Navigation – Plan du site

AccueilNuméros62Positional skipgrams for Bambara:...

Positional skipgrams for Bambara: a resource for corpus-based studies

Les skipgrams positionnels pour le bambara : une ressource pour la recherche linguistique orientée corpus
Позиционные скип-граммы для бамана: ресурс для корпусных исследований
Kirill Maslinsky

Résumés

L’article présente un nouveau paquet de données linguistiques de fréquences de n‑grams pour le bambara, basé sur le sous-corpus désambiguïsé du Corpus bambara de référence. Les n‑grams sont des skipgrams positionnels qui capturent l’information sur la co-occurrence des lexèmes avec des catégories grammaticales à des positions différentes. Ces n‑grams ont été conçus pour tirer profit de ce type d’informations disponibles dans le corpus bambara morphologiquement annoté, vu le volume limité des données textuelles. La discussion de la méthodologie et les données utilisées pour la construction des n‑grams pour le bambara est suivie par quelques illustrations d’utilisation des skipgrams positionnels dans des recherches linguistiques basées sur un corpus.

Haut de page

Texte intégral

Introduction

1N-grams — fixed-length sequences of adjacent tokens collected from textual data — have been widely used in computational linguistics and natural language processing for decades (Rosenfeld, 2000). A frequency list of n‑grams obtained from a corpus has proven to be a simple yet powerful tool to represent contextual information and sequential phenomena in natural language. Publishing n‑gram frequencies is also a way to share statistics on word distribution in a corpus, the most notable example being the Google Ngrams corpus (Brants & Franz, 2006).

2The idea that is key to the practical success of n‑grams in a wide variety of language modeling tasks (from spelling correction to speech recognition) is to extract the information encoded in the relative positioning of linguistic units into a list of easily quantified atomic “co-occurrence events”. When used in a general sense, the approach leaves room for flexibility in choosing how to build n‑grams, and what to include in them. Adjacency constraints can be relaxed to include items occurring anywhere within a fixed-width context window, thus producing skipgrams. There is also no need to limit the scope to the lexical or graphical level, as in the traditional word n‑grams and letter n‑grams, or even to the surface level in general. In cases when linguistic annotation is available for the text, it may be used for building n‑grams. The most common example of the latter is to make n‑grams from part of speech tags of subsequent words that reveal recurrent word order patterns. The part-of-speech n‑grams are used in diverse fields, for instance, in text-to-speech generation (Taylor & Black, 1998) or in sentiment analysis (Jaggi, Uzdilli & Cieliebak, 2014). Thus n‑grams can represent phenomena other than plain lexical co-occurrence.

  • 1 The corpus search interface as well as general info about the corpus are available online at: http: (...)

3This article presents a new online dataset of linguistically rich n‑gram frequency data for Bambara based on the disambiguated part of the Corpus bambara de référence1 (Vydrin, 2013). N‑grams in this dataset were constructed with the aim to capture those types of information that are available in the morphologically annotated corpus of Bambara. Beyond the usual lexical focus, n‑grams were supplemented with paradigmatic grammatical information and positional features that should allow for inferences to be made about various aspects of morphosyntax.

4Making this dataset publicly available is a way to provide access to the linguistic data derived from the full annotated corpus for a wider audience of students and researchers without disclosing copyright-protected texts. The data format has to be general enough to allow open-ended exploration and use of the data in broad areas of linguistic research, language learning, and downstream NLP tasks. In my view, the n-grams list format matches this objective and has the additional benefit of retaining readability by a human as well as a machine. While simple tabular format makes data easily quantifiable for research and engineering tasks, for a human reader, a frequency-ordered n‑grams list preserves meaningful linguistic categories such as lexemes, grammatical tags, and relative word positions in a sentence.

5The article is structured as follows. Sections 2–4 explain the methodology and data used for constructing n‑grams for Bambara, followed by section 5 with brief illustrations of how the n‑gram data presented here may be employed in corpus-based linguistic research. A discussion of the advantages that positional skipgrams provide in the low-resourced setting is presented in section 6.

Positional Skipgrams

6The approach used in this article to combine lexical, grammatical, and positional information in a single n‑gram framework is tentatively labeled here positional skipgrams. To make sense of this framework, consider a sentence in Bambara that has part of speech tags defined for each token.

