Two baselines for unsupervised dependency parsing

Results in unsupervised dependency parsing are typically compared to branching baselines and the DMV-EM parser of Klein and Manning (2004). State-of-the-art results are now well beyond these baselines. This paper describes two simple, heuristic baselines that are much harder to beat: a simple, heuristic algorithm recently presented in Søgaard (2012) and a heuristic application of the universal rules presented in Naseem et al. (2010). Our first baseline (RANK) outperforms existing baselines, including PR-DVM (Gillenwater et al., 2010), while relying only on raw text, but all submitted systems in the Pascal Grammar Induction Challenge score better. Our second baseline (RULES), however, outperforms several submitted systems. 1 RANK: a simple heuristic baseline Our first baseline RANK is a simple heuristic baseline that does not rely on part of speech. It only assumes raw text. The intuition behind it is that a dependency structure encodes something related to the relatively salience of words in a sentence (Søgaard, 2012). It constructs a word graph of the words in a sentence and applies a random walk algorithm to rank the words by salience. The word ranking is then converted into a dependency tree using a simple heuristic algorithm. The graph over the words in the input sentence is constructed by adding directed edges between the ∗ word nodes. The edges are not weighted, but multiple edges between nodes will make transitions between them more likely. The edge template was validated on development data from the English Penn-III treebank (Marcus et al., 1993) and first presented in Søgaard (2012): • Short edges. To favor short dependencies, we add links between all words and their neighbors. This makes probability mass flow from central words to their neighboring words. • Function words. We use a keyword extraction algorithm without stop word lists to extract function or non-content words. The algorithm is a crude simplification of TextRank (Mihalcea and Tarau, 2004) that does not rely on linguistic resources, so that we can easily apply it to low-resource languages. Since we do not use stop word lists, highly ranked words will typically be function words. For the 50-most highly ranked words, we add additional links from their neighboring words. This will add additional probability mass to the function words. This is relevant to capture structures such as prepositional phrases where the function words take content words as complements. • Morphological inequality. If two words wi, wj have different prefixes or suffixes, i.e. the first two or last three letters, we add an edge between them. Given the constructed graph we rank the nodes using the algorithm in Page and Brin (1998), also known as PageRank. The input to the PageRank algorithm is any directed graph G = 〈E,V 〉 and the output is an assignment PR : V → R of a score, also referred to as PageRank, to each node in the graph, reflecting the probability of ending up in that node in a random walk.