diff --git a/requirements.txt b/requirements.txt
index aae0f9388..fa1a3e6d3 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -3,7 +3,7 @@ pathlib
 numpy>=1.7
 cymem>=1.30,<1.32
 preshed>=1.0.0,<2.0.0
-thinc>=6.8.0,<6.9.0
+thinc>=6.8.1,<6.9.0
 murmurhash>=0.28,<0.29
 plac<1.0.0,>=0.9.6
 six
diff --git a/setup.py b/setup.py
index 0a3384ed5..73c38b14a 100755
--- a/setup.py
+++ b/setup.py
@@ -193,7 +193,7 @@ def setup_package():
                 'murmurhash>=0.28,<0.29',
                 'cymem>=1.30,<1.32',
                 'preshed>=1.0.0,<2.0.0',
-                'thinc>=6.8.0,<6.9.0',
+                'thinc>=6.8.1,<6.9.0',
                 'plac<1.0.0,>=0.9.6',
                 'pip>=9.0.0,<10.0.0',
                 'six',
diff --git a/spacy/_ml.py b/spacy/_ml.py
index 0d67ce01e..d6116cfe5 100644
--- a/spacy/_ml.py
+++ b/spacy/_ml.py
@@ -8,7 +8,8 @@ import random
 
 from thinc.neural._classes.convolution import ExtractWindow
 from thinc.neural._classes.static_vectors import StaticVectors
-from thinc.neural._classes.batchnorm import BatchNorm
+from thinc.neural._classes.batchnorm import BatchNorm as BN
+from thinc.neural._classes.layernorm import LayerNorm as LN
 from thinc.neural._classes.resnet import Residual
 from thinc.neural import ReLu
 from thinc.neural._classes.selu import SELU
@@ -19,7 +20,9 @@ from thinc.api import FeatureExtracter, with_getitem
 from thinc.neural.pooling import Pooling, max_pool, mean_pool, sum_pool
 from thinc.neural._classes.attention import ParametricAttention
 from thinc.linear.linear import LinearModel
-from thinc.api import uniqued, wrap
+from thinc.api import uniqued, wrap, flatten_add_lengths
+from thinc.neural._classes.rnn import BiLSTM
+
 
 from .attrs import ID, ORTH, LOWER, NORM, PREFIX, SUFFIX, SHAPE, TAG, DEP
 from .tokens.doc import Doc
@@ -192,17 +195,17 @@ def Tok2Vec(width, embed_size, preprocess=None):
         suffix = get_col(cols.index(SUFFIX)) >> HashEmbed(width, embed_size//2, name='embed_suffix')
         shape = get_col(cols.index(SHAPE))   >> HashEmbed(width, embed_size//2, name='embed_shape')
 
-        embed = (norm | prefix | suffix | shape )
+        embed = (norm | prefix | suffix | shape ) >> LN(Maxout(width, width*4, pieces=3))
         tok2vec = (
             with_flatten(
                 asarray(Model.ops, dtype='uint64')
-                >> embed
-                >> Maxout(width, width*4, pieces=3)
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)),
-            pad=4)
+                >> uniqued(embed, column=5)
+                >> drop_layer(
+                    Residual(
+                        (ExtractWindow(nW=1) >> BN(Maxout(width, width*3)))
+                    )
+                ) ** 4, pad=4
+            )
         )
         if preprocess not in (False, None):
             tok2vec = preprocess >> tok2vec
diff --git a/spacy/pipeline.pyx b/spacy/pipeline.pyx
index 947f0a1f1..b9f59681e 100644
--- a/spacy/pipeline.pyx
+++ b/spacy/pipeline.pyx
@@ -654,14 +654,6 @@ cdef class NeuralEntityRecognizer(NeuralParser):
 
     nr_feature = 6
 
-    def predict_confidences(self, docs):
-        tensors = [d.tensor for d in docs]
-        samples = []
-        for i in range(10):
-            states = self.parse_batch(docs, tensors, drop=0.3)
-            for state in states:
-                samples.append(self._get_entities(state))
-
     def __reduce__(self):
         return (NeuralEntityRecognizer, (self.vocab, self.moves, self.model), None, None)
 
