Yoshikawa 2009 jointly identifying temporal relations with markov logic

Citation

Jointly Identifying Temporal Relations with Markov Logic, by K. Yoshikawa, J. NAIST, S. Riedel, M. Asahara, Y. Matsumoto. In Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, 2009.

Online version

This Paper is available online [1].

Summary

This paper is a follow up paper to Chambers and Jurafsky (2008) that focuses on using global inference to improve local, pairwise temporal ordering of events in text. Similarly in this paper, instead of predicting the three types of temporal relations: between events in adjacent sentences, between events and time expressions in the same sentence, and between events in a document and the document creation time (DCT), in isolation; the paper proposes to use Markov Logic Networks to jointly identify relations of all three relation types simultaneously while respecting logical constraints between these temporal relations.

The experiment is done on the TempEval-07 data, for the task of classifying temporal relations into one of the 6 classes: BEFORE (e.g. event A is before event B), OVERLAP, AFTER, BEFORE-OR-OVERLAP, OVERLAP-OR-AFTER, and VAGUE (unknown). The paper shows an accuracy increase of 2% for all three types of relations: event-event, event-time, event-DCT, compared to other machine learning based approaches.

Brief description of the method

The paper uses Markov Logic Network to represent constraints of temporal consistency. Three hidden predicates corresponding to the temporal relations to be predicted are:

• relE2T(e,t,r) representing the temporal relation r between an event e and a time expression t
• relDCT(e,r) representing the temporal relation r between an event e and DCT
• relE2E(e1,e2,r) representing the relation r between two events of adjacent sentences e1 and e2

The observed predicates, corresponding to information that is given are:

• words, syntactic and lexical feature predicates. For example, predicates to denote POS tags, grammatical aspect, tense, etc; e.g. the predicate tense(e,t) denotes the tense t for an event e
• relT2T(t1,t2,r) denoting the temporal relation r between two time expressions t1 and t2
• dctOrder(t,r) representing the temporal relation r beetween a time expression t and DCT.

The illustration of all temporal predicates are given in the figure below

For example, if an event X happens BEFORE DCT, and another event Y happens AFTER DCT, then X should have happened BEFORE Y. This intuition about the task can be incorporated using weighted first order logic formulae:

$beforeDCT(x)\land \lnot beforeDCT(y)\Rightarrow before(x,y)\,\!$

There are other rules that are probable but do not always hold. For example, it is reasonable to assume that if an event X happens BEFORE an event Y and the event Y OVERLAPS with an event Z, then an event X has a good chance of happening BEFORE Z; but this is not guaranteed. For such rule:

$before(x,y)\land overlap(y,z)\Rightarrow before(x,z)\,\!$

Markov Logic Networks allows to model the uncertainty with regard to this formula as a weight w associated with it. The larger the weight, the more likely/often the formula holds. These weights are learned from the training data.

Once the formulae are defined, the training and inference method

In the first component, Support Vector Machine (SVM) classifier is used. Using features varying from POS tags and lexical features surrounding the event to tense, grammatical aspect features of the events, probabilities of temporal relations between pairwise events are computed. These scores are then used as confidence scores to choose an optimal global ordering.

In the second component, the ILP uses the following objective function:

$max\sum _{i}\sum _{j}p_{ij}x_{ij}\,\!$

with the constraints:

$\forall _{i}\forall _{j}{x_{ij}}\in {0,1}$

$\forall _{i}{x_{i1}}+{x_{i2}}+...+{x_{im}}=1\,\!$

${x_{ia}}+{x_{jb}}-{x_{kc}}<=1\,\!$

where ${x_{ij}}\,\!$ represents the ith pair of events classified into the jth relation of m relations.

The first constraint simply says that each variable must be 0 or 1. The second constraint says that a pair of events cannot have two relations at the same time. The third constraint is added for connected pairs of events $i,j,k\,\!$, for each transitivity condition that infers relation $c\,\!$ given $a\,\!$ and $b\,\!$.

Prior to running the two components, the set of training relations is expanded to create a more well-connected network of events. One way to expand it is to perform temporal reasoning over the document's time expression (e.g. yesterday is before today) to add new relations between times. Once new time-time relations are added, transitive closure is conducted through transitive rules that creates new connections in the network, such as:

A simultaneous B $\land$ A before C $\to$ B before C

Experimental Result

The experiment is done using the data from TempEval temporal ordering challenge, with the tasks of classifying temporal relations between events and time expressions in the same sentence (Task A), between events in a document and the document creation time (Task B), and between events in two consecutive sentences (Task C). Temporal relations are classified into one of six classes: BEFORE, OVERLAP, AFTER, BEFORE-OR-OVERLAP, OVERLAP-OR-AFTER, and VAGUE.

The paper shows that by incorporating global constraints that hold between temporal relations predicted in Task A, B, and C, the accuracy for all three tasks can be improved significantly. For two out of the three tasks, the approach in this paper achieves the best accuracy by at least 2% more than other approaches. For task B, the approach's accuracy is less than that of rule-based approach; however it is better than all other machine learning approaches.

Related papers

The approach in this paper is similar to that of an earlier work by Chambers and Jurafsky (2008) that proposes to use global framework based on Integer Linear Programming (ILP) to jointly infer temporal relations between events. Chambers and Jurafsky (2008) show that adding global inference improves the accuracy of the inferred temporal relations. However they only focus on event-event temporal relations while this paper also jointly predicts temporal order between events and time expressions, and between events and document creation time.

Secondly, Chambers and Jurafsky (2008) combines the output of local classifiers using ILP framework while this paper uses Markov Logic Networks which represents global constraints through the addition of weighted first order logic formulae. The advantage is that it allows for representation of non-deterministic rules that tend to hold between temporal relations but do not always have to. For example, if A happens before B and B overlaps with C, then there is a good chance that A also happens before C, but this is not always the case. The learned weights of the rules allow for soft enforcement of the constraints.