{ "paper_id": "2006", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:20:19.305517Z" }, "title": "Overview of the IWSLT 2006 Evaluation Campaign", "authors": [ { "first": "Michael", "middle": [], "last": "Paul", "suffix": "", "affiliation": { "laboratory": "ATR Spoken Language Communication Research Labs", "institution": "Keihanna Science City", "location": { "addrLine": "Hikaridai 2-2-2", "postCode": "619-0288", "settlement": "Kyoto" } }, "email": "michael.paul@atr.jp" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper gives an overview of the evaluation campaign results of the International Workshop on Spoken Language Translation (IWSLT) 2006 1. In this workshop, we focused on the translation of spontaneous speech. The translation directions were Arabic, Chinese, Italian, or Japanese into English. In total, 21 translation systems from 19 research groups participated in this year's evaluation campaign. Both automatic and subjective evaluations were carried out in order to investigate the impact of spontaneity aspects on automatic speech recognition (ASR) and machine translation (MT) system performance as well as the robustness of stateof-the-art MT systems towards speech recognition errors. (' * ' indicates late run submissions that were submitted after the official submission period) ASR Output MT Correct Recognition Result official evaluation additional evaluation Engine official evaluation additional evaluation BLEU4 NIST METEOR BLEU4 NIST METEOR BLEU4 NIST METEOR BLEU4 NIST METEOR", "pdf_parse": { "paper_id": "2006", "_pdf_hash": "", "abstract": [ { "text": "This paper gives an overview of the evaluation campaign results of the International Workshop on Spoken Language Translation (IWSLT) 2006 1. In this workshop, we focused on the translation of spontaneous speech. The translation directions were Arabic, Chinese, Italian, or Japanese into English. In total, 21 translation systems from 19 research groups participated in this year's evaluation campaign. Both automatic and subjective evaluations were carried out in order to investigate the impact of spontaneity aspects on automatic speech recognition (ASR) and machine translation (MT) system performance as well as the robustness of stateof-the-art MT systems towards speech recognition errors. (' * ' indicates late run submissions that were submitted after the official submission period) ASR Output MT Correct Recognition Result official evaluation additional evaluation Engine official evaluation additional evaluation BLEU4 NIST METEOR BLEU4 NIST METEOR BLEU4 NIST METEOR BLEU4 NIST METEOR", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The International Workshop on Spoken Language Translation (IWSLT) is an evaluation campaign organized by the Consortium for Speech Translation Advanced Research (C-STAR) 2 , that provides a common framework to compare and improve current state-of-the-art speech-to-speech translation technologies. Previous IWSLT workshops focused on the establishment of evaluation metrics for multilingual speech-to-speech translation [1] and the translation of automatic speech recognition results from read-speech input [2] .", "cite_spans": [ { "start": 170, "end": 171, "text": "2", "ref_id": "BIBREF1" }, { "start": 420, "end": 423, "text": "[1]", "ref_id": "BIBREF0" }, { "start": 507, "end": 510, "text": "[2]", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "The focus of this year's IWSLT was the translation of spontaneous-speech input. The evaluation campaign was carried out using a multilingual spoken language corpus including Arabic, Chinese, Italian, Japanese, and English sentences from the travel domain. The input to the machine translation (MT) engines was either the output of an automatic speech recognition (ASR) system applied to spontaneous-speech and read-speech input or the correct recognition result (CRR). The translation was carried out from Arabic, Chinese, Italian, or Japanese into English.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "Participants were supplied with in-domain resources, but were free to use additional resources as well. Depending on the amount of in-domain training data, two different data tracks (OPEN, CSTAR) were distinguished. In total, 21 MT systems from 19 research groups participated in this year's evaluation campaign. A total of 73 MT engines were built to cover different combinations of language pairs and data tracks. The translation quality of all official run submissions was evaluated using automatic evaluation metrics. In addition, human assessments were carried out for the most popular track, i.e., the translation of Chinese ASR output into English. Based on the evaluation results, the impact of the spontaneity aspects of speech on the ASR and MT systems performance as well as the robustness of state-of-the-art MT systems towards speech recognition errors were investigated.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "The IWSLT 2006 evaluation campaign was carried out using a multilingual spoken language corpus. The Basic Travel Expression Corpus (BTEC ) contains tourism-related sentences similar to those that are usually found in phrase books for tourists going abroad [3] . Parts of this corpus were already used in previous IWSLT evaluation campaigns [1, 2] . In addition to the sentence-aligned training corpus, the evaluation data sets of previous workshops including multiple reference translations were provided to the participants as a development corpus.", "cite_spans": [ { "start": 4, "end": 14, "text": "IWSLT 2006", "ref_id": null }, { "start": 256, "end": 259, "text": "[3]", "ref_id": "BIBREF2" }, { "start": 340, "end": 343, "text": "[1,", "ref_id": "BIBREF0" }, { "start": 344, "end": 346, "text": "2]", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "IWSLT 2006 Spoken Language Corpus", "sec_num": "2.1." }, { "text": "The evaluation data set of IWSLT 2006 consisted of spontaneous answers to questions in the tourism domain. This \"Challenge Task 2006\" differed greatly from the translation tasks of previous workshops. In addition to the spontaneous speech data, read-speech recordings of the cleaned transcripts were also used for evaluation purposes. ASR engines provided by the C-STAR partners were applied to the speech input and produced word lattices from which NBEST/1BEST lists were extracted automatically using publicly available tools. Word segmentations according to the output of the ASR engines were also provided for all supplied resources.", "cite_spans": [ { "start": 27, "end": 37, "text": "IWSLT 2006", "ref_id": null }, { "start": 123, "end": 133, "text": "Task 2006\"", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "IWSLT 2006 Spoken Language Corpus", "sec_num": "2.1." }, { "text": "For this year's evaluation campaign, parts of the Arabic (A), Chinese (C), Italian (I), Japanese (J), and English (E) subsets of the BTEC corpus were used. The participants were supplied with a training corpus of 40K sentence pairs for CE/JE, and 20K sentence pairs for AE/IE and three development data sets (dev1, dev2, dev3; 500 sentences each) consisting of the evaluation data sets of previous IWSLT evaluation campaigns including up to 16 English reference translations for evaluation purposes. Table 1: The IWSLT 2006 spoken language corpus type lang sentence count avg. word word uage total unique length tokens types training C/E 39,953 37,559 / 39,633 8.6 / 9.2 342,362 / 367,265 11,174 / 7,225 J/E 39,953 37,173 / 39,633 10.0 / 9.2 398,498 / 367,265 11,407 / 7,225 A/E 19,972 19,777 / 19,880 Details of the IWSLT 2006 spoken language corpus are given in Table 1 . The total sentence counts show the number of bilingual sentence pairs and the unique sentence counts refer to the number of unique monolingual sentences. The average length column shows the average number of words per training sentence where the word segmentation for the source language was the one given by the output of the ASR engines. The English target sentences were tokenized according to the evaluation specifications used for this year's evaluation campaign. Word token refers to the number of words in the corpus and word type refers to the vocabulary size.", "cite_spans": [], "ref_spans": [ { "start": 500, "end": 818, "text": "Table 1: The IWSLT 2006 spoken language corpus type lang sentence count avg. word word uage total unique length tokens types training C/E 39,953 37,559 / 39,633 8.6 / 9.2 342,362 / 367,265 11,174 / 7,225 J/E 39,953 37,173 / 39,633 10.0 / 9.2 398,498 / 367,265 11,407 / 7,225 A/E 19,972 19,777 / 19,880", "ref_id": "TABREF1" }, { "start": 881, "end": 888, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "Supplied Resources", "sec_num": "2.1.1." }, { "text": "In order to obtain speech input with a certain level of spontaneity, question/answer conversations between Chinese speakers were recorded by the C-STAR partners. In the preparation phase, around 1000 questions were extracted manually from the original BTEC corpus, avoiding redundancy and an attempt was made to maximize the diversity of the topics addressed. In addition, answer keys, i.e. short phrases providing hints on the answer contents, were added to each question. For recording, the questions were split into 20 subsets and pairs of native Chinese speakers 3 were asked to carryout a \"one-turn\" role play. A brief scene description (outline of the role-play) was given to both speakers. Speaker SQ obtained a list of questions and asked one question after the other. Speaker SA obtained a list of answer keys and answered to each question using the following