diff --git a/spacy/syntax/_beam_utils.pyx b/spacy/syntax/_beam_utils.pyx
new file mode 100644
index 000000000..4d90fe23b
--- /dev/null
+++ b/spacy/syntax/_beam_utils.pyx
@@ -0,0 +1,286 @@
+# cython: infer_types=True
+# cython: profile=True
+cimport numpy as np
+import numpy
+from cpython.ref cimport PyObject, Py_INCREF, Py_XDECREF
+from thinc.extra.search cimport Beam
+from thinc.extra.search import MaxViolation
+from thinc.typedefs cimport hash_t, class_t
+from thinc.extra.search cimport MaxViolation
+
+from .transition_system cimport TransitionSystem, Transition
+from .stateclass cimport StateClass
+from ..gold cimport GoldParse
+from ..tokens.doc cimport Doc
+
+
+# These are passed as callbacks to thinc.search.Beam
+cdef int _transition_state(void* _dest, void* _src, class_t clas, void* _moves) except -1:
+    dest = <StateClass>_dest
+    src = <StateClass>_src
+    moves = <const Transition*>_moves
+    dest.clone(src)
+    moves[clas].do(dest.c, moves[clas].label)
+
+
+cdef int _check_final_state(void* _state, void* extra_args) except -1:
+    return (<StateClass>_state).is_final()
+
+
+def _cleanup(Beam beam):
+    for i in range(beam.width):
+        Py_XDECREF(<PyObject*>beam._states[i].content)
+        Py_XDECREF(<PyObject*>beam._parents[i].content)
+
+
+cdef hash_t _hash_state(void* _state, void* _) except 0:
+    state = <StateClass>_state
+    if state.c.is_final():
+        return 1
+    else:
+        return state.c.hash()
+
+
+cdef class ParserBeam(object):
+    cdef public TransitionSystem moves
+    cdef public object states
+    cdef public object golds
+    cdef public object beams
+    cdef public object dones
+
+    def __init__(self, TransitionSystem moves, states, golds,
+            int width, float density):
+        self.moves = moves
+        self.states = states
+        self.golds = golds
+        self.beams = []
+        cdef Beam beam
+        cdef StateClass state, st
+        for state in states:
+            beam = Beam(self.moves.n_moves, width, density)
+            beam.initialize(self.moves.init_beam_state, state.c.length, state.c._sent)
+            for i in range(beam.width):
+                st = <StateClass>beam.at(i)
+                st.c.offset = state.c.offset
+            self.beams.append(beam)
+        self.dones = [False] * len(self.beams)
+
+    def __dealloc__(self):
+        if self.beams is not None:
+            for beam in self.beams:
+                if beam is not None:
+                    _cleanup(beam)
+
+    @property
+    def is_done(self):
+        return all(b.is_done or self.dones[i] for i, b in enumerate(self.beams))
+
+    def __getitem__(self, i):
+        return self.beams[i]
+
+    def __len__(self):
+        return len(self.beams)
+
+    def advance(self, scores, follow_gold=False):
+        cdef Beam beam
+        for i, beam in enumerate(self.beams):
+            if beam.is_done or not scores[i].size or self.dones[i]:
+                continue
+            self._set_scores(beam, scores[i])
+            if self.golds is not None:
+                self._set_costs(beam, self.golds[i], follow_gold=follow_gold)
+            if follow_gold:
+                beam.advance(_transition_state, NULL, <void*>self.moves.c)
+            else:
+                beam.advance(_transition_state, _hash_state, <void*>self.moves.c)
+            beam.check_done(_check_final_state, NULL)
+            if beam.is_done and self.golds is not None:
+                for j in range(beam.size):
+                    state = <StateClass>beam.at(j)
+                    if state.is_final():
+                        try:
+                            if self.moves.is_gold_parse(state, self.golds[i]):
+                                beam._states[j].loss = 0.0
+                            elif beam._states[j].loss == 0.0:
+                                beam._states[j].loss = 1.0
+                        except NotImplementedError:
