2.1 Searching for Finnish language example sentences

For the FinnFN project, around 80,000 English sentences representing 866 frames were chosen from the FrameNet version 1.5 corpusFootnote ⁶ (https://framenet.icsi.berkeley.edu/), after which the annotated parts of the sentences were translated into Finnish by professional translators. The Finnish translations with their English original sentences constitute a separate corpus (the English–Finnish TransFrame corpus) which is further analyzed in a separate article (Lindén et al. to appear). The main finding of that article is that 92% of the frames were transferred with their meaning intact when the frame elements were translated from English to Finnish. However, in the current article, we focus on the frame universality hypothesis using attested natural language as our starting point. We also focus on the modifications needed in the FrameNet annotation conventions in order to annotate authentic Finnish example sentences.

After translating the English FEEs and the FEs into Finnish, the Finnish translations were used in an automated search for similar sentences in the National Library Newspaper corpus,Footnote ⁷ a corpus of Finnish newspapers and magazines compiled by the National Library of Finland comprising five billion words of running text.Footnote ⁸ The text was preprocessed with the Turku Dependency Parser.Footnote ⁹ The corpus search was carried out in order to find attested natural language examples of frame evoking sentences to ensure that the translated frames were not mere ‘translationese’. The goal was to find out whether there exist example sentences in Finnish for all the frames in BFN included in the parsed FrameNet 1.5. corpus. Because we sampled from authentic Finnish language, we could examine whether the frame elements defined by BFN also occurred in authentic Finnish sentences. In addition, it was possible to see if and how the syntax and semantics of the annotated Finnish sentences differed from English, for example, whether the annotated Finnish example sentences contained semantic roles that could not be covered by the frame elements of the original BFN frame.

In our attempt to find authentic Finnish example sentences, different search strategies were tried using Corpus Workbench queries in Korp (Borin, Forsberg & Roxendal Reference Borin, Forsberg and Roxendal2012) in order to find which automatic searches gave the best results. The best search strategy (in which 36.7% of the search results were accepted for manual frame annotation) was to use the word form, lemma and the syntactic dependence label of the FEE. The second best search strategy (in which less than 30% of the search results were accepted for manual frame annotation) was to determine the word form, lemma and the syntactic dependence label of the FEE as well as the dependence label of the FEs immediately to the right or left of the FEE. The worst search strategy (only 0.1% good examples) was to narrow down the search too much by using the word form, lemma and dependence labels of both the FEs and the FEE. Only for 14 frames did we have to manually retrieve example sentences as the automatic search strategies did not provide any suitable results. These problematic frames are further discussed in Section 4.

The low success rate of some of the search strategies mainly indicates that although five billion words is a big corpus, it is not big enough for fully automating the procedure to find authentic examples. The results of the automatic search strategies were therefore further pruned manually. All the search results were read and the best examples were selected for annotation. An example was considered good if it evoked the same frame as the original English example and preferably included the same frame elements. The main focus was on the FEE and other core elements. The discarded sentences are relevant examples of ambiguous lexemes as FEEs, which are disambiguated into specific frames by the core FEs in the context. Some of the discarded examples sentences were collected for future use as they are relevant negative examples when training a machine to automatically identify examples of frames in running text.

2.2 Annotating the FinnFN corpus

Each example sentence deemed to represent its intended frame was manually annotated with the corresponding FEE and FEs. The annotation method that was used in the FinnFN project can be described as a frame - by - frame approach as opposed to the full-text annotation or the lemma - by - lemma approach used for instance in the German SALSA project (Burchardt et al. Reference Burchardt, Erk, Frank, Kowalski, Padó and Pinkal2006) and Japanese FrameNet project (Ohara et al. Reference Ohara, Fujii, Ohori, Suzuki, Saito and Ishizaki2004, Ohara Reference Ohara2012). The frame-by-frame approach assumes that the sentences to be annotated contain a lexical unit evoking the intended frame, in contrast to the lemma-by-lemma approach, which takes a sentence and determines which frames its lemmas might evoke. If an FEE in a sentence context clearly evoked an unintended frame, the sentence was left unannotated. For more ambiguous cases, the adopted approach probably resulted in more varying FEEs for certain frames than the lemma-by-lemma approach would have yielded, as a large sample of sentences from different topics were evaluated. As stated above, we created no new frames at this point, instead the frame-by-frame approach made it easier to account for the diversity of FEEs that may evoke a frame. Because the FinnFN project used the frames and annotation conventions originally created for English, some special characteristics of the Finnish language became readily observable when applying the BFN annotation conventions to Finnish.