guidelines:", "cite_spans": [ { "start": 567, "end": 568, "text": "3", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Challenge Task 2006", "sec_num": "2.1.2." }, { "text": "\u2022 answer in a natural way based on the answer keys Examples of questions, answer keys, and recorded answers are given in Table 2 . The obtained Challenge Task 2006 data sets were split into two subsets: dev4 (489 sentences, development corpus) and eval (500 sentences, evaluation corpus). The difficulty of this year's evaluation data set is illustrated in Table 3 . It lists the target language perplexity of all translation tasks according to the supplied resources of IWSLT 2006. Compared to last year's evaluation data sets, the language perplexities of dev4 and eval were three times higher. In addition to the Chinese spontaneous-speech recordings, read-speech recordings of the Challenge Task 2006 were produced for all source languages. The cleaned transcriptions of the Chinese spontaneous-speech recordings were translated into English, Japanese, Arabic, and Italian by human translators. For English, two native speakers produced three alternative translations each resulting in a total of seven reference translations for the dev4 and eval data set, respectively. The source language texts were read aloud by 20 native speakers of the respective source language 4 and recognition results were obtained using ASR engines provided by the C-STAR partners. Table 4 summarizes the out-of-vocabulary (OOV) rates of the obtained data sets. The OOV rates are listed for all source languages and input conditions (CRR, 1BEST, NBEST) and for the English reference translations using the 20K/40K training corpus. In general, the OOV rates of CRR are higher than the OOV rates of the 1BEST data sets, because unknown words might either be ignored or mis-recognized as known words by the ASR engine. For NBEST lists, OOV rates are naturally higher than those of the 1BEST data sets. 2.7 / 1.9 Cs: spontaneous speech, Cr: read speech The lowest OOV rates for the CRR data are found for Japanese and Chinese (1.2-2.6%). The figures for Italian are twice as high. However, very large OOV rates of 13-17% are obtained for Arabic which are caused mainly by word segmentation issues (prefix/postfix attachment) and spelling variations in Arabic. The spontaneous speech data sets have slightly lower OOV rates than the read speech data.", "cite_spans": [ { "start": 471, "end": 482, "text": "IWSLT 2006.", "ref_id": null }, { "start": 1174, "end": 1175, "text": "4", "ref_id": "BIBREF3" } ], "ref_spans": [ { "start": 121, "end": 128, "text": "Table 2", "ref_id": "TABREF1" }, { "start": 357, "end": 364, "text": "Table 3", "ref_id": "TABREF2" }, { "start": 1265, "end": 1272, "text": "Table 4", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Challenge Task 2006", "sec_num": "2.1.2." }, { "text": "The recognition accuracies of the utilized ASR engines for the Challenge Task 2006 data sets are summarized in Table 5. The lattice accuracy figures show the percentage of correct recognition results contained in the lattices, where the 1BEST accuracy is the accuracy of the best path 5 extracted from each lattice. Besides for Italian, the word accuracies of the read-speech recordings ranged between 82%-90% (lattice) and 74%-85% (1BEST), where the percentages of correctly recognized sentences (sentence accuracy) ranged between 30%-50% (lattice) and 20%-40% (1BEST). However, a large difference can be seen between the different source languages. The lattice accuracies of Chinese were 5%-8% lower than those obtained for Japanese and Arabic. For Chinese and Arabic, a large drop in recognition performance was seen when comparing lattice and 1BEST accuracies.", "cite_spans": [ { "start": 63, "end": 82, "text": "Challenge Task 2006", "ref_id": null }, { "start": 285, "end": 286, "text": "5", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Challenge Task 2006", "sec_num": "2.1.2." }, { "text": "Concerning different speech types, a drop in recognition performance of 3%-6% in word accuracy and 5%-8% in sentence accuracy was seen for the spontaneous-speech data compared to the read-speech results.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Challenge Task 2006", "sec_num": "2.1.2." }, { "text": "In order to investigate the effects of recognition errors on the MT performance, the participants were asked to translate two types of input using the same MT engine:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Input Conditions", "sec_num": "2.2." }, { "text": "1. speech input (wave forms) or ASR output (lattices, NBEST/1BEST lists) 2. correct recognition results (plain text)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Input Conditions", "sec_num": "2.2." }, { "text": "The translation of the correct recognition results was mandatory for all participants. For the ASR output, most of the participants applied their MT engines to the 1BEST recognition results. Three research groups reported a gain in translation performance by translating NBEST lists and combining the obtained translation hypotheses. In addition, three groups exploited the ASR lattices directly to obtain its translation results. Concerning the speech input, the participants were allowed to use their own ASR engine, however none of the participants took this opportunity.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Input Conditions", "sec_num": "2.2." }, { "text": "For training purpose, the spoken language corpus described in Section 2.1 was provided to all participating research groups. In addition, the participants were free to use additional resources 6 as well.", "cite_spans": [ { "start": 193, "end": 194, "text": "6", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Data Track Conditions", "sec_num": "2.3." }, { "text": "The past IWSLT workshop results have shown that the amount of BTEC sentence pairs used for training has a dominant effect on the performance of the MT systems on the given task. However, only C-STAR partners have access to the full BTEC corpus 7 consisting of 172K sentence pairs. In order to allow a fair comparison between the systems, the following two data tracks were distinguished:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Track Conditions", "sec_num": "2.3." }, { "text": "\u2022 Open Data Track:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Track Conditions", "sec_num": "2.3." }, { "text": "Any resources, except for the full BTEC corpus and proprietary data, can be used as the training data for the MT engines. Concerning the BTEC and proprietary data, only the Supplied Resources (see Section 2.1.1) were allowed to be used for training purposes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Track Conditions", "sec_num": "2.3." }, { "text": "\u2022 C-STAR Data Track: Any resources (including the full BTEC corpus and proprietary data) can be used as the training data of MT engines.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Track Conditions", "sec_num": "2.3." }, { "text": "The supplied resources of IWSLT 2006 were released three months ahead of the official run submissions. The organizers also set-up an online evaluation server that could be used to evaluate system performance on the provided development data sets using automatic scoring metrics (see Section 2.5.1).", "cite_spans": [ { "start": 26, "end": 36, "text": "IWSLT 2006", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Run Submissions", "sec_num": "2.4." }, { "text": "The official run submission period was limited to three days during which the automatic scoring result feedback to the participant via email was made unavailable in order to avoid any system tuning towards the eval data. The schedule of the evaluation campaign is summarized in Table 6 . In total, 19 research groups took part in this year's evaluation campaign and two groups registered multiple translation systems. Information on the organisations and the utilized translation systems is summarized in Appendix A. Most participants used statistical machine translation (SMT) systems. In addition, example-based MT (EBMT) systems, rule-based MT (RBMT) systems and hybrid approaches combining multiple MT engines were also exploited. Five of the MT systems were applied to all input conditions. Each participant submitted at least one run. In total, 73 official and 83 contrastive runs were submitted for the eval. The distribution of run submissions for the respective data track/input condition is summarized in Table 7 . After the official run submission period, the participants still had access to the evaluation server and in order to do additional experiments.", "cite_spans": [], "ref_spans": [ { "start": 278, "end": 285, "text": "Table 6", "ref_id": "TABREF5" }, { "start": 1015, "end": 1022, "text": "Table 7", "ref_id": "TABREF6" } ], "eq_spans": [], "section": "Run Submissions", "sec_num": "2.4." }, { "text": "In order to deliver more usable translations, both for reading and for listening, and to make the IWSLT evaluation 7 http://cstar.atr.jp/cstar-corpus campaign results more comparable to outcomes of other MT evaluation workshops like those organized by NIST 8 or TC-STAR 9 , the official evaluation specifications of this year's IWSLT were defined as:", "cite_spans": [ { "start": 115, "end": 116, "text": "7", "ref_id": "BIBREF6" }, { "start": 257, "end": 258, "text": "8", "ref_id": "BIBREF7" }, { "start": 270, "end": 271, "text": "9", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "\u2022 case-sensitive", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "\u2022 with punctuation marks ('.' ',' '?' '!' '\"') tokenized", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "However, in order to be able to compare this year's IWSLT results to the outcomes of previous IWSLT workshops, the evaluation specifications of last year were also applied as an additional evaluation:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "\u2022 case-insensitive (lower-case only)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "\u2022 no punctuation marks (remove '.' ',' '?' '!' '\"')", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "\u2022 no word compounds (replace hyphens '-' with space)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "The focus of this year's evaluation campaign was the translation of speech data. Therefore, all input data files (see Section 2.2) were case-insensitive and without punctuation information. However, true-case and punctuation information was provided for all training data sets that could be used for recovering case/punctuation information according to the official evaluation specifications. Instructions 10 on how to build a baseline tool for case/punctuation insertions using the SRI Language Modeling Toolkit was provided to all participants.