.041), ('grmm', 0.041), ('hakkanitur', 0.041), ('iask', 0.041), ('qmark', 0.041), ('reparameterization', 0.041), ('reservations', 0.041), ('inserting', 0.041), ('icassp', 0.041), ('datasets', 0.038), ('dependencies', 0.038), ('reordering', 0.037), ('efforts', 0.037), ('boundaries', 0.037), ('mccallum', 0.037), ('segmentation', 0.036), ('graphical', 0.036), ('matusov', 0.035), ('clique', 0.035), ('dialogs', 0.035), ('efron', 0.035), ('favre', 0.035), ('gravano', 0.035), ('linearchain', 0.035), ('pitch', 0.035), ('sarawagi', 0.035), ('spontaneous', 0.035), ('sequence', 0.033), ('hmm', 0.032), ('english', 0.032), ('kirchhoff', 0.032), ('kfk', 0.032), ('ostendorf', 0.032), ('pause', 0.032), ('tseng', 0.032), ('welcome', 0.032), ('official', 0.031)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000006 25 emnlp-2010-Better Punctuation Prediction with Dynamic Conditional Random Fields
Author: Wei Lu ; Hwee Tou Ng
Abstract: This paper focuses on the task of inserting punctuation symbols into transcribed conversational speech texts, without relying on prosodic cues. We investigate limitations associated with previous methods, and propose a novel approach based on dynamic conditional random fields. Different from previous work, our proposed approach is designed to jointly perform both sentence boundary and sentence type prediction, and punctuation prediction on speech utterances. We performed evaluations on a transcribed conversational speech domain consisting of both English and Chinese texts. Empirical results show that our method outperforms an approach based on linear-chain conditional random fields and other previous approaches.
2 0.19478771 75 emnlp-2010-Lessons Learned in Part-of-Speech Tagging of Conversational Speech
Author: Vladimir Eidelman ; Zhongqiang Huang ; Mary Harper
Abstract: This paper examines tagging models for spontaneous English speech transcripts. We analyze the performance of state-of-the-art tagging models, either generative or discriminative, left-to-right or bidirectional, with or without latent annotations, together with the use of ToBI break indexes and several methods for segmenting the speech transcripts (i.e., conversation side, speaker turn, or humanannotated sentence). Based on these studies, we observe that: (1) bidirectional models tend to achieve better accuracy levels than left-toright models, (2) generative models seem to perform somewhat better than discriminative models on this task, and (3) prosody improves tagging performance of models on conversation sides, but has much less impact on smaller segments. We conclude that, although the use of break indexes can indeed significantly im- prove performance over baseline models without them on conversation sides, tagging accuracy improves more by using smaller segments, for which the impact of the break indexes is marginal.
3 0.13608755 41 emnlp-2010-Efficient Graph-Based Semi-Supervised Learning of Structured Tagging Models
Author: Amarnag Subramanya ; Slav Petrov ; Fernando Pereira
Abstract: We describe a new scalable algorithm for semi-supervised training of conditional random fields (CRF) and its application to partof-speech (POS) tagging. The algorithm uses a similarity graph to encourage similar ngrams to have similar POS tags. We demonstrate the efficacy of our approach on a domain adaptation task, where we assume that we have access to large amounts of unlabeled data from the target domain, but no additional labeled data. The similarity graph is used during training to smooth the state posteriors on the target domain. Standard inference can be used at test time. Our approach is able to scale to very large problems and yields significantly improved target domain accuracy.
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.
5 0.10387012 30 emnlp-2010-Confidence in Structured-Prediction Using Confidence-Weighted Models
Author: Avihai Mejer ; Koby Crammer
Abstract: Confidence-Weighted linear classifiers (CW) and its successors were shown to perform well on binary and multiclass NLP problems. In this paper we extend the CW approach for sequence learning and show that it achieves state-of-the-art performance on four noun phrase chucking and named entity recognition tasks. We then derive few algorithmic approaches to estimate the prediction’s correctness of each label in the output sequence. We show that our approach provides a reliable relative correctness information as it outperforms other alternatives in ranking label-predictions according to their error. We also show empirically that our methods output close to absolute estimation of error. Finally, we show how to use this information to improve active learning.
