Over the past year, the University of Virginia Library's Japanese Text Initiative (JTI), in partnership with the Electronic Text Center (Etext), has been developing an online multimedia dictionary of haiku. In this paper, I will describe the scope and content of the project, and how it combines text and multimedia materials in unique ways. I will also discuss the project from a technical perspective, describing its encoding, interface, and search and delivery mechanisms, with an emphasis on new problems encountered in the course of development. Haiku is a well-known and increasingly ubiquitous poetic form in the West and around the world. But while many non-Japanese are familiar with the metrical structure of haiku poetry, few have a sophisticated knowledge of its traditional themes, which are based upon deeply nuanced kidai and kigo (season-words). At the heart of the JTI project is an English translation of portions of the Nyumon Saijiki, a standard reference of kidai and kigo produced by the Museum of Haiku Literature in Tokyo. The Nyumon Saijiki contains entries for the people, places, and things found in haiku, with examples of modern and classical haiku under each entry. Our translation of this work, completed by world-renowned haiku expert William Higginson, consists of kanji, hiragana, romaji, and English renderings for each of the 3,000 kidai and kigo, along with a definition and brief explication. For the 100 most important terms, Mr. Higginson has translated the full entry, which includes a discussion of proper usage and several example poems. In addition, we have supplemented these full entries with image and sound files harvested from online archives. Audio and visual illustrations of these recurrent themes (the cries of autumnal insects, the hazy moon of a spring night, and so forth) will greatly enhance the user's understanding of poetry that is so deeply rooted in sensory perceptions of nature; this is doubly true for Western audiences who live half a world away from the particular plants, animals, rituals, and seasonal phenomena discussed in the poems. Although Etext has produced and maintained full-text SGML/XML collections for over a decade, and the JTI's work with Japanese language materials dates back to 1997, the current project has presented us with many new challenges. In the past, the use of images in our collections was limited primarily to page images and book illustrations; sound files were all but non-existent. The project's web designer has strived to integrate the multimedia components into the project in a way that supplements but does not overwhelm the textual content. The potential audience for this project is incredibly broad; we anticipate that it will be used around the world by everyone from grade-school students to professional scholars of Japanese literature. Different users will require different views of the data and different tools for manipulating it. A K-12 audience might want to see just the English translations of the poems, accompanied by the relevant images, while someone learning the Japanese language would be better served by a poem in its kanji and romanized versions, along with an audio recitation and highlighted links from the keywords in the poem to their full dictionary entries. More expert users might want to run sophisticated search queries on the dictionary, and a student of comparative literature might want to cross-query the dictionary against the English Poetry Database or a similar collection. In order to present these multiple views of the project, we have encoded the data in XML-compliant TEI and have generated much of the output with XSL stylesheets. This represents a major transition for the Etext Center, which since 1992 has stored data in SGML and converted it "on the fly" to HTML with CGI scripts. The nature of the saijiki data, with its deeply nested hierarchies and the database-like entry structure, combined with our need to sort, recombine and re-present this data very nimbly, led us to an XML/XSL workflow. However, we made some sacrifices with this approach. Our on-the-fly delivery of the data through XSL stylesheets has been considerably slower than a comparable CGI transformation. For this reason, we currently deliver large chunks of data through CGI or as XSL-generated static HTML pages while we work to reduce the delivery time. A further complication has arisen from our need to use UTF-8 encoding for this project, instead of the EUC encoding we have used in the rest of the JTI collection. Unicode allows us to easily represent kana and hiragana alongside Western characters, including macron vowels that aren't part of the EUC character set. However, our OpenText search tool does not support Unicode; until current work on Unicode compatibility for OpenText is completed, we need to convert the Unicode to EUC prior to indexing and essentially maintain two parallel versions of the same data. In the coming year, we look forward to translating the remaining portions of the Nyumon Saijiki, adding to the supplementary multimedia materials, and integrating this haiku collection both "vertically" (with our other Japanese-language materials) and "horizontally" (with other poetry collections and dictionary projects across our holdings).