Rough notes from Thursday’s presentations at COLING\/ACL. <\/p>\n
<\/p>\n
Exploiting Syntactic Patterns as Clues in Zero-Anaphora Resolution
\nRyu Iida, Kentaro Inui, Yuji Matsumoto<\/p>\n
A nominee for the best asian paper award. A fairly interesting
\napproach to anaphoric identification for “missing” zero-anaphora. I
\ncan’t really say that I know much about anaphoric resolution research,
\nbut they have a fairly sophisticated approach that uses dependency
\nparse trees (retaining only relevant sub-trees and reducing them to
\npart of speech information to avoid data sparsity.) <\/p>\n
—<\/p>\n
Challenges in NLP: Some New Perspectives from the East<\/p>\n
A very interesting panel I think. Three points:<\/p>\n
* Are Asian language special?<\/p>\n
* What are some salient linguistic differences in Asian languages, and
\n what are the implications for NLP?<\/p>\n
* Can the availability of detailed linguistic information (e.g.,
\n morphology) help ameliorate problems with the scarcity of large
\n annotated corpora?<\/p>\n
(1 Tsuji Junichi)<\/p>\n
For Asian languages, there is some separation between sentence grammar
\nand discourse grammar. Asian languages are discourse oriented
\nlanguages. Topic markers (wa) and particles that are very sensitive
\nto discourse (sure, dake, mo, koso…) so they have a real impact on
\ndiscourse understanding. Similarly for Chinese, there does not seem
\nto be a packing device in English that corresponds to Chinese topic
\nmarking. The concept of zero-anaphora, pronouns that are just dropped
\nand have to be recovered, and in some way are quite related to
\ndiscourse. Context is very important to understand underspecified
\nmeaning (\u50d5\u306f\u3046\u306a\u304e\u3060.)
\n(2 Kenjamin K. Tsou City University of Hong Kong)<\/p>\n
Writing system in asian languages is very broad. Alphabetic and
\nideographic and mixtures of them. What are some differences? Entropy
\nof European writing systems is about 4, but with Asian languages can
\ngo up to 9. <\/p>\n
Structurally when you have two nouns together, what is their
\nrelationship for (for MT etc.) The love of God vs. the love of
\nMoney. (subjective & objective vs objective only.) There is some
\nculture basis about how to interpret these things as well. Intrinsic
\ndifferences in the writing systems are seen in the differences in the
\nentropy of the writing systems. Different relationships between
\nconstituents, sometimes marked, sometimes you need culture to
\ndisambiguate. <\/p>\n
(3 Pushpak Bhattacharyya)<\/p>\n
Annotated corpora is a scarce resource. Only about 10 languages have
\nlarge annotated corpora out of about 7000 languages. English has
\nrelatively low morphology compared to other languages, so can we
\nunderstand rich morphology? Exploiting morphological richness can
\nreally help. We should build high quality morphological analyzers. <\/p>\n
—<\/p>\n
After the panel on Asian NLP I decided to go shopping for new shoes.
