April 23, 2004
Thomas Lux, the Bourne chair of poetry within LCC, has organized a series of poetry readings, performances and discussions at Georgia Tech. I recently attended a reading and discussion featuring British poet George Szirtes, and the US poet laureate (2001-2003) Billy Collins. Their discussion of their own creative process as poets led me to think about poetry generation, and particularly my discomfort with purely statistical approaches to poetry generation employed by systems such as gnoetry (1 2 3).
It was great to hear George and Billy both read poems and discuss the process of writing poetry, using their poems as examples. One issue they discussed was the problem of finding a balance between revealing and concealing. A poem that conceals too much from the reader becomes private language, something the reader is completely unable to enter. But a poem that reveals too much, that wears all of its meanings on its sleeve, in some sense fails to be poetry, fails to lead the reader to meanings not capturable in everyday language, fails to underlay meaning with mystery. One analogy they used for this was eye charts. On an eye chart, everyone can read the big “E” at the top of the chart. Eventually you get lines that are hard, and then impossible to read. A poem shouldn’t consist of only big Es or tiny small lines, but, like the eye chart, should have layers.
George read a love poem, and talked about the difficulty of writing love poems. On the one hand, the poet is trying to evoke an emotion that feels uniquely their own, on the other hand billions of people have felt the same emotion and tens of thousands have written poetry about it. How do you handle a topic like this without becoming trite and cliched? Billy commented that a poet is defining ideas and experiences, almost in the sense of dictionary definition, for which there are no dictionary terms. The poet is defining specific and precise senses of sadness, or love, or of an idea (e.g. of being young or old) or a response to an event, etc. An audience member commented that she was only able to write poetry during moments of excessive emotion, and asked if it’s possible to write poetry in other states. Both of them responded that it is not only possible, but necessary – that, when trying to capture a precise experience, they both work in a state of somewhat abstracted calm, sitting above the potentially traumatic or ecstatic experience they are capturing (offering a definition for), driving their minds around to understand the contours of the specific something (“Is it this, no. A bit more of this, yes…”).
Both of them talked about the difficulty of figuring out where to end a poem. Often, after getting down the main body of the poem, there something more it needs, something to truly complete it. This is related to the layering of meaning (the eye chart); if you stop too soon the poem doesn’t have enough depth, if you stop too late, the poem starts becoming confused, incoherent.
Also striking was the shear volume of poetry they could quote from memory. During the discussion they would often quote poems as examples. Obviously their own poems are deeply informed by broad and deep reading of poetry.
Now, returning to poetry generation, notice that a probablistic walk of a learned n-gram model (word co-occurrence probabilities) bears no relation to the strategies and problems described by George and Billy. What would it mean to build a poetry generator that actively reasons about the layering of meaning, about what’s been revealed and what’s been concealed? First of all, such a generator would need a notion of meaning, which n-gram models don’t have. What would it mean to have a generator that actively tries to capture and define a precise experience? Such a generator would have to have a model of experience, certainly an emotion model. But the emotion model would need to be more complex than the typical “small collection of real-valued knobs” emotion architecture in AI systems. If the system’s subjective state consisted of only a few real-valued knobs with labels like “anger”, “sadness” and so forth, then, in a model of generation in which a poem tries to capture a subjective state, what could the system say?
With a sadness of 3.2
I feel so blue and yet
goal failure has induced in me
an anger of 5.3, undirected
as I can not infer the cause
of this unlikely turn
What you’d really want is a model in which subjective state (including emotional state) is globally distributed throughout the architecture, in the manner that Aaron Sloman has been talking about for years.
Probabilistic models do capture, to some degree, the idea that good poets have read a lot of poetry. In the case of n-gram models, the word co-occurrence probabilities are learned from a corpus of poems. But unlike George and Billy, who have extracted strategies, topics, a sense of historical progression, turns of phrase and style, from the poems they’ve read, probabilistic generators at best have extracted a sense of style, but a sense of style divorced from meaning.
This is not to say that a probabilistic word selection model wouldn’t potentially have a place within a larger poetry generation architecture. But, without additional processes and structures that actively manipulate revealing and concealing, that have the goal of expressing a precise experience, generated poems won’t reliably mean anything, won’t have interesting layers of meaning, won’t have a distinct voice (which is more than style). A larger, more heterogeneous architecture is also an interesting procedural portrait of poetry creation.