After an initial round of manual annotation, which was done by four linguists working in parallel on the automatically retrieved sentences representing all frames, a second round of annotation was carried out, in which one linguist checked all the sentences representing each frame. During this second round, with the experience of having annotated the full corpus, examples could be reassigned to more suitable frames, which resulted in example sentences representing 938 frames. Typically the examples would be assigned to more specific frames during the second round, as new frames had been added to BFN during the course of the FinnFN project.

The FinnFN corpus currently (June 2017) consists of 40,721 annotated example sentences (with approximately one million lexemes) that represent 938 frames. The FinnFrame corpus is thus one of the biggest FrameNet projects so far (together with the BFN and the Swedish and the German FrameNet projects) and can therefore readily be utilised in corpus-linguistics for frame semantic research. The applicability of the English frames and frame elements in Finnish will be further discussed in Section 4. In most cases, both the frames and the frame elements based on English apply to Finnish without modifications. Each FE and its use in the FinnFN corpus has been described in detail on the FinnFN wiki page.Footnote ¹⁰

As in BFN, in addition to semantic criteria, morphosyntactic criteria may also apply to frame elements in the FinnFN annotation. For example, in Finnish, the frame element Topic is often expressed with the elative suffix (-stA), and Message is typically a subordinate clause beginning with the conjunction että ‘that’. This is exemplified in examples (1) and (2).

1. (1)

   

2. (2)

   

The examples above demonstrate the notational conventions we have assumed in this paper: FEEs in example sentences are written in capital letters and the intended frame is stated after the example sentence in parentheses in a monospaced font. Other frame elements, when relevant, are enclosed in square brackets, annotated with the element's label in small capitals. If no language is specified, the examples are in Finnish or English.

Table 2 provides an example of the annotation layout used in the FinnFN project. The table shows an example sentence annotated with the frame Artificiality, in which an entity is evaluated for whether it evokes a particular reference category. In our annotation work, we used a three-layered annotation layout with tags on separate layers in order to express all relevant frame elements for a word. The separate layers are numbered for clarity so that the FEE, is on Layer 1, other FEs on Layer 2 and, when frame elements overlap, additional FEs are on Layer 3. The FEE in this example is epäaitoja ‘fake’. The whole noun phrase epäaitoja koruja ‘fake jewellery’ is annotated as Entity since the adjective epäaito ‘fake’ is a dependent of the head (koruja ‘jewellery’) of the NP. On the third layer, koruja ‘jewellery’ is marked with the tag Reference_category as it also indicates the type relative to which the entity is evaluated. Not shown in Table 2 is the fact that the FinnFN corpus also provides morphosyntactic and dependency annotations on additional layers to make the corpus useful for researching the interface between syntax and semantics.

Table 2.

A corpus example Sekä aitoja että epäaitoja koruja kannattaa hoitaa hyvin ‘Both real and fake jewellery should be treated well’ with annotations for the frame Artificiality using epäaitoja ‘fake.a’ as FEE on Layer 1, the entity epäaitoja koruja ‘ARTIFICIAL jewellery’ on Layer 2 and koruja ‘jewellery’ as the reference category on Layer 3.

Annotation of so-called null instantiations is a feature of BFN that was not used in the FinnFN project. Null instantiation is a category for annotating absent semantic constituents or frame elements (Fillmore et al. Reference Fillmore, Johnson and Petruck2003:245). The term null refers to the fact that a core frame element is absent. Conceptually salient FEs are not always realized as lexical or phrasal material in a sentence, but they have nevertheless been annotated in BFN since they provide relevant insight into the omissibility of semantic material (Ruppenhofer et al. Reference Ruppenhofer, Ellsworth, Petruck, Johnson and Scheffczyk2010:24–26).

3.1 Compounds in FinnFN

In English, the few compounds that are written as single words are handled in BFN as indivisible units. In example (3), the modifying part bath of bathrobe is not annotated with the semantic role Use, even though it denotes the use of the garment. We use a typographical convention for annotating an FEE with capital letters, as in BNF, but we also annotate a frame element, e.g. Garment, with a subscript in small caps, which is useful when the two overlap as in example (3).

1. (3)

   

Compounds consisting of a head noun and a relational modifying part that are not written as a single indivisible word have been annotated separately in BFN, e.g. as in example (4).

1. (4)

   

In Finnish, the matter is somewhat more complicated. In Finnish, compounds are practically always written as single indivisible words, and in theory an unlimited number of new compounds can be formed. In FinnFN, we sometimes need to label the parts of a compound with different semantic roles as the part of a compound that modifies the head most often carries a separate semantic meaning. In the FinnFN annotation, we therefore separated the modifying part from the head of the compound and annotated it as illustrated in examples (5) and (6) by the equivalent compounds riding helmet and ratsastuskypärä. Note the word spacing in English and the lack of spacing in Finnish, which is also reflected in the annotation.

1. (5)

   

2. (6)

   

The FinnFN database is not fully uniform in the annotation of compounds and exhibits some variation in what parts are deemed semantically important enough to be labelled separately. Parts are annotated separately in cases where one part of the compound is the FEE and the other is not, similar to the English head noun with a relational modifier as in examples (7) and (8).

1. (7)

   

2. (8)

   

Some lexicalized compounds have been annotated as a whole in Finnish: aluspaita lit.: ‘undershirt’, ‘camisole’; or rintaliivit, lit.: ‘breast vests’, ‘bra’ as in examples (9) and (10).

1. (9)

   

2. (10)

   

The part of the compound that evokes the frame is annotated as the FEE, and the other relevant parts of the compound are annotated according to their semantic roles. In example (11), only the head is annotated as the FEE.

1. (11)

   

The general idea has been to annotate all the relevant and clearly transparent frame elements of the compound. If the compound is opaque, it has been treated as a single unit. However, drawing the line between the two is not always straightforward. Ruppenhofer et al. (Reference Ruppenhofer, Ellsworth, Petruck, Johnson and Scheffczyk2010:41) admit that distinguishing between conventionalized and non-conventionalized compounds is not always clear in English either. Often it was left for the BFN annotator to decide and evaluate.

Finnish compounding resembles German and Swedish compounding more than that of English, so the comparison with German and Swedish FrameNet-related projects regarding how they have dealt with compound annotation is useful. In the German SALSA project, the compounds have been divided into two units and can thus be annotated separately.Footnote ¹¹ In the frame Hostile_encounter the German word Machtkämpfe (lit.: ‘power struggles’, ‘struggles for power’) consists of the FEE Kämpfe and the part denoting the Topic of the encounter, Macht. The compound is structurally very similar in Finnish; the compound valtataistelut consists of lexemes with a similar meaning as in German (valta ‘power’, taistelut ‘struggles, battles’).

The Swedish FrameNet project has developed different strategies for annotating compounds.Footnote ¹² As in Finnish, non-compositional, fully lexicalized compounds are not given any internal analysis in the lexical source database, which treats solid compounds, i.e. single orthographic words, just as it treats single-word items, and does not normally define their formal structure explicitly (Friberg Heppin & Petruck Reference Friberg Heppin and Petruck2014:68). However, in the Swedish FrameNet project compositional compounds are annotated as far as they are compatible with the frame they evoke. The compound parts are annotated separately especially in the cases where one of the parts is the FEE. Usually the whole compound and the compound head are FEEs evoking the same frame. In these cases, the whole Swedish compound is tagged as an FEE, while the modifier is tagged as the appropriate FE, as in example (12).

1. (12)

   

Like in the Swedish project, FinnFN treats the whole compound as an FEE when it has transparent compositional parts evoking the same frame.

A common compound type expresses an item and the material it is made of. In example (13) in Swedish illustrating the Accoutrements frame, guldring ‘gold ring’ has been separated into FEE (ring) and FE denoting Material (guld).

1. (13)

   

In Finnish, the compound villasukat ‘wool socks’ is also split into head and modifier, each part being annotated separately as in example (14) and only the head acts as the FEE.

1. (14)

   

We note that the Finnish compound annotation is crafted to be similar to that of German and Swedish extending some of the annotation conventions that were available in the BFN.

3.2 Finnish morphology

Agglutination, a central feature in Finnish, presented a challenge to be solved in FinnFN. To deal with morphemes attached to Finnish word stems, most typically possessive and verbal suffixes and comparatives, an annotation convention resembling that of compounding was further developed in which words were split into morphemes if needed. Morphemes have been annotated separately to apply a BFN frame when a core element of the frame would otherwise have been left out and the meaning of an FE was carried by a morpheme, e.g. in the case of the Possessor being marked only by a possessive suffix, see example (15). This convention allowed all the core elements of the frames to be accounted for when annotating Finnish.

Hebrew is another FrameNet language with rich morphology, but the Hebrew FrameNet has not annotated possessives and other bound morphemes as FEs. Instead, the elements have been marked as externally instantiated, i.e. no single word or phrase is marked as representing the frame (Hayoun & Elhadad Reference Hayoun and Elhadad2015) following the annotation conventions of the Spanish project, which did not mark semantic roles such as subject markers expressed by bound morphemes. The Finnish project, by contrast, decided to annotate some of the key morphological units because of the agglutinating nature of the language, i.e. the morphology was seen as too important to be omitted: sometimes the FEE, the most essential element evoking a frame, is expressed only by a suffix, which will be discussed later in this section. In Finnish, a suffix may, for example, correspond to an English preposition as the FEE. We also claim that in order to develop FrameNet into a truly universal semantic annotation system, semantic role labeling for agglutinative and polysynthetic languages should be further developed. The assumption that semantic content can only be carried by independent FEEs is rather limiting from a universal point of view, since semantic content can also be carried by bound morphemes as demonstrated by Finnish. If the FrameNet annotations had originally been developed for West Greenlandic, for instance, it would have been essential to account for bound morphemes typical for polysynthetic languages. This is a central point that we wish to emphasize and develop further in the future.

However, in our current project our annotation deals primarily with independent words and phrases: the annotation has been marked on an independent word if one is present, but on a bound morpheme if not. This is illustrated in examples (15) and (16) with regard to both a possessive suffix and a subject marker.

1. (15)

   

2. (16)

   

In at least some of the other languages for which a FrameNet currently exists, the agent is visible in the verb form. It seems, however, that in most of these languages it is not possible to leave the overt subject out, and a situation similar to example (16) above will never arise. One of the exceptions is Japanese, where the person is generally unmarked altogether, whereas in Spanish, Italian and Brazilian Portuguese there is often a person marker on the verb. The Japanese FrameNet has dealt with this by marking the agent as a null instantiation similar to the marking of an absent agent in passive sentences in English. In Finnish, where the subject is always visible (often only) on the verb form, it was deemed relevant to mark the FE on the particular morpheme.

Possessives and comparatives are among the relevant phenomena for some of the frames although they are not core elements. Of these, possessives as affixes can be found only in Finnish and Hebrew of the current FrameNet languages. Comparative and superlative endings, however, are used for indicating degree in other languages as well. In BFN, one can of course claim that the element Degree marks only the intensity of the comparison and not the comparison as such, so there is no need to mark the comparative ending -er alone. In example (17), the expression of Degree of comparison is only visible in the ending and has been left unmarked in BFN.

1. (17)

   

The element Degree of comparison is a peripheral element in almost every frame of BFN, so omitting its annotation does not leave out core information of the frame. In some individual cases, BFN has marked Degree on an entire word denoting comparison, as in example (18), in which the word older (earlier in the same sentence) contrasts with the word newer.

1. (18)

   

Note that the word ‘more’ is often annotated in BFN with Degree as in example (19) annotated with the Obscurity frame. In Finnish, it is mandatory to use comparative morphemes for expressing the phenomenon represented by the English word ‘more’, so if the Finnish comparative morphemes were left unannotated, this information would have to be omitted in the Finnish annotation. Comparative morphemes have therefore often been annotated in FinnFN.

1. (19)

   

Having introduced the annotation of a bound comparative morpheme for Degree, we can use this for annotating the comparative ending (-mpi) in ärsyttävämpi ‘more annoying’ along with tavallista ‘than usual’ as Degree in example (20), while the rest of the word (ärsyttävä-) is the FEE.

1. (20)

   

Not all Finnish endings, such as enclitic particles like the interrogative particle -kO, carry meanings that are relevant for a frame. Polarity or tense do not affect the frames, and for instance the negation verb did not have any particular role or meaning evoking a frame in FrameNet v1.5. Consequently, particles like -pAs/-hAn (various meanings; contradiction or mirative, among others) and -kin ‘too, also’ have not been annotated in any frame in FinnFN.

The original BFN database dealt only with lexical units (Fillmore & Baker Reference Fillmore, Baker, Heine and Narrog2010:338). However, there are also grammatical structures which evoke semantic, pragmatic or interactional frames on their own (Fillmore & Baker Reference Fillmore, Baker, Heine and Narrog2010:338), so the current BFN deals with much more than FEEs. The present study shows that in Finnish even the FEE itself can be a bound morpheme or a morphosyntactic structure, even though there are only a few examples of this kind in the current corpus. In the initial translations of annotated English examples, each English FE was matched by a Finnish one, so the meanings conveyed solely through a specific grammatical structure or bound morpheme are underrepresented. The following examples (21)–(23) demonstrate three instances in which FEEs might be annotated in grammatical structures or bound morphemes.

1. (21)

   

2. (22)

   

3. (23)

   

The annotation system created for annotating parts of compound words makes it possible to mark bound morphemes as FEEs in a similar way. We annotated text examples with the frame Partitive, as in example (23b) above. In this frame, the sought meaning in Finnish can only be conveyed through the elative suffix (-stA).

Altogether, the instances where the FinnFN corpus currently deviates from the BFN annotation are in the marking of FEs on personal, possessive and comparative markers, and the marking of the FEE on the case suffix in the sole case of the Partitive frame. In future development, it might be justified to expand the range of data to include more varied annotations for different types of morphologically encoded linguistic phenomena to the extent that they encode language-independent meanings.