Detecting Causally Embedded Structures Using an Evolutionary Algorithm

Causality is an important relation among events and entities. Embedded causal structures represent an important class, expressing complex causal chains; but they are traditionally difficult to uncover automatically. In this paper we propose a method for the efficient identification and extraction of embedded causal relations with minimal supervision, by combining a representation of structured language data with modified prototype theory specifically suited to the data type. We then utilize a form of genetic algorithm specifically adapted for our purpose to locate the likely candidate linguistic structures that contain causal chains. With this procedure, we were able to identify many embedded structures with complex causal chains in two corpora of different genres, applying this algorithm as a ranking procedure for all structures in the data. We obtained 79.5% percision for top quantiles of both of our datasets (BNC & novels). 1 Embedded Causality Long chains of causal relations are frequently denoted by a complex embedding of multiple clauses through lexico-syntactic structures, structures which are causally linked. Following previous approaches (Menzies 2009, Beamer & Girju 2009), we define a causal relation as e1 cause −−−−→ e2, where e1 precedes e2 temporally and, had e1 failed to take place, e2 would also not have taken place, or more generally, P (e2|e1) > P (e2|¬e1). This is a general and agreed upon definition of causality which encompasses various classes of causal types of interest (if one chooses to go deeper into this problem). Our unit of representation (for both the cause and the effect) is a semantic frame, given by a predicate and a list of arguments in the form φ(ARGi, ARGj , ARGk, ....). This corresponds to a clause. Such clauses ocurring in embedded structures can form a causal chain. For example (from Little Women): 1. a smart shower at eleven had evidently quenched the enthusiasm of the young ladies who were to arrive at twelve for nobody came and at two the exhausted family sat down in a blaze of sunshine to consume the perishable portions of the feast (prepared in anticipation of the guests) that nothing might be lost (Alcott, 1868) (a) a smart shower at eleven had evidently quenched the enthusiasm of the young ladies who were to arrive at twelve (b) cause −−−−→ nobody came (c) cause −−−−→ the exhausted family sat down in a blaze of sunshire (d) cause −−−−→ consume the perishable portions of the feast (e) cause −−−−→ nothing might be lost In this paper we focus on causal relations between clauses (marked or not by discourse markers). 1.1 Distinct characteristics Each embedded causal structure has a causer entity identified by the main clause, and an effect event identified by the embedded (i.e. subordinate) clause. A class of semantically rich verbs is often present, that convey some notion of causation, coloring the causing event with additional manner of causation – verbs such as inspire, suggest, prompt, bribe, incite, bully, force, compel, etc. We call this class MCCverbs. Other verbs such as cause, bring-about, however, are just simple causatives (Girju 2003). Depending on its complexity, there may be one or more intermediate clausal structures that represent links in the causal chain, along with intermediate causal agents whose presence could have little specific semantic information, e.g. “...caused the circumstances to line up in such a way as to...”, but informs of its properties as a causal chain. Due to the complexity of these elements and the intervening structures, there are many combinatorial possibilities, and the depths of such structures are potentially unbounded. So rather than finding a comprehensive set of exemplars that cover all cases, it is better to assemble patterns that represent a diffuse prototype, finding characteristic structures common in embedded causal frames, such as: 43 i ENTITYcauser caused it to come about that ENTITYcausee [PREDemb ....] ii ENTITYcauser arranged the events so that it comes about that ENTITYcausee [PREDemb ....] iii ENTITYcauser had the forsight to prepare the circumstances so that it comes about that ENTITYcausee [PREDemb....] For all examples above, we can see that a subtree producing the terminals would be “to come about that ....”. A subtree like this can be used to further identify larger embedded structures as causal, and each embedded causative construction thus identified would contain one or more such subtrees.