邢唷��>� 9;��8��欹�� 3)bjbj螝螝 .03!��LL��0000<08Pfffffff��,W� ��fffff�ff�fff�f�:b�� y�`椡��R��0Z��bb fff��fff��fffffffffL��: Word Sense and Sensibility by Susan Windisch Brown, ICS and the Dept. of Linguistics A ham sandwich walks into a bar and says, 揑抣l have a beer.� The bartender replies, 揑抦 sorry. We don抰 serve food here.� Groaners like the one above rely on our ability to match the correct meaning of a word to the current context. Or, actually, they depend on our initial mismatching. What is really astonishing, however, is how many words have multiple meanings and how infrequently we misinterpret the meaning. Of the words listed in WordNet, a dictionary-like lexical resource, nouns have an average of 3 meanings and verbs an average of 4 meanings. These numbers suggest that a very simple sentence with 2 nouns and a verb would have, on average, 36 possible meanings. Of course, the context allows us to eliminate the inappropriate ones. But how exactly do we do this, and how do we do it so quickly and easily? Not only is the question intriguing, it also has larger implications for a wide variety of fields. Knowing how the brain stores and accesses the meanings of words could improve treatment for certain kinds of language disorders or improve word learning for children with delayed vocabulary development. In addition, a clearer understanding of this process in human beings could give us some insight into improving computer imitation of this process. A common theory in linguistics and psychology sees our mental lexicon as a kind of dictionary, with each word connected to mental representations of meaning, one for each sense of the word. According to this theory, when we hear a certain sequence of sounds or read a sequence of letters, we 搇ook up� the word in our mental lexicon, do a quick search through the connected meaning representations, and settle on the correct one. This theory has many challengers, however. Walter Kintsch, former director of ICS and Professor Emeritus of the Psychology Department has advocated a far different view of the mental lexicon. Rather than each word having a 搇ist� of separate senses, he proposes that a word has one complex semantic representation, with the correct nuances of meaning emerging from the word抯 combination with other words in an utterance. He has investigated this theory in numerous ways using Latent Semantic Analysis, which represents word meaning with vectors in a high-dimensional space (Kintsch, 2001). For instance, the verb serve would be represented as a very long list of numbers, indicating the frequency that serve occurs with other words, including food words, and the words tennis and ball. Kintsch抯 theory梩he Construction-Integration model梪ses a highly sophisticated mathematical algorithm in combination with these vectors to determine which sense of serve is appropriate for a given context (Kintsch, 1998). In addition to fancy mathematics and computer algorithms, brain imaging techniques are increasingly used to investigate these questions. A recent MEG (magnetoencephalography) study (Beretta et al., 2005) found neural correlates to the 損olysemy effect.