To construct BioProp, a biomedical proposition bank, we adopted GENIA 22 as the underlying corpus. It is a collection of 2,000 MEDLINE abstracts selected from the search results for queries using the keywords "human", "blood cells", or "transcription factors". GENIA is often used as a biomedical text mining test bed 23. In its officially released version, it is annotated with various levels of linguistic information, such as parts-of-speech, named entities, and conjunctions. In the summer of 2005, Tateisi 24 published full parsing information for the corpus that basically follows the Penn Treebank II (PTB) annotation scheme 25 encoded in XML. The GENIA corpus annotated with full parsing information is called GENIA Treebank (GTB). Currently, GTB is a beta version containing 500 abstracts.

As noted earlier, we chose thirty verbs because of their high frequency or important usage in describing molecular events. To select the verbs, we calculated the frequency of each verb based on its occurrence in GENIA, our underlying corpus, rather than in MEDLINE. It is noteworthy that some verbs that occur frequently in MEDLINE are rarely found in GENIA. Since we focus on molecular events, only sentences containing protein or gene names are used to calculate a verb's frequency. We listed verbs according to their frequency and removed generally used verbs such as is, have, show, use, do, and suggest. We then selected the verbs with highest frequencies and added some verbs of biological importance. The thirty verbs with their characteristics and frequency of occurrence in BioProp are listed in Table 1.

To build our SRL system, we followed the PropBank I 20 standard. PropBank I, with more than ten years of development history, has a large verb lexicon, and contains more annotated examples than other standards 26. In PropBank I, a PAS is annotated on top of a Penn-style full parsing tree. Figure 1 illustrates such a tree with syntactic and semantic role information. The semantic roles Arg0, Arg1, and ArgM-LOC are annotated on top of the words or phrases labelled as noun phrase subjects (NP-SBJ), noun phrases (NP), and prepositional phrases (PP), respectively. A proposition bank is a collection of full parsing trees annotated with propositions or PAS'. The first annotated PropBank-style proposition bank was the Wall Street Journal (WSJ) newswire corpus, which has 24 sections. Before being annotated with propositions, it was annotated with Penn-style full parsing trees. Sections 2 to 21 are conventionally used as a training set, Section 24 is used as a development set, and Section 23 is used as a test set in several NLP tasks 27.

PropBank I inherits verb senses from VerbNet, but the semantic arguments of individual verbs in PropBank I are numbered from 0. For a specific verb, Arg0 is usually the argument corresponding to the agent 28, while Arg1 usually corresponds to the patient or theme. For higher-numbered arguments, however, there is no consistent generalization for their roles. In addition to the main arguments, ArgMs refer to adjunct arguments. Table 2 details all the semantic role categories of arguments and their descriptions. The possible set of roles for a distinct sense of a verb is called a roleset, which can be paired with a set of syntactic frames that show all the acceptable syntactic expressions of those roles. A roleset with its associated frames is called a frameset 20. Verbs may have different rolesets and framesets for different senses, which are numbered .01, .02, etc. An example of the frameset is given by the verb activate shown below.

Frameset activate.01 "make active"

Arg0: Activator

Arg1: Thing activated

Arg2: Activated-from

Arg3: Attribute

Ex1: [_Arg0 IL4 and IL13 receptors] activate [_Arg1 STAT6, STAT3, and STAT5 proteins] [_ArgM-LOC in the human B cells].

Ex2: [_Arg1 The simian virus 40 early promoter] is [_ArgM-DIS also] [_ArgM-MNR synergistically] activated [_Arg0 by the Z/c-myb combination].

Basically, the annotation of BioProp is based on PropBank's framesets, which were originally designed for newswire texts. We further customize the framesets of biomedical verbs, since some of them may have different usages in biomedical texts. Table 1 indicates whether each verb has the same usage in the newswire and biomedical domains.

For verbs with the same usage in both domains, we adopt the newswire definitions and framesets. However, we need to make adjustments for other cases because some verbs have different usages and rarely appear in newswire texts. Thus, they are not defined in PropBank I. For example, "phosphorylate" is not defined in PropBank I, but it has been found increasingly in PubMed abstracts describing the experiment results of phosphorylated events 29. Therefore, after analyzing every sentence in our corpus containing such verbs, we added the latter to our list and defined framesets for them. For verbs not found in PropBank I, but with similar usages to other verbs in the proposition bank, we borrowed the PropBank I definitions and framesets. For instance, "transactivate" is not found in PropBank I, but we can apply the frameset of "activate" to it.

