HOME PAGE of PATRICK SCHONE (Last Update: 05/24/2008)

Speech Signal

SPEECH AND HUMAN LANGUAGE TECHNOLOGIES RESEARCH

The following are publications (including pre-prints):

Human Language Technology
Paper Title	Authors	Subjects	Reference	Synopsis
1996
A dictionary-based method for determining topics in text and transcribed speech	Schone, P., Nelson, D.	Topic identification	Proceedings of the 1996 IEEE International Conference on Acoustics, Speech, & Signals Processing, Atlanta, GA. Vol. 1, pp. 295-298.	We describe an algorithm which mines electronic dictionaries and applies that information to the task of tagging incoming documents for topics. Topics in this case are not some pre-described word set, but rather, they are lists of words and generalizations about the content of the text.
1997
Text Retrieval via Semantic Forests	Schone, P., Townsend, J., Crystal, T. Olano, C.	Speech retrieval	The 6th Text Retrieval Conference (TREC-6), Gaithersburg, MD. NIST Special Publication 500-240, pp. 761-773	This paper describes using topical description of documents as the indexing terms for information retrieval. The topics were generated using electronic dictionaries, which have a taxonomic structure which can be exploited similarly to the structures of ontologies. There is additional work in how the system was modified to tackle the challenge of known-item retrieval...an SDR task in TREC-6.
1998
Text Retrieval via Semantic Forests: TREC7.	Hendrickson, G., Schone, P., Crystal, T.	Speech retrieval	The 7th Text Retrieval Conference (TREC-7), Gaithersburg, MD. NIST Special Publication 500-242, pp. 583-593	This paper shows some additional improvements over our first try (TREC-6) at information retrieval. Our precision-in-top30 increased by 50% relative using semantic classes and pseudo-relevance feedback in addition to topics only.
1999
Automatically generating a topic description for text and searching and sorting text by topic using the same	Nelson, D., Schone, P., Bates, R.	Topic identification	U.S. PATENT 5,937,422	This is a patent on the above-mentioned (1996) topic algorithm.
2000
Knowledge-Free Induction of Morphology Using Latent Semantic Analysis	Schone, P., Jurafsky, D.	Morphology induction	Conference on Natural Language Learning 2000 (CoNLL-2000), Lisbon, Portugal, September 2000	In brief, we identify the morphological conflation sets for each word in English and compare our results to those of the hand-developed CELEX lexicon. As a first step, we identify plausible word suffixes by frequency and then we compare pairs of potential morphological variants (for example, "dog" to "dogs") using latent semantic analysis to see if the word pairs are semantically similar. We use this and other distributional information to make a judgment as to whether words are true conflations of each other.
2001
Mandarin-English Information (MEI): investigating translingual speech retrieval	Meng, H., Chen, B., Khudanpur, S., Levow G-A., Lo, W-K, Oard, D., Schone, P., Tang, K., Wang, H-M, Wang J.	Speech retrieval	Proceedings of the 2001 Human Language Technology (HLT) Conference, San Diego, 2001	This paper describes the results of the JHU Summer Workshop 2000 group working on Mandarin-English Information Retrieval. The task was to use English news text as query exemplars and to retrieve Broadcast News Mandarin Chinese on the same topics.
Multi-scale audio indexing for translingual spoken document retrieval	Hsin-min WANG, Helen MENG, Patrick SCHONE, Berlin CHEN and Wai-Kit LO	Speech retrieval	Proceedings of the 26th International Conference on Acoustics, Speech, and Signal Processing (ICASSP-2001), vol. 1, pp.605-608, Salt Lake City, 2001	This paper describes a subcomponent of the JHU Summer Workshop group working on Mandarin-English Information Retrieval. The particular focus was to create multiscale Chinese word recognitions and searches thereon.
Multi-scale retrieval in MEI: an English-Chinese translingual speech retrieval system	Wai-Kit LO, Patrick SCHONE and Helen MENG	Speech retrieval	Proceedings of the Seventh European Conference on Speech Communication and Technology (EUROSPEECH), vol. 2, pp.1303-1306, Aalborg, 2001	This paper describes a separate subcomponent of the JHU Summer Workshop group working on Mandarin-English Information Retrieval.