� That is, words with multiple related meanings are recognized as words more quickly than single-meaning words or words with unrelated meanings (homonyms). This imaging technique indicates that brain activity differs in these two circumstances. Here in ICS, graduate student Vicky Lai from the Department of Linguistics is using event-related potentials (ERP) technique to study the comprehension of literal meaning and conventional and novel metaphors, which rely heavily on polysemy. Getting computers to successfully distinguish between multiple word senses has been a vexing problem in natural language processing. The most common approach has been to label the words in a large corpus of text with the appropriate senses and then train a machine learning program on that data. This method begs the question of what are the appropriate senses of a word. Most projects of this sort have used the senses of words given in WordNet, the freely available online resource describe above. WordNet makes very fine distinctions in meaning for each word. For example, break in 慴reak the glass� is considered a different sense from break in 慴reak the radio,� or draw in 慸raw a line� is a different sense from draw in 慸raw a tree.� Fine-grained sense distinctions such as these mean that only one or two instances of a sense may be found in a very large corpus of text, making it difficult for machine learning programs to discern any patterns. In addition, people annotating texts with these senses disagree quite often on which sense is meant in a particular context. For most, if not all tasks, we do not need computers to distinguish a word抯 meaning more precisely than we ourselves do. Successful word sense disambiguation by computers depends on deciding which senses of a word are distinct enough to warrant a separate representation for the task. These problems and issues have lead researchers to explore the usefulness of coarser-grained sense distinctions. Martha Palmer, ICS Faculty Fellow and professor in the Department of Linguistics, has led an ambitious project to merge the fine-grained senses of WordNet into coarser grained senses (2001, 2006). These senses are then used to label the words in a large collection of text. Several ICS students, including Jill Duffield, Jena Hwang, Sarah Vieweg and myself, have contributed to the sense-merging portion of this project (Duffield et al., forthcoming). My own research investigates mental representations of word meaning through psycholinguistic experimentation. Subjects are asked to judge the semantic coherence of short phrases, that is, whether or not the phrases 搈ake sense.� Subjects are primed with a phrase using one meaning of a verb and then see another phrase with a different meaning of the same verb. Because the meanings of polysemous words often blend from one to the other, a polysemous word can have meanings that are quite distantly related and some that are quite closely related. The pairs in this experiment fall into several categories of relatedness: unrelated (toast the bread/toast the host); distantly related (served the soup/served the country); very closely related (drew a line/drew a tree); syntactically related (rang the bell/the bell rang); and same sense (cleaned the shirt/cleaned the cup). Previous research (Klein and Murphy, 2001) with a similar task found no difference in the cognitive processing between homonyms and polysemous words. The authors took this as support for the theory that we have separate semantic representations for every sense of a word. By splitting the category of polysemy into several categories representing degrees of relatedness, and comparing these categories to both homonym and same-sense pairs, I hope to discover differences in the processing of meaning between them. Differences in processing time may indicate differences in the semantic representations, with closely related senses sharing part or all of their semantic representations and distantly related senses sharing little to no portion of their semantic representations. My hope is both to refine our understanding of the storage and processing of word meaning and also to improve the selection of reasonable sense distinctions for word sense discrimination by computers. Beretta, A., R. Fiorentino, and D. Poeppel. 