Using the large volumes of information contained in biodiversity literature, we aim to provide scientists with rich computable data so they can build a more complete tree of life, predict the causal genes for wider ranges of diseases/conditions, derive better models of the decline of species populations and improve climate change predictions. Biodiversity literature contains various descriptive information on extinct and extant taxa, including habitat, distribution, phenology, ecology, physiology and morphology. In this work, we are primarily concerned with making the information contained in morphological descriptions more accessible to machines. One way to accomplish this is to extract character information from descriptions into a structured format, such as taxon-by-character matrices. This work involves linking a structure/organ (e.g. leaf or stem) with its properties (e.g. colour, shape, orientation etc., often called “characters” in systematics biology), as described in the literature (Dececchi et al. 2015). While developing the Explorer of Taxon Concepts (ETC) toolkit web application (Cui et al. 2016), we encountered the challenge of resolving “orphaned” parts extracted by the ETC Text Capture tool. The orphaned parts occur because taxonomic descriptions often include a large set of terms, representing non-specific structural parts, such as base, apex and tip. In this paper, we refer to such terms as non-specific structure (NSS) terms. They are non-specific because these terms stand alone and do not correspond to any identifiable organ; instead, they could be part of a number of different organs. For example, leaf, stem and petal could each have an apex, therefore apex is a NSS. When a NSS is linked to its parent organ, as in leaf apex, it represents an independent entity and we say it is resolved. Extracting character information about such structures requires the machine to bridge the non-specific structures with their anchor organs. Following Poesio and Vieira’s work (Poesio and Vieira 1998), the term anchor is used as a generalisation of the term antecedent to indicate the discourse entity to which an anaphoric expression is related, regardless of its relative location. For example, associating apex acuminate with its appropriate organ, leaf, produces a complete and accurate character leaf apex accuminate.

In computational linguistics, this problem is treated as a subclass of the bridging reference problem, whose resolution requires not only finding the anchor in the text to which the bridging reference is related, but also identifying the relationship between them (Sindner 1979, Poesio and Vieira 1998, Clark 1975, Grosz 1977, Hawkins 1978, Fraurud 1990). Characterised by the relationship between the reference and the anchor, a bridging reference may be classified as an anaphoric reference or associative reference; in the former case, the reference and the anchor denote the same entity (e.g. Mr. Smith and he), while in the latter case the two entities are associated. In the case of associative references, further distinction can be made based on the relation, for example, synonyms, hypernyms, hyponyms or meronyms (Poesio and Vieira 1998). In our study, we focus on mereological bridging references because the relationships between bridging references and anchors in this case are part-of relationships. Resolving part-of relationships amongst anatomical structures is already needed to extract phenotypic characters from vast amounts of taxonomic descriptions continuously being published in human languages.

An illustrative example of our task is given below:

Leaflets articulated, inserted near the edges of the rhachis towards the adaxial side, lacking a differently coloured basal gland; stomata on lower surface only or on both surfaces; epidermal cells elongated parallel to long axes of leaflets.

In this example, the goal is to associate edges, adaxial side, lower surface, surfaces, cells and long axes with their anchors (parents) rhachis, leaflets, leaflets, leaflets, epidermal and leaflets, respectively.

This paper reports the evaluation of two methods we developed to resolve NSS in taxonomic descriptions. We start by giving a formal task definition and then, in the Data Resources section, we describe the corpus, datasets and ontologies used to develop and evaluate the methods. After introducing the two methods, we present evaluation results. We conclude the paper with future directions for research.

Morphological descriptions of the biodiversity domain pose a number of challenges to existing natural language processing (NLP) algorithms. Thessen et al. (2012) identified three characteristics about morphological descriptions that make them difficult to parse using existing methods: 1) Specialised language is used and a large number of terms are not contained in pre-compiled lexica. Ambiguity is also more prevalent in biological text than texts from other domains (Chen et al. 2004). In addition, there is a heavy use of abbreviations in life science literature. 2) Morphological descriptions are very diverse amongst and within taxon corpora. 3) The syntax of morphological descriptions often differs from standard syntactic structure in English, i.e. a telegraphic sub-language that lacks verbs is used. The ETC toolkit (http://etc.cs.umb.edu/etcsite) is the only web-based, publicly accessible application to our knowledge that extracts character data from organismal morphological descriptions to support biological research. The toolkit has been used by biology research projects, for example, the extraction of phenotypic traits for the tree of life (Endara et al. 2018), the Building a Comprehensive Evolutionary History of Flagellate Plants project (US NSF DEB-1541506), the Botanical Knowledge Portal project (co-sponsored by Agriculture and Agri-Food Canada) and the Flora of North America project.

