Notes from the second machine translation session at COLING\/ACL in Sydney Australia. If you aren’t a computational linguist, this will probably not interest you. Even if you are, I am not making any promises…<\/p>\n
<\/p>\n
Machine Translation Session I:<\/p>\n
Combining Arabic Preprocessing Schemes for Statistical Machine Translation
\nFatiha Sadat, Nizar Habash<\/p>\n
The presentation was not very polished, and slides had lots of text on
\nthem. The content seemed very nice though. <\/p>\n
Nizar is not here, but I did run into Owen Rambow later. <\/p>\n
—<\/p>\n
Going Beyond AER: An Extensive Analysis of Word Alignment and Their Impact on MT
\nNecip Fazil Ayan, Bonnie J. Dohr<\/p>\n
I thought this was a good presentation, but Bonnie went very quickly
\non some of the slides. Overall if you are using BLEU or Alignment
\nError Rate (AER) for evaluation, you have to look at the actual
\napplication to be able to make a good evaluation. AER doesn’t seem to
\nbe a very good metric for word alignment evaluation for phrase
\ntranslation. <\/p>\n
—<\/p>\n
Discriminative Word Alignment with Conditional Random Fields
\nPhil Blunsom, Trevor Cohn (The University of Melbourne)<\/p>\n
Because they do forward and backwards models, they can do many to many
\n(since one direction can do one to many) alignments. <\/p>\n
They add features to the alignment process that you can’t use in
\nGiza. Orthographic features like (stems, prefix and suffix match,
\nword length difference.) Markov feature, lets model learn that one to
\nmany are not good? Relative sentence position, POS tags, bilingual
\ndictionary, some features on null. <\/p>\n
They conclude that Conditional Random Fields are a very good model for
\nword alignment even with small number of aligned words per sentence.
\nThey are tractable to train and use. Using multiple features are
\nuseful, especially when combined with using other models (GIZA \/ IBM
\nmodel 4) are very useful. Markov sequence features improve
\nalignments. <\/p>\n
—<\/p>\n
Named Entity Transliteration with Comparable Corpora
\nRichard Sproat, Tao Tao, ChengXiang Zhai (University of Illinois at
\nUrbana-Champaign)<\/p>\n
Using data from Xinhau in English and Chinese from the same news
\nagency. Start with an English name, and try to identify Chinese
\ncharacter n-grams. Score the candidates via phonetic scoring, and a
\nfrequency profile looking at how the English and Chinese candidate
\nshow up over time together. Looks also at scores based on
\ncooccurrences in document pairs over corpora. <\/p>\n
Phonetic model is probabilistic noisy source-channel model that maps
\nfrom Chinese characters into sequences of English phonemes. Uses
\nGood-Turning smoothing for unseen mappings. <\/p>\n
Identify texts with potential names by searching for strings with
\nsequences of characters (about 500) that are used for transliteration
\ncommonly. <\/p>\n
If you have documents that contain other potential transliteration
\npairs that are in the same document, you increase your weight for the
\ncandidate. They introduce a new parameter to account for scores being
\nreinforced by co-occurring terms also being likely translation pairs.
\nSeems to be similar to PageRank (they say, and it makes sense.) <\/p>\n
Both the phonetic and time based model are known I think, but the
\nscore propagation is novel, and it is shown to help. <\/p>\n
At the talk, Pascale Fung was sitting next to me, but I didn’t notice
\nuntil mid-way through the talk. She suggests using the time frequency
\nmeasure as a pre-selection step for generating candidates. <\/p>\n
Also asked a question on whether they have compared their approach to
\nusing a large dictionary. <\/p>\n
—<\/p>\n
Extracting Parallel Fragments from Comparable Corpora
\nDragos Munteanu, Daniel Marcu<\/p>\n
Parallel data is almost entirely from the Political domain
\n(parliamentary data) and news. Looks like they are doing comparable
\ncorpora extraction from the web. They are focusing on finding
\nparallel bits from the document pairs, but are not talking about
\nfinding candidate documents. <\/p>\n
They use a bilingual lexicon in their system that has three features:
\nhigh precision, probability of translation, probability of not being
\ntranslations. They learn this dictionary, using log likelyhood ratio
\nscores, and put a positive or negative association on them based on
\noccurrence patterns. <\/p>\n
Using this learned resource they find sub-fragments of a sentence that
\nare translations of each other. They look at a smoothed signal
\n(average translation probabilities of word + five surrounding areas)
\nto see which parts of the sentence are translation. <\/p>\n
They evaluate comparing MT systems that are training on initial data,
\nor data + extracted parallel fragments, data + extracted sentences. <\/p>\n
(In general, the font was so small on the slides that I could not read
\nany of the examples at all.)<\/p>\n
Question from Pascale about using Mutual Information with T-score to
\nsolve the problem with fewer steps instead of log likelyhood ratio. <\/p>\n
—<\/p>\n
Applications I session<\/p>\n
Automated Japanese Essay Scoring System based on Articles written by
\nExperts
\nTsunenori Ishioka, Masayuki Kameda<\/p>\n
Their system rates on rhetoric, organization, and contents. <\/p>\n
Rhetorical: They have some metrics for rhetorical judgments: shorter
\nsentence length, clause length, number of clauses, kanji \/ kana ratio,