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "2.5." }, { "text": "The automatic evaluation of run submissions was carried out using an online evaluation server. The participants had to upload two translation files (see Section 2.2). Text preprocessing was carried out automatically according to the evaluation specification described above. For the official evaluation, an English tokenizer tool, that was made available to all participants, was applied. For the additional evaluation all punctuation marks were removed and the text was converted to lower-case. For development purposes, the participants had access to the online evaluation server of the dev4 data set three weeks before the eval run submission period. For the official evaluation results 11 of the IWSLT 2006 workshop, we utilized the following three metrics: ", "cite_spans": [ { "start": 690, "end": 692, "text": "11", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Automatic Evaluation", "sec_num": "2.5.1." }, { "text": "Human assessments of translation quality were carried out with respect to the fluency and adequacy of the translation. Fluency indicates how the evaluation segment sounds to a native speaker of English. For adequacy, the evaluator was presented with the source language input as well as a \"gold standard\" translation and has to judge how much of the information from the original translation is expressed in the translation. The fluency and adequacy judgments consist of one of the grades listed in Table 9 . The subjective evaluation was carried out only for the most popular track, i.e., the translation of Chinese ASR output into English. In order to compare different translation input conditions (CE spont, CE read, CE CRR), 7 MT systems that were applied to all input conditions were selected according to the automatic scoring results. In total, 21 run submissions were evaluated by humans. The human assessment was limited to the translation outputs of 400 input sentences selected randomly from the eval data. In order to reduce the costs further, all translation results were pooled, i.e., in case of identical translations of the same source sentence by multiple MT engines, the translation was graded only once, and the respective rank was assigned to all MT engines with the same output.", "cite_spans": [], "ref_spans": [ { "start": 499, "end": 506, "text": "Table 9", "ref_id": "TABREF8" } ], "eq_spans": [], "section": "Subjective Evaluation", "sec_num": "2.5.2." }, { "text": "Each translation of a single MT engine was evaluated by three judges where each system score is calculated as the median of the assigned grades. For fluency, only native speakers of English were used. The adequcay evaluation was carried out by native speakers and non-native speakers with sufficient knowledge of English. In total, 12 English native speakers and 11 non-native speakers were involved in this year's evaluation task. A total of 38,198 grading operations were performed.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Subjective Evaluation", "sec_num": "2.5.2." }, { "text": "The evaluation results of the IWSLT 2006 workshop are summarized in Appendix B (human assessment) and Appendix C (automatic evaluation). For each translation condition/evaluation metric, the best score is marked in bold-face.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Results", "sec_num": "3." }, { "text": "Based on the obtained evaluation results, the respective MT engines were ranked. In order to decide whether the translation output of one MT engine is significantly better than another one, we used the bootStrap 12 method that (1) performs a random sampling with replacement from the eval data set, (2) calculates the respective evaluation metric score of each engine for the sampled test sentences and the difference between the two MT system scores, (3) repeats the sampling/scoring step iteratively 13 , and (4) applies the Student's t-test at a significance level of 95% confidence to test whether the score differences are significant [9] . In this paper, we omit a horizontal line between two MT engines in the MT engine ranking tables, if the system performances do not differ significantly according to the bootStrap method.", "cite_spans": [ { "start": 502, "end": 504, "text": "13", "ref_id": "BIBREF12" }, { "start": 640, "end": 643, "text": "[9]", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Evaluation Results", "sec_num": "3." }, { "text": "Each sentence was evaluated by three human judges. Due to different levels of experience and background of the evaluators, variations in judgments were to be expected. The grader consistency is investigated in Section 3.1.1.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Subjective Evaluation Results", "sec_num": "3.1." }, { "text": "The subjective evaluation results of the MT outputs for the CE translation tasks are summarized in Appendix B.1. where the MT engines are in descending order with respect to the adequacy score. Some general findings are given in Section 3.1.2.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Subjective Evaluation Results", "sec_num": "3.1." }, { "text": "In order to investigate the degree of grading consistency between the human evaluators, the Kappa statistics for the agreement of fluency and adequacy ratings were calculated. Only low agreement levels (fluency: 0.24, adequacy: 0.31) were obtained for both metrics. In addition, the average grading difference between two graders was 0.80 points for fluency and 0.68 points for adequacy.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Grader Consistency", "sec_num": "3.1.1." }, { "text": "In order to check the self-consistency of subjective evaluations, each grader had to evaluate a set of 100 sentences a second time. Based on these grades, the average difference between the first and second grade (fluency: 0.50, adequacy: 0.40) and the probability that the grader will assign a different grade (fluency: 0.44, adequacy: 0.39) were calculated.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Grader Consistency", "sec_num": "3.1.1." }, { "text": "The grader consistency figures are slightly worse than those obtained in the previous IWSLT workshops, which might be partly caused by the lower translation quality of the MT outputs. In order to minimize the impact of grader inconsistencies, the median of the three assigned grades was selected as the final judgment for each sentence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Grader Consistency", "sec_num": "3.1.1." }, { "text": "The highest fluency and adequacy scores were obtained for the translation of the correct recognition results (1.67 for adequacy, 1.59 for fluency). Speech recognition errors for read speech input led to a drop in MT performance of 0.33-0.47 points for adequacy and 0.12-0.35 points for fluency. This indicates that recognition errors affected primarily the information content of the translation output. Moreover, only minor degradations in both metrics can be seen when comparing read-speech with spontaneous speech results.