Some verbs have unique biomedical meanings not defined in PropBank I; however, their usage is similar to verbs in Propbank I. In most cases, we borrow framesets from synonyms. For example, "modulate" is defined as "change, modify, esp. of music" in the PropBank I frame files. However, its usage is very similar to "regulate" in the biomedical domain. Thus, we can use the frameset of "regulate" for "modulate". Table 3 shows the framesets and corresponding examples of "modulate" in the newswire and biomedical domains, as well as those of "regulate" in PropBank I.

Some other verbs have different primary senses in the newswire and biomedical domains. "Bind", for example, is common in the newswire domain and usually means "to tie" or "restrain with bonds". In the biomedical domain, however, its intransitive use, "attach or stick to", is far more common. A Google search for the phrase "glue binds to" only returns 21 results, while the same search replacing "glue" with "protein" yields 197,000 hits. For such verbs, we just select the appropriate alternative meanings and corresponding framesets.

Once the framesets for the verbs have been defined, we use a semi-automatic strategy to annotate BioProp. We used our newswire SRL system SMILE, which achieved an F-score of over 86% on Section 24 of PropBank I, to annotate the GENIA treebank automatically. Then, we asked three biologists to verify the automatically tagged results. One of the biologists has three years experience in biomedical text mining research, and he managed the task. The other two biologists received three months of linguistic training for this task. After annotating BioProp, we evaluated the performance of SMILE on BioProp. The F-score was approximately 65%, which is 20% lower than its performance on PropBank I. Even so, this semi-automatic approach substantially reduces the annotation effort.

We performed a preliminary consistency test on 1,982 instances of biomedical propositions by having two of the biologists annotate the results, while the third checked the annotations for consistency. Following the procedure used to calculate the inner-annotator agreement of PropBank 20, we measured the agreement between the two annotations before the adjudication step using the kappa statistic 30. Kappa is defined with respect to the probability of inter-annotator agreement, P(A), and the agreement expected by chance, P(E), as follows:

κ = P ( A ) − P ( E ) 1 − P ( E ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaaiiGacqWF6oWAcqGH9aqpdaWcaaqaaiabdcfaqjabcIcaOiabdgeabjabcMcaPiabgkHiTiabdcfaqjabcIcaOiabdweafjabcMcaPaqaaiabigdaXiabgkHiTiabdcfaqjabcIcaOiabdweafjabcMcaPaaaaaa@3E07@

The inter-annotator agreement probability P(A) is the number of nodes that the annotators agree on the annotation divided by the total number of nodes considered. To calculate P(E), for each category c, let p_1c denote the probability of c annotated by annotator 1, and p_2c denote the probability annotated by annotator 2. Then P(E) is the summation of p_1c * p_2c over all categories c of the semantic role labels. However, the calculation of P(A) and P(E) is distinguished into two cases that correspond to role identification (role vs. non-role) and role classification, since the vast majority of arguments are located on a small number of nodes near the verb and we need to avoid inflating the kappa score artificially. For role identification, the denominator of P(A) and P(E) the total number of nodes considered, is given by the number of nodes in each parse tree multiplied by the number of predicates annotated in the sentence, and the numerator is given by the number of nodes that are labeled as arguments (without considering whether a correct argument is assigned). For the role classification kappa, we only consider nodes marked as arguments by both annotators, which yields the denominator of P(A) and P(E), and compute kappa over the choices of possible argument labels. Furthermore, for both role identification and role classification, we compute kappa to process ArgM labels in two ways. The first (denoted as "Including ArgM in Table 4) processes ArgM labels as arguments like any other type of argument, such that ArgM-TMP, ArgM-LOC and so on are considered as separate labels for the role classification kappa. In the second scenario (denoted as "Excluding ArgM in Table 4), we ignore ArgM labels, treating them as unlabeled nodes, and calculate the agreement for identification and classification of numbered arguments only.

The kappa statistics for the above decisions are shown in Table 4. Given the large number of obviously irrelevant nodes, agreement on role identification is very high (.97 for both treatments of ArgM). The kappas for the more difficult role classification task are also high, .95 for all types of ArgM and .98 for numbered arguments only.

Wattarujeekrit et al. 26 developed PASBio, which has become a standard for annotating predicate-argument structures in the biomedical domain. It contains analyzed PAS's for over 30 verbs and is publicly available. Using predicate argument structures to analyze molecular biology information, PASBio is specifically designed for annotating molecular events and defines a core argument as one that is important for completing the meaning of an event. If a locative argument appears in a specific molecular PAS with a frequency greater than 80%, it is considered necessary and is therefore a main argument. To describe molecular events in greater detail, PASBio places biomedical constraints on main arguments. For example, considering the verb "express" (see figure 2), its Arg1, which is defined as named entity being expressed, is limited to a gene or gene products.

