The AFRL IWSLT 2018 Systems: What Worked, What Didn’t

The AFRL IWSLT 2018 Systems: What Worked, What Didn’t Brian Ore, Eric Hansen, Katherine Young, Grant Erdmann and Jeremy Gwinnup Air Force Research Laboratory {brian.ore.1, eric.hansen.5, katherine.young.1.ctr, grant.erdmann, jeremy.gwinnup.1}@us.af.mil Abstract This report summarizes the Air Force Research Labora- tory (AFRL) machine translation (MT) and automatic speech recognition (ASR) systems submitted to the spoken language translation (SLT) and low-resource MT tasks as part of the IWSLT18 evaluation campaign. 1. Introduction As part of the evaluation campaign for the 2018 International Workshop on Spoken Language Translation (IWSLT18) [1], the AFRL Human Language Technology team applied and improved techniques from previous workshops [2] and Con- ference on Machine Translation efforts [3] to the Spoken Lan- guage Translation and Low-Resource Machine Translation tasks. 2. Spoken Language Translation 2.1. Automatic Speech Recognition This section describes the ASR systems that were developed for the baseline condition of the Speech Translation task. We trained two different English systems and performed system combination to obtain the final hypothesis for translation. Section 2.1.1 describes that language models (LMs) that were used for decoding and rescoring. Section 2.1.2 discusses the Kaldi ASR system, and Section 2.1.3 describes the Hidden Markov Model ToolKit (HTK) Tensorflow system. Finally, Section 2.1.4 describes how we segmented the test data and performed system combination. 2.1.1. Language Models LMs were estimated on the provided TED data and subsets of News Crawl 2007-2017 and News Discussions versions 1- 3. The subset of each news corpus was selected using cross- entropy difference scoring [4] with TED as the in-domain text, and selection thresholds were chosen to use 1/8 of each corpus to train N-gram LMs, and 1/16 of each corpus to train a recurrent neural network (RNN) LM. Interpolated bigram, trigram, and 4-gram LMs were estimated using the SRILM Toolkit,1 and a RNN maximum entropy LM was trained using the RNNLM Toolkit.2 The RNN included 160 hidden units, 1http://www.speech.sri.com/projects/srilm 2http://www.fit.vutbr.cz/$\sim$imikolov/rnnlm Table 1: Kaldi WER. Decoding was performed using a tri- gram LM trained on TED. Acoustic Training Data dev2010 tst2010 tst2013 Speech-Translation TED 19.8 19.6 30.5 TEDLIUM 16.9 14.8 22.3 Combined 16.6 15.1 23.6 300 classes in the output layer, 4-gram features for the di- rect connections, and a hash size of 109. The LM vocabulary included 100,000 words that were chosen using the select- vocab tool from SRILM. 2.1.2. Kaldi System The acoustic training data available for this year’s evalua- tion included the Speech-Translation TED corpus and the TEDLIUM corpus. Based on a preliminary analysis of the Speech-Translation TED corpus, we removed all segments longer than 15 seconds from this corpus. The devtest and off- limit talks were sequestered from TEDLIUM, and a third data set was created by searching the Speech-Translation TED and TEDLIUM corpora for non-overlapping time segments. Next, an initial set of ASR systems were trained on each data set using the Kaldi open source speech recognition toolkit [5]. All Kaldi models discussed in this paper are based on the chain time delay neural network (TDNN)-rectified linear unit (ReLU) setup using i-vectors.3 Standard data …

Content truncated to 500 words