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "System Performance", "sec_num": "3.1.2." }, { "text": "However, the degree of degradation varies between MT engines. The smallest drop in performance can be seen for the JHU WS06 system [16] . Although it does not achieve the best performance on the CRR task, it seems to be quite robust against recognition errors. One reason might be that it does not restrict its input to 1BEST ASR outputs, instead it uses information from the word lattice to overcome recognition problems. In contrast, the MIT-LL AFRL system [18] achieved the highest adequacy score on the CRR task, but performance became worse on the CE spont task. Curiously, its fluency score for spontaneous speech is higher than its read speech score.", "cite_spans": [ { "start": 131, "end": 135, "text": "[16]", "ref_id": "BIBREF15" }, { "start": 459, "end": 463, "text": "[18]", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "System Performance", "sec_num": "3.1.2." }, { "text": "Such system specific phenomena lead to quite different MT engine rankings depending on which metric is used (see Appendix B.2.). In order to get an idea on the \"overall\" performance of each system, MT engine rankings of multiple metrics are combined by simply calculating the average rank for each MT engine. If no significant difference between two MT engine scores could be determined, the same rank was assigned to both MT engines. Table 10 summarizes the MT engine rankings when combining fluency and adequacy results. An omitted horizontal line between MT engines indicates the systems were not significantly different. ", "cite_spans": [], "ref_spans": [ { "start": 435, "end": 443, "text": "Table 10", "ref_id": "TABREF9" } ], "eq_spans": [], "section": "System Performance", "sec_num": "3.1.2." }, { "text": "The automatic evaluation results of all MT engines using the official as well as the additional evaluation specifications are listed in Appendix C.1. The MT systems are ordered according to the BLEU4 metrics. The correct recognition results of all MT systems that were applied to the CE spont as well as the CE read translation task are identical and they are listed redundantly for the convenience of the reader. The MT engine rankings are summarized in Appendix C.2. Similar to the subjective evaluation results, the rankings vary with respect to the utilized automatic evaluation metrics. In order to get an idea of how closely the respective metrics are related, the Pearson correlation coefficients were calculated for all automatic evaluation metric combinations. For each translation direction, we used the official run submissions for both (ASR, CRR) input conditions, i.e., 24 runs for CE spont, 28 runs for CE read, 28 for JE, 22 runs for AE, and 24 runs for IE, respectively. The correlation coefficients are given in Table 11 . On the left hand side of the table, the BLEU4 metric is compared to the NIST and METEOR metric. The correlation between NIST and METEOR is given on the right hand side. With the exception of the CE translation task, very high correlation coefficients were obtained, but large differences can be seen for each source language. BLEU4 seems to correlate better with NIST for JE, but better with METEOR for AE. These characteristics also affect the MT engine rankings (see Appendix B.2.). Analogous to the subjective evaluation, an \"overall\" MT engine ranking combining all automatic evaluation metrics for the translation of ASR output is given in Table 12 . Again, an omitted horizontal line between MT engines indicates the systems were not significantly different.", "cite_spans": [], "ref_spans": [ { "start": 1029, "end": 1037, "text": "Table 11", "ref_id": "TABREF10" }, { "start": 1685, "end": 1693, "text": "Table 12", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Automatic Evaluation Results", "sec_num": "3.2." }, { "text": "The evaluation metrics listed in Table 8 were selected because the outcomes of last year's IWSLT workshop showed that these metrics were closely related to human judgement. Table 13 summarizes the Pearson correlation coefficients between BLEU4/NIST/METEOR and adequacy/fluency for this year's CE translation task. The results confirm previous findings that fluency correlates best with BLEU4 and that adequacy correlates best with METEOR. The NIST metric has only moderate correlation to both subjective evaluation metrics.", "cite_spans": [], "ref_spans": [ { "start": 33, "end": 40, "text": "Table 8", "ref_id": "TABREF7" }, { "start": 173, "end": 181, "text": "Table 13", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Correlation between Subjective and Automatic Evaluation Results", "sec_num": "3.3." }, { "text": "Interestingly, the correlation coefficients are much higher for correct recognition results than for the translation of ASR outputs. This is especially so for the spontaneous speech translation task where only low correlations were obtained for adequacy. This indicates that standard evaluation metrics might not be appropriate for dealing with spontaneous speech translation tasks. Future investigations should focus on how to measure the impact of spontaneity aspects on the MT translation quality in order to improve the reliability of automatic evaluation metrics.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Correlation between Subjective and Automatic Evaluation Results", "sec_num": "3.3." }, { "text": "As indicated by the English language perplexity figures listed in Table 3 , the Challenge Task 2006 of this year's evaluation campaign was much more difficult than the translation tasks of previous IWSLT workshops. The MT performance for all translation conditions on this year's evaluation set was much lower compared to the results of previous IWSLT evaluation campaigns.", "cite_spans": [], "ref_spans": [ { "start": 66, "end": 73, "text": "Table 3", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Challenge Task 2006", "sec_num": "4.1." }, { "text": "One of the reasons is the discrepancy between the supplied resources and this year's evaluation data set. The supplied resources contain mainly short sentences, whereas the evaluation data sentences were much longer. In addition, the OOV rate is quite high for this year's IWSLT 2006 evaluation data. Hence, current state-of-the-art MT systems need to improve their capability to deal with input sentences having characteristics not covered by the training corpus or containing phrases never seen before. Further research on automatic text preprocessing techniques (sentence splitting, word segmentation, etc,), model adaptation and the translation of unknown words is necessary to fill the gap.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Challenge Task 2006", "sec_num": "4.1." }, { "text": "Comparing the Open Data Track with the CSTAR Data Track results improvements of up to 4%-5% in BLEU as well as METEOR and 0.5-0.7 points in NIST were obtained when using additional in-domain training data for CE and JE.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Additional Resources", "sec_num": "4.2." }, { "text": "In addition, some participants also investigated in the utilization of additional out-of-domain training resources [14, 29] and reported mixed success depending on the input condition and translation task.", "cite_spans": [ { "start": 115, "end": 119, "text": "[14,", "ref_id": "BIBREF13" }, { "start": 120, "end": 123, "text": "29]", "ref_id": "BIBREF28" } ], "ref_spans": [], "eq_spans": [], "section": "Additional Resources", "sec_num": "4.2." }, { "text": "When comparing the results of the official and additional evaluation specification, the utilized evaluation metrics showed quite different phenomena. The NIST scores are generally lower for the evaluation taking into account punctuation and case information.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "4.3." }, { "text": "Very similar scores were obtained for METEOR. However, the current version of this metric is not compatible with the official evaluation specifications. The script removes punctuation/case information and separates word compounds, differing from the additional evaluation specifications and therefore resulting in slightly different scores.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "4.3." }, { "text": "An unexpected effect, however, was seen for the BLEU metric. In contrast to NIST, many MT engines achieved higher BLEU scores for the official evaluation specifications, despite punctuation/case errors in the MT output. The extent of this phenomenon, however, differed between the language pairs (JE: 50%, AE: 30%, CE: 30% of the utilized MT engines). Interestingly, this phenomenon was not found for the translation of Italian where the BLEU scores of the additional evaluation specifications were always higher.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Specifications", "sec_num": "4.3." }, { "text": "For the IWSLT 2006 evaluation data, the same set of English reference translations were used for the evaluation of all translations outputs. Therefore, the translation results of MT engines using different source languages as the input can be directly compared.", "cite_spans": [ { "start": 8, "end": 18, "text": "IWSLT 2006", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Language Dependency", "sec_num": "4.4." }, { "text": "Looking at the automatic evaluation results of the Open Data Track, the highest scores were obtained on the IE translation task for the CRR and the ASR output translation conditions. The latter was surprising given Italian had the worst recognition accuracies. One reason might be the close relationship between these two languages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Language Dependency", "sec_num": "4.4." }, { "text": "For Arabic, the high OOV rate caused problems for MT systems that relied on the supplied word segmentations. However, resegmenting the data set proved to be effective for increasing the vocabulary coverage and improving translation quality [14] .", "cite_spans": [ { "start": 240, "end": 244, "text": "[14]", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Language Dependency", "sec_num": "4.4." }, { "text": "For Japanese, the highest recognition accuracy was obtained. However, due to large differences in syntactic structure and word order, the JE translation task seems to be one of the most difficult tasks and the best performing systems obtained lower scores compared to the AE and IE results. Interestingly, the JE task featured the largest number of non-SMT engines including a commercial system that achieved quite good performances (see [24, 17] ).", "cite_spans": [ { "start": 438, "end": 442, "text": "[24,", "ref_id": "BIBREF23" }, { "start": 443, "end": 446, "text": "17]", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Language Dependency", "sec_num": "4.4." }, { "text": "For Chinese, the recognition