Clippers: A computational linguistics discussion group (Meurers, 795Y, Winter 2005)

Clippers is our forum for informal discussion of all issues related to computational linguistics: from work in progress of visitors and people in the department, over presentation of new papers, to practical concerns such as hints on the use of CL related software tools.

Everyone with an interest in computational linguistics is most welcome!

To see what happened in previous quarters of Clippers, you can check out the pages of some previous quarters: Autumn 04, Spring 04, Winter 04, Autumn 03, Spring 03, Winter 03, Autumn 02, Spring 02, Autumn 01

When and where:Tuesdays at 17³⁰-18⁴⁸ in 340 Central Classrooms.

Important: Please be sure to subscribe to our local computational linguistics mailing list on which all Clippers sessions and talks are announced.

The plan, as usual, is to start each session with 5-10 minutes on whatever someone wants to bring up and then to continue with the following topics:

1. Tue, 4. Jan.: Organization
2. Tue, 11. Jan.: no meeting, in support of ACL paper writing ;-)
3. Tue, 18. Jan.: no meeting
4. Tue, 25. Jan.: Laura Stoia on Populating Semantic Classes using Large Scale Corpora
5. Tue, 1. Feb.: Markus Dickinson and Detmar Meurers on Detecting errors in discontinuous structural corpus annotation
6. Tue, 8. Feb.: Tianfang Xu on Finding Landmarks
7. Tue, 15. Feb.: Yang Shao and Soundararajan Srinivasan on On building large-vocabulary speech recognition systems
   Abstract: CSE 888R04 was started in Spring 2004 to provide experience in building large-vocabulary continuous speech recognition systems. In this talk we will describe the motivation for and the challenges in building such systems and then detail our preliminary work on the Wall Street Journal corpus. We will conclude with some thoughts on future systems/approaches that we intend to pursue.
8. Tue, 22. Feb.: Mona Diab (Computational Linguistics Center for Computational Learning Systems (CCLS), Columbia University) on Bootstrapping an Arabic WordNet: Issues of scale and representation

   NOTE the unusual time and place: 3:30 in 122 Oxley Hall

   Abstract: I propose the automatic bootstrapping of a Modern Standard Arabic WordNet on the lexeme level using Arabic English parallel corpora and an English WordNet. I address the feasibility of such an endeavor and present a qualitative evaluation of the meaning correspondences cross linguistically between Arabic and English. I further present an automatic means of performing this task using an unsupervised Word Sense Disambiguation System. I test the feasibility of the bootstrapping by qualitatively evaluating the meaning definition projection of English words onto their Arabic translations. I manually evaluate 447 word instances of the Arabic words that correspond to correctly sense tagged English words using English WordNet 1.7. from the SENSEVAL 3 data. The words evaluated correspond to Nouns, verbs, adjectives in English. I find that for Arabic verbs, adjectives and nouns, on average 52.3% of all the words examined, the corresponding English WordNet set of definitions are sufficient as definitions for the Arabic translation word; 39.96% of the Arabic words correspond to specific subsets of the WordNet definitions; and finally, 7.8% of the Arabic words comprise supersets of their corresponding English WordNet translation definitions. These results are very encouraging as they are similar to those obtained by researchers building EuroWordNet. Moreover we present an evaluation of the automatic creation of an Arabic WordNet using our unsupervised system SALAAM utilizing Senseval 2 data. Finally, I will discuss the appropriateness of the lexeme level as the granularity of representation for an Arabic WordNet
9. Tue, 1. Mar.: Wilbert Heeringa (University of Groningen) on Measuring Norwegian Dialect Distances using Word Pronunciation Transcriptions and Acoustic Word Samples
   Abstract: The term "dialectometry" was coined by Jean Séguy, who was director of the Atlas linguistique de la Gascogne. On the basis of the material of this atlas he measured linguistic distances between dialects, defining distance as the number of items on which two dialects disagree. A more refined approach was introduced by Kessler in 1995. He used Levenshtein distance as a tool for measuring distances between Irish Gaelic. Levenshtein distance is equal to the minimum cost of changing one word pronunciation into another. The minimum cost is based on the weights of insertions, deletions and substitutions, the three operations allowed to be used for changing the word pronunciation.

   In this paper we apply Levenshtein distance to 15 Norwegian dialects, using a data set compiled by Jørn Almberg. First, we apply Levenshtein distance to word pronunciation transcriptions, just as Kessler did. However, a novel element here is the weighting of insertions, deletions and substitutions by using acoustic segment distances. We examine different acoustic segment distances: the Barkfilter, the cochleagram and the formant-track representation. Second, we apply Levenshtein distance directly to a acoustic representations of the word samples, nearly without using information provided by the transcription. The minimum cost is based on the weights of insertions, deletions and substitutions of spectra or formant bundles, rather than phonetic transcription segments. Again, the Barkfilter, the cochleagram and the formant-track representation are examined.

   Finally, we validate our results by comparing them with the results of a perception experiment, carried out by Charlotte Gooskens in the spring of 2000. She measured perceptual distances between the 15 Norwegian dialects by presenting recordings of the same dialects to a class of school children in each of the 15 dialect locations. The pupils rated the distances of the dialects to their own dialect on a scale of 1 (=similar) to 10 (=completely different). We found significant correlations between the Levenshtein distance measurements and the perceptual measurements.

10. Tue, 8. Mar.:
    • Stacey Bailey on On (not quite) recognizing textual entailment
    • Ilana Bromberg on An update on Systematicity in the Arabic Lexicon