Báo cáo khoa học: "Beyond Log-Linear Models: Boosted Minimum Error Rate Training for N-best Re-ranking" docx

... 37–40, Columbus, Ohio, USA, June 2008. c 2008 Association for Computational Linguistics Beyond Log-Linear Models: Boosted Minimum Error Rate Training for N-best Re-ranking Kevin Duh ∗ Dept. of Electrical ... algorithms for machine translation rely on log-linear models, which have the potential problem of underfitting the training data. We present BoostedMERT, a novel...

Ngày tải lên: 23/03/2014, 17:20

... read- ability of machine generated summaries. We also introduce two new feature sources to enhance the model with hierarchical and Explicit/Non-Explicit information, and demonstrate that they improve ... are also evaluated for their ability to assess sum- mary readability, i.e., to measure how linguistically readable a machine generated summary is. Sub- mitted metrics that perform consisten...

Ngày tải lên: 07/03/2014, 18:20

... that integrates local and global coherence concerns. Summary sentences are grouped before ordering is applied on two levels: group-level and sentence-level. Different algorithms for grouping ... explicitly stated that sentence ordering for summarization is primarily driven by coherence. For example, Barzilay et al. (2002) use lexical cohesion information to model local coherenc...

Ngày tải lên: 23/03/2014, 16:20

... we have also demonstrated perplexity re- ductions of 20% and statistically significant im- provement in speech recognition error rate. Fur- thermore, each method has generated some dis- cussion ... convergent iterative reestimation process for p(glc), p(YlC ) and p(C). These distributions form the model for the given/3. It is easy to see that for/ 3 = 0, p(nlc ) does not...

Ngày tải lên: 20/02/2014, 18:20

... concept. Phrasets can provide useful information for different kind of NLP tasks, both in a monolin- gual and multilingual environment. For instance, phrasets can be useful for knowledge-based word alignment ... WordNet is con- stantly updated and extended with different kinds of information such as domain information, syn- tactic information, topic signatures, syntactic parsing and PoS...

Ngày tải lên: 22/02/2014, 02:20

... different stack contents. 5.2 Training Method We randomly selected roughly 90% of the dialogs for training, and used the remainder for testing. We separately trained models for: user dia- log act classication ... to improved performance for any method. One utterance of context is best for shift-reduce and start-join; three is best for the connection path method. The shift- redu...

Ngày tải lên: 22/02/2014, 02:20

... again form a lattice and we can calculate the best path for layer 2. The Markov Model for layer 1 operates on the output of the Markov Model for part-of-speech tagging, the model for layer ... off the treebank. Training on annotated data is straight forward. First, we number the layers, starting with 0 for the part-of-speech layer. Subsequently, informa- tion for the dif...

Ngày tải lên: 22/02/2014, 03:20

... use E M for infer- ence can be very sensitive to the number of latent variables chosen. For example, the performance of ROOTH-EM worsens quickly if the number of clusters is overestimated; for the ... indicates that it does not perform well in this context. The most likely explanation is that LinkLDA generates its two arguments in- dependently, which may be suitable for distinct argum...

Ngày tải lên: 07/03/2014, 22:20

... the probability of the current word. We henceforth call the previous n − 1 words plus the current word as forward n-grams and a language model built 1288 on forward n-grams as forward n-gram language model. ... (w i |w m i+1 ) ≈ m  i=1 P (w i |w i+n−1 i+1 ) (2) 3.1 Training For the convenience of training, we invert the or- der in each sentence in the training data, i.e., from the ori...

Ngày tải lên: 07/03/2014, 22:20

... process illustrates how to generate a docu- ment d in pLSI: 1. Pick a topic mixture distribution (|). 2. For each word w i in d a. Pick a latent topic z with the probabil- ity (|) for w i ... “topics”. To further improve efficiency, we also perform preprocess- ing (refer to Section 3.4 for details) before build- ing topic models for C R (q), where some low- frequency items...

Ngày tải lên: 08/03/2014, 00:20