accuracy for read speech is similar to the Arabic recognition results, but the automatic evaluation scores obtained for the top-scoring MT engines are much lower. The complexity of the CE translation task seems to be similar to JE. Altogether, the complexity 14 of the translation tasks of this year's IWSLT evaluation campaign can be summarized as: CE \u2248 JE > AE IE", "cite_spans": [ { "start": 288, "end": 290, "text": "14", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Language Dependency", "sec_num": "4.4." }, { "text": "When comparing the results of the ASR Output condition and the CRR data sets, all MT engines achieved lower scores for the translation of ASR output. The complexity of the translation input conditions can be summarized as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Robustness towards ASR Output", "sec_num": "4.5." }, { "text": "ASR spont > ASR read CRR", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Robustness towards ASR Output", "sec_num": "4.5." }, { "text": "The impact of recognition errors, however, differs between the languages and is closely related to the recognition accuracy obtained for the respective speech input. A moderate degradation was seen for the JE/AE/CE translation tasks (0.5-3% for BLEU, 0.3-0.7 points for NIST, 3-7% for ME-TEOR). However, the low recognition performance for Italian caused the largest difference (5-8% for BLEU, 0.9-1.2 points for NIST, 6-12% for METEOR) for IE. In addition, MT engines that were only applied to the first-best recognition output showed a larger drop in performance than MT engines that directly used information from the word lattice. In order to make MT systems more robust against speech recognition errors and to tap the full potential of the ASR systems, more research on how to directly exploit word lattices efficiently is required. The results on using confusion networks reported by IWSLT 2006 participants [15, 16, 29] are promising and lead into the right direction.", "cite_spans": [ { "start": 891, "end": 901, "text": "IWSLT 2006", "ref_id": null }, { "start": 915, "end": 919, "text": "[15,", "ref_id": "BIBREF14" }, { "start": 920, "end": 923, "text": "16,", "ref_id": "BIBREF15" }, { "start": 924, "end": 927, "text": "29]", "ref_id": "BIBREF28" } ], "ref_spans": [], "eq_spans": [], "section": "Robustness towards ASR Output", "sec_num": "4.5." }, { "text": "This year's workshop provided a testbed for applying novel ideas on how to deal with problems in the area of spontaneous speech translation. Various innovative ideas were explored, most notably the usage of out-of-domain training data [14, 29] , new methods for distortion modeling [15, 26] , topic-dependent model adaptation [20, 23] , efficient decoding of word lattices [16] , and rescoring/reranking methods of NBEST list [22, 23, 29] . Although not all ideas proved to be effective, new insights into the complexity of combining speech recognition and machine translation technologies were obtained that will help to advance the current state-ofthe-art in speech translation.", "cite_spans": [ { "start": 235, "end": 239, "text": "[14,", "ref_id": "BIBREF13" }, { "start": 240, "end": 243, "text": "29]", "ref_id": "BIBREF28" }, { "start": 282, "end": 286, "text": "[15,", "ref_id": "BIBREF14" }, { "start": 287, "end": 290, "text": "26]", "ref_id": "BIBREF25" }, { "start": 326, "end": 330, "text": "[20,", "ref_id": "BIBREF19" }, { "start": 331, "end": 334, "text": "23]", "ref_id": "BIBREF22" }, { "start": 373, "end": 377, "text": "[16]", "ref_id": "BIBREF15" }, { "start": 426, "end": 430, "text": "[22,", "ref_id": "BIBREF21" }, { "start": 431, "end": 434, "text": "23,", "ref_id": "BIBREF22" }, { "start": 435, "end": 438, "text": "29]", "ref_id": "BIBREF28" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "5." }, { "text": "I thank the C-STAR partners for their accomplishments during the preparation of this workshop and the subjective evaluation task. In particular, I would like to thank Roldano Cattoni, Roger Hsiao, Gen Itoh, Shigeki Matsuda, Jinsong Zhang, Shuwu Zhang for their support in recording the speech data sets and generating the ASR outputs. Special thanks to Matthias Eck for his extensive technical support in setting-up and maintaining the online evaluation servers. I also thank the program committee members for reviewing a large number of MT system descriptions and technical paper submissions. Last, but not least, I thank all research groups for their active participation in the IWSLT 2006 evaluation campaign and for making the IWSLT 2006 workshop a success.", "cite_spans": [ { "start": 681, "end": 691, "text": "IWSLT 2006", "ref_id": null }, { "start": 731, "end": 741, "text": "IWSLT 2006", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": "6." }, { "text": "http://www.slc.atr.jp/IWSLT2006 2 http://www.c-star.org/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "An exception was Arabic with only one native speaker.5 We used the lattice-toolkit of the SRI Language Modeling Toolkit (http://www.speech.sri.com/projects/srilm) to automatically extract NBEST lists from ASR lattices.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Please refer to the MT system descriptions of each participant for details on what kind of additional resources were used.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://www.nist.gov/speech/tests/{mt|gale} 9 http://www.elda.org/en/proj/tcstar-wp4/index.htm 10 http://www.slc.atr.jp/IWSLT2006/downloads/case+punc tool using SR ILM.instructions.txt", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "In addition to the official evaluation metrics used for IWSLT 2006, the word error rate (WER)[4] and position-independent WER (PER)[5] were also calculated by the evaluation server and provided to the participants for the analysis of their systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "http://projectile.is.cs.cmu.edu/research/public/tools/bootStrap/tutorial.htm13 2000 iterations were used for the analysis of the IWSLT 2006 results", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "\u2248 : \"similar\", > : \"more difficult\", : \"much more difficult\".", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "' * ' indicates late run submissions that were submitted after the official run submission period.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "Open", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "AE -read speech", "sec_num": null }, { "text": "Research", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Appendix A. MT System Overview", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Overview of the IWSLT04 evaluation campaign", "authors": [ { "first": "Y", "middle": [], "last": "Akiba", "suffix": "" }, { "first": "M", "middle": [], "last": "Federico", "suffix": "" }, { "first": "N", "middle": [], "last": "Kando", "suffix": "" }, { "first": "H", "middle": [], "last": "Nakaiwa", "suffix": "" }, { "first": "M", "middle": [], "last": "Paul", "suffix": "" }, { "first": "J", "middle": [], "last": "Tsujii", "suffix": "" } ], "year": 2004, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "1--12", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Akiba, M. Federico, N. Kando, H. Nakaiwa, M. Paul, and J. Tsujii, \"Overview of the IWSLT04 evaluation campaign,\" in Proc. of the International Workshop on Spoken Language Translation, Kyoto, Japan, 2004, pp. 1-12.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Overview of the IWSLT 2005 evaluation campaign", "authors": [ { "first": "M", "middle": [], "last": "Eck", "suffix": "" }, { "first": "C", "middle": [], "last": "Hori", "suffix": "" } ], "year": 2005, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "11--32", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Eck and C. Hori, \"Overview of the IWSLT 2005 evaluation campaign,\" in Proc. of the International Workshop on Spoken Language Translation, Pittsburgh, USA, 2005, pp. 11-32.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Creating corpora for speech-to-speech translation", "authors": [ { "first": "G", "middle": [], "last": "Kikui", "suffix": "" }, { "first": "E", "middle": [], "last": "Sumita", "suffix": "" }, { "first": "T", "middle": [], "last": "Takezawa", "suffix": "" }, { "first": "S", "middle": [], "last": "Yamamoto", "suffix": "" } ], "year": 2003, "venue": "Proc. of the EUROSPEECH03", "volume": "", "issue": "", "pages": "381--384", "other_ids": {}, "num": null, "urls": [], "raw_text": "G. Kikui, E. Sumita, T. Takezawa, and S. Yamamoto, \"Creating corpora for speech-to-speech translation,\" in Proc. of the EUROSPEECH03, Geneve, Switzerland, 2003, pp. 381-384.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "An evaluation tool for machine translation: Fast evaluation for machine translation research", "authors": [ { "first": "S", "middle": [], "last": "Niessen", "suffix": "" }, { "first": "F", "middle": [ "J" ], "last": "Och", "suffix": "" }, { "first": "G", "middle": [], "last": "Leusch", "suffix": "" }, { "first": "H", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2000, "venue": "Proc. of the 2nd LREC", "volume": "", "issue": "", "pages": "39--45", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Niessen, F. J. Och, G. Leusch, and H. Ney, \"An evaluation tool for machine translation: Fast evaluation for machine translation research,\" in Proc. of the 2nd LREC, Athens, Greece, 2000, pp. 39-45.