(1) í k’ à dɔ́n nàta mùso ɲùman tóbi .
  pers pm pers v cop n n pm n adj v c
    pm:-5 pers:-4 v:-3 cop:-2 n:-1 n:0 pm:1 n:2 adj:3 v:4 c:5
‘You should know that a greedy woman won’t cook a good sauce’
  • 2 The term n‑gram presupposes a variable number of co-occurring words, but in the data and in the exa (...)
  • 3 Depending on the task at hand, it may be convenient to record reverse co-occurrence events (pm:1 —  (...)

7For this sentence, the list of regular word n‑grams (bigrams)2 would include the pairs of consecutive words: í‑k’, k’‑à, à‑dɔ́n, etc. This is the most common (“default”) reading of the term word n‑grams in the literature. In case of skipgrams, pairs of all words that fall within the fixed-width context window (five words on each side in our example) are considered a co-occurrence. The list of skipgrams would include pairs that are up to five words apart in the sentence: í — à, í — dɔ̀n, í — kó, k’ — mùso, etc. In contrast to regular word skipgrams, in positional skipgrams a numeric index is appended to the second item in the pair that reflects its relative position in respect to the first item: 1 indicates the next word to the right, ‑3 indicates the third word on the left, and 0 is the word itself. Besides that, in the variant of positional skipgrams presented here part-of-speech tag is used instead of the co-occurring word as the second item of the skipgram. In our example, for the word mùso the following positional skipgrams will be generated: mùso — n:0, mùso — pm:1, mùso — n:‑1, etc.3 Essentially, this list may be read as a set of statements equivalent to: “in this sentence, mùso co-occurs with a noun in a previous position, with an auxiliary (predicative marker) in the following position, and is itself tagged as noun”.

8Instead of tracking word co-occurrence events, positional skipgrams record the information on the occurrence of the word in a certain position in the surface syntactic structure, to the extent that syntactic information is reflected in the sequence of part of speech tags. As usual with n‑grams, this positional occurrence is represented as a series of atomic “co-occurrence events”. In this representation, the structure of the context is lost, but the disparate events (words and sentences) thus become comparable. For example, two occurrences of a word can share a significant part of their positional skipgrams while not sharing that many context words. The same principle makes it possible to compare different words by the similarity of their syntactic contexts (in terms of the relative frequencies of their positional collocates).

9While the idea of appending the positional index to the collocate is all that is needed to define positional skipgrams in general, several other constraints should be observed to make them more relevant as linguistic data and to make sure that they are tractable in downstream computational tasks.

  1. Note that in the examples above words are never included as positional collocates to other words. While technically nothing prevents us from doing so, the focus of the method is to relate words to the underlying linguistic categories, and more generally, to recurrent phenomena at the non-lexical level. Essentially, what we are interested in is the type of contexts that words are likely to share. Moreover, in a less-resourced setting where lexical data are already sparse, multiplying the lexicon size by the positional dimension would be clearly detrimental for statistical inference of any kind.
  2. Since the positional part of speech tags are included as a proxy for syntactic structure, it is reasonable to require that n‑grams do not cross sentence boundaries. At the same time, punctuation tokens could be recorded as collocates to keep track of the relative positions of the word in respect to sentence and clause boundaries (for instance, the final stop in the example (1) that would produce mùso — c:5).
  3. To further compensate for lexical sparsity, it makes sense to include n‑grams consisting of two positional tags alongside positional skipgrams with words. For instance, accumulating frequency counts for n:0‑pm:1 would help track the fact that nouns tend to fill the position before predicative markers as an integral part of the data.

Related work

10In such a long and rich tradition as application of n‑grams in natural language processing hardly anything can be truly novel. But to summarize, compared to other n‑gram building methods positional skipgrams are characterized by the two distinctive features: they combine information from different levels of annotation, and they incorporate positional information into the n‑gram in the form of a positional index.

11Positional skipgrams as implemented in this article combine features from two different levels of annotation in the form of n‑grams, namely lexical items and grammatical categories. This simple cross-level setup seems to be uncommon in practical n‑gram applications in recent literature on natural language processing. This could be due to the fact that in the history of language modeling with n‑grams grammatical categories (part-of-speech tags and the like) were primarily seen as a desired result to be produced by the model or at least as a latent variable for better word prediction, but definitely not as input data (see, for example, [Brown et al., 1992]).