+                            break
+
+    def _set_scores(self, Beam beam, float[:, ::1] scores):
+        cdef float* c_scores = &scores[0, 0]
+        cdef int nr_state = min(scores.shape[0], beam.size)
+        cdef int nr_class = scores.shape[1]
+        for i in range(nr_state):
+            state = <StateClass>beam.at(i)
+            if not state.is_final():
+                for j in range(nr_class):
+                    beam.scores[i][j] = c_scores[i * nr_class + j]
+                self.moves.set_valid(beam.is_valid[i], state.c)
+            else:
+                for j in range(beam.nr_class):
+                    beam.scores[i][j] = 0
+                    beam.costs[i][j] = 0
+
+    def _set_costs(self, Beam beam, GoldParse gold, int follow_gold=False):
+        for i in range(beam.size):
+            state = <StateClass>beam.at(i)
+            if not state.c.is_final():
+                self.moves.set_costs(beam.is_valid[i], beam.costs[i], state, gold)
+                if follow_gold:
+                    for j in range(beam.nr_class):
+                        if beam.costs[i][j] >= 1:
+                            beam.is_valid[i][j] = 0
+
+
+def get_token_ids(states, int n_tokens):
+    cdef StateClass state
+    cdef np.ndarray ids = numpy.zeros((len(states), n_tokens),
+                                      dtype='int32', order='C')
+    c_ids = <int*>ids.data
+    for i, state in enumerate(states):
+        if not state.is_final():
+            state.c.set_context_tokens(c_ids, n_tokens)
+        else:
+            ids[i] = -1
+        c_ids += ids.shape[1]
+    return ids
+
+nr_update = 0
+def update_beam(TransitionSystem moves, int nr_feature, int max_steps,
+                states, tokvecs, golds,
+                state2vec, vec2scores, 
+                int width, float density,
+                sgd=None, losses=None, drop=0.):
+    global nr_update
+    cdef MaxViolation violn
+    nr_update += 1
+    pbeam = ParserBeam(moves, states, golds,
+                       width=width, density=density)
+    gbeam = ParserBeam(moves, states, golds,
+                       width=width, density=0.0)
+    cdef StateClass state
+    beam_maps = []
+    backprops = []
+    violns = [MaxViolation() for _ in range(len(states))]
+    for t in range(max_steps):
+        if pbeam.is_done and gbeam.is_done:
+            break
+        # The beam maps let us find the right row in the flattened scores
+        # arrays for each state. States are identified by (example id, history).
+        # We keep a different beam map for each step (since we'll have a flat
+        # scores array for each step). The beam map will let us take the per-state
+        # losses, and compute the gradient for each (step, state, class).
+        beam_maps.append({})
+        # Gather all states from the two beams in a list. Some stats may occur
+        # in both beams. To figure out which beam each state belonged to,
+        # we keep two lists of indices, p_indices and g_indices
+        states, p_indices, g_indices = get_states(pbeam, gbeam, beam_maps[-1], nr_update)
+        if not states:
+            break
+        # Now that we have our flat list of states, feed them through the model
+        token_ids = get_token_ids(states, nr_feature)
+        vectors, bp_vectors = state2vec.begin_update(token_ids, drop=drop)
+        scores, bp_scores = vec2scores.begin_update(vectors, drop=drop)
+
+        # Store the callbacks for the backward pass
+        backprops.append((token_ids, bp_vectors, bp_scores))
+
+        # Unpack the flat scores into lists for the two beams. The indices arrays
+        # tell us which example and state the scores-row refers to.
+        p_scores = [numpy.ascontiguousarray(scores[indices], dtype='f') for indices in p_indices]
+        g_scores = [numpy.ascontiguousarray(scores[indices], dtype='f')  for indices in g_indices]
+        # Now advance the states in the beams. The gold beam is contrained to