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Minimum error rate training in statistical machine translation", "authors": [ { "first": "F", "middle": [ "J" ], "last": "Och", "suffix": "" } ], "year": 2003, "venue": "Proc. of the 41st ACL", "volume": "", "issue": "", "pages": "160--167", "other_ids": {}, "num": null, "urls": [], "raw_text": "F. J. Och, \"Minimum error rate training in statistical machine translation,\" in Proc. of the 41st ACL, Sapporo, Japan, 2003, pp. 160-167.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "BLEU: a method for automatic evaluation of machine translation", "authors": [ { "first": "K", "middle": [], "last": "Papineni", "suffix": "" }, { "first": "S", "middle": [], "last": "Roukos", "suffix": "" }, { "first": "T", "middle": [], "last": "Ward", "suffix": "" }, { "first": "W", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2002, "venue": "Proc. of the 40th ACL", "volume": "", "issue": "", "pages": "311--318", "other_ids": {}, "num": null, "urls": [], "raw_text": "K. Papineni, S. Roukos, T. Ward, and W. Zhu, \"BLEU: a method for automatic evaluation of machine trans- lation,\" in Proc. of the 40th ACL, Philadelphia, USA, 2002, pp. 311-318.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Automatic evaluation of machine translation quality using n-gram co-occurrence statistics", "authors": [ { "first": "G", "middle": [], "last": "Doddington", "suffix": "" } ], "year": 2002, "venue": "Proc. of the HLT 2002", "volume": "", "issue": "", "pages": "257--258", "other_ids": {}, "num": null, "urls": [], "raw_text": "G. Doddington, \"Automatic evaluation of machine translation quality using n-gram co-occurrence statis- tics,\" in Proc. of the HLT 2002, San Diego, USA, 2002, pp. 257-258.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "METEOR: An automatic metric for MT evaluation with improved correlation with human judgments", "authors": [ { "first": "S", "middle": [], "last": "Banerjee", "suffix": "" }, { "first": "A", "middle": [], "last": "Lavie", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization", "volume": "", "issue": "", "pages": "65--72", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Banerjee and A. Lavie, \"METEOR: An automatic metric for MT evaluation with improved correlation with human judgments,\" in Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Mea- sures for Machine Translation and/or Summarization, Ann Arbor, Michigan, 2005, pp. 65-72.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Interpreting Bleu/NIST Scores: How Much Improvement do We Need to Have a Better System", "authors": [ { "first": "Y", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "S", "middle": [], "last": "Vogel", "suffix": "" }, { "first": "A", "middle": [], "last": "Waibel", "suffix": "" } ], "year": 2004, "venue": "Proc of the LREC", "volume": "", "issue": "", "pages": "2051--2054", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Zhang, S. Vogel, and A. Waibel, \"Interpreting Bleu/NIST Scores: How Much Improvement do We Need to Have a Better System?\" in Proc of the LREC, 2004, pp. 2051-2054.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Finite-State Transducer-based Statistical Machine Translation using Joint Probabilities", "authors": [ { "first": "S", "middle": [], "last": "Bangalore", "suffix": "" }, { "first": "S", "middle": [], "last": "Kanthak", "suffix": "" }, { "first": "P", "middle": [], "last": "Haffner", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "16--22", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Bangalore, S. Kanthak, and P. Haffner, \"Finite-State Transducer-based Statistical Machine Translation using Joint Probabilities,\" in Proc. of the International Work- shop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 16-22.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "IWSLT-06: experiments with commercial MT systems and lessons from subjective evaluations", "authors": [ { "first": "C", "middle": [], "last": "Boitet", "suffix": "" }, { "first": "Y", "middle": [], "last": "Bey", "suffix": "" }, { "first": "M", "middle": [], "last": "Tomokiyo", "suffix": "" }, { "first": "W", "middle": [], "last": "Cao", "suffix": "" }, { "first": "H", "middle": [], "last": "Blanchon", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "23--30", "other_ids": {}, "num": null, "urls": [], "raw_text": "C. Boitet, Y. Bey, M. Tomokiyo, W. Cao, and H. Blan- chon, \"IWSLT-06: experiments with commercial MT systems and lessons from subjective evaluations,\" in Proc. of the International Workshop on Spoken Lan- guage Translation, Kyoto, Japan, 2006, pp. 23-30.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "MATREX: DCU Machine Translation System for IWSLT 2006", "authors": [ { "first": "N", "middle": [], "last": "Stroppa", "suffix": "" }, { "first": "A", "middle": [], "last": "Way", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "31--36", "other_ids": {}, "num": null, "urls": [], "raw_text": "N. Stroppa and A. Way, \"MATREX: DCU Machine Translation System for IWSLT 2006,\" in Proc. of the International Workshop on Spoken Language Transla- tion, Kyoto, Japan, 2006, pp. 31-36.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Toward Integrating Word Sense and Entity Disambiguation into Statistical Machine Translation", "authors": [ { "first": "M", "middle": [], "last": "Carpuat", "suffix": "" }, { "first": "Y", "middle": [], "last": "Shen", "suffix": "" }, { "first": "X", "middle": [], "last": "Yu", "suffix": "" }, { "first": "D", "middle": [], "last": "Wu", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "37--44", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Carpuat, Y. Shen, X. Yu, and D. Wu, \"Toward In- tegrating Word Sense and Entity Disambiguation into Statistical Machine Translation,\" in Proc. of the Inter- national Workshop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 37-44.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "IBM Arabic-to-English Translation for IWSLT 2006", "authors": [ { "first": "Y.-S", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "45--52", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y.-S. Lee, \"IBM Arabic-to-English Translation for IWSLT 2006,\" in Proc. of the International Workshop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 45-52.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "The ITC-irst SMT System for IWSLT 2006", "authors": [ { "first": "B", "middle": [], "last": "Chen", "suffix": "" }, { "first": "R", "middle": [], "last": "Cattoni", "suffix": "" }, { "first": "N", "middle": [], "last": "Bertoldi", "suffix": "" }, { "first": "M", "middle": [], "last": "Cetello", "suffix": "" }, { "first": "M", "middle": [], "last": "Federico", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "53--58", "other_ids": {}, "num": null, "urls": [], "raw_text": "B. Chen, R. Cattoni, N. Bertoldi, M. Cetello, and M. Federico, \"The ITC-irst SMT System for IWSLT 2006,\" in Proc. of the International Workshop on Spo- ken Language Translation, Kyoto, Japan, 2006, pp. 53- 58.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "The JHU Workshop 2006 IWSLT System", "authors": [ { "first": "W", "middle": [], "last": "Shen", "suffix": "" }, { "first": "R", "middle": [], "last": "Zens", "suffix": "" }, { "first": "N", "middle": [], "last": "Bertoldi", "suffix": "" }, { "first": "M", "middle": [], "last": "Federico", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "59--63", "other_ids": {}, "num": null, "urls": [], "raw_text": "W. Shen, R. Zens, N. Bertoldi, and M. Federico, \"The JHU Workshop 2006 IWSLT System,\" in Proc. of the International Workshop on Spoken Language Transla- tion, Kyoto, Japan, 2006, pp. 59-63.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Example-based Machine Translation based on Deeper NLP", "authors": [ { "first": "T", "middle": [], "last": "Nakazawa", "suffix": "" }, { "first": "K", "middle": [], "last": "Yu", "suffix": "" }, { "first": "D", "middle": [], "last": "Kawahara", "suffix": "" }, { "first": "S", "middle": [], "last": "Kurohashi", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "64--70", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Nakazawa, K. Yu, D. Kawahara, and S. Kurohashi, \"Example-based Machine Translation based on Deeper NLP,\" in Proc. of the International Workshop on Spo- ken Language Translation, Kyoto, Japan, 2006, pp. 64- 70.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "The MIT-LL/AFRL IWSLT-2006 MT System", "authors": [ { "first": "W", "middle": [], "last": "Shen", "suffix": "" }, { "first": "B", "middle": [], "last": "Delaney", "suffix": "" }, { "first": "T", "middle": [], "last": "Anderson", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "71--76", "other_ids": {}, "num": null, "urls": [], "raw_text": "W. Shen, B. Delaney, and T. Anderson, \"The MIT- LL/AFRL IWSLT-2006 MT System,\" in Proc. of the International Workshop on Spoken Language Transla- tion, Kyoto, Japan, 2006, pp. 71-76.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Phrase Reordering for Statistical Machine Translation Based on Predicate-Argument Structure", "authors": [ { "first": "M", "middle": [], "last": "Komachi", "suffix": "" }, { "first": "M", "middle": [], "last": "Nagata", "suffix": "" }, { "first": "Y", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "77--82", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Komachi, M. Nagata, and Y. Matsumoto, \"Phrase Reordering for Statistical Machine Translation Based on Predicate-Argument Structure,\" in Proc. of the Inter- national Workshop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 77-82.