12In contrast, in more basic research, where the goal is language description rather than solving applied tasks, there is a rich tradition of looking for patterns that combine lexical items with syntactically defined slots. In corpus linguistics, the constructs that encompass both lexical and grammatical components in a single pattern were used to identify idiomatic constructions, and to make inferences about lexical meaning (e.g. polysemy) based on usage. These include behavioral profiles suggested by Hanks (1996) as a generalization of verb complimentation patterns; collostructions (Stefanowitsch & Gries, 2003) that track co-occurrence between words and constructions; colligations as “co-occurrence of word forms with grammatical phenomena” (Gries & Divjak, 2009); and more ad‑hoc instruments, like gapped patterns used to identify grammatical constraints in multi-word expressions (Kopotev et al., 2013). A common methodological feature of all the above approaches is that to collect data, researchers have to pre-define a specific construction or pattern they are looking for. The positional skipgrams approach is different from all the above constructs in that it does not specify a particular construction, but rather captures any constructions that can be reduced to the set of lexical items and grammatical categories positioned in text at some fixed interval in respect to each other.

13Positional skipgrams explicitly record the position of a collocate relative to the current word. Common skipgram-based models may incorporate positional information implicitly. In particular, it has been shown that word2vec actually benefit from taking distances between words into account by using the decreasing weight coefficient for more distant words (Levy, Goldberg & Dagan, 2015). The closest to our approach is the work by Ling et al. (2015) that included “what words go where” type of information in addition to “what words go together” in word2vec by creating separate models for each position of a context word relative to the current word. The idea to have positional information as a part of term in n‑gram itself was motivated by the example rythmical n‑grams in the work of Petr Plecháč in quantitative analysis of poetry. He uses n‑grams to represent the structural position of sounds in the verse line (Plecháč, 2019: 38).

Dataset description

14The dataset presented in this article was built using the manually disambiguated part of the Bambara Reference Corpus (corbama-net). As of December 2019, the disambiguated subcorpus contains 1.3M words in 1650 documents. The corpus provides token-level morphological annotation as well as document-level metadata on the author, the source of the text, and several tags categorizing the medium, genre, and theme of the text (on metadata, see for details: Davydov, 2010). The annotation provided in the corpus was obtained using the morphological processor Daba based on a dictionary and a set of rules (Maslinsky, 2014), followed by manual disambiguation by Bambara-proficient operators.

  • 4 See the full list of the glosses for grammatical morphemes and auxiliaries for Bambara at: http://c (...)

15The annotation layers available in this subcorpus include the orthographically normalized token (part of the corpus is in the old Bambara orthography), lemma, part of speech tag, and a gloss (lexical equivalent) in French. For multi-morpheme words there is also a recursive structure that annotates each morpheme with the similar attributes of a form, a part of speech tag, and a gloss. Grammatical morphemes, as well as standalone function words are assigned a Leipzig-style formal gloss from a standard list of glosses for Bambara4 instead of the French equivalent.

  • 5 See information on the dictionary at http://cormand.huma-num.fr/bamadaba.html.

16The main objective of publishing this dataset is to present quantitative data on morphosyntactic regularities and variation in the corpus. Hence other types of variation that are attested in the corpus are not represented, namely orthographic variation, dialectal variation, and tonal variation. To eliminate these types of variation only orthographically normalized forms are used throughout the dataset. All variants of the same lemma (dialectal, tonal, phonetic, etc.) were transformed to the canonical form, which is operationalized as the first variant listed for a lexical entry in the Bamadaba dictionary5.

17To make the most of the structural information available in the annotation, the basic positional skipgrams model presented above is supplemented with the n‑grams based on the morpheme-level grammatical information. To keep data sparsity at a manageable level, the principle of limiting the right-hand side of the n‑grams to the closed-class and frequent phenomena was observed (see section 2 for details). Thus out of the morpheme-level annotation layer only morphological tags from a standard list of glosses were taken into account. The resulting list of skipgrams includes the pairs of the following form:

  • wordform (or lemma) — part of speech tag + position
  • part of speech tag — part of speech tag + position
  • wordform (or lemma) — standard gloss + position
  • standard gloss — standard gloss + position

18Numerals and punctuation are not included as the left-hand side items in the n‑grams, but may appear as positional collocates on the right-hand side. The context window width for building skipgrams is defined to be five tokens on each side of the word, but is not allowed to cross sentence boundaries. Sentence boundaries are included in the list of positional collocates using a conventional SENT tag. The choice of five tokens as a context window width is arbitrary, although it is in accord with the common practice in other n‑gram-based models. It is reasonable to expect that clause length in Bambara will not frequently exceed this width, so that not much useful statistics could be collected with a wider context window.

