为何基于多类别感知机的依存解析器反复选择同一动作？

我正在从零开发一个基于多类别感知机的小型依存解析器，解析器包含START_ARC、MOVE_RIGHT、END_ARC三种动作用于处理待处理token。训练器可以通过输入学习到的动作准确复现树库中的依存弧，但ParserRunner.parse()方法中，感知机始终预测END_ARC动作，每个token都只会执行该动作，完全不选择其他动作。

class DependencyParser1:

    """
    self.token_index must be manually increased along with the sentence,
    as each token has been processed. Do it as soon as self.start_arc()
    has been called. self.move_right() only moves the move pointer to the
    next token, while self.token_index keeps track of the token that is
    to be assigned a head.

    Each arc is a tuple [x, y] where x is the start, and y is the head.
    """
    def __init__(self):
        self.move_index = 0
        self.token_index = -1
        self.arcs = []
        self.arc = [0, 0]

    def move_right(self):
        self.move_index += 1

    def start_arc(self):
        self.arc[0] = self.token_index + 1

    def end_arc(self):
        self.arc[1] = 0 if self.move_index == 0 else self.move_index + 1
        self.arcs.append(self.arc)
        self.arc = [0, 0]
        self.move_index = 0

    def reset(self):
        self.arcs = []
        self.arc = [0, 0]
        self.token_index = - 1
        self.move_index = 0

class Trainer:

    def __init__(self, parser):
        self.parser = parser
        self.features = []
        self.classes = []

    def train(self, sentence):
        arc_start_index = 0   # 0 is the first token. The root
                              # is numbered 0, the rest of the
                              # tokens are numbered as the index
                              # plus 1.

        sentence_length = len(sentence)

        while arc_start_index < sentence_length:
            self.parser.token_index += 1
            self.parser.start_arc()
            self.features.append(FeaturesRecorder.record(locals()))
            self.classes.append(START_ARC)

            arc_end_index = 0
            head = int(sentence[arc_start_index].head) - 1

            while arc_end_index < head:
                arc_end_index += 1
                self.parser.move_right()
                self.features.append(FeaturesRecorder.record(locals()))
                self.classes.append(MOVE_RIGHT)

            self.parser.end_arc()
            self.features.append(FeaturesRecorder.record(locals()))
            self.classes.append(END_ARC)
            arc_start_index += 1

class ParserRunner:

    def __init__(self, parser, perceptron):
        self.parser = parser
        self.perceptron = perceptron

    def cheat(self, sentence, actions):
        actions_copy = actions.copy()
        arc_start_index = 0

        sentence_length = len(sentence)

        while actions_copy:
            action = actions_copy.pop(0)
            if action == START_ARC:
                self.parser.token_index += 1
                self.parser.start_arc()
            elif action == MOVE_RIGHT:
                self.parser.move_right()
            elif action == END_ARC:
                self.parser.end_arc()

        return self.parser.arcs

    def parse(self, sentence):
        arc_start_index = 0
        sentence_length = len(sentence)
        arc_is_started = 0

        while arc_start_index < sentence_length:
            recording = FeaturesRecorder.record(locals())
            print(recording)
            action = self.perceptron.predict([recording])
            print(action)
            #print(action)
            if action == START_ARC:
                self.parser.token_index += 1
                self.parser.start_arc()
                arc_is_started = 1
            elif action == MOVE_RIGHT:
                self.parser.move_right()
            elif action ==  END_ARC:
                self.parser.end_arc()
            arc_start_index += 1

        return self.parser.arcs

• 特征一致性问题：检查FeaturesRecorder.record(locals())在训练和预测时提取的特征是否一致。训练时locals()包含arc_start_index、head、arc_end_index等变量，而预测时的locals()只有arc_start_index、arc_is_started等，缺少关键特征会导致感知机无法区分动作类别。
• 训练数据分布：统计训练集中三类动作的样本占比，如果END_ARC样本占比过高，感知机会倾向于预测该类别。
• 感知机训练有效性：确认是否用Trainer收集的features和classes完成了完整训练，权重更新逻辑是否正确执行。
• 解析逻辑不匹配：训练时每个token的处理流程是START_ARC→若干MOVE_RIGHT→END_ARC，但预测时parse()方法中arc_start_index每次循环都会强制+1，不管当前执行的是什么动作。比如执行MOVE_RIGHT后，arc_start_index依然递增，导致后续token的处理状态和训练时完全不一致，特征错位。
• 状态重置问题：每次调用parse()前，必须调用parser.reset()重置解析器状态，否则残留的状态会影响特征提取和动作预测。
• 动作编码一致性：确认START_ARC、MOVE_RIGHT、END_ARC的编码（比如整数映射）在训练和预测时完全一致，编码不匹配会导致预测结果混乱。

内容的提问来源于stack exchange，提问作者Nils Blomqvist