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "The NiCT-ATR Statistical Machine Translation System for IWSLT 2006", "authors": [ { "first": "R", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "H", "middle": [], "last": "Yamamoto", "suffix": "" }, { "first": "M", "middle": [], "last": "Paul", "suffix": "" }, { "first": "H", "middle": [], "last": "Okuma", "suffix": "" }, { "first": "K", "middle": [], "last": "Yasuda", "suffix": "" }, { "first": "Y", "middle": [], "last": "Lepage", "suffix": "" }, { "first": "E", "middle": [], "last": "Denoual", "suffix": "" }, { "first": "D", "middle": [], "last": "Mochihashi", "suffix": "" }, { "first": "A", "middle": [], "last": "Finch", "suffix": "" }, { "first": "E", "middle": [], "last": "Sumita", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "83--90", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. Zhang, H. Yamamoto, M. Paul, H. Okuma, K. Ya- suda, Y. Lepage, E. Denoual, D. Mochihashi, A. Finch, and E. Sumita, \"The NiCT-ATR Statistical Machine Translation System for IWSLT 2006,\" in Proc. of the International Workshop on Spoken Language Transla- tion, Kyoto, Japan, 2006, pp. 83-90.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "NLPR Translation System for IWSLT 2006 Evaluation Campaign", "authors": [ { "first": "C", "middle": [], "last": "Chai", "suffix": "" }, { "first": "J", "middle": [], "last": "Du", "suffix": "" }, { "first": "W", "middle": [], "last": "Wei", "suffix": "" }, { "first": "P", "middle": [], "last": "Liu", "suffix": "" }, { "first": "K", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "Y", "middle": [], "last": "He", "suffix": "" }, { "first": "C", "middle": [], "last": "Zong", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "91--94", "other_ids": {}, "num": null, "urls": [], "raw_text": "C. Chai, J. Du, W. Wei, P. Liu, K. Zhou, Y. He, and C. Zong, \"NLPR Translation System for IWSLT 2006 Evaluation Campaign,\" in Proc. of the Interna- tional Workshop on Spoken Language Translation, Ky- oto, Japan, 2006, pp. 91-94.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "NTT Statistical Machine Translation for IWSLT 2006", "authors": [ { "first": "T", "middle": [], "last": "Watanabe", "suffix": "" }, { "first": "J", "middle": [], "last": "Suzuki", "suffix": "" }, { "first": "H", "middle": [], "last": "Tsukada", "suffix": "" }, { "first": "H", "middle": [], "last": "Isozaki", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "95--102", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Watanabe, J. Suzuki, H. Tsukada, and H. Isozaki, \"NTT Statistical Machine Translation for IWSLT 2006,\" in Proc. of the International Workshop on Spo- ken Language Translation, Kyoto, Japan, 2006, pp. 95- 102.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "The RWTH Statistical Machine Translation System for the IWSLT 2006 Evaluation", "authors": [ { "first": "A", "middle": [], "last": "Mauser", "suffix": "" }, { "first": "R", "middle": [], "last": "Zens", "suffix": "" }, { "first": "E", "middle": [], "last": "Matusov", "suffix": "" }, { "first": "S", "middle": [], "last": "Hasan", "suffix": "" }, { "first": "H", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "103--110", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Mauser, R. Zens, E. Matusov, S. Hasan, and H. Ney, \"The RWTH Statistical Machine Translation System for the IWSLT 2006 Evaluation,\" in Proc. of the Inter- national Workshop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 103-110.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "The SLE Example-Based Translation System", "authors": [ { "first": "P", "middle": [], "last": "Whitelock", "suffix": "" }, { "first": "V", "middle": [], "last": "Poznanski", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "111--115", "other_ids": {}, "num": null, "urls": [], "raw_text": "P. Whitelock and V. Poznanski, \"The SLE Example- Based Translation System,\" in Proc. of the Interna- tional Workshop on Spoken Language Translation, Ky- oto, Japan, 2006, pp. 111-115.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "The TALP Ngram-based SMT System for IWSLT 2006", "authors": [ { "first": "J", "middle": [ "M" ], "last": "Crego", "suffix": "" }, { "first": "A", "middle": [], "last": "Gispert", "suffix": "" }, { "first": "P", "middle": [], "last": "Lambert", "suffix": "" }, { "first": "M", "middle": [], "last": "Khalilov", "suffix": "" }, { "first": "M", "middle": [ "R" ], "last": "Costa-Juss\u00e0", "suffix": "" }, { "first": "J", "middle": [ "B" ], "last": "Mari\u00f1o", "suffix": "" }, { "first": "R", "middle": [], "last": "Banchs", "suffix": "" }, { "first": "J", "middle": [ "A" ], "last": "Fonollosa", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "116--122", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. M. Crego, A. de Gispert, P. Lambert, M. Khalilov, M. R. Costa-juss\u00e0, J. B. Mari\u00f1o, R. Banchs, and J. A. Fonollosa, \"The TALP Ngram-based SMT System for IWSLT 2006,\" in Proc. of the International Workshop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 116-122.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "TALP Phrase-Based System and TALP System Combination for IWSLT 2006", "authors": [ { "first": "M", "middle": [ "R" ], "last": "Costa-Juss\u00e0", "suffix": "" }, { "first": "J", "middle": [ "M" ], "last": "Crego", "suffix": "" }, { "first": "A", "middle": [], "last": "Gispert", "suffix": "" }, { "first": "P", "middle": [], "last": "Lambert", "suffix": "" }, { "first": "M", "middle": [], "last": "Khalilov", "suffix": "" }, { "first": "J", "middle": [ "A" ], "last": "Fonollosa", "suffix": "" }, { "first": "J", "middle": [ "B" ], "last": "Mari\u00f1o", "suffix": "" }, { "first": "R", "middle": [], "last": "Banchs", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "123--129", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. R. Costa-juss\u00e0, J. M. Crego, A. de Gispert, P. Lam- bert, M. Khalilov, J. A. Fonollosa, J. B. Mari\u00f1o, and R. Banchs, \"TALP Phrase-Based System and TALP System Combination for IWSLT 2006,\" in Proc. of the International Workshop on Spoken Language Transla- tion, Kyoto, Japan, 2006, pp. 123-129.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "The UKA/CMU Statistical Machine Translation System for IWSLT 2006", "authors": [ { "first": "M", "middle": [], "last": "Eck", "suffix": "" }, { "first": "I", "middle": [], "last": "Lane", "suffix": "" }, { "first": "N", "middle": [], "last": "Bach", "suffix": "" }, { "first": "S", "middle": [], "last": "Hewavitharana", "suffix": "" }, { "first": "M", "middle": [], "last": "Kolss", "suffix": "" }, { "first": "B", "middle": [], "last": "Zhao", "suffix": "" }, { "first": "A", "middle": [ "S" ], "last": "Hildebrand", "suffix": "" }, { "first": "S", "middle": [], "last": "Vogel", "suffix": "" }, { "first": "A", "middle": [], "last": "Waibel", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "130--137", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Eck, I. Lane, N. Bach, S. Hewavitharana, M. Kolss, B. Zhao, A. S. Hildebrand, S. Vogel, and A. Waibel, \"The UKA/CMU Statistical Machine Translation Sys- tem for IWSLT 2006,\" in Proc. of the International Workshop on Spoken Language Translation, Kyoto, Japan, 2006, pp. 130-137.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "The CMU-UKA Syntax Augmented Machine Translation System for IWSLT-06", "authors": [ { "first": "A", "middle": [], "last": "Zollmann", "suffix": "" }, { "first": "A", "middle": [], "last": "Venugopal", "suffix": "" }, { "first": "S", "middle": [], "last": "Vogel", "suffix": "" }, { "first": "A", "middle": [], "last": "Waibel", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "138--144", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Zollmann, A. Venugopal, S. Vogel, and A. Waibel, \"The CMU-UKA Syntax Augmented Machine Trans- lation System for IWSLT-06,\" in Proc. of the Interna- tional Workshop on Spoken Language Translation, Ky- oto, Japan, 2006, pp. 138-144.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "The University of Washington Machine Translation System for IWSLT 2006", "authors": [ { "first": "K", "middle": [], "last": "Kirchhoff", "suffix": "" }, { "first": "K", "middle": [], "last": "Duh", "suffix": "" }, { "first": "C", "middle": [], "last": "Lim", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "145--152", "other_ids": {}, "num": null, "urls": [], "raw_text": "K. Kirchhoff, K. Duh, and C. Lim, \"The University of Washington Machine Translation System for IWSLT 2006,\" in Proc. of the International Workshop on Spo- ken Language Translation, Kyoto, Japan, 2006, pp. 145-152.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "The XMU Phrase-Based Statistical Machine Translation System for IWSLT 2006", "authors": [ { "first": "Y", "middle": [], "last": "Chen", "suffix": "" }, { "first": "X", "middle": [], "last": "Shi", "suffix": "" }, { "first": "C", "middle": [], "last": "Zhou", "suffix": "" } ], "year": 2006, "venue": "Proc. of the International Workshop on Spoken Language Translation", "volume": "", "issue": "", "pages": "153--157", "other_ids": {}, "num": null, "urls": [], "raw_text": "Y. Chen, X. Shi, and C. Zhou, \"The XMU Phrase-Based Statistical Machine Translation System for IWSLT 2006,\" in Proc. of the International Workshop on Spo- ken Language Translation, Kyoto, Japan, 2006, pp. 153-157.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "text": "avoid direct recitation of answer keys \u2022 in case of Yes/No-questions, try to explain the reason 3 20 speakers, gender: 10x female/male each, age: 21 -32 (avg: 25.7)", "uris": null, "num": null }, "TABREF1": { "type_str": "table", "html": null, "num": null, "content": "
scene: [airplane] passenger asks flight attendance for help
question: Okay. Where can I put my luggage? Is it here okay?