19For the convenience of dataset users, the skipgram frequency data is presented in several variants. First, the data is split according to the basic lexical item used for building skipgrams that is either an orthographically normalized wordform, or a canonical lemma. Second, frequency data on both wordfrom-based and lemma-based skipgrams are presented in two forms: an aggregated variant showing total counts for a whole corpus, and a disaggregated variant showing document-level frequencies.

20The data is presented in a text-based tabular format. Skipgram frequency tables are in the TSV (tab separated values) format and contain the following columns:

  • lexical item, tag or standard gloss;
  • its positional collocate;
  • total frequency of the lexical item/tag/gloss;
  • total frequency of the collocate;
  • frequency of the co-occurrence of the item with the collocate (n‑gram frequency);
  • a label indicating the type of the collocate (word–tag, tag–tag, etc.) to facilitate filtering.

21The document-level frequency data has an additional column with the document ID that precedes the list. Document-level metadata are provided as a separate CSV file that can be linked to the document-level skipgram frequency tables based on the value of the document ID field.

Possible applications

22This section presents a few examples of the ways in which information contained in the positional skipgrams can be rearranged and explored to address linguistic queries. The statistical processing of the data in the examples is intentionally kept to a minimum, in order to demonstrate conceptual simplicity and interpretability that the lists of positional skipgrams can offer by themselves. The examples presented in this section are neither an exhaustive list of the uses for positional skipgrams, nor a set of finished linguistic case-studies in Bambara; they are meant to serve just as illustrations of possible applications.

Lexical comparison

23Let’s start with a simple query on lexical semantics where the application of the positional skipgrams is quite straightforward. In Bambara, there is a pair of moderately frequent verbs, gòsi and bùgɔ, both of which mean ‘to hit’. Having corpus data at hand, we may make inferences about the semantic differences of these two verbs based on the differences in their context distributions. In addition to the traditional reading of the concordance for both verbs, positional skipgrams can offer a summary of morphosyntactic positions of each verb together with frequency statistics (see table 1).

Table 1. Top 8 frequent positional skipgrams for bùgɔ and gòsi. Columns indicate: freq1—the frequency of the verb itself; freq2—total frequency of a collocate in a corpus; ngram—frequency of co-occurrence of a collocate with the verb

item collocate freq1 freq2 ngram
bùgɔ_v v:0 187 188431 187
bùgɔ_v pm:-2 187 146505 126
bùgɔ_v pers:-1 187 179651 76
bùgɔ_v c:1 187 130483 64
bùgɔ_v pers:-3 187 145917 55
bùgɔ_v 3SG:-1 187 81640 43
bùgɔ_v INF:-2 187 49982 42
bùgɔ_v n:-1 187 309187 38
gòsi_v v:0 285 188431 285
gòsi_v pm:-2 285 146505 179
gòsi_v n:-1 285 309187 91
gòsi_v PFV.TR:-2 285 21125 84
gòsi_v num:3 285 20081 75
gòsi_v n.prop:-1 285 41012 74
gòsi_v conj:2 285 45496 73
gòsi_v pers:-1 285 179651 71

24The data in table 1 essentially presents an excerpt from the unaltered table of aggregated counts of positional skipgrams on the whole Bambara corpus. The only operations needed to get this view are just proper filtering (all lines including gòsi_v and bùgɔ_v) and sorting (in the descending order of skipgram frequency). Yet even this simple frequency list immediately reveals differences in use that point to the semantic contrast between these two lexical items. While the first two positions in the list for both verbs are trivial in that they just reflect the part of speech and the position of the verb in a clause (S AUX O V), the third position is of particular interest because it reflects the position of the direct object, and is different for the two verbs. Taken together, all n‑grams that refer to that position in the top of the lists indicate, that for bùgɔ, personal pronouns (especially 3SG) dominate over nouns in the position of the direct object, while for gòsi the position of a direct object is more equally distributed among nouns, proper nouns, and personal pronouns. Thus a hypothesis may be formulated that bùgɔ is preferred when talking about hitting people, while gòsi is more general and probably more suitable in talking about hitting objects.

25Interpretation of raw frequency data may be suggestive, but it is misleading in many cases. While frequencies of the two verbs in question are on the same order of magnitude, they still differ by a factor of 1.5, which makes numbers in the two lists not directly comparable. A more principled way to identify differences in usage would require some sort of a statistical model that takes into account the differences in frequencies. There are plenty of approaches to this task in natural language processing. For the purposes of this demonstration we adopt a weighted log-odds model suggested in Monroe, Colaresi & Quinn (2008).

26To put it simply, the weighted log odds method is used to compare relative frequencies of two events. For the sake of example, let’s consider the frequency of occurrence of the personal pronoun before the verbs bùgɔ and gòsi, respectively. The values of these frequencies can be found in the rows for pers:‑1 collocate in table 1. To decide which verb personal pronouns co-occur with more often, the overall frequencies of the verbs should be taken into account. This can be done by transforming frequencies into odds, that is the ratio of the number of cases when there is a pronoun in that position to the number of cases when there is something else. This gives us 76:(187-76)=0.68 for bùgɔ, and 71:(285-71)=0.33 for gòsi. By taking the ratio of these two values (the odds ratio), we immediately find that personal pronouns are approximately two times more likely to occur before bùgɔ than before gòsi. It is conventional to take the logarithm of the odds ratio (log-odds) to make the measurement symmetrical with respect to the order of values. In the example above, if we were to divide odds for gòsi by odds for bùgɔ, the result would be close to 1/2. But the logarithm of 2 is 0.69 while the logarithm of 1/2 is –0.69, which reflects the fact that the magnitude of the difference is the same in both cases, and the sign indicates whether the feature in question is preferred or avoided by the verb that is on top of the ratio. The important intuition behind the weighted log odds is that for rare events we may observe the frequencies 2 and 1 that produce the same ratio, but this observation is much less reliable compared to the case of observed frequencies of, for instance, 100 and 50. The magnitude and even the direction of difference in the former case is more likely to be due to sampling error. Hence the method includes a correction term in the formula that puts more weight on those frequency differences that are supported by more evidence (examples). The values of the weighted log odds for gòsi vs. bùgɔ are shown in the last column of table 2. Positive values indicate the prevalence of the collocate with gòsi, and negative values correspond to higher co-occurrence rate with bùgɔ.

Table 2. Collocates for the two preceding positions for gòsi and bùgɔ, ordered by weighted log-odds. Positive log-odds indicate prevalence of a collocate with gòsi, negative — with bùgɔ. Only collocates with overall frequency of 10 or more are included in the list

collocate ngram_bùgɔ_v ngram_gòsi_v log_odds_gòsi_v
TOP:-1 2 67 3.69
n.prop:-1 13 74 2.87
PFV.TR:-2 28 84 2.03
n:-1 38 91 1.59
v:-2 4 20 1.40
prn:-1 25 9 -2.16
RECP:-1 13 2 -2.00
NOM.F:-1 10 1 -1.87
pers:-1 76 71 -1.48
PFV.NEG:-2 11 4 -1.43
  • 6 The computation was performed using the tidylo R package (Schnoebelen & Silge, 2019).

27Table 2 shows a list of the positional collocates in the two preceding positions for both verbs, ranked by the magnitude of the frequency difference as evaluated by weighted log-odds.6 These data support the observation that pronouns preferentially occur in the position of the direct object with bùgɔ. The list also demonstrates that most of the proper nouns that fill the position of the direct object for gòsi are toponyms.

28The above example demonstrates that positional skipgrams may serve as a tool to focus the attention and guide the analysis of differences in lexical usage, though they cannot directly show what the difference is. In particular, it helps to construct specific hypotheses in terms of the positional collocates. The tentative hypotheses built using positional skipgrams may be further explored with a classical concordance or more sophisticated statistical modeling.

Subcorpora comparison

29Analysis of positional skipgrams need not be limited to the individual lexical units. The n‑gram frequency easily lends itself to aggregation by any relevant metatextual properties. As a result, it is easy to obtain a frequency list of positional skipgrams for a subcorpus of texts that are comparable in some respect. In effect, this method allows for comparison of positional distributions of lexical items and grammatical tags across genres, time periods, regions, etc.