\nDespite my best efforts, my shoes stank from Wednesday’s incident
\n(thanks for owning up to that Kris!) and I didn’t think I could take
\nthe smell much longer myself. So I missed out on the morning sessions
\nin exchange for a bout of shopping. <\/p>\n
—<\/p>\n
Discourse session (Daniel Marcu chair)<\/p>\n
Proximity in Context: an empirically grounded computational model of
\nproximity for processing topological spatial expressions
\nJohn Kelleher, Geert-Jan Kruijff, Fintan Costello<\/p>\n
This paper so far wins my “Best author list for creating a D&D
\nparty from”. <\/p>\n
They want to be able to talk to robots about the space they are in and
\ntheir surroundings. This means they need to understand spatial
\nreferences. They have looked at previous (human) research on
\nproximity and determine that proximity is a smooth function that is
\ninversely related to distance, depends on the absence or presence of
\n“distractor” objects and the size of the reference object. They use
\nfunctions to approximate proximity and merge the fields of the
\ndistractors to determine overall proximity. They used human
\nexperiments to verify their model. Presence of a distractor has a
\nstatistically significant influence. Their relative proximity
\nequation better predicted results than just absolute proximity
\n(without the normalization with the distractor.) They had a funny
\nvideo with a robot that answered questions about where things are
\n(cows, pigs, and donkeys.) <\/p>\n
They have another paper on how the descriptions are chosen to describe
\nthe scene – that would have been more interesting for me I think. The
\ntalk was very understandable though, and well-presented I thought. <\/p>\n
—<\/p>\n
Machine Learning of Temporal Relations
\nInderjeet Mani, Marc Verhagen, Ben Wellner, Chong Min Lee, James
\nPustejovsky<\/p>\n
Inderjeet Mani presenting, joint work with MITRE, Georgetown
\nUniversity, and Brandeis University. <\/p>\n
There is an annotation scheme, TIMEX, to annotate time-based
\nexpressions and some relations between them. TLINKS in TIMEX3 are
\nused to express time links using Allen’s Interval Relations. MITRE
\nhas some tools for annotating text (Tango? probably available for
\nresearch use.) TimeBank corpus and they have an Opinion corpus with
\nabout 100,000 documents!? I need to look into this. Inter-annotator
\nagreement is pretty low on TLINKs (.55F links and labels) . So they
\nare looking at, given a human-determined link, can they learn the
\ncorrect link type?<\/p>\n
They have a tool for enriching link chains with all links types
\nbetween all entities (called closure). So a small amount of markup is
\ndone, then a full graph is created that allows for temporal reasoning.
\nLearning over the full graphs improves results greatly.<\/p>\n
Had a baseline of human-created rules, and rules mined from the web
\nusing ISI’s VerbOcean. Their approach of learning with a maximum
\nentropy system over the closed data is the best. <\/p>\n
—<\/p>\n
You Can’t Beat Frequency (Unless You Use Linguistic Knowledge) — A
\nQualitative Evaluation of Association Measures for Collocation and
\nTerm Extraction
\nJoachim Wemter, Udo Hahn<\/p>\n
Looking at collocation and technical term identification. Wants to
\nsee how to determine if one statistical method can be said to be
\nbetter than another or not. Only looked at candidates that had
\nfrequency higher than 10 to avoid low frequency data. In English they
\nlooked at trigram noun phrases, in German they looked at PP-Verb
\ncombinations. <\/p>\n
Compared to t-test for collocations and terms. In each case their
\nlinguistically motivated approach compared to how often the candidates
\ncould be modified, or how much variation they showed. For their
\ncomparison they computed their terms, then split the n-best lists in
\nhalf and looked to see how much other statistical tests would change
\nthe ranking with respect to true negatives and true positives. <\/p>\n
His result tables and graphs were a little hard to follow because he
\nwent by them very quickly. I think the scatter plots were easier to
\nunderstand than the tables, and on the tables I’m not sure how
\nsignificant the differences (in percentages) were. They seemed pretty
\nclose. <\/p>\n
The overall conclusion is that frequency counting does about as good
\nas the t-test for identifying these entities. The linguistically
\nmotivated measures that are presented here do perform better than
\nfrequency counting (and the t-test) though. <\/p>\n
Question from Pascale: The linguistically motivated measure also
\nseemed to be statistical though. Doesn’t agree with the conclusion
\n(talks about Frank Smadja’s collocation work) and saying that nobody
\nproposed to use just t-score. <\/p>\n