+        # to follow only gold analyses.
+        pbeam.advance(p_scores)
+        gbeam.advance(g_scores, follow_gold=True)
+        # Track the "maximum violation", to use in the update.
+        for i, violn in enumerate(violns):
+            violn.check_crf(pbeam[i], gbeam[i])
+    histories = []
+    losses = []
+    for violn in violns:
+        if violn.p_hist:
+            histories.append(violn.p_hist + violn.g_hist)
+            losses.append(violn.p_probs + violn.g_probs)
+        else:
+            histories.append([])
+            losses.append([])
+    states_d_scores = get_gradient(moves.n_moves, beam_maps, histories, losses)
+    return states_d_scores, backprops[:len(states_d_scores)]
+
+
+def get_states(pbeams, gbeams, beam_map, nr_update):
+    seen = {}
+    states = []
+    p_indices = []
+    g_indices = []
+    cdef Beam pbeam, gbeam
+    assert len(pbeams) == len(gbeams)
+    for eg_id, (pbeam, gbeam) in enumerate(zip(pbeams, gbeams)):
+        p_indices.append([])
+        g_indices.append([])
+        for i in range(pbeam.size):
+            state = <StateClass>pbeam.at(i)
+            if not state.is_final():
+                key = tuple([eg_id] + pbeam.histories[i])
+                assert key not in seen, (key, seen)
+                seen[key] = len(states)
+                p_indices[-1].append(len(states))
+                states.append(state)
+        beam_map.update(seen)
+        for i in range(gbeam.size):
+            state = <StateClass>gbeam.at(i)
+            if not state.is_final():
+                key = tuple([eg_id] + gbeam.histories[i])
+                if key in seen:
+                    g_indices[-1].append(seen[key])
+                else:
+                    g_indices[-1].append(len(states))
+                    beam_map[key] = len(states)
+                    states.append(state)
+    p_idx = [numpy.asarray(idx, dtype='i') for idx in p_indices]
+    g_idx = [numpy.asarray(idx, dtype='i') for idx in g_indices]
+    return states, p_idx, g_idx
+
+
+def get_gradient(nr_class, beam_maps, histories, losses):
+    """
+    The global model assigns a loss to each parse. The beam scores
+    are additive, so the same gradient is applied to each action
+    in the history. This gives the gradient of a single *action*
+    for a beam state -- so we have "the gradient of loss for taking
+    action i given history H."
+
+    Histories: Each hitory is a list of actions
+    Each candidate has a history
+    Each beam has multiple candidates
+    Each batch has multiple beams
+    So history is list of lists of lists of ints
+    """
+    nr_step = len(beam_maps)
+    grads = []
+    nr_step = 0
+    for eg_id, hists in enumerate(histories):
+        for loss, hist in zip(losses[eg_id], hists):
+            if loss != 0.0 and not numpy.isnan(loss):
+                nr_step = max(nr_step, len(hist))
+    for i in range(nr_step):
+        grads.append(numpy.zeros((max(beam_maps[i].values())+1, nr_class), dtype='f'))
+    assert len(histories) == len(losses)
+    for eg_id, hists in enumerate(histories):
+        for loss, hist in zip(losses[eg_id], hists):
+            if loss == 0.0 or numpy.isnan(loss):
+                continue
+            key = tuple([eg_id])
+            # Adjust loss for length
+            avg_loss = loss / len(hist)
+            loss += avg_loss * (nr_step - len(hist))
+            for j, clas in enumerate(hist):
+                i = beam_maps[j][key]
+                # In step j, at state i action clas
+                # resulted in loss
+                grads[j][i, clas] += loss
+                key = key + tuple([clas])
+    return grads
+
+
diff --git a/spacy/syntax/_state.pxd b/spacy/syntax/_state.pxd
index c06851978..b2a4e9eee 100644
--- a/spacy/syntax/_state.pxd
+++ b/spacy/syntax/_state.pxd
@@ -73,7 +73,16 @@ cdef cppclass StateC:
         free(this.shifted - PADDING)
 