Although natural language processing tools and algorithms have been successfully implemented in various tasks, bridging references are considered one of the most challenging problems in discourse analysis since the knowledge about the events and natural objects required to identify each relation goes beyond the text itself (Clark 1975). While various work has been published on this subject (Poesio et al. 2004, Bunescu 2003, Fan et al. 2005), most of those efforts used traditional linguistic methods rather than machine-learning approaches. We are not aware of open source algorithms that resolve meronym associative reference problems in taxonomic descriptions. Our work represents a small but solid contribution to this area.

Given a short piece of text (one or a few sentences) from a morphological description that contains some NSS terms, our task was to resolve the NSSs, i.e. identify the anchor (parent) structure for each NSS. More formally, given a set of NSS-terms A = {a₁, a₂, a₃,…, a_n} and a text segment W = {w₁, w₂, w₃, ..., w_n} consisting of n tokens, for any given w_i=a_j and, amongst all semantically valid part-of relations (w_i, w_k), our task was to find the specific w_k that would resolve w_i as the intended independent entity described by the description.

A total of 7562 sentences from 3876 morphological description files of 11 taxon groups were extracted from various taxonomic data sources we have collected in prior years, including Plazi.org and the Flora of North America (FNA). A list of 39 NSS terms pertinent to these descriptions was created by domain experts. Table 1 shows the NSS terms along with the number of occurrences in the corpus.

An initial dataset consisting of 169 sentences were randomly selected from taxonomic descriptions of ants, mushrooms and plants with 389 NSS term occurrences. This was used as the dataset to develop the two methods reported in this paper. To expand the taxon and NSS terms coverage, another 167 sentences were sampled in a stratified random manner from the corpus to form the test dataset; sentences were randomly selected from all 11 taxon groups and each sentence contained at least one NSS term. 366 NSS term occurrences were in the test dataset. Sentences containing target anchors were added if they were not sampled in the first round. After the data collection procedure, sentences that contained both NSS term(s) and target anchor(s) were treated as statements. The composition of the training and test datasets is shown in Table 2. Statements in the development dataset and test dataset did not overlap.

Generation of the annotated evaluation datasets

The evaluation datasets (development and test) needed to be annotated with NSSs and their anchor organs. This was done semi-automatically. The fine-grained morphological character extraction system ETC Text Capture tool (Cui et al. 2016) was used to annotate all structures (i.e. biological entities, e.g. elytron), characters (e.g. shape, rounded) and relationships, including preposition and verb phrases in the statements (e.g. on primary stem). Since the raw annotations did not resolve NSSs, the annotations contained NSS structures as independent biological entities without their anchor structures. Statements with raw annotations were the input for the algorithms.

For performance evaluation, NSS-terms in both datasets were manually resolved by adding anchor organs to the NSS-terms. Fig. 1 shows an example of raw annotations and the annotated answer key (gold standard). The annotations were completed by one author and the gold standard was hidden from algorithm developers until the algorithms were ready for evaluation.

Figure 1.  

An example showing raw annotations (top) and gold standard annotations (bottom).

