DEEPER: A Full Parsing Based Approach to Protein Relation Extraction

Abstract

Abstract. Lexical variance in biomedical texts poses a challenge to automatic protein relation mining. We therefore propose a new approach that relies only on more general language structures such as parsing and dependency information for the construction of feature vectors that can be used by standard machine learning algorithms in deciding whether a sentence describes a protein interaction or not. As our approach is not dependent on the use of specic interaction keywords, it is applicable to heterogeneous corpora. Evaluation on benchmark datasets shows that our method is competitive with existing state-of-the-art algorithms for the extraction of protein interactions.

The role of syntactic features in protein interaction extraction

“…Using a C4.5 and a BayesNet classifier relying only on deep syntactic information, in [9] we were able to obtain results comparable with state of the art approaches for protein interaction mining, including in a modest cross-dataset experiment involving the LLL [16] and AIMed [2] datasets. This observation encouraged us to engage in a more systematic study of the individual effects of lexical and syntactic features, which we report on in this paper.…”

“…As can be seen from Table 1 and 2 all the approaches mentioned above (except our own previous work [9]) utilize lexical information (words) along with syntactic information. The main research question tackled in this paper is how much performance can still be achieved when using no lexical information whatsoever.…”

“…Named entity recognition is a whole separate area of text mining which is commonly treated separately from the relation mining task (see e.g. [3,9,10,14]). Example 1.…”

“…One of the interesting aspects of the work of Kim et al is that they consider two formal representations of a sentence that contains a protein pair, namely a subtree of the dependency tree and a flat feature vector. Unlike in [21] where the role of lexical versus shallow and deep syntactic features in Kim et al's feature vector based approach was explored further, in this paper we focus on the use of the dependency tree format like in [9].…”

“…Method Information Algorithm Datasets [3] lexical SVM AIMed [10] lexical Hand-built rules HPRD50 shallow deep [9] shallow C4.5 AIMed deep BayesNet LLL [11] lexical SVM AIMed shallow LLL [14] lexical deep BayesNet AIMed deep NaiveBayes K-nearest neighbour Ensembles [21] lexical SVM AIMed shallow HPRD50 deep IEPA LLL [22] lexical Maximum entropy IEPA shallow deep [23] lexical SVM AIMed shallow deep Table 1: General approaches for protein interaction extraction…”