R. Schwartz, T. Imai, F. Kubala, L. Nguyen, and J. Makhoul. A maximum likelihood model for topic classification of broadcast news. In Proceeding of the European Conference On Speech Communication and Technology, Rhodes, Greece, 1997.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....can be realized using topic identification (TID) which is our aim in this paper. The main objective of TID is to assign one or several topic labels to a flow of textual data. Labels are chosen from a set of topics fixed a priori. Several approaches have already been proposed in the literature [5, 8] which are generally based on a specific metric. This paper deals with the evaluation problem of TID algorithms on two kinds of textual corpora: newspaper and e mails. A comparison study of several TID methods is presented. In the literature, another comparison has been performed by Yang and ....

....of them share the same vocabulary. In what follows, a sequence to be classified is made up of the first words of a document. Each document is associated with one single label among predefined topics. 3. 1 Topic unigram language model The topic unigram language model [5, 8] is one of the most classical and standard ones. It is based on a counting of the number of occurrences of each word for each topic and involves all words of each topic vocabulary. The posterior probability is expressed as: ....

R. Schwartz, T. Imai, F. Kubala, L. Nguyen, and J. Makhoul. A maximum likelihood model for topic classification of broadcast news. In Proceeding of the European Conference On Speech Communication and Technology, Rhodes, Greece, 1997.

....list. Each topic is modeled with a discrete single state HMM using the word index numbers as observations. The vocabulary is extracted from the test and training set and from the vocabulary of the speech recognizer in such a way that all words can be assigned an index. This system is motivated by [11]. BeginAVcutN topic N AVcut AVcut Weather forecast NAVCutEnd N topic R topic upto20 times N topic= AVcut R ed eff N AVcutN Winchange AVcut Winchange AVcut [Winchange]InterviewWinchange R topic= R ed eff R ed eff= Cut AVcut LRwipe Dissolve [Winchange]InterviewWinchange N ....

Richard Schwartz, Toru Imai, Francis Kubala, Long Nguyen, and John Makhoul, "A maximum likelihood model for topic classification of broadcast news," in Proceedings of Eurospeech '97, Rhodes, Greece, Sept. 1997, pp. 1455--1458.

....can be realized using topic identification (TID) which is our aim in this paper. The main objective of TID is to assign one or several topic labels to a flow of textual data. Labels are chosen from a set of topics fixed a priori. Several approaches have already been proposed in the literature [5, 8] which are generally based on a specific metric. This paper deals with the evaluation problem of TID algorithms on two kinds of textual corpora: newspaper and e mails. A comparison study of several TID methods is presented. In the literature, another comparison has been performed by Yang and ....

....of them share the same vocabulary. In what follows, a sequence W N 1 = w 1 ; wN to be classified is made up of the first N words of a document. Each document is associated with one single label among J predefined topics. 3. 1 Topic unigram language model The topic unigram language model [5, 8] is one of the most classical and standard ones. It is based on a counting of the number of occurrences of each word for each topic and involves all words of each topic vocabulary. The posterior probability P (T j j W N 1 ) is expressed as: P (T j j W N 1 ) P (T j )P (W N 1 j T j ) P J ....

R. Schwartz, T. Imai, F. Kubala, L. Nguyen, and J. Makhoul. A maximum likelihood model for topic classification of broadcast news. In Proceeding of the European Conference On Speech Communication and Technology, Rhodes, Greece, 1997.

.... and the benchmark evaluations[1, 6] An increasing number of information retrieval and machine learning techniques have been applied, including k Nearest Neighbor (kNN) classification, Decision Tree induction, a variety of Language Modeling (LM) approaches and relevance based filtering methods[2, 24, 22, 18, 15, 16], and systematic analysis of their behavior on the event tracking task have just begun. In this paper, we report our new research findings with kNN and Rocchio in event tracking. Our KNN methods were among the two top performing tracking systems in the official TDT1 and TDT3 evaluations, ....

R. Schwartz, T. Imai, L. Nguyen, and J. Makhoul. A maximum likelihood model for topic classification of broadcast news. In Proceedings of Eurospeech, Rhodes, Greece, 1997.

....the somewhat narrow topic Labor Unions should go under both Economics and Politics . Topic Categorization Model There are many methods for determining the topics in a document. The OnTopic TM system at BBN uses a Hidden Markov Model (HMM) to model multiple topics in documents explicitly [1]. The model is pictured in Figure 1. We assume (make believe) that the story is generated by this model. According to this model, when an author decides to write a story, the first thing he does is pick a set of topics. The topics are chosen according to the prior distribution for topics. There is ....

R. Schwartz, T. Imai, F. Kubala, L. Nguyen, J. Makhoul, "A Maximum Likelihood Model for Topic Classification of Broadcast News", Eurospeech-97, Rhodes, Greece, September, 1999.

....offer a formal way of expressing computed quantities. A useful set of metrics for topic detection is the class of metrics that calculate P #CjS#.Weshallan alyze one particular example of such a metric, the BBN topic spotting metric. The BBN topic spotting metric is derived from Bayes Rule [4]: p#CjS#= p#C##p#SjC# p#S# ; 1) where p#C# is the aprioriprobability that any new story will be relevant to cluster C. If we assume that the story words sn are conditionally independent, we get: p#CjS# # p#C# # Y n p#sn jC# p#sn# ; 2) where p#sn jC# is the probability that a word in a ....

.... n jC# #1,###p#s n # (3) English Topic p(GE) General story words p(S C) Figure 3: BBN topic spotting metric two state model for a topic The estimates for the general English state distribution and topic state distributions can be refined using the ExpectationMaximization (EM) algorithm [4]. This process allows new words to be added to the distributions and emphasizes topic specific words. Therefore, the EM algorithm automatically assigns higher probabilities to words that are specific to the topic. 4. Clustering Metrics There are two types of metrics that are useful for the ....

R. Schwartz, T. Imai, L. Nguyen, and J. Makhoul, "A Maximum Likelihood Model for Topic Classification of Broadcast News," In Proc. Eurospeech, Rhodes, Greece, September 1997.

....task. 1 Introduction Hidden Markovmodelshave been applied successfully over the last two decades in a wide variety of speechand language related recognition problems including speech recognition [9] named entity finding [2] optical character recognition [10] and topic identification [19]. In the present work, we describe an application of this technology to the problem of ad hoc information retrieval. In all HMM applications, the observed data (e.g. audio recording, image bitmap) is modeled as being the output produced by passing some unknown key (e.g. words, letters) through a ....

R. Schwartz, T. Imai, F. Kubala, L. Nguyen, J. Makhoul, "A maximum likelihood model for topic classification of broadcast news." Proc. Eurospeech '97, Rhodes, Greece, pp. 1455-1458 (1997).

....on TREC 6. 1 Introduction Hidden Markov models have been applied successfully over the last two decades in a wide variety of speech and language related recognition problems including speech recognition [8] named entity finding [2] optical character recognition [9] and topic identification [18]. For TREC 7, we applied this technology for the first time to the problem of ad hoc information retrieval. On the TREC 7 ad hoc task, our entry ranked among the top tier of systems in average non interpolated precison [22] Moreover, our strong development results on TREC 6 hold out the promise ....

R. Schwartz, T. Imai, F. Kubala, L. Nguyen, J. Makhoul, "A maximum likelihood model for topic classification of broadcast news." Proc. Eurospeech '97, Rhodes, Greece, pp. 1455-1458 (1997).

No context found.

Schwartz R. et al. : "A Maximum Likelihood Model for Topic Classification of Broadcast News, " in Proceedings of Eurospeech Conference, 1997.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC