This paper focuses on the automated extraction of concrete entities from a specialized-domain corpus. Then, in a bootstrapping phase, the candidates are used to extract new candidates. Concrete entities are automatically identified by a set of spatial features. In a spatial scene something is located by virtue of the spatial properties associated with a reference object. The axial properties are represented by place adverbs. Additionally, for identifying referent objects in a sentence we consider syntactical patterns extracted by chunking. In order to reduce noise in results, we take into account a corpus comparison approach and linguist heuristics. Results show high precision in candidates with high weights.
In recent years, the automatic mining of relevant knowledge in the biomedical domain
has become in an interesting research area, particularly in tasks related to the
generation of taxonomies and ontologies
Therefore, in this work we propose a methodology for the automatic extraction of concrete entities implicit in medical documents. Then, in a bootstrapping phase, these candidates are used for extracting a larger set of new candidates.
Linguistically speaking, a main concern is those noun phrases (NP) whose modifiers are relational adjectives and where the noun head is a concrete entity, because relational adjectives introduce semantic features which describe specific properties such as formal, constitutive, telic and agentive qualities (Fábregas, 2007). The identification of this type of NP contributes to delimit the number of possible semantic relations. For testing our method, we work with a corpus of medical texts in Spanish.
We organize our paper as follows: in section 2 we define what a concrete entity is, taking into account the description proposed by Fellbaum (1998) for classifying names in WordNet. Then, in section 3, we show a brief explanation about the representation of space in natural language, according to a cognitive framework. In section 4, we describe the most common deverbal nominalizations in specialized texts. In section 5 we explain the relation noun + relational adjective in order to delineate a set of linguistic heuristics useful for filtering non-relevant adjectives. In section 6 we describe our methodology. In section 7 we offer a description of preliminary results. Finally, in section 8, we give our conclusions. 2.
We understand all that exists in the world as a concrete entity which something can be
predicated (in Aristotle’s categories: substance). For example, concrete entities can be
artifactual categories like vehicles, clothing and weapons, or natural kinds like birds,
fruits and vegetables
Levinson (2004) points out that the spatial thinking is a crucial feature in our lives: we constantly consult our spatial memories in events such as finding our way across town, giving route directions, searching for lost keys, and so on. This importance is mirrored in real discourse where knowledge about formal, agentive, constitutive and telic features, as well as spatial features, are found in specialized domains.
There are three frames of reference lexicalized in language: intrinsic, relative and absolute frame. Intrinsic frame involves an object-centred coordinate system, where the coordinates are determined by the “inherent features”, sidedness or facets of the objet to be used as the ground (i.e., he’s in front of the house). Relative frame of reference presupposes a viewpoint where a perceiver is located, a figure and ground distinct from the viewpoint. Thus, it offers a triangulation of three points, and utilizes coordinates fixed on viewpoint to assign directions to figure and ground (i.e., the ball is to the left of the tree). Finally, absolute frame refers to the fixed direction provided by gravity (i.e., he’s north of the house). 3.1.
Mani et al. (2010) focused on the problem of extracting information about places, considering both absolute and relative references. Their goal was on grounding such references to precise positions that can be characterized in terms of geo-coordinates. These authors use a supervised approach to mark up PLACE tags in documents. SpatialML is an annotation scheme derived from this work and which has been applied to annotated corpora in English and Mandarin Chinese. An automatic tagger for SpatialML extents scores 86.9 F-measure, which is a reasonable performance. On the other hand, Clementini et al. (1997) propose a unified framework for the qualitative representation of positional information in a two-dimensional space in order to perform spatial reasoning. The orientation and distance relations for objects modeled as points can determine positional information. The implicit characteristics of an object are its topology and its extension, while, with respect to other objects, topological, orientation, and distance relations have to be considered.
Evans (2007) explains that a spatial scene is a linguistic unit containing information based on our spatial experience. This space is structured according to four parameters: a figure (or trajector), a referent object (that is, a landmark), a region and —in certain cases— a secondary reference object. These two reference objects configure a reference frame. We can understand this configuration by considering the following example: un auto está estacionado detrás de la escuela (Eng.: “a car is parked behind the school”). In this sentence, un auto is the figure and la escuela is the referent object. The region is established by the combination of the adverb detrás1 which sketches a spatial relation with the referent object. This relation encodes the location of the figure.
Moreover, Evans (2007) points out the existence of axial properties, that is, a set of spatial features associated to a specific referent object. Considering again the sentence a car is parked near to the school, we can identify the location of the car searching for it in the region near to the school. Therefore, this search can be performed because the referent object (the school) has a set of axial divisions: front, back and side areas.
Axial properties are linguistically represented by place adverbs. In this experiment we
only consider adverbs functioning in Spanish with preposition de
Enfrente, delante (Engl. In front to/of); Detrás, atrás (Engl. Behind); sobre, encima (Engl. On); abajo, debajo (Engl. under); dentro, adentro (Engl. In/inside); fuera, afuera (Engl. Out/outside); arriba (Engl. Above/ over).
Additionally, we use some synonymous nouns such as exterior (outside) and interior (in), as well as side nouns synonymous with the dimensions left and right.
According to Martin (1993: 203-220) and Vivanco (2006), from a linguistic perspective, the discourse neutrality in science and technology is presented by means of im1 In English, behind is a preposition. In contrast, in Spanish is an adverb. personation: missing second person, low presence of first person, abundance of impersonal verbs and passive voice, as well as nominalizations hiding actions made by the subject. These nominalizations are used by scientists to support their arguments, coining new terms by means of nouns and summarizing information previously provided in a text.
In line with the frequent use of nominalization in specialized texts, in the case of Spanish, Cademártori, Parodi and Venegas (2006) show data concerning the use of deverbal nominalizations in three domains: commercial, maritime and industrial. The most used suffixes for constructing nouns are: -ción, -miento, -sión, and -dor. 5.
An adjective is a grammatical category whose function is to modify nouns
If we consider the internal structure of adjectives, we can identify two types:
permanent and episodic adjectives
Almost all descriptive adjectives derived of participles belong to this latter class as
well all adjectival participles (i.e., harto/“jaded”). Spanish is one of the few languages
that in its syntax represent this difference in the meaning of adjectives. In many
languages this difference is only recognizable through interpretation. In Spanish,
individual properties can be predicated with the verb ser, and episodic properties with the
verb estar, which is an essential test to recognize what class an adjective belongs to.
In this sense, with the goal of identifying and extracting non-relevant adjectives, we
propose extracting adjectives predicated with the verb estar
Another linguistic heuristic for identifying descriptive adjectives is that only these kinds of adjectives accept degree adverbs or are part of comparative constructions, e.g., muy alto/“very high”, Juan es más alto que Pedro/“John is taller than Peter”. Finally, only descriptive adjectives can precede a noun because —in Spanish— relational adjectives are always postposed (e.g., la antigua casa/“the old house”). 5.2.
According to Bosque (1993) relational adjectives such as salivary in the noun phrase salivary gland belong to a kind of relational adjectives which do not occupy positions in the argument structure of the predicate, but they denote entities which establish a specific relation with the head noun. Bosque refers to these relational adjectives as classification relational adjectives, while the term thematic relational adjectives is left for the other group, e.g., the case of renal infection, where infection is derived from a verb. 6.
In this paper we propose a methodology for extracting concrete entities from a specialized domain corpus with part-of-speech tags. 6.1.
Part-of-Speech (POS) tagging is the process of assigning a grammatical category to
each word in a corpus. The most common taggers used for Spanish are TreeTagger
Chunking is the process of identifying and classifying segments of a sentence by grouping the major parts-of-speech that form basic non-recursive phrases.