Is Knowledge-Free Induction of Multiword Unit Dictionary Headwords A Solved Problem?	Schone, P., Jurafsky, D.	Lemmatization, Multiword-Units	Empirical Methods in Natural Language Processing, Pittsburgh, PA, 2001	This paper explores the issue of identifying which n-grams from a corpus of text have the appropriate properties of a multiword unit dictionary headword. We first compare nine different collocation-finding algorithms and test these using both WordNet and a large compendium of Internet dictionaries. We then attempt to see whether Latent Semantic Analysis (LSA) can help to isolate better multiword headwords. Multiword units should be either non-compositional, non-substitutable, or non-modifiability, so we use LSA to look for the first and the second of these. We are able to make some performance gains using LSA to find non-substitutability (though LSA does not help find non-compositionals).
Knowledge-Free Induction of Inflectional Morphologies	Schone, P., Jurafsky, D.	Morphology induction	Proceedings of the North American Chapter of the Association of Computational Linguistics (NAACL-2001), Pittsburgh, PA June 2001	This work is an extension of the work described in Portugal. In particular, we look explicitly for circumfixes and prefix/affix combinations to begin with rather than just suffixes. We then incorporate induced syntactic, orthographic, and transitive properties to decrease the error over our original algorithm by 25% relative. Additionally, we show performance on German and Dutch using CELEX again as the gold standard.
Language Independent Induction of Part of Speech Class Labels Using only Language Universals	Schone, P., Jurafsky, D.	Part of speech induction	"Machine Learning: Beyond Supervision," Workshop at IJCAI-2001, Seattle, WA., August 2001	This work looks at the question of whether it is possible to induce part of speech labels for syntactic clusters. This approach starts with no lexicon, no hand-marked corpus -- nothing of this kind. Instead, we assume we start with some (perfect) syntactic clusters from which we extract a number of features: openness, boundedness, numeracy, optionality, affixation, adn cluster ordering. We then appeal to Bayesian networks and to the body of work that exists in linguistic typology and language universals to estimate which tag is most befitting for each cluster.
2002
Toward Knowledge-Free Induction of Machine-Readable Dictionaries	Schone, P.	Dissertation	University of Colorado at Boulder, December, 2001 (Copyright 2002). Advisors: Daniel S. Jurafsky and James H. Martin	The goal of this dissertation was to see how far one could go toward knowledge-free induction of an electronic dictionary. By "knowledge-free," I mean that there is no human input ... only a corpus of text. Where this is not possible, my requirement was that whatever limited human input is used, it must at least be language independent. This work involved multiword-unit induction, Chinese and phonetic segmentation, inflectional morphologies, and parts of speech. Also, the dissertation describes what was then the state of the art in automatic induction of word hierarchies.
2003
Novel Approaches to Arabic Speech Recognition: Report from the 2002 Johns-Hopkins Workshop	K. Kirchhoff, J. Bilmes, S. Das, N. Duta, M. Egan, G. Ji, F. He, J. Henderson, D. Liu, M. Noamany, P. Schone, R. Schwartz and D. Vergyri	Arabic speech-to-text	Proceedings of the International Conference on Acoustics, Speech and Signal Processing, Hong Kong, April 2003	This paper describes efforts of the Conversational Arabic speech-to-text team at the JHU Summer Workshop of 2002. The big ideas here were (1) the notion of trying to improve recognition using factored language models which incorporate various syntactics components of Arabic language like morphology; and (2) trying to induce diacritization for Arabic and determine whether or not this will aid in recognition of Arabic.
Language-reconfigurable universal phone recognition	Walker, B., Lackey, B., Muller, J., Schone, P.	Phonetic recognition	EUROSPEECH-2003, pp. 153-156, Geneva, Switzerland	This paper describes a universal phone (phonetic) recognizer for conversational telephone-quality speech. The recognizer automatically reconfigures itself to apply the strongest language model in its inventory to whatever language it is used on. We describe the system and performance measurements for it using extensive testing material both from languages in its training set as well as from a language it has never seen. The recognizer produces near-equivalent performance between the two types of data thus showing its true universality and represents a solution for processing conversational, telephone-quality speech in any language - even in low-resourced languages.