\nnumber of embeddings, vocabulary diversity, converts kanji to kana
\nreading then counts lengths where longer is more complicated, and
\npassive sentences. They compute statistical distributions from
\neditorials and text from Mainichi news. They compare input text to
\nthe distribution over these variables.<\/p>\n
Organization: They have a list of (125) connective conjunctive phrases
\nthat are used to segment the document, they are classified into 8
\ncategories. They look at the number of connective conjunctions, and
\nlook at the order of the appearance with a trigram and unigram model. <\/p>\n
Latent Semantic Indexing is used to evaluate contents. Their system
\nworks on essays from 400-800 characters in Japanese. <\/p>\n
—<\/p>\n
A Feedback-Augmented Method for Detecting Errors in the Writing of
\nLearners of English
\nRyo Nagata, Atsuo Kawaii, Koichiro Morihiro, Naoki Isu
\n(Hyogo University of Teacher Education and Mie University)<\/p>\n
They are focusing on errors of articles and countability. <\/p>\n
They generate WIC lists for a word with context to get examples for
\nrules, then have some rule based system for classifying into countable
\nand uncountable instances. They then learn rules to predict
\ncountability, and learn a log liklihood ratio to determine when to
\napply the rules. The rules are predicated on the context of the word. <\/p>\n
Augmenting the decision lists with feedback. They use a corpus of
\nteachers’ corrections to learn rules, and a probabilistic model to mix
\nthem with the rules from the general corpus. Since the feedback
\ncorpus is small, they looked at many types of models for mixing the
\nresults. <\/p>\n
I think it would have been nice to see an example of the feedback in
\nthe feedback corpus. <\/p>\n
Question about the statistical significance of the results – they were
\nnot statistically significant across different types of the system.
\nAnother question about corpus choice, perhaps should download texts on
\nthe specific topic to improve lexical coverage. Comment that he
\nshould look at Bond’s work on AltJ\/E and some of his papers on
\ntranslation and countability. <\/p>\n
—<\/p>\n
Correcting ESL Errors using Phrasal SMT Techniques
\nChris Brokett, William B. Dolan, Michael Gamon
\n(Microsoft Research)<\/p>\n
The idea of ESL error correction as statistical machine translation.
\nThey model the learner errors as a noisy channel model and try to find
\nthe best translation sentence with the input error sentence. This
\npaper is a small pilot study with an off the shelf SMT system
\ncorrecting errors in mass nouns. Used “Tree to string” system
\n(Menezes and Quirk 2005) uses unlabeled dependency parse. <\/p>\n
Created artificial error-full sentences. With 45K words of training
\ndata, they were able to correct completely about 55% of the testing
\ndata. <\/p>\n
Overall I really liked this presentation. <\/p>\n
—<\/p>\n
Dialogue II Session<\/p>\n
Learning to Generate Naturalistic Utterances Using Reviews in Spoken
\nDialogue Systems
\nRyuichiro Higashinaka, Rashimi Prasad, Marilyn Walker<\/p>\n
I enjoyed this paper. The basic approach was to mine review websites
\n(food in this case) for reviews and their associated scores (1-5) on a
\nvariety of features (food quality, ambiance, value, etc.) They
\ndetermined which sentences are related to the scores using a
\nhand-crafted lexicon of words that connect to the various scored
\nfeatures, and built a template that realized the scores. One template
\ncan have multiple value associated with it, and will only be used for
\ngeneration when all the values are available to generate. He
\ndemonstrated that the system could be plugged into a dialogue system.
\nThere were some questions that were not quite on point I think (can
\nyou use multiple websites? — of course, if you write a front-end
\nparser to scrape it, and have a good lexicon for the feature mapping.
\nAnother was an interesting question, but not on point either really.
\nThe question was a comment that the learned templates have words
\n(adjectives, descriptions) that are used for different scores. For
\nexample, 1 and 2 for food both had “bad food.” Of course, that is a
\nquestion of human interpretation, and since humans used these words to
\ndescribe the food, of course it is valid.)<\/p>\n
—<\/p>\n
Poster Sessions<\/p>\n
I saw a good poster from Satoshi Sekine from NYU. <\/p>\n","protected":false},"excerpt":{"rendered":"
Notes from the second machine translation session at COLING\/ACL in Sydney Australia. If you aren’t a computational linguist, this will probably not interest you. Even if you are, I am not making any promises… Machine Translation Session I: Combining Arabic Preprocessing Schemes for Statistical Machine Translation Fatiha Sadat, Nizar Habash The presentation was not very […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[10],"tags":[],"_links":{"self":[{"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/posts\/122"}],"collection":[{"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/comments?post=122"}],"version-history":[{"count":0,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/posts\/122\/revisions"}],"wp:attachment":[{"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/media?parent=122"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/categories?post=122"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/tags?post=122"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}