Shah et al. 31 successfully applied PASBio in the construction of the LSAT system for extracting information about alternative transcripts from the same gene, while Cohen et al. 32 showed that the suitability of the PAS representational model of representation for biomedical text. They concluded that PAS representations work well for biomedical text. Kogan et al. 33 built a domain-specific set of PAS for the medical domain. Their work agrees a bit more with ours in terms of their assessment of the match between PropBank's representations and the biomedical domain.

Unlike PASBio, BioProp is not a standard for annotating the PAS' of biomedical verbs. The main goal of BioProp is to port the proposition bank to the biomedical domain for training a biomedical SRL system. Thus, BioProp follows PropBank guidelines and uses the latter's framesets with further customization for some biomedical verbs. Subsequently, we use PropBank I as our initial training corpus for the construction of BioProp, and then ask annotators to refine the automatically tagged results. This semi-automatic approach substantially reduces the annotation effort so that Bioprop can be used for training SRL systems in the biomedical domain.

We use PropBank I and BioProp, which are associated with the general English and biomedical domains, respectively, as the sources of our data. The PropBank I corpus contains 950,028 words, 39,892 sentences, and 18,737 PAS'. However, only 1,449 of the PAS use the 30 biomedical verbs on our list as their predicates. BioProp currently contains 1,982 PAS'. The numbers and ratios of each argument type in the PAS' of the selected 30 biomedical verbs in PropBank I and BioProp are listed in Tables 5 and 6, respectively.

We use two SRL systems: SMILE 19 and BIOSMILE 34. The main difference between them is that SMILE is trained on PropBank I, while BIOSMILE is trained on BioProp. In addition, BIOSMILE has additional biomedical-specific features. Details of the features and the statistical models used in the two systems will be introduced in the Methods section.

We design two experiments: one to compare the performance of SMILE and BIOSMILE on biomedical applications by testing them on BioProp, and the other to measure the effects of using biomedical-specific features on the system's performance.

To improve the performance of SRL on biomedical literature, we add two domain specific features, NE features and argument-template features (denoted as BIOSMILE_NE and BIOSMILE_Template respectively) to BIOSMILE. This experiment tests the effectiveness of adding the features to BIOSMILE_Baseline and uses the same datasets as BIOSMILE_Baseline.

Bio-specific NE features are created for each of the following five primary named entity (NE) categories in the GENIA ontology³: protein, nucleotide, other organic compounds, source, and others. When a constituent (node on the full-parsing tree) matches an NE exactly, the corresponding NE feature is enabled.

Additionally, we integrate argument-template features. Usually, each argument type has its own patterns. For example, in the biomedical domain, the regular expression "in * cell" is a locative argument pattern (ArgM-LOC). We automatically generate argument templates, which are composed of words, NEs, and POS's, to represent the patterns of each argument. These templates are generated by using the Smith and Waterman local alignment algorithm 35 to align all instances of a specific argument type. The template feature is enabled if a constituent matches a template exactly. NE features and argument template features are discussed further in the Methods section.

The results are given as F-scores using the CoNLL-05 evaluation script and defined as F = (2PR)/(P + R), where P denotes the precision and R denotes the recall. The formulas for calculating precision and recall are as follows:

Precision = the number of correctly recognized arguments the number of recognized arguments Recall = the number of correctly recognized arguments the number of true arguments MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@05E3@

Table 8 shows the results of Experiment 1. We find that BIOSMILE outperforms SMILE by 21.45% on the F-score when tested on BioProp. It also outperforms SMILE for all argument types with a statistically significant difference. Specifically, ArgM-LOC, ArgM-MNR, and Arg2 achieve the biggest performance improvement, followed by ArgM-ADV, ArgM-DIS, and Arg0.

Table 9 compares the performance of BIOSMILE_Baseline and BIOSMILE_NE. Initially, we expected that the NE features would improve the recognition of adjunct arguments (ArgM), such as ArgM-LOC. However, they failed to do so.

Argument template features, on the other hand, boost the system's performance. Table 10 compares the performance of argument templates on BIOSMILE_Baseline and BIOSMILE_Template. The overall F-score is only improved slightly (0.52%). However, we achieve 3.33%, 2.27%, and 1.44% increases in the F-scores of ArgM-ADV, ArgM-LOC, and ArgM-MNR, respectively. These improvements are statistically significant. Although the increase in ArgM-TMP's F-score is not statistically significant, it is still appreciable. Figure 3 gives a clearer illustration of the improved performance of these argument types.

