Impact of Automatic Comma Prediction on POS/Name Tagging of Speech

DMCA

Impact of Automatic Comma Prediction on POS/Name Tagging of Speech (2006)

Cached

Download Links

by D. Hillard , Z. Huang , H. Ji , R. Grishman , D. Hakkani-tur , M. Harper , M. Ostendorf , W. Wang

Venue:	Proc. of the IEEE/ACL 2006 Workshop on Spoken Language Technology
Citations:	13 - 8 self

Citations

1217	SRILM an extensible language modeling toolkit - Stolcke - 2002 (Show Context) Citation Context ...de various measures and normalizations of pause duration, phone duration, fundamental frequency and energy. The posteriors from the two models are interpolated using equal weights. The SRI-LM toolkit =-=[14]-=- (with Kneser-Ney smoothing) is used for training the hidden-event model for sentence boundary prediction, with the same data sources as for training the Mandarin ASR language models, including broadc...
667	Boostexter: A boosting-based system for text categorization - Schapire, Singer (Show Context) Citation Context ...ce boundary vs. non-sentence boundary. The classifier uses a combination of hidden-event language models (5-gram) to exploit lexical information and sequence dependencies, and a boostexter classifier =-=[13]-=- to exploit lexical cues (word triples) in combination with prosodic and speaker change information. Prosodic features include various measures and normalizations of pause duration, phone duration, fu...
333	Discriminative reranking for natural language parsing - Collins, Koo - 2004 (Show Context) Citation Context ...). The second POS tagger, which utilizes the N-best extraction of the first tagger, incorporates various higher order N-gram features in a reranking method based on the boosting approach described in =-=[5]-=-. This tagger improvesstagging accuracy to 94.84% (76.98% on unknown words). The name tagger is based on an HMM that generally follows the Nymble model [6]. It identifies names of three classes: peopl...
82	A second-order hidden markov model for part-of-speech tagging,” in - Thede, Harper - 1999 (Show Context) Citation Context ...he broadcast news development set used in these experiments, character error rate is 5.6%. We utilized two part-of-speech taggers. The first is a Viterbi tagger that builds on the tagger developed in =-=[4]-=- that uses trigram transition probability P (Ti\|Ti−1Ti−2) and trigram emission probability P (Wi\|TiTi−1), where Ti and Wi represent the i-th tag and word. When a word was not observed during training ...
65	Maximum entropy model for punctuation annotation from speech, in: - Huang, Zweig - 2002 (Show Context) Citation Context ...h access to the reference transcription. 4. PUNCTUATION PREDICTION Previous work on punctuation detection to enhance the output of automatic speech recognizers mostly focuses on sentence segmentation =-=[7, 8, 9]-=-. For intra-sentence punctuation insertion for text, Beeferman et al. [10] use lexical information in the form of trigram language models. Zhang et. al. [11] use decision trees with linguistically sop...
18	et al., “Recent Innovations in Speech-to-Text Transcription at SRI-ICSI-UW - Stolcke - 2006 (Show Context) Citation Context ... tasks, and finally conclude with a summary of the key findings. 2. COMPONENT SYSTEMS The speech recognition system used in this work is a state-ofthe-art system, based on the SRI Decipher recognizer =-=[2]-=- and trained/tuned specifically for the Mandarin broadcast news task. Training texts for the system from a variety of sources were automatically word-segmented, using a maximum n-gram probability crit...
15	et al., Prosody-based automatic segmentation of speech into sentences and topics - Shriberg - 2000 (Show Context) Citation Context ...h access to the reference transcription. 4. PUNCTUATION PREDICTION Previous work on punctuation detection to enhance the output of automatic speech recognizers mostly focuses on sentence segmentation =-=[7, 8, 9]-=-. For intra-sentence punctuation insertion for text, Beeferman et al. [10] use lexical information in the form of trigram language models. Zhang et. al. [11] use decision trees with linguistically sop...
4	et al., “The Effects of Speech Recognition and Punctuation on Information Extraction performance - Makhoul - 2005 (Show Context) Citation Context ...nces between speech and text sources, other than the potential for transcription errors, is that typical ASR systems do not output punctuation cues, which are used in most text processing systems. In =-=[1]-=-, researchers from BBN showed that missing commas can have a dramatic impact on information extraction performance compared to using hand-transcribed commas, with performance losses typically bigger t...
4	Enriching speech recognition with sentence boundaries and disfluencies - Liu - 2006 (Show Context) Citation Context ...h access to the reference transcription. 4. PUNCTUATION PREDICTION Previous work on punctuation detection to enhance the output of automatic speech recognizers mostly focuses on sentence segmentation =-=[7, 8, 9]-=-. For intra-sentence punctuation insertion for text, Beeferman et al. [10] use lexical information in the form of trigram language models. Zhang et. al. [11] use decision trees with linguistically sop...
2	et al., “Nymble: A high-performance learning namefinder - Bikel - 1997 (Show Context) Citation Context ... based on the boosting approach described in [5]. This tagger improvesstagging accuracy to 94.84% (76.98% on unknown words). The name tagger is based on an HMM that generally follows the Nymble model =-=[6]-=-. It identifies names of three classes: people, organizations, and locations. Nymble used an HMM with a single state for each name class, plus one state for nonname tokens. To take advantage of the st...
2	et al., “An efficient repair procedure for quick transcriptions - Venkataraman (Show Context) Citation Context ... not give a gain, and some other sources did not distinguish between commas and caesura.) The boostexter model is trained on a subset of the TDT4 Chinese news data (40 shows) using flexible alignment =-=[15]-=- to obtain word transcripts from the closed-captions. For performance analysis, reference punctuation is determined by first aligning the ASR words to reference words, and then choosing the best punct...
2	et al., “Structural meta-data research - Liu - 2005 (Show Context) Citation Context ...h access to the reference transcription. 4. PUNCTUATION PREDICTION Previous work on punctuation detection to enhance the output of automatic speech recognizers mostly focuses on sentence segmentation =-=[7, 8, 9]-=-. For intra-sentence punctuation insertion for text, Beeferman et al. [10] use lexical information in the form of trigram language models. Zhang et. al. [11] use decision trees with linguistically sop...
1	et al., “CYBERPUNC: A lightweight punctuation annotation system for speech - Beeferman - 1998 (Show Context) Citation Context ...k on punctuation detection to enhance the output of automatic speech recognizers mostly focuses on sentence segmentation [7, 8, 9]. For intra-sentence punctuation insertion for text, Beeferman et al. =-=[10]-=- use lexical information in the form of trigram language models. Zhang et. al. [11] use decision trees with linguistically sophisticated features for enriching natural language generation output, and ...
1	et al., “Intra-sentence punctuation insertion in natural language generation - Zhang - 2002 (Show Context) Citation Context ...stly focuses on sentence segmentation [7, 8, 9]. For intra-sentence punctuation insertion for text, Beeferman et al. [10] use lexical information in the form of trigram language models. Zhang et. al. =-=[11]-=- use decision trees with linguistically sophisticated features for enriching natural language generation output, and obtain better results than using n-gram language models. In this work, we use the I...
1	et al., “The ICSI+ multilingual sentence segmentation system - Zimmerman - 2006 (Show Context) Citation Context ...icated features for enriching natural language generation output, and obtain better results than using n-gram language models. In this work, we use the ICSI+ multi-lingual sentence segmentation tools =-=[12]-=- for both comma and sentence boundary detection. The sentence boundary detection is treated as a classification problem, where every word boundary can be of one of two classes: sentence boundary vs. n...