Case Studies — Olga Velichkina¶
Three explorations testing Quagga Agent on Irish traditional music topics across The Session, MusicBrainz, GoTriple, and Wikidata. Each study includes a critical assessment of agent reasoning and knowledge-graph coverage.
Exploration 1: Down by the Salley Gardens¶
| Resource | Link |
|---|---|
| Archived conversation | PDF export / Markdown export |
The Irish folk song Down by the Salley Gardens is known to most musicians through the version arranged by Benjamin Britten on the poem by W. B. Yeats (see this annex for additional information about this song). I wanted to learn more about the origin of this song, its recorded and notated versions, its sources, and its poetic texts. An important point is that folk songs frequently exist in many melodic and textual versions, and their titles do not always match. Britten chose a version associated with Yeats\'s poem, but the same melody is also known with another text, The Maids of the Mourne Shore/Moorlough Shore.
The search process was rather slow, and some important sources were absent from the knowledge graph (KG), such as the Roud Folk Song Index, one of the most important sources for song texts. The agent was able to retrieve information about editions of Britten's arrangement, which appeared in two anthologies: Folksong Arrangements, Volume 1: British Isles and Britten: Complete Folksong Arrangements (High) in the UTSI graph. It then claimed that TheSession database does not contain this song (under either title, The Salley Gardens or Moorlough Shore), which is entirely incorrect. A simple search on TheSession.org shows otherwise.
In a second round, the agent acknowledged that the song is present in TheSession and identified two entries, nos. 1314 and 98, asserting that they represent the same tune. This is, again, incorrect: a simple musical analysis reveals that these are two completely different melodies, coincidentally associated with the same title (although spelled slightly differently: Salley and Sally). What is troubling here is that the agent attempted to justify its conclusion:
"The key finding is that tune 1314 (https://thesession.org/tunes/1314) carries both 'Down By The Sally Gardens' and 'The Maids of the Mourne Shore' as names for the same tune --- confirming they are the same air/melody under different titles. This is consistent with the well-known tradition that the air 'The Maids of the Mourne Shore' was repurposed by W. B. Yeats for the poem 'Down by the Salley Gardens.'"
"Tune 1314 has 9 settings (melodic variants, linked via P747), and tune 98 has 8 settings. Both are in the The Session graph."
The wording is suspicious. A human would never say "consistent with a well-known tradition" in this context; one may speak of a fact here, but not of a tradition. Moreover, the air was not "repurposed" by Yeats, who had nothing to do with the use of his poem with this melody.
On a subsequent query, the agent offered another false argument:
"There is no direct link between the two tunes, but there is an indirect relationship: they co-appear in three shared tunesets."
According to the agent, this reflected "their real-world melodic kinship."
The problem with this argument is that tunesets are compiled by users, and the logic or motivation behind these compilations is entirely subjective. No conclusions about relationships between tunes can be drawn simply from their co-occurrence in user-created sets.
Finally, the agent explained that it could not draw conclusions about melodic relationships between tunes on the basis of the data available to it, since it had access only to metadata and not to incipits or musical notation. This is correct, but it could have been stated much earlier.
Looking back on this interaction, I realize that my questions were formulated rather awkwardly, since I myself knew little about the topic and was merely exploring it. I asked about the origin of the song and its traditional version, meaning the earlier (pre-Yeats) version associated with the title The Mourne Shore. The agent searched for the phrase "Irish traditional" and found it in a recent work by the American composer John Corigliano (Three Irish Folksong Settings), who produced his own arrangement of Britten's song. The designation "Irish traditional" appeared in the "author" field, which the agent interpreted as an explicit reference to a traditional version. In fact, however, the answer was unsatisfactory, since this was simply another arrangement of the same tune.
In short, the limitations of the original KG (lack of data, such as musical notation or incipits, and inconsistencies in field completion), the agent's faulty reasoning, and my own lack of experience in formulating effective prompts all contributed to the limited results of this search.
Exploration 2: Music for the Irish (Celtic) Harp¶
| Resource | Link |
|---|---|
| Archived conversation | PDF export / Markdown export |
The task was to identify as many sources as possible relating to the historical and modern repertoire of the Irish (Celtic) harp (clarsach). Initially, I requested a list of composers who used the Irish harp in their music. QUAGGA compiled a list of 53 composers and performers based on the MusicBrainz KG. This list included several groups as well as a number of kora players (or so it appeared). On the other hand, some well-known composers and performers whom I expected to find, such as Turlough O'Carolan (an eighteenth-century Irish composer) and Alain Stivell (a modern Breton interpreter of Celtic harp music), were absent.
O'Carolan: a second query specifically requesting a list of his works returned a substantial number of entries from the KG. However, because QUAGGA had not identified him as a composer who wrote specifically for the Irish harp, he was omitted from the first list. It should also be noted that searching simply for "Carolan," rather than "Turlough O'Carolan," produced unrelated results (e.g., The Session tunes nos. 5842 and 78830, and probably many others, which would require manual verification).
This search was not completed because the system crashed when I asked a question requiring both GoTriple and Wikidata. The error message was:
"\<Response payload is not completed: \<TransferEncodingError: 400, message=\'Not enough data to satisfy transfer length header.\'>"
Exploration 3: Tunes in 7/8 Time¶
| Resource | Link |
|---|---|
| Live conversation | TODO |
| PDF export |
I requested a complete list of pieces in 7/8 time in order to examine their geographical and historical distribution across different European musical traditions. The agent identified 170 tunes, all of them from The Session. This result seemed counterintuitive from the outset, since it is well known that this meter is highly untypical of Irish traditional music.
The agent explained that:
"The time signature is encoded via the Wikidata property P3440 (time signature) pointing to entity Q50353331 (7/8)."
However, checking the listed examples revealed that most of them were actually in 3/2 rather than 7/8. One exception was the Greek tune Kalamatianos, which indeed uses a 7/8 meter, but it is clearly an exception within The Session corpus. We went through several rounds of questioning, but the incorrect answer persisted.
When I eventually asked what entity Q50353331 actually represented in Wikidata, the agent suddenly arrived at the following conclusion:
"The entity Q50353331 is labeled '3/2' (three-two time signature) in Wikidata --- not '7/8'. This confirms your earlier observation that the 170 tunes previously listed under Q50353331 are actually in 3/2 meter, not 7/8. The Wikidata URI Q50353331 represents the 3/2 time signature, meaning the data in The Session graph that links those tunes to Q50353331 is consistent with 3/2, not 7/8."
I cannot determine the reason for this persistent error.
Overall, the agent needs to improve both its speed and the reliability of its answers. Part of the problem appears to stem from deficiencies in the underlying KG, but some issues are clearly attributable to the agent's reasoning processes.