According to the results of Experiment 1, verbs with different framesets in the newswire and biomedical domain exhibit larger differences in performance between SMILE and BIOSMILE_Baseline. Table 11 details the average F-score difference between verbs with different framesets in both domains and verbs with the same framesets in both domains; the verb types are defined in the Background section (Verb Selection). These performance differences suggest that variations in cross-domain framesets and the usage of specific verbs influence SRL performance.

In the newswire domain, NE features have proven effective in improving SRL performance. However, the results of Experiment 2 show that we did not have the same success in the biomedical domain. Table 12 lists the five NE classes used and the number of NEs that occur in the main arguments. Though arguments frequently contain NEs, according to our findings, the converse does not hold. We find that most NEs match the NULL arguments, i.e., they match the nodes that are not labelled by BIOSMILE and, equivalently, do not correspond to the argument types of interest to us. Thus, trained on this data, a machine learning model would give so much weight to the NULL class that it would render all NE features ineffective for argument classification.

Only five argument types have a sufficient number of generated templates to be useful: ArgM-DIS, ArgM-MNR, ArgM-ADV, ArgM-TMP, and ArgM-LOC. We do not generate templates for arguments like ArgM-MOD and ArgM-NEG as they are usually composed of single words, such as "can" (ArgM-MOD) and "not" (ArgM-NEG). Baseline features, such as headword features, can generally recognize these arguments with a high degree of accuracy (F-scores of 95.82% for ArgM-MOD and 96.17% for ArgM-NEG). Templates for Arg0 and Arg1 are difficult to implement because a phrase pattern tagged as Arg0 may be tagged as Arg1 elsewhere, which results in all generated patterns being filtered out. Table 13 lists the performance improvement generated by template features on the five specified argument types, based on the improvement of the F-score over the baseline BIOSMILE. It also reports the number of templates, the number of instances, and the template density (# of templates/# of instances) for each type.

Figures 4 and 5 further illustrate the positive logarithmic correlation between template density and F-score difference (ΔF) for each argument type. ΔF initially increases with template density, but then appears to taper off. Note that the R-squared between ΔF and the logarithmic template density is 0.5046. However, after removing ArgM-ADV from the data, the R-squared increases to 0.8562, as shown in Figure 5. There are two possible explanations for ArgM-ADV's exceptionally high ΔF. First, its baseline F-score value is the lowest (56.60%) among the five adjunct arguments. Thus, increasing its F-score would be the easiest among the five arguments. Second, its average length is the longest so that its templates longer and possibly more precise.

Here, we give an example to illustrate how template features recognize an ArgM-TMP. Table 14 shows the features of each word in the sentence: "NAC not only blocks the effect of TPCK, but also enhances mitogenesis and cytokine production (>2.5-fold in some cases) upon activation of unsuppressed T cells." The phrase "upon activation of unsuppressed T cells" matches the ArgM-TMP template "IN NN IN SRC SRC cells". Each template slot indicates the allowable real words, NE tags, or POS tags. (Please refer to the Named Entity Features section to find all the abbreviations of NE tags.) However, the baseline BIOSMILE can not recognize this phrase by itself (shown in column five). Turning template features on, however, correctly identifies the ArgM-TMP, as shown in column six.

Like POS tagging, chunking, and named entity recognition, SRL can be formulated as a sentence tagging problem. A sentence can be represented by a sequence of words, a sequence of phrases, or a parsing tree; the basic units of a sentence are words, phrases, and constituents (a node on a full parsing tree) arranged in the above representations, respectively. Hacioglu et al. 36 showed that tagging phrase-by-phrase (P-by-P) is better than word-by-word (W-by-W). However, Punyakanok et al. 15 showed that constituent-by-constituent (C-by-C, or node-by-node) tagging is better than P-by-P. Therefore, we adopt C-by-C tagging for SRL.

In the following subsections, we first describe the maximum entropy model used for argument classification, and then illustrate the basic features of our SMILE and BIOSMILE systems.

The maximum entropy (ME) model is a flexible statistical framework that assigns an outcome for each instance based on the instance's history, which is made up of all the conditioning data that enables one to assign probabilities to the space of all outcomes. In SRL, a history can be viewed as all the information related to the current token that is derivable from the training corpus. ME computes the probability, p(o|h), for any o from the space of all possible outcomes, O, and for every h from the space of all possible histories, H.