key: (not here, overhead compartement)
answer: \"sorry you'd better put it in the overhead comparte-
ment\"
scene: [airport] asking directions
question: Take me to this address. How long will it take?
key: (depending on traffic condition, around 20 minutes)
answer: \"it's hard to say it depends on the traffic conditions
it should take only twenty minutes or so if there's no
traffic jam\"
", "text": "Data preparation of Challenge Task 2006" }, "TABREF2": { "type_str": "table", "html": null, "num": null, "content": "
: English language perplexity of IWSLT 2006 transla-
tion tasks
translationtraining data
typetask 40K (CE/JE) 20K (AE/IE)
development dev127.532.6
dev231.436.7
dev332.938.8
dev485.698.3
evaluation eval105.9113.9
", "text": "" }, "TABREF3": { "type_str": "table", "html": null, "num": null, "content": "
type langOOV rates (%)
uage CRR 1BEST NBEST
dev4 Cs2.02.02.3
Cr2.01.72.3
J1.71.31.3
A13.114.215.4
I3.62.12.2
E 71.7 / 1.4
eval Cs2.62.12.4
Cr2.62.42.5
J2.21.62.3
A14.316.017.1
I4.32.52.6
E 7
", "text": "OOV rates ofIWSLT 2006 spoken language corpus" }, "TABREF4": { "type_str": "table", "html": null, "num": null, "content": "
80
", "text": "Recognition accuracy of IWSLT 2006 spoken language corpus type lang word (%) sentence (%) uage lattice 1BEST lattice 1BEST dev4 Cs 76.95 67.38 22.49 18.00 Cr 83.24 74.78 30.47 23.31 J 88.95 84.35 50.31 40.08 A 86.71 73.36 41.10 19.84 I 76.02 74.10 15.34 13.91 eval Cs 79.08 68.11 22.80 16.60 Cr 82.07 73.64 28.40 22." }, "TABREF5": { "type_str": "table", "html": null, "num": null, "content": "
EventDate
Training Corpus ReleaseMay 12, 2006
Development Corpus Release June 30, 2006
Evaluation Corpus ReleaseAugust 7, 2006
Result Submission DueAugust 9, 2006
", "text": "IWSLT 2006 evaluation campaign schedule" }, "TABREF6": { "type_str": "table", "html": null, "num": null, "content": "
TranslationOpen Data TrackCSTAR Data Track
InputResearchOfficialResearchOfficial
Condition Groups (Contrastive) Groups (Contrastive)
CE spont1212 (11)23 (3)
read1214 (17)23 (3)
JE read1214 (14)22 (3)
AE read911 (14)11 (1)
IE read1012 (14)11 (3)
TOTAL1963 (70)210 (13)
", "text": "Distribution of run submissions" }, "TABREF7": { "type_str": "table", "html": null, "num": null, "content": "
Scores range between 0 (worst) and 1 (best) [6]
NIST:a variant of BLEU4 using the arithmetic mean of
weighted n-gram precision values. Scores are posi-
tive with 0 being the worst possible [7]
METEOR: calculates unigram overlaps between a translations
and reference texts using various levels of matches
(exact, stem, synonym) are taken into account.
Scores range between 0 (worst) and 1 (best) [8]
", "text": "Automatic evaluation metrics BLEU4: the geometric mean of n-gram precision by the system output with respect to reference translations." }, "TABREF8": { "type_str": "table", "html": null, "num": null, "content": "
FluencyAdequacy
4 Flawless English4 All Information
3 Good English3 Most Information
2 Non-native English2 Much Information
1 Disfluent English1 Little Information
0 Incomprehensible0 None
", "text": "Human assessment" }, "TABREF9": { "type_str": "table", "html": null, "num": null, "content": "
CE spontCE readCRR
JHU WS06JHU WS06MIT-LL AFRL
RWTHMIT-LL AFRLRWTH
NTTRWTHNTT
MIT-LL AFRLNTTJHU WS06
UKACMU SMTNiCT-ATRNiCT-ATR
NiCT-ATRUKACMU SMTUKACMU SMT
", "text": "Combination of Subjective Evaluation Rankings" }, "TABREF10": { "type_str": "table", "html": null, "num": null, "content": "
BLEU4NIST METEORNISTMETEOR
CE spont0.780.86CE spont0.86
CE read0.690.73CE read0.72
JE0.950.88JE0.91
AE0.850.98AE0.90
IE0.980.95IE0.97
", "text": "Correlation between Automatic Evaluation Metrics" }, "TABREF11": { "type_str": "table", "html": null, "num": null, "content": "
CE spontCE readJE readAE readIE read
RWTHRWTHRWTHIBMWashington-U
JHU WS06MIT-LL AFRLNiCT-ATRNiCT-ATRNiCT-ATR
NiCT-ATRNiCT-ATRUKACMU SMTTALP tuplesTALP tuples
UKACMU SMTJHU WS06NTTTALP combMIT-LL AFRL
HKUSTITC-irstMIT-LL AFRLNTTTALP comb
ITC-irstTALP tuplesITC-irstUKACMU SMTITC-irst
MIT-LL AFRLTALP phrasesSLETALP phrasesTALP phrases
NTTUKACMU SMTHKUSTITC-irstNTT
Xiamen-UHKUSTTALP tuplesDCUDCU
ATTTALP combNAISTHKUSTUKACMU SMT
NLPRNTTKyoto-UCLIPSHKUST
CLIPSXiamen-UTALP combCLIPS
NLPRTALP phrases
ATTCLIPS
", "text": "Combination of Automatic Evaluation Rankings" }, "TABREF12": { "type_str": "table", "html": null, "num": null, "content": "
CE spon BLEU4 NIST METEOR
Fluency0.88 0.550.72
Adequacy 0.34 0.570.54
CE read BLEU4 NIST METEOR
Fluency0.89 0.630.66
Adequacy 0.83 0.640.89
CE CRR BLEU4 NIST METEOR
Fluency0.96 0.840.93
Adequacy 0.95 0.820.96
", "text": "Correlation between Subjective and Automatic Evaluation Metrics" } } } }