4. SyntaxNetII (from Google)
● Major upgrade to SyntaxNet (Mar/15, 2017)
– Multilingual language understanding
– Joint modeling of multiple levels of linguistic structure
– Neuralnetwork architectures to be created dynamically during processing of a
sentence or document
● Implementation*
: combination of below 2.
– Recurrent multitask parsing model (with DRAGNN)
– Characterbased representation (with LSTM) as input to DRAGNN
● DRAGNN as the new core
– Architecture as a series of modular unit(TBRU)s
– Connections between modules are unfolded dynamically.
→ Dynamic Recurrent Acyclic Graphcal Neural Networks
* Alberti et al. [2017] SyntaxNet Models for the CoNLL 2017 Shared Task ,
(paper) https://arxiv.org/abs/1703.04929 (code) https://github.com/tensorflow/models/tree/master/syntaxnet
7. Transition System
● Transition System : T
– T = {S, A, t}
– S(x) : Set of States ( s+
∈ S(x) : Start State )
– x : input (eg. sentence)
– A(s, x) : Set of allowed decision for any s ∈ S
– transition function t(s, d, x)
● s’ = t(s, d, x) : new state s’ for any decision d ∈ A(s, x)
● s’ = t(s, d) for brevity
● Complete structure : sequence of state/decision pairs
– (s1
, d1
)...(sn
, dn
)
– s1 = s+
, di
∈ A(s, x)
– si+1
= t(si
, di
)
8. Transition Based Recurrent Unit (TBRU)
● m(s) : Input embedding function
– eg. lookup op m(s) : S R→ k
● r(s) : recurrence function (Connection to previous states)
– mapping : states set of previous time steps→
– r(si
) : S P{1, …, i1} ( P : power set, so n(P) variable)→
● (RNN) Network Cell : computes new hidden representation
– hs
= RNN( m(s), {hi
| i ∈ r(s)} )
( Logically : r,m h d )→ →
di
argmax← d A(si)∈
wd
T
hi