The computation of p(o|h) in an ME depends on a set of binary features, which are useful for making predictions about the outcome. For instance, a node in the parsing tree that ends with "cell" is very likely to be an ArgM-LOC. Formally, we can represent this feature as follows:

f ( h , o ) = { 1 : if c u r r e n t _ n o d e _ e n d s _ w i t h _ c e l l ( h ) = true and o = ArgM-LOC 0 : otherwise . MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@88CD@

Here, "current_node_ends_with_cell(h)" is a binary function that returns a true value if the current node in the history, h, ends with "cell". Given a set of features and a training corpus, the ME estimation process produces a model in which every feature f_i has a weight α_i. Like 25, we compute the conditional probability as follows:

The probability is calculated by multiplying the weights of the active features (i.e., those with f_i (h,o) = 1); and α_i is estimated by a procedure called Generalized Iterative Scaling (GIS) 37. The ME estimation technique guarantees that, for every feature f_i, the expected value of α_i will equal the empirical expectation of α_i in the training corpus. We use Zhang's MaxEnt toolkit and the L-BFGS 38 method of parameter estimation for our ME model.

Table 15 shows the features used in our baseline argument classification model. Their effectiveness has been shown previously in 13141739. Detailed descriptions of the features can be found in 19.

In the English newswire domain, Surdeanu et al. 39 used NE features to indicate whether a constituent contains NEs of interest, such as personal names, organization names, location names, time expressions, and quantities of money. After adding these NE features to their system, the F-score improved by 2.12%. However, because NEs in the biomedical domain are quite different from English newswire NEs, we create specific biological NE features using the following five primary NE categories found in the GENIA ontology: protein, nucleotide, other organic compounds, source, and others. Table 16 lists the definitions of these five categories. When a constituent matches an NE exactly, the corresponding NE feature is enabled.

Although a few NEs tend to belong almost exclusively to certain argument types (e.g., "...cell" tends to be mainly an ArgM-LOC), this information alone is not sufficient for argument-type classification for two reasons: 1) most NEs appear in a variety of argument types; and 2), many appear in more than one constituent (a node in a parsing tree) in the same sentence. Take the sentence "IL4 and IL13 receptors activate STAT6, STAT3, and STAT5 proteins in the human B cells" for example. The NE "the human B cells" is found in two constituents ("the human B cells" and "in the human B cells") as shown in Figure 1. Yet only "in the human B cells" is annotated as ArgM-LOC because here "human B cells" is preceded by the preposition "in" and the determiner "the", which matches the template–"IN the SRC". Templates composed of NEs, words, and POS tags can be helpful for identifying the argument type of a constituent. In this section, we first describe our template generation algorithm, and then explain how we use the generated templates to improve SRL performance.

Our template generation (TG) algorithm, which extracts general patterns for all argument types using Smith and Waterman's local alignment algorithm 35, starts by pairing all arguments belonging to the same type according to their similarity. Closely matched pairs are then aligned word-by-word and a template satisfying the alignment result is created. Each slot in the template is given by the corresponding constraint information expressed in the form of a word (e.g. "the"), NE type, or POS. The preference of the constraint information is word > NE type > POS. If two aligned arguments have nothing in common for a given slot, the TG algorithm puts a wildcard in the position. Figure 6 shows a pair of aligned arguments, from which the TG algorithm generated the template "AP-1 CC PTN" (PTN: protein name) because, in the first position, both arguments have "AP-1"; in the second position, they have the same POS CC; and in the third position, they share a common NE type, PTN. The complete TG algorithm is described by pseudo code in Algorithm 1. The similarity function used to compare the similarity of two tokens in Smith and Waterman's algorithm is defined as:

S i m ( x , y ) = max ⁡ { 1 , x = y 1 , N E ( x ) = N E ( y ) 0.8 , P O S ( x ) = P O S ( y ) 0 , o t h e r w i s e , MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@7115@

where x and y are tokens in arguments a_i and a_j, respectively. The similarity of two arguments is calculated by the Smith and Waterman algorithm based on this token-level similarity function.

Input: A set of Arguments A = {a₁,..., a_k},

Output: A set of templates T = {t₁,...,t_k}.

1: T = {};

2: for each argument a_i from a₁ to a_n-1 do

3:    for each argument a_j from a_i to a_n do

4:       if the similarity of a_i and a_j calculated by an alignment is above the threshold τ

5:       then generate a common template t for a_i and a_j;

6:       T←T⋃t;

7:    end;

8: end;

After the templates have been generated, we filter out any template that matches at least two kinds of argument.

The generated templates may match with constituents exactly or partially. In our experience, exact matches are more useful for argument classification. For example, constituents that perfectly match the template "IN a * SRC" ('*' means wildcards) are overwhelmingly ArgM-LOCs. Therefore, we only accept exact template matches. In other words, if a constituent matches a template t exactly, then the feature corresponding to t will be enabled.