2005. The effects of homonomy and polysemy on lexical access: An MEG study. Cognitive Brain Research 24: 57-65. Duffield, C.J., Hwang, J.D., Brown, S.W., Vieweg, S.E., Davis, J., & Palmer, M. (accepted). Criteria for the manual grouping of verb senses. In Proceedings of the Linguistic Annotation Workshop, Prague, June 2007. Association for Computational Linguistics. Kintsch, W. 2001. Predication. Cognitive Science 25: 173-202. ------. (1998) Comprehension: A paradigm for cognition. New York: Cambridge University Press. Klein, D., and G. Murphy. 2001. The representation of polysemous words. Journal of Memory and Language 45: 259-82. Martha Palmer. 2000. HYPERLINK "http://verbs.colorado.edu/~mpalmer/papers/chum.ps.gz" Consistent Criteria for Sense Distinctions. Computers and the Humanities, SENSEVAL98: Evaluating Word Sense Disambiguation Systems, Kluwer, 34: 1-2, 2000. Palmer, Martha, Hoa Trang Dang, and Christiane Fellbaum. 2006. Making fine-grained and coarse-grained sense distinctions, both manually and automatically. Natural Language Engineering, July. uz��"��05pu��k%�%�%m&�&|'�'�'�'�'�'(((( (r(),)3)栎｜湐嚁噠嚁渢滭鼽h�.7h@�0Jh�.7h@�0J5�jh�.7h@�5乁h�.7h@�5�h�.7h@�h6h@�'h@�0JB*CJOJQJ^JaJph-h6h@�0JB*CJOJQJ^JaJphhZ\�h@�6�huJdh@�B*phh@�B*phhuJdh@� h@�6�h@�*UV��Z [ 67�� $�$k%m&��勑dh`勑gd@�dhgd@�m&�& '|'s(3)��dhgd@�勑dh`勑gd@�,1恏靶/ 班=!�"�#悹$悹%�靶靶愋��666666666vvvvvvvvv666666>666666666666666666666666666�6666666666�666666666666hH66666666666666666666666666666666666666666666666666666666666666666�66666666662�� 0@P`p€��2(�� 0@P`p€�� 0@P`p€�� 0@P`p€�� 0@P`p€�� 0@P`p€�� 0@P`p€�8X�V~_HmH nH sH tH @`�@NormalCJ_HaJmH sH tH DA`��DDefault Paragraph FontRi��RTable Normal�4� l4�a�(k ��(No List*W@��*�8Strong5乗�.X@�.�8Emphasis6乚丯g@�N6HTML TypewriterCJOJPJQJ^JaJPK!涜pO�[Content_Types].xml瑧薺�0E鲄䞍卸豶�(ヘ微廬嬼yl嬝#!MB蝼;.�卬姨谨台\艫�1&绌耀夹生芉W橄蚄v疷b�OX�#&疆释1`RⅵT闉9<搇�#ぼ$┐>幚r崫 `沸」-�;c=1g词'縸�虫笔咓颈叠$篒�翱�1脢办9报滨鬰尝�涢贬�隮<�;*v�7'榓瓻\�h>嗟�=青镑,�*彿阗�8辩赌;湷镑�*4髑槿?�奥辩调苍阍塷驳淎撙>曧�8蹿�2�*<")QH裤麆xK┟� 梶ㄐ]奪榸)橛亾MS欯\&>!�7;飞笔�3熥摆淅贰叠鲍缚�1綩楥�5VD翕� 豖靉?p墾勠 S喧4坻曐[驨S叫28;嶻慵[戢櫌�,荇撘骉1|筅n�;�+衡/蕰j�蟎阉\,E:!莒宼�4.T獭e1 �謢�; [鄗^鵳縧�銨�ok侮0e g@礕GHPXNT,賲禿诧e鮸楸*YdT臷橸�鋩熬顏+�(員7�$ow2�茜紓#紾┲浻墒ビ掷?絨迸N澵K椟-�/M,W鏖阦鍗x盕V�/Q吴范朞鯕&渆刀c蛒几\Q滜她LW@H��!�+魝阽{衃|{包!汝K韩A�i `遚m2iU栚|纾Y+�蕤栿� [[v杧煋靣錘缭釫挐3靝m擎R �=Y�04,�!&0�+W魿軆@o習�樫OS2�'S佼0�5嘌$�嗓]pm珠3孎瞭�崕婓G蓜-!鹹へ"�訅V .� 甡幾�,�O.%烫胁Kas楁S球骗弙€稭謟瓔玚姤叢3铵蘽�9+e鰻@�e觇諗錖y硥�7啠W萧_Xtl斴��PK! 褠煻'theme/theme/_rels/themeManager.xml.rels剰M �0匃倃oo雍�&輬协�勪5 6?$Q祉 �,.嘺緳i粭澤c2�1h�:闀q毩m胳嶡RN壻;d癭値o7�g慘(M&$R(.1榬'J摐袏T鶂�8V�"＆A然蠬鱱}狇�|�$絙{�朠�除8塯/]As賲(⑵锑#洩L蔥汉倪��PK-!涜pO�[Content_Types].xmlPK-!ブх�6-_rels/.relsPK-!ky��theme/theme/themeManager.xmlPK-!!Z!��theme/theme/theme1.xmlPK-! 褠煻'( theme/theme/_rels/themeManager.xml.relsPK]# 3!0��3)m&3)�� 3!X��$,>EM T # , � � ��T[ CM|��mt?I� � � � � � 5!��>K��Gj��@A2!5!::::::::::[�襣l��勑剺��^勑`剺�.€劆剺��^劆`剺�.€刾剺��p^刾`剺�.€凘剺��@^凘`剺�.€�剺��^�`剺�.€勦剺��^勦`剺�.€劙剺��^劙`剺�.€剙剺��€^剙`剺�.€凱剺��P^凱`剺�.[��Be@�3!5!�@襂{3!�@��Unknown��G��*郃x� �Times New Roman5�€Symbol3��*郈x� �Arial?� �*郈x� �Courier NewA��$B�Cambria Math"q�鹦h-� �-� ��?<��?<�!��亖24#!#!� 3僎��HP)�?��,2!xx�,��ICS newsletter articleSW BrownWendy Turnbull� 鄥燆鵒h珣+'迟0�� ( LX dpx€�'ICS newsletter articleSW BrownNormal.dotmWendy Turnbull2Microsoft Macintosh Word@@^� `椡@^� `椡�?� 胀諟.摋+,D胀諟.摋+,d hp�� '$Got a Pen & I Know how to use it<�#!ICS newsletter articleTitle� 8@_PID_HLINKS'A�MO5http://verbs.colorado.edu/~mpalmer/papers/chum.ps.gz �� !"#$%&'��)*+,-./��1234567��:��Root Entry�� F4坞`椡<�€1Table��WordDocument��.0SummaryInformation(��(DocumentSummaryInformation8��0CompObj��`�� F Microsoft Word 97-2004 Document��NB6WWord.Document.8