     void set_context_tokens(int* ids, int n) nogil:
-        if n == 13:
+        if n == 8:
+            ids[0] = this.B(0)
+            ids[1] = this.B(1)
+            ids[2] = this.S(0)
+            ids[3] = this.S(1)
+            ids[4] = this.H(this.S(0))
+            ids[5] = this.L(this.B(0), 1)
+            ids[6] = this.L(this.S(0), 2)
+            ids[7] = this.R(this.S(0), 1)
+        elif n == 13:
             ids[0] = this.B(0)
             ids[1] = this.B(1)
             ids[2] = this.S(0)
diff --git a/spacy/syntax/arc_eager.pyx b/spacy/syntax/arc_eager.pyx
index 29e8de0aa..25a6ca42e 100644
--- a/spacy/syntax/arc_eager.pyx
+++ b/spacy/syntax/arc_eager.pyx
@@ -351,6 +351,20 @@ cdef class ArcEager(TransitionSystem):
         def __get__(self):
             return (SHIFT, REDUCE, LEFT, RIGHT, BREAK)
 
+    def is_gold_parse(self, StateClass state, GoldParse gold):
+        predicted = set()
+        truth = set()
+        for i in range(gold.length):
+            if gold.cand_to_gold[i] is None:
+                continue
+            if state.safe_get(i).dep:
+                predicted.add((i, state.H(i), self.strings[state.safe_get(i).dep]))
+            else:
+                predicted.add((i, state.H(i), 'ROOT'))
+            id_, word, tag, head, dep, ner = gold.orig_annot[gold.cand_to_gold[i]]
+            truth.add((id_, head, dep))
+        return truth == predicted
+
     def has_gold(self, GoldParse gold, start=0, end=None):
         end = end or len(gold.heads)
         if all([tag is None for tag in gold.heads[start:end]]):
diff --git a/spacy/syntax/beam_parser.pyx b/spacy/syntax/beam_parser.pyx
index e96e28fcf..e585e2ccb 100644
--- a/spacy/syntax/beam_parser.pyx
+++ b/spacy/syntax/beam_parser.pyx
@@ -107,7 +107,7 @@ cdef class BeamParser(Parser):
             # The non-monotonic oracle makes it difficult to ensure final costs are
             # correct. Therefore do final correction
             for i in range(pred.size):
-                if is_gold(<StateClass>pred.at(i), gold_parse, self.moves.strings):
+                if self.moves.is_gold_parse(<StateClass>pred.at(i), gold_parse):
                     pred._states[i].loss = 0.0
                 elif pred._states[i].loss == 0.0:
                     pred._states[i].loss = 1.0
@@ -213,7 +213,7 @@ def _check_train_integrity(Beam pred, Beam gold, GoldParse gold_parse, Transitio
         if not pred._states[i].is_done or pred._states[i].loss == 0:
             continue
         state = <StateClass>pred.at(i)
-        if is_gold(state, gold_parse, moves.strings) == True:
+        if moves.is_gold_parse(state, gold_parse) == True:
             for dep in gold_parse.orig_annot:
                 print(dep[1], dep[3], dep[4])
             print("Cost", pred._states[i].loss)
@@ -227,7 +227,7 @@ def _check_train_integrity(Beam pred, Beam gold, GoldParse gold_parse, Transitio
         if not gold._states[i].is_done:
             continue
         state = <StateClass>gold.at(i)
-        if is_gold(state, gold_parse, moves.strings) == False:
+        if moves.is_gold(state, gold_parse) == False:
             print("Truth")
             for dep in gold_parse.orig_annot:
                 print(dep[1], dep[3], dep[4])
diff --git a/spacy/syntax/nn_parser.pyx b/spacy/syntax/nn_parser.pyx
index 0b39e2216..9e7cb9a0e 100644
--- a/spacy/syntax/nn_parser.pyx
+++ b/spacy/syntax/nn_parser.pyx
@@ -37,7 +37,14 @@ from preshed.maps cimport MapStruct
 from preshed.maps cimport map_get
 