2004
Mandarin English Information (MEI): Investigating translingual speech retrieval	Helen M. MENG, Berlin CHEN, Sanjeev KHUDANPUR, Gian-Anne LEVOW, Wai-Kit LO, Douglas OARD, Patrick SCHONE, Karen TANG, Hsin-min WANG and Jianqiang WANG	Speech retrieval	Computer, Speech and Language, vol. 18, iss 2, pp. 163-179, Elsevier Press, Apr 2004	This paper is an expanded overview of the JHU 2000 Summer Workshop group working on Mandarin-English Information Retrieval. The task was to use English news text as query exemplars and to retrieve Broadcast News Mandarin Chinese on the same topics.
2005
Question Answering with QACTIS at TREC 2004	Schone, P., Ciany, G., McNamee, P., Mayfield, J., Kulman, A., Bassi, T.	Question answering	The 13th Text Retrieval Conference (TREC-13), Gaithersburg, MD. NIST Special Publication 500-261	This describes a strategy for performing automatic question answering of factoid, list, and definitional style questions using a two-way strategy. A top-down approach uses induced attributed object- relationship graphs which are mined to discover answers with graphical properties akin to those of the question. It is top-down in that "one size fits all": all questions are tackled in much the same whay, regardless of whether they are who, what, where, when, etc. questions. The second strategy is bottom-up and focuses on only a few types of questions, but it handles them each very well. It uses a cascade of filters to identify the answer word/phrase which survives after a number of filtering processes are applied. Lastly, we use Web validation to further reduce spurious answers, which is particularly beneficial for list type questions.
Searching Conversational Telephone Speech in Any of the World's Languages	Schone, P., McNamee, P., Morris, G., Ciany, G., Lewis, S.	Speech retrieval	International Conference on Intelligence Analysis. McLean, VA.	This effort uses rule-based transliteration in 90+ languages, universal phonetic recognition, and speech-to-text processing to provide speech retrieval in potentially any language. Results are provided in five languages, including one for which the system had virtually no prior training.
2006
QACTIS-based Question Answering at TREC 2005	Schone, P., Ciany, G., Cutts, R., McNamee, P., Mayfield, J., Smith, T.	Question answering	The 14th Text Retrieval Conference (TREC-14), Gaithersburg, MD. NIST Special Publication 500-261	Significant advances were made to the base system by teaching the system about word categorization.
Low-Resource Autodiacritization of Abjads for Speech Keyword Search	Schone, P.	Rule-based transliteration, speech retrieval	Interspeech 2006 -- ICSLP, Pittsburgh, PA, September 2006	This paper described an effort to learn the vowelization of abjadic languages (i.e., those that generally do not include vowels) for a context-obly-based rule-based transliteration system. Results are shown in five abjadic languages (Arabic, Farsi, Hebrew, Pashto, and Urdu) with as much as 31.2% relative improvement.
2007
QACTIS Enhancements in TREC QA 2006	Schone, P., Ciany, G., Cutts, R., McNamee, P., Mayfield, J., Smith, T.	Question Answering	The 13th Text Retrieval Conference (TREC-15), Gaithersburg, MD. NIST Special Publication 500-272	Additional improvements were made to the base system to bring up results further. In particular, interest this year was on subcategorization of class types. For example, for a question such as 'What team what the 1988 World Series?' the system would determine that team could be sports, comedy, or various other teams, but 'world series' is a reference to baseball. So only a baseball team could satisfy the question.
2008
Learning Named Entity Hyponyms for Question Answering	P. McNamee, R. Snow, P. Schone and J. Mayfield	Question answering, Hyponym induction	Third International Joint Conference on Natural Language Processing, Hyderabad, India, January 2008	Since hyponym dictionaries have a major effect on the performance of question-answering system, this system shows how induction of hyponyms can be applied to the same question answering and the results thereof.
Mining Wiki Resources for Multilingual Named Entity Recognition	Richman, A., Schone, P.	Content Extraction	ACL-2008, Columbus, OH	This paper provides a process whereby one can mine freely-accessible Wikipedia pages as a mechanism of automatically obtaining content-tagged text in potentially any language for the prupose of training a statistical content extractor.

Return to Main Page