Section: New Results
Transition-based constituency parsing with HyParse
Participants : Benoît Crabbé, Maximin Coavoux.
Transition-based parsing reduces the parsing task to predict a sequence of atomic decisions. These decisions are taken while sequentially reading words from a buffer and combining them incrementally into syntactic structures. The resulting structures are often dependency structures but can also be constituents, as is the case for our parser HyParse. Such an approach is therefore linear in the length of the input sentence, making transition-based parsing computationally efficient relative to other approaches. The challenge in transition-based parsing is modelling which action should be taken in each state it encounters as it progresses in a sentence provided as an input.
Training of a transition-based parser therefore consists in training a function that maps each of the unboundedly many states the parser might encounter to the best possible action, or transition, it should take. This function generally relies on a huge set of features, often conveniently grouped in the form of more abstract feature templates. Yet selecting the optimal subset of feature( template)s remains a challenge.
The training procedure therefore requires the help of an “oracle”, that is a function that returns the action that the parser should take in a given parser state given the gold parse. If the oracle assumes that the next action is necessarily the one given in the gold parse, it is said to be "static’’ and the oracle is deteminist. In order to train the parser to take relevant decisions when in an erroneous state, we can introduce some non determinism in the oracle in order to explore not only gold transition sequences but also near-gold transition sequences. This is the purpose of a dynamic oracle. Dynamic oracle training has shown substantial improvements for dependency parsing in various settings, but had not previously been explored for constituent parsing.
The two research directions we have investigated reflect the two above-mentioned challenges.
First, in collaboration with Rachel Bawden, now PhD student at LIMSI, we resumed our work on developing an efficient, language-independent model selection method for our parser HyParse [61]. It is designed for model selection when faced with a large number of possible feature templates, which is typically the case for morphologically rich languages, for which we want to exploit morphological information. The method we proposed uses multi-class boosting for iterative selection in constant time, using virtually no a priori constraints on the search space. We did however use a pre-ranking step before selection in order to guide the selection process. Our experiments have illustrated the feasibility of the method for our working language, French and resulted in high-performing, compact models much more efficiently than naive methods [22].
Second, we developed a dynamic oracle for HyParse. First, we replaced the traditional feature-based approach used in the above-described experiments by a neural approach. This is a way to overcome the feature selection issue addressed in the above-described work. The neural network weighting function we developed uses a non-linear hidden layer to automatically capture interactions between variables, and embeds morphological information in a vector space, as is usual for words and other symbols. Then, we developed our dynamic oracle based on this neural function and conducted experiments on the 9 languages of the SPMRL dataset in order to assess the impact of this oracle [25]. The experiments have shown that a neural greedy parser with morphological features, trained with a dynamic oracle, leads to accuracies comparable with the best currently available non-reranking and non-ensemble parsers.