A canonical semi-deterministic transducer

Transcription

1 A canonical semi-deterministic transducer Achilles A. Beros Joint work with Colin de la Higuera Laboratoire d Informatique de Nantes Atlantique, Université de Nantes September 18, 2014

2 The Results There are learning algorithms for deterministic transducers. How far toward non-determinism can we extend learning? We consider a natural extension of deterministic transducers called semi-deterministic transducers (SDTs), and we find three results: There is a canonical form for SDTs. Even with domain knowledge, they are not learnable regardless of efficiency bounds. They are learnable with very weak additional queries (with polynomial bounds). It is possible that these weak queries can be simulated through statistical methods.

3 The Model A picture to illustrate the definition. a:aa,bb q a a:aa,bb q λ b:λ q aa a:a b:ab,ba q b b:ab,ba

4 The Model A picture to illustrate the definition. aa:aa,bb q a a:aa,bb q λ b:λ q aa a:a b:ab,ba q b b:ab,ba For a transition, e, the input (i(e)) is a member of the input alphabet (Σ).

5 The Model A picture to illustrate the definition. a:aa,bb q a a:aa,bb q λ b:λ q aa a:a b:a,ab q b b:ab,ba For a transition, e, the input (i(e)) is a member of the input alphabet (Σ). The output (o(e)) is a finite set of incomparable strings over the output alphabet (Ω).

6 The Model A picture to illustrate the definition. a:aa,bb q a a:aa,bb q λ b:λ q aa a:a b:ab,ba q b b:ab,ba b:a For a transition, e, the input (i(e)) is a member of the input alphabet (Σ). The output (o(e)) is a finite set of incomparable strings over the output alphabet (Ω). From each state, there is at most one transition with input x Σ.

7 The Model A picture to illustrate the definition. a:aa,bb q a a:aa,bb q λ b:λ q aa a:a b:ab,ba q b b:ab,ba The translations of aa are AAAA, AABB, BBAA and BBBB. The number of translations of a string can grow exponentially in the length of the string, for a fixed transducer.

8 The Model One final assumption: there is a unique initial state. This is for convenience. A few useful observations: λ is never the input of transition. If λ is in a set of outputs, it is the only member. Every string has a unique path through an SDT. Every string has a finite number of translations. These observations allow us to form a tree of outputs analogous to one for a sequential or subsequential transducer. In our case it is a tree of output trees.

9 The Tree of Output Trees Consider the tree of inputs that will not be rejected. λ a b aa bb aaa aab bba bbb aaaa aabb aaab aaba bbaa bbbb bbab bbba

10 The Tree of Output Trees The tree is the domain of a function to P(Ω ). λ {λ} a {AA, BB} b {AB, BA} aa {AAAA, AABB, BBAA, BBBB} bb {ABAB, ABBA, BAAB, BABA} (further outputs omitted) aaa aab bba bbb aaaa aabb aaab aaba bbaa bbbb bbab bbba

11 The Tree of Output Trees By looking one level higher in the ranked universe, a relation that was not well-defined has become a well-defined function. This function is the semantic representation of the transducer. In OSTIA and other state-merging algorithms, there are two phases: onwarding and merging. Onwarding is performed for deterministic transducers by finding common prefixes for translations comparing outputs in a much simpler tree of outputs. We demonstrate that this is possible in the much more general case of SDTs.

12 Onwarding The existing technique: Fix a dataset. λ : λ a : AB ba : AAB aaa : ABBA abb : ABAA bba : AABB bbb : AABA

13 Onwarding The existing technique: Fix a dataset. Consider inputs with a common prefix. λ : λ a : AB b :? ba : AAB aaa : ABBA abb : ABAA bba : AABB bbb : AABA

14 Onwarding The existing technique: Fix a dataset. Consider inputs with a common prefix. List the outputs of those inputs. λ : λ a : AB b :? ba : AAB aaa : ABBA abb : ABAA bba : AABB bbb : AABA

15 Onwarding The existing technique: Fix a dataset. Consider inputs with a common prefix. List the outputs of those inputs. Find the prefixes common to all the outputs. λ : λ a : AB b : A, AA or AAB ba : AAB aaa : ABBA abb : ABAA bba : AABB bbb : AABA

16 Onwarding The existing technique: Fix a dataset. Consider inputs with a common prefix. List the outputs of those inputs. Find the prefixes common to all the outputs. Choose a best common prefix to onward. λ : λ a : AB b : AAB ba : AAB aaa : ABBA abb : ABAA bba : AABB bbb : AABA

17 Onwarding To understand our new technique, we examine sets of prefixes. Consider the translations of aabb using the earlier example: {AAAAAB, AAAABA, AABBAB, AABBBA, BBAAAB, BBAABA, BBBBAB, BBBBBA} AAAAAB AABBAB AAAABA AABBBA BBAAAB BBBBAB BBAABA BBBBBA

18 Onwarding To understand our new technique, we examine sets of prefixes. Consider the translations of aabb using the earlier example: {AAAAAB, AAAABA, AABBAB, AABBBA, BBAAAB, BBAABA, BBBBAB, BBBBBA} AA BB AAAAAB AABBAB AAAABA AABBBA BBAAAB BBBBAB BBAABA BBBBBA

19 Onwarding {AA, BB} is a maximal antichain of the tree. We call it a valid antichain because the tree below AA is identical to the tree below BB. This is a consequence of AA and BB being outputs of the same transition. For a tree T and x T, T x is the tree of suffixes of x in T. Definition Given a tree, T, a subset S of T is a valid antichain if S is a maximal antichain and ( x, y S)(T x = T y ). Definition For P and Q, sets of strings over some common alphabet, we say that P < ac Q (P is antichain less than Q) if either 1. P < Q, or 2. P = Q and, for all x P and y Q, if x y, then x y.

20 Onwarding Theorem For a finite set of strings, S, the valid antichains of S are a finite linear order under < ac. Consequence: A canonical way of decomposing trees into initial segments. The new technique: Consider all inputs that have a common prefix. List the trees of outputs for those inputs. Find the valid antichains of the trees. Pick a valid antichain common to all of the trees.

21 Conclusion Given access to translation queries, the canonical form can be found and SDTs can be learned from a polynomial size characteristic sample. In other words, there is a single algorithm that learns every SDT under these conditions. Without translation queries, this is not possible. One possible avenue to pursue is to determine if translation queries can be replaced with statistical analysis: both provide negative information. This will require a careful development of probabilistic model.

22 Thank You