 from thinc.api import layerize, chain, noop, clone
+<<<<<<< HEAD
 from thinc.neural import Model, Affine, ELU, ReLu, Maxout
+=======
+from thinc.neural import Model, Affine, ReLu, Maxout
+from thinc.neural._classes.batchnorm import BatchNorm as BN
+from thinc.neural._classes.selu import SELU
+from thinc.neural._classes.layernorm import LayerNorm
+>>>>>>> feature/nn-beam-parser
 from thinc.neural.ops import NumpyOps, CupyOps
 from thinc.neural.util import get_array_module
 
@@ -61,6 +68,10 @@ from ..strings cimport StringStore
 from ..gold cimport GoldParse
 from ..attrs cimport TAG, DEP
 
+<<<<<<< HEAD
+=======
+USE_FINE_TUNE = True
+>>>>>>> feature/nn-beam-parser
 
 def get_templates(*args, **kwargs):
     return []
@@ -248,6 +259,7 @@ cdef class Parser:
                         nI=token_vector_width)
 
         with Model.use_device('cpu'):
+<<<<<<< HEAD
             if depth == 0:
                 upper = chain()
                 upper.is_noop = True
@@ -257,6 +269,12 @@ cdef class Parser:
                     zero_init(Affine(nr_class, drop_factor=0.0))
                 )
                 upper.is_noop = False
+=======
+            upper = chain(
+                clone(Maxout(hidden_width), (depth-1)),
+                zero_init(Affine(nr_class, drop_factor=0.0))
+            )
+>>>>>>> feature/nn-beam-parser
         # TODO: This is an unfortunate hack atm!
         # Used to set input dimensions in network.
         lower.begin_training(lower.ops.allocate((500, token_vector_width)))
@@ -294,6 +312,10 @@ cdef class Parser:
             self.moves = self.TransitionSystem(self.vocab.strings, {})
         else:
             self.moves = moves
+        if 'beam_width' not in cfg:
+            cfg['beam_width'] = util.env_opt('beam_width', 1)
+        if 'beam_density' not in cfg:
+            cfg['beam_density'] = util.env_opt('beam_density', 0.0)
         self.cfg = cfg
         if 'actions' in self.cfg:
             for action, labels in self.cfg.get('actions', {}).items():
@@ -316,7 +338,7 @@ cdef class Parser:
         if beam_width is None:
             beam_width = self.cfg.get('beam_width', 1)
         if beam_density is None:
-            beam_density = self.cfg.get('beam_density', 0.001)
+            beam_density = self.cfg.get('beam_density', 0.0)
         cdef Beam beam
         if beam_width == 1:
             states = self.parse_batch([doc], [doc.tensor])
@@ -332,7 +354,7 @@ cdef class Parser:
             return output
 
     def pipe(self, docs, int batch_size=1000, int n_threads=2,
-             beam_width=1, beam_density=0.001):
+             beam_width=None, beam_density=None):
         """
         Process a stream of documents.
 
@@ -345,6 +367,10 @@ cdef class Parser:
         Yields (Doc): Documents, in order.
         """
         cdef StateClass parse_state
+        if beam_width is None:
+            beam_width = self.cfg.get('beam_width', 1)
+        if beam_density is None:
+            beam_density = self.cfg.get('beam_density', 0.0)
         cdef Doc doc
         queue = []
         for docs in cytoolz.partition_all(batch_size, docs):
@@ -396,6 +422,7 @@ cdef class Parser:
         c_is_valid = <int*>is_valid.data
         cdef int has_hidden = not getattr(vec2scores, 'is_noop', False)
         while not next_step.empty():
+<<<<<<< HEAD
             if not has_hidden:
                 for i in cython.parallel.prange(
                         next_step.size(), num_threads=6, nogil=True):
@@ -415,6 +442,21 @@ cdef class Parser:
                         &c_scores[i*nr_class], &c_is_valid[i*nr_class], nr_class)
                     action = self.moves.c[guess]
                     action.do(st, action.label)
+=======
+            for i in range(next_step.size()):
+                st = next_step[i]
+                st.set_context_tokens(&c_token_ids[i*nr_feat], nr_feat)
+                self.moves.set_valid(&c_is_valid[i*nr_class], st)
+            vectors = state2vec(token_ids[:next_step.size()])
+            scores = vec2scores(vectors)
+            c_scores = <float*>scores.data
+            for i in range(next_step.size()):
+                st = next_step[i]
+                guess = arg_max_if_valid(
+                    &c_scores[i*nr_class], &c_is_valid[i*nr_class], nr_class)
+                action = self.moves.c[guess]
+                action.do(st, action.label)
+>>>>>>> feature/nn-beam-parser
             this_step, next_step = next_step, this_step
             next_step.clear()
             for st in this_step:
@@ -422,7 +464,7 @@ cdef class Parser:
                     next_step.push_back(st)
         return states
 
-    def beam_parse(self, docs, tokvecses, int beam_width=8, float beam_density=0.001):
+    def beam_parse(self, docs, tokvecses, int beam_width=3, float beam_density=0.001):
         cdef Beam beam
         cdef np.ndarray scores
         cdef Doc doc
@@ -484,6 +526,13 @@ cdef class Parser:
         free(token_ids)
 