In this work, we concern the automated extraction of concrete entities. Concrete
entities relevant to a domain are terms and the most productive patterns of terms
consist of a noun and zero or more adjectives
Concrete entities can be located in spatial scenes as figures or reference objects. In this experiment, only reference objects are extracted with their axial properties that can be linguistically represented as:
<RG|NC><PDEL><DA>?<NC><AQ>* The regular expressions used to extract non-relevant adjectives according to the linguistic heuristics mentioned in section 5.1 are: <RG><AQ> <VAE><AQ> < D.*|P.*|F.* |S.*><AQ><NOUN> Where RG, AQ and VAE as tagged with FreeLing, correspond to adverbs, adjectives and the verb estar, respectively. Tags <D.*|P.*|F.*|S.*> correspond to determinants, pronouns, punctuation signs and prepositions. The expression <D.*|P.*|F.*|S.*> is a restriction to reduce noise, since elements wrongly tagged by FreeLing as adjectives are extracted without this restriction. 6.3.
We use the candidates to concrete entities obtained in the first step as seeds for extracting more candidates. On the one hand, we assume that coordinating phrases where a good candidate occurs have a high probability of containing other good candidates for a concrete entity:
<NC><AQ>*<CC><NC><AQ>* Where <CC> tag corresponds to the disjunction (i.e.: kidney or liver) and conjunction (i.e.: kidney and liver).
On the other hand, noun phrases with at least an adjective take advantage of the noun head of candidates for a concrete entity for finding more specific candidates (i.e., artery-femoral artery):
<NC><AQ>+ 6.4.
We sought to remove non-relevant words from noun phrases before ranking candidates for concrete entities. After the chunking phase, noise was reduced by removing non-relevant open-class words. One of our goals consists of building this stopword list as automatically as possible.
Since concrete entities are terms in the domain, a list of non-relevant words from the domain (i.e., stopword list) can be used to refine the terminology obtained from an automatic process. We considered a list constructed with high frequency words in a reference corpus to have drawbacks because, apart from the selection by occurrence frequency (in the domain corpus, words with high frequency can be terms), human supervision is required in order to determine whether a word is relevant to the domain.
Given the above, we consider that linguistic heuristics operating in a specific
language can be taken into account in order to automate the selection of non-relevant
words. One of the disadvantages, however, is that this leads to language dependence.
For the case of adjectives, in Spanish, characteristic features have been proposed in
order to distinguish between descriptive and relational adjectives as mentioned in
section 5. On the other hand, with a corpus comparison approach, we obtain both
nouns and adjectives where the relative frequency in a reference corpus is greater or
equal than in the domain corpus. These words can be used as part of the stopword list.
Additionally, we take into account empirical evidence concerning the use of deverbal
nominalizations in specialized discourse
Adjectives from the reference corpus can be used as a fixed-size list where nonrelevant adjectives automatically extracted from the domain can be added. These can be obtained taking into account the three heuristics mentioned in section 5.1. Then, these adjectives can be manually reviewed in order to determine their relevance to any specialized knowledge domain (i.e., adjectives as relevant, important, necessary, appropriate, and so on can be considered for the stopword list). This is a fixed-size list and can be the base-list where non-relevant adjectives automatically extracted from the domain can be added. 6.5.
We evaluate termhood of simple words by means of rank difference
Kit and Liu (2008) only focus on extracting single-word term candidates, so they only weigh words occurring in both the domain and the general corpus. In our experiment we also consider words that only occur in the domain corpus. We assumed that the reference corpus is large enough to filter out non-relevant words, hence words only occurring in the domain corpus have a higher probability of being relevant and the word’s frequency reflects its importance: We consider that the larger the reference corpus, the higher the exhaustivity3 of open class words of general usage, as well as a higher probability that specialty terms occur at least one time (the reference corpus was collected from an online newspaper where news about science and technology are published too), so that we would expect a higher precision in ranking. 6.6.
Formally, if a candidate noun phrase (np) has a length of n words, w1 w2 …wn, where n>1, then the ranking of the candidate np is the sum of the frequency of np as a whole plus the weights of all the individual words wi: (2) (3) 7.
This section presents the results of our experiment considering a subset of 1,200,000 tokens of the MedLineplus corpus. 7.1.