30The idea that a frequency list of n‑grams for a certain corpus characterizes the language variety used in the texts is not new. In the literature on natural language processing and on stylometry it is known as n‑gram profile. N‑gram profiles can be used to formally distinguish between different language varieties, provided that corresponding textual corpora are available to collect n‑grams. It was successfully applied, for instance, in tasks to detect language by script (with character n‑grams) (Cavnar, Trenkle et al., 1994), and in authorship attribution (Koppel & Schler, 2003).

31In the following example, two subsets of the Bambara corpus are contrasted using n‑gram profiles built from positional skipgrams: folkloric texts versus news articles. These two broad genres can be reasonably expected to differ in many respects of language use, some of which should clearly manifest itself in the prevailing syntactic patterns as well as in frequency distributions of part of speech tags, grammatical categories, and lexical items. The point is not to use positional skipgrams in a statistical classification setting (predictive modeling), but to employ them as a guide in the search for linguistically meaningful contrasts in language use.

32The disambiguated part of the Bambara corpus contains 148 files classified as folklore (0.25M words in total), and 834 files of news articles (0.36M words). The n‑gram profiles for the two subcorpora were built using the file-level positional-skipgrams data and the metadata table. Even a quick inspection of the top-frequency skipgrams lists for the two genres shows an appreciable difference in the syntactic patterns of the two subcorpora. The folkloric subcorpus has the first person singular pronoun  as the most frequent feature and the top 10 is dominated by n‑grams involving verbs and personal pronouns. Contrariwise, all top 10 n‑grams for news include a noun, and most of them consist of two nouns in some positional relationship. The third person singular occurs only on the 13th line. This clearly attests to the higher frequency of nouns and longer noun groups. When the weighted log-odds test discussed in the previous section is applied to the folklore/news dichotomy, it confirms that the syntactic differences in the narrative and reported speech versus noun groups is the most prominent contrast (see table 3).

Table 3. Skipgrams most characteristic of folklore and news subcorpora, ordered by weighted log-odds

skipgram log_odds_folk log_odds_news f_folk f_news
pers – pers:-2 30.58 –30.58 7046 3716
pers – pm:1 30.11 –30.11 10572 7216
kó_cop – pers:-1 28.69 –28.69 2640 463
pers – v:3 28.23 –28.23 7694 4752
3SG – QUOT:1 27.19 –27.19 2156 286
kó_cop – 3SG:-1 27.18 –27.18 2155 286
n – n:1 –28.66 28.66 5435 17621
n.prop – n.prop:1 –25.79 25.79 608 5200
n – num:1 –25.56 25.56 1780 8243
n.prop – n.prop:2 –24.17 24.17 339 3978
n – n:4 –23.66 23.66 7075 18918

33The same data and method may be used to explore subtler differences between these subcorpora, and to test more specific hypotheses about their differences. As an example, nominalized forms can be taken, since they are expected to be much more prominent in news. To get an overview of the differences between folklore and news in respect to nominalizations, it suffices to filter the skipgrams list to get the lines containing a reference to the nominalization marker (standard gloss — NMLZ). The differences here are not so pronounced, but they do exist (see table 4).

Table 4. Skipgrams that include nominalization, ordered by the weighted log-odds difference between folklore and news. Items with overall frequency less than 10 are omitted

skipgram log_odds_folk log_odds_news f_folk f_news
sɔ̀sɔli_n – NMLZ:0 2.48 –2.48 19 3
dún_n – NMLZ:0 2.03 –2.03 141 139
nà_v – NMLZ:1 1.73 –1.73 13 4
kòlijí_n – NMLZ:0 1.53 –1.53 10 3
IPFV.NEG – NMLZ:1 1.47 –1.47 20 12
PL – NMLZ:1 –11.83 11.83 18 741
NMLZ – PP:1 –7.83 7.83 36 419
yé_pp – NMLZ:-1 –7.83 7.83 36 419
lá_pp – NMLZ:-1 –7.53 7.53 81 526
NMLZ – ADR:1 –7.33 7.33 27 353
ni_conj – NMLZ:2 –6.74 6.74 3 230

Discussion

  • 7 See Church (2011) for a discussion of the n‑gram based language models in a wider context of ration (...)

34N-grams are among the earliest and most widely used methods in statistical language processing. Despite the criticism by Chomsky (Chomsky, 1956) who showed that n‑grams (along with other finite-state models) cannot fully model