The NSS Ontologies

Anatomical and phenotype character ontologies are being created by bioinformatics communities to generate computable data and support integration of data silos (Köhler et al. 2016, Hoehndorf et al. 2016). Ontologies carry useful knowledge often not explicitly stated in morphological descriptions (Huang et al. 2015). To enable intelligent organisation and use of organism-based morphological information, ontologies need to be created on a per domain basis since research on different taxon groups employ different vocabularies and new terminology is constantly encountered (Cui 2010). In terms of non-specific structures resolution, different domains also use various non-specific structures. For example, in the domain of botany, body is a non-specific structure, while in insects or some single-cell organisms, body is a specific structure as there is only one body. Following previous research on domain ontology building (Chong et al. 2017), we used the Ontology Building tool in the ETC toolkit to create domain ontologies. The reason for creating task-specific ontologies, instead of using existing ontologies, such as Biological Spatial Ontology or Plant Ontology, is that existing formal ontologies often do not include non-specific structures. For part-of relationships to be included in a formal ontology, the related entities must meet the so-called 'all some' rules. To say apex is part of leaf in a formal ontology, all apices must be part of some leaves. However, this is not true for all cases (e.g. an apex could be a stem apex), so this relationship cannot be included in a formal ontology (e.g. the Plant Ontology). In general, non-specific structures are not included in OBO Foundry (Open Biological and Biomedical Ontology Foundry) ontologies for this reason. The ontologies we created in our task only specify that an apex could be part of a leaf and an apex could be part of a stem. The soft constraints in our ontologies express the potential of part-of relationships and the algorithms to act upon those possibilities and select the most likely parent structure.

In this experiment, the tool was used by the same author who subsequently annotated the answer keys to create ontologies holding part-of relationships amongst various biological structures for the development and test data, respectively. Taking the statement in Fig. 1 as an example, we used the tool to first add the structures elytron, side, tip, elytron side and elytron tip into the ontologies and then linked the terms with part-of and subclass relationships (e.g. elytron side is a subclass of side and a part-of elytron). A two-month gap separated each of the ontology building events and gold standard annotation to control for any potential confounding between the ontologies and the gold standard. A total of 488/540 unique structures (including organ and NSS-terms) and 392/351 part-of relationships were created in the development/test versions of the NSS ontologies.

Syntactic rule method

Support vector machine

Two baselines: subject entity and closest entity

Evaluation metrics

We compared the different methods described in the Methods section. Table 3 shows the best performance results using the development dataset and the final results using the test dataset. First, it is evident that naive approaches such as the baseline methods were not sufficient to solve the problem, as they achieved F1 scores of 42.3% (Baseline 1) and 33.2% (Baseline 2).

Second, we found the syntactic approach outperformed the SVM method by a large margin. For example, the best syntactic approach achieved an F1 score of 92.1%, while the best SVM model achieved an F1 of only 80.7%. This may be due to the small training size provided to the SVM. From a practical point of view, it was not realistic to request that domain scientists annotate large sets of training examples. In contrast, they are more likely to construct an ontology that can be reused to process future descriptions and solve other related problems, such as extracting phenotypes.

Third, we found ontologies were reliable knowledge sources in resolving orphaned parts in morphological descriptions. Using an ontology alone, the syntactic rule-based approach achieved an F1 score of 91.4%. Using an ontology also improved SVM performance by 20 percentage points for the F1 score. The syntactic rules (of-phrases and possession words) seem to be useful for the rule-based method, but using these rules with an ontology only improved the F1 score by 0.7% and, when considering the 366 test examples, their effects on reference resolution were unclear since the results may have been confounded by usage of an ontology. We are planning to evaluate a second variation of the rule-based method using only the rules and not an ontology to further examine the effects of these rules.

Our error analysis showed that the syntactic method and the SVM method were complementary to each other, as the mistakes made by the two methods were largely disjoint. Fig. 2 shows the error distribution between the best performances of the SVM and syntactic methods. Amongst 366 NSS terms to be resolved, the SVM method made 65 mistakes while the syntactic method made 32 mistakes, but they were both incorrect in only 7 of the same cases. 

Figure 2.  

Error distribution between the best performances of the SVM and syntactic methods.

With the results from the dataset covering a wide range of descriptions, we can tentatively conclude that the syntactic rule method is what the ETC toolkit will adopt, given that ETC toolkit already provides a user-friendly ontology building tool for domain scientists and students to create ontologies. However, in the immediate future, it is worthwhile for additional testing on the development version of ETC toolkit before we make the functionality publicly available and to further investigate and leverage the complementary nature of the syntactic rule method and the SVM method to minimise the error rate, as in practical application NSS, terms are commonly used in morphological descriptions and even a 1% error rate could lead to numerous errors.

This material is based upon work supported by the National Science Foundation under Grant No. NSF DBI-1147266. The authors also thank anonymous reviewers for their constructive suggestions that helped to improve this paper.