6 0.10090657 26 emnlp-2010-Classifying Dialogue Acts in One-on-One Live Chats
7 0.096119776 33 emnlp-2010-Cross Language Text Classification by Model Translation and Semi-Supervised Learning
8 0.093955293 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks
9 0.082639158 84 emnlp-2010-NLP on Spoken Documents Without ASR
10 0.081649609 64 emnlp-2010-Incorporating Content Structure into Text Analysis Applications
11 0.08054582 108 emnlp-2010-Training Continuous Space Language Models: Some Practical Issues
12 0.075224087 20 emnlp-2010-Automatic Detection and Classification of Social Events
13 0.069170855 36 emnlp-2010-Discriminative Word Alignment with a Function Word Reordering Model
14 0.06766884 46 emnlp-2010-Evaluating the Impact of Alternative Dependency Graph Encodings on Solving Event Extraction Tasks
15 0.06732516 17 emnlp-2010-An Efficient Algorithm for Unsupervised Word Segmentation with Branching Entropy and MDL
16 0.066742577 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction
17 0.066032261 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice
18 0.064765058 63 emnlp-2010-Improving Translation via Targeted Paraphrasing
19 0.06091392 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
20 0.06046965 97 emnlp-2010-Simple Type-Level Unsupervised POS Tagging
topicId topicWeight
[(0, 0.241), (1, 0.036), (2, 0.019), (3, -0.052), (4, -0.235), (5, 0.004), (6, 0.092), (7, -0.017), (8, -0.114), (9, 0.072), (10, -0.216), (11, -0.241), (12, 0.011), (13, 0.16), (14, -0.047), (15, -0.027), (16, 0.02), (17, 0.016), (18, 0.031), (19, 0.04), (20, -0.051), (21, -0.026), (22, 0.024), (23, -0.055), (24, -0.072), (25, -0.061), (26, 0.01), (27, -0.072), (28, -0.134), (29, -0.212), (30, 0.044), (31, 0.018), (32, 0.023), (33, 0.119), (34, 0.001), (35, -0.147), (36, -0.078), (37, 0.043), (38, -0.05), (39, -0.222), (40, -0.025), (41, 0.089), (42, 0.018), (43, 0.017), (44, -0.036), (45, 0.122), (46, 0.227), (47, 0.054), (48, 0.036), (49, -0.026)]
simIndex simValue paperId paperTitle
same-paper 1 0.96155816 25 emnlp-2010-Better Punctuation Prediction with Dynamic Conditional Random Fields
Author: Wei Lu ; Hwee Tou Ng
Abstract: This paper focuses on the task of inserting punctuation symbols into transcribed conversational speech texts, without relying on prosodic cues. We investigate limitations associated with previous methods, and propose a novel approach based on dynamic conditional random fields. Different from previous work, our proposed approach is designed to jointly perform both sentence boundary and sentence type prediction, and punctuation prediction on speech utterances. We performed evaluations on a transcribed conversational speech domain consisting of both English and Chinese texts. Empirical results show that our method outperforms an approach based on linear-chain conditional random fields and other previous approaches.
2 0.63688242 75 emnlp-2010-Lessons Learned in Part-of-Speech Tagging of Conversational Speech
Author: Vladimir Eidelman ; Zhongqiang Huang ; Mary Harper
Abstract: This paper examines tagging models for spontaneous English speech transcripts. We analyze the performance of state-of-the-art tagging models, either generative or discriminative, left-to-right or bidirectional, with or without latent annotations, together with the use of ToBI break indexes and several methods for segmenting the speech transcripts (i.e., conversation side, speaker turn, or humanannotated sentence). Based on these studies, we observe that: (1) bidirectional models tend to achieve better accuracy levels than left-toright models, (2) generative models seem to perform somewhat better than discriminative models on this task, and (3) prosody improves tagging performance of models on conversation sides, but has much less impact on smaller segments. We conclude that, although the use of break indexes can indeed significantly im- prove performance over baseline models without them on conversation sides, tagging accuracy improves more by using smaller segments, for which the impact of the break indexes is marginal.