     def update(self, docs_tokvecs, golds, drop=0., sgd=None, losses=None):
+<<<<<<< HEAD
+=======
+        if self.cfg.get('beam_width', 1) >= 2 and numpy.random.random() >= 0.5:
+            return self.update_beam(docs_tokvecs, golds,
+                    self.cfg['beam_width'], self.cfg['beam_density'],
+                    drop=drop, sgd=sgd, losses=losses)
+>>>>>>> feature/nn-beam-parser
         if losses is not None and self.name not in losses:
             losses[self.name] = 0.
         docs, tokvec_lists = docs_tokvecs
@@ -540,7 +589,64 @@ cdef class Parser:
                 break
         self._make_updates(d_tokvecs,
             backprops, sgd, cuda_stream)
+<<<<<<< HEAD
         return self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs])
+=======
+        d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs])
+        if USE_FINE_TUNE:
+            bp_my_tokvecs(d_tokvecs, sgd=sgd)
+        return d_tokvecs
+
+    def update_beam(self, docs_tokvecs, golds, width=None, density=None,
+            drop=0., sgd=None, losses=None):
+        if width is None:
+            width = self.cfg.get('beam_width', 2)
+        if density is None:
+            density = self.cfg.get('beam_density', 0.0)
+        if losses is not None and self.name not in losses:
+            losses[self.name] = 0.
+        docs, tokvecs = docs_tokvecs
+        lengths = [len(d) for d in docs]
+        assert min(lengths) >= 1
+        tokvecs = self.model[0].ops.flatten(tokvecs)
+        if USE_FINE_TUNE:
+            my_tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop)
+            my_tokvecs = self.model[0].ops.flatten(my_tokvecs)
+            tokvecs += my_tokvecs
+
+        states = self.moves.init_batch(docs)
+        for gold in golds:
+            self.moves.preprocess_gold(gold)
+
+        cuda_stream = get_cuda_stream()
+        state2vec, vec2scores = self.get_batch_model(len(states), tokvecs, cuda_stream, 0.0)
+
+        states_d_scores, backprops = _beam_utils.update_beam(self.moves, self.nr_feature, 500,
+                                        states, tokvecs, golds,
+                                        state2vec, vec2scores,
+                                        width, density,
+                                        sgd=sgd, drop=drop, losses=losses)
+        backprop_lower = []
+        for i, d_scores in enumerate(states_d_scores):
+            if losses is not None:
+                losses[self.name] += (d_scores**2).sum()
+            ids, bp_vectors, bp_scores = backprops[i]
+            d_vector = bp_scores(d_scores, sgd=sgd)
+            if isinstance(self.model[0].ops, CupyOps) \
+            and not isinstance(ids, state2vec.ops.xp.ndarray):
+                backprop_lower.append((
+                    get_async(cuda_stream, ids),
+                    get_async(cuda_stream, d_vector),
+                    bp_vectors))
+            else:
+                backprop_lower.append((ids, d_vector, bp_vectors))
+        d_tokvecs = self.model[0].ops.allocate(tokvecs.shape)
+        self._make_updates(d_tokvecs, backprop_lower, sgd, cuda_stream)
+        d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, lengths)
+        if USE_FINE_TUNE:
+            bp_my_tokvecs(d_tokvecs, sgd=sgd)
+        return d_tokvecs
+>>>>>>> feature/nn-beam-parser
 
     def _init_gold_batch(self, whole_docs, whole_golds):
         """Make a square batch, of length equal to the shortest doc. A long
@@ -585,6 +691,7 @@ cdef class Parser:
         xp = get_array_module(d_tokvecs)
         for ids, d_vector, bp_vector in backprops:
             d_state_features = bp_vector(d_vector, sgd=sgd)
+<<<<<<< HEAD
             active_feats = ids * (ids >= 0)
             active_feats = active_feats.reshape((ids.shape[0], ids.shape[1], 1))
             if hasattr(xp, 'scatter_add'):
@@ -593,6 +700,12 @@ cdef class Parser:
             else:
                 xp.add.at(d_tokvecs,
                     ids, d_state_features * active_feats)
+=======
+            mask = ids >= 0
+            d_state_features *= mask.reshape(ids.shape + (1,))
+            self.model[0].ops.scatter_add(d_tokvecs, ids * mask,
+                d_state_features)
+>>>>>>> feature/nn-beam-parser
 
     @property
     def move_names(self):
@@ -608,7 +721,7 @@ cdef class Parser:
                         lower, stream, drop=dropout)
         return state2vec, upper
 
-    nr_feature = 13
+    nr_feature = 8
 
     def get_token_ids(self, states):
         cdef StateClass state
diff --git a/spacy/syntax/transition_system.pyx b/spacy/syntax/transition_system.pyx
index 27b375bba..647f5e24a 100644
--- a/spacy/syntax/transition_system.pyx
+++ b/spacy/syntax/transition_system.pyx
@@ -99,6 +99,9 @@ cdef class TransitionSystem:
     def preprocess_gold(self, GoldParse gold):
         raise NotImplementedError
 
+    def is_gold_parse(self, StateClass state, GoldParse gold):
+        raise NotImplementedError
+
     cdef Transition lookup_transition(self, object name) except *:
         raise NotImplementedError