The source of textual information is constituted by a set of documents of the medical domain, basically human body diseases and related topics (surgeries, treatments, and so on). These documents were collected from MedlinePlus in Spanish.
The size of the corpus is 1.2 million tokens, but we carried out our experiment with a subset of 200,000 words in order to determine manually the number of concrete entities present in the results. As an ongoing work, we are manually determining how many concrete entities are present in the complete corpus. We chose a medical domain due to the availability of textual resources in digital format. Finally, we assume that the choice of domain does not suppose a very strong constraint for generalizing the results to other domains.
With the goal of ranking words relevant to the domain by means of their relative
frequency ratio, a large reference corpus was collected from an online newspaper4 with
new articles from 2014 (the size of corpus is about 5 million tokens). URLs from the
3 Exhaustivity of a document description is the coverage it provides for the main topics of the
document. So, if we add new vocabulary terms to a document, the exhaustivity of the
document description increases
The first phase of extraction of candidates to concrete entity without filters achieves a global precision of 56%. The tables 2 and 3 show precision with different thresholds of candidates starting with the better ranked candidates. With the stopword list built as mentioned in section 6.4, we achieve a global precision of 76%. Global precision with a stopword list reflects an improvement of 20%, but a significant loss of 17% of true candidates. As can be seen from these tables, the ranking of words and noun phrases is useful for sorting results from the most relevant to the least relevant results.
The bootstrapping phase taking into account coordinating phrases achieves a set of 1248 candidates, of which 262 are new true candidates. The global precision with this second phase is of 47%, with a precision by thresholds as shown in table 3. The advantage of this phrase structure is that single-word candidates can be extracted.
On the other hand, the bootstrapping phase considering noun phrases achieves a set of 2796 candidates, of which 1534 are good candidates. The global precision of this phase is of 55%, with a precision by thresholds as shown is table 3. One disadvantage of this structure is that only candidates with at least one adjective can be selected.
Table 3 shows a better performance with noun phases. The identification of the concrete entities present in corpus is an ongoing task that will let us evaluate in terms of recall too.
The candidates in a bootstrapping phase give us insight about the kind of semantic relations implicit in noun phrases of the type <NC><AQ>. Given the phase of reduction of non-relevant adjectives, we have a great deal of relational adjectives where it is possible to find different relations. For example, salivary gland has implicit a telic relation. On the other hand, testicular gland has a part-whole or locative relation. Finally, meibomian gland may be considered as a specific type of gland.
With respect to the extraction of lexical relations, specifically
hyponymy-hypernymy relations
We discussed a methodology for extracting concrete entities in the medical domain. Concrete entities have been studied since Aristotle’s works, particularly in his biological and zoological descriptions. According to Aristotle’s categories (the first category), many things can be predicated of substances. We assume that substances are concrete entities, with a more extended meaning, i.e.: the eight tangible categories formulated by Fellbaum for WordNet (1998). Thus, we consider that the automated identification and extraction of this kind of information is an important advance in further NLP tasks.
Cognitive abilities as the spatial knowledge and his representation in natural language are important for our extraction methodology. We observe that spatial descriptions are frequent in specialized discourses. Additionally, we propose a further step of bootstrapping in order to find a great number of candidates for concrete entities. Candidates with a concrete entity as a noun head and a relational adjective show semantic relations as part-whole, locative, agentive and telic, which can be interpreted, at first, as hyponymy/hyperonymy relations.
On the other hand, to assign relevance to words is an important step for ranking candidates, according to our exposed results. In this sense, as ongoing work, we are collecting more information about science and technology at the same electronic journal in order to improve the results in the ranking process.
Finally, it is necessary to mention that POST taggers as FreeLing and TreeTagger fail in the task of identifying nouns, adjectives and verbs closely related with the domain. This failure has a negative impact on the results. We believe it is important to face this problem in future extraction tasks.
This paper has been supported by the National Commission for Scientific and Technological Research (CONICYT) of Chile, Project Numbers: 3140332 and 11130565.