3 0.43075648 41 emnlp-2010-Efficient Graph-Based Semi-Supervised Learning of Structured Tagging Models
Author: Amarnag Subramanya ; Slav Petrov ; Fernando Pereira
Abstract: We describe a new scalable algorithm for semi-supervised training of conditional random fields (CRF) and its application to partof-speech (POS) tagging. The algorithm uses a similarity graph to encourage similar ngrams to have similar POS tags. We demonstrate the efficacy of our approach on a domain adaptation task, where we assume that we have access to large amounts of unlabeled data from the target domain, but no additional labeled data. The similarity graph is used during training to smooth the state posteriors on the target domain. Standard inference can be used at test time. Our approach is able to scale to very large problems and yields significantly improved target domain accuracy.
4 0.41361624 26 emnlp-2010-Classifying Dialogue Acts in One-on-One Live Chats
Author: Su Nam Kim ; Lawrence Cavedon ; Timothy Baldwin
Abstract: We explore the task of automatically classifying dialogue acts in 1-on-1 online chat forums, an increasingly popular means of providing customer service. In particular, we investigate the effectiveness of various features and machine learners for this task. While a simple bag-of-words approach provides a solid baseline, we find that adding information from dialogue structure and inter-utterance dependency provides some increase in performance; learners that account for sequential dependencies (CRFs) show the best performance. We report our results from testing using a corpus of chat dialogues derived from online shopping customer-feedback data.
5 0.4102428 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.
7 0.39250499 30 emnlp-2010-Confidence in Structured-Prediction Using Confidence-Weighted Models
8 0.34206152 40 emnlp-2010-Effects of Empty Categories on Machine Translation
9 0.31029683 17 emnlp-2010-An Efficient Algorithm for Unsupervised Word Segmentation with Branching Entropy and MDL
10 0.29644379 46 emnlp-2010-Evaluating the Impact of Alternative Dependency Graph Encodings on Solving Event Extraction Tasks
11 0.29096609 84 emnlp-2010-NLP on Spoken Documents Without ASR
12 0.28333905 20 emnlp-2010-Automatic Detection and Classification of Social Events
13 0.27781597 64 emnlp-2010-Incorporating Content Structure into Text Analysis Applications
14 0.25479043 69 emnlp-2010-Joint Training and Decoding Using Virtual Nodes for Cascaded Segmentation and Tagging Tasks
15 0.25390068 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
16 0.24787411 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice
17 0.24570642 5 emnlp-2010-A Hybrid Morpheme-Word Representation for Machine Translation of Morphologically Rich Languages
18 0.24538004 116 emnlp-2010-Using Universal Linguistic Knowledge to Guide Grammar Induction
19 0.24016142 4 emnlp-2010-A Game-Theoretic Approach to Generating Spatial Descriptions
20 0.23796658 53 emnlp-2010-Fusing Eye Gaze with Speech Recognition Hypotheses to Resolve Exophoric References in Situated Dialogue
topicId topicWeight
[(3, 0.018), (10, 0.018), (11, 0.306), (12, 0.036), (29, 0.089), (30, 0.011), (32, 0.025), (52, 0.027), (56, 0.088), (62, 0.015), (66, 0.112), (72, 0.062), (76, 0.038), (77, 0.016), (79, 0.027), (87, 0.036), (89, 0.01)]
simIndex simValue paperId paperTitle
same-paper 1 0.7534132 25 emnlp-2010-Better Punctuation Prediction with Dynamic Conditional Random Fields
Author: Wei Lu ; Hwee Tou Ng
Abstract: This paper focuses on the task of inserting punctuation symbols into transcribed conversational speech texts, without relying on prosodic cues. We investigate limitations associated with previous methods, and propose a novel approach based on dynamic conditional random fields. Different from previous work, our proposed approach is designed to jointly perform both sentence boundary and sentence type prediction, and punctuation prediction on speech utterances. We performed evaluations on a transcribed conversational speech domain consisting of both English and Chinese texts. Empirical results show that our method outperforms an approach based on linear-chain conditional random fields and other previous approaches.
2 0.50363308 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.
3 0.5009554 82 emnlp-2010-Multi-Document Summarization Using A* Search and Discriminative Learning
Author: Ahmet Aker ; Trevor Cohn ; Robert Gaizauskas
Abstract: In this paper we address two key challenges for extractive multi-document summarization: the search problem of finding the best scoring summary and the training problem of learning the best model parameters. We propose an A* search algorithm to find the best extractive summary up to a given length, which is both optimal and efficient to run. Further, we propose a discriminative training algorithm which directly maximises the quality ofthe best summary, rather than assuming a sentence-level decomposition as in earlier work. Our approach leads to significantly better results than earlier techniques across a number of evaluation metrics.
4 0.49994469 86 emnlp-2010-Non-Isomorphic Forest Pair Translation
Author: Hui Zhang ; Min Zhang ; Haizhou Li ; Eng Siong Chng
Abstract: This paper studies two issues, non-isomorphic structure translation and target syntactic structure usage, for statistical machine translation in the context of forest-based tree to tree sequence translation. For the first issue, we propose a novel non-isomorphic translation framework to capture more non-isomorphic structure mappings than traditional tree-based and tree-sequence-based translation methods. For the second issue, we propose a parallel space searching method to generate hypothesis using tree-to-string model and evaluate its syntactic goodness using tree-to-tree/tree sequence model. This not only reduces the search complexity by merging spurious-ambiguity translation paths and solves the data sparseness issue in training, but also serves as a syntax-based target language model for better grammatical generation. Experiment results on the benchmark data show our proposed two solutions are very effective, achieving significant performance improvement over baselines when applying to different translation models.
5 0.49796507 105 emnlp-2010-Title Generation with Quasi-Synchronous Grammar
Author: Kristian Woodsend ; Yansong Feng ; Mirella Lapata
Abstract: The task of selecting information and rendering it appropriately appears in multiple contexts in summarization. In this paper we present a model that simultaneously optimizes selection and rendering preferences. The model operates over a phrase-based representation of the source document which we obtain by merging PCFG parse trees and dependency graphs. Selection preferences for individual phrases are learned discriminatively, while a quasi-synchronous grammar (Smith and Eisner, 2006) captures rendering preferences such as paraphrases and compressions. Based on an integer linear programming formulation, the model learns to generate summaries that satisfy both types of preferences, while ensuring that length, topic coverage and grammar constraints are met. Experiments on headline and image caption generation show that our method obtains state-of-the-art performance using essentially the same model for both tasks without any major modifications.
6 0.49477968 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
7 0.49399346 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice
8 0.49343622 49 emnlp-2010-Extracting Opinion Targets in a Single and Cross-Domain Setting with Conditional Random Fields
9 0.49273747 120 emnlp-2010-What's with the Attitude? Identifying Sentences with Attitude in Online Discussions
10 0.49168247 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification
11 0.49079853 107 emnlp-2010-Towards Conversation Entailment: An Empirical Investigation
12 0.49075064 58 emnlp-2010-Holistic Sentiment Analysis Across Languages: Multilingual Supervised Latent Dirichlet Allocation
13 0.48812234 60 emnlp-2010-Improved Fully Unsupervised Parsing with Zoomed Learning
14 0.4877055 103 emnlp-2010-Tense Sense Disambiguation: A New Syntactic Polysemy Task
15 0.48725313 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
16 0.48563367 84 emnlp-2010-NLP on Spoken Documents Without ASR
17 0.48527774 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment
18 0.48497635 48 emnlp-2010-Exploiting Conversation Structure in Unsupervised Topic Segmentation for Emails
19 0.48259717 45 emnlp-2010-Evaluating Models of Latent Document Semantics in the Presence of OCR Errors
20 0.48115233 20 emnlp-2010-Automatic Detection and Classification of Social Events