Another question: does he do stop-listing? Yes, they do take out
\nstop words for terms. <\/p>\n
Another question: What exactly do they show in this study? What
\nwouldn’t I learn from just looking at a precision recall graph or
\ntable from X’s study (some other study cited, but I didn’t catch it.) <\/p>\n
—<\/p>\n
Weakly Supervised Named Entity Transliteration and Discovery from
\nMultilingual Comparable Corpora
\nAlexandre Klementiev, Dan Roth (UIUC Cognitive Computation Group)<\/p>\n
Use an English NE tagger to tag text, in a comparable corpus use the
\nEnglish NE tags and identify counterparts in another language
\n(Russian) and move tags across. <\/p>\n
They use temporal alignment as a supervision signal to learn a
\ntransliteration model. Identify a NE in the source language, ask
\ntransliteration model M for best candidates in Target language, then
\nselect candidate that has good temporal match to candidate NE, add to
\ntraining data for model. Re-train model M and repeat until it stops
\nchanging, or some stability has been reached. <\/p>\n
Linear discriminative approach for transliteration model, is a
\nperceptron M(Es, Et) gives a score for transliteration pair. Features
\nare substring pairs of letters in source and target. Feature-set
\n(sets of pairs) grows as they see more examples. Model is initialized
\nwith 20 NE pairs. They also look at how many \/ few pairs you can use
\nto seed model with. 5 candidates worked as well as 20, but required a
\nlot longer to converge. <\/p>\n
For their temporal similarity model they group word variants into
\nequivalence class for Russian. English isn’t as diverse so just uses
\nexact strings. Use DFT then do euclidean distance in Fourier space.
\nThey use about 5 years of short BBC articles with loose Russian
\ntranslations. <\/p>\n
Using temporal feature along with learned transliteration model they
\nget accuracy that is pretty high, up to 70% for top-5 list for a NE
\nterm. About 66% accuracy on multi-word NEs. <\/p>\n
I really liked this presentation and think it is a very novel look at
\ntransliteration. Would it work with Chinese? It is similar to
\nPascale’s early work on time warping with comparable corpora to learn
\na bilingual dictionary, but the feature set is novel to me – on the
\nfly learning of character transliteration pairs. <\/p>\n
I asked a question about application to Chinese and Japanese, and got
\nabout the answer that I expected. <\/p>\n
—<\/p>\n
A Composite Kernel to Extract Relations between Entities with both
\nFlat and Structured Features
\nMin Zhang, Jie Zhang, Jian Su, Guodong Zhou<\/p>\n
They elect to use a kernel method for a learning system (SVM?) because
\nit can easily take hierarchical features as input. Their contribution
\nis that they designed a composite kernel that combines flat features
\nand syntactic features (hierarchical) for classification. <\/p>\n
So they did a lot of work choosing which sub-trees from the parse
\ninformation to throw at their machine learning system, and ways to
\nprune the data to make it more tractable. Trained on ACE data (’03
\nand ’04.) They report that a specific type of tree kernel (mixed both
\nfeatures and syntax trees with a polynomial) has the best performance,
\nbeating state-of-the-art systems. <\/p>\n
I guess they have a nice result, but I didn’t like the talk much. The
\nspeaker focused a lot on details that I don’t think are important.
\nWhat I took away from this is that it is possible to use a tree kernel
\nto mix in numeric feature values with syntax type hierarchical
\nfeatures and achieve good performance. This seems like it should have
\nbeen a poster paper to me. (Of course, I don’t work in this area so
\nmy opinion probably isn’t worth much…) <\/p>\n","protected":false},"excerpt":{"rendered":"
Rough notes from Thursday’s presentations at COLING\/ACL. Exploiting Syntactic Patterns as Clues in Zero-Anaphora Resolution Ryu Iida, Kentaro Inui, Yuji Matsumoto A nominee for the best asian paper award. A fairly interesting approach to anaphoric identification for “missing” zero-anaphora. I can’t really say that I know much about anaphoric resolution research, but they have a […]<\/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\/124"}],"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=124"}],"version-history":[{"count":0,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/posts\/124\/revisions"}],"wp:attachment":[{"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/media?parent=124"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/categories?post=124"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fugutabetai.com\/blog\/wp-json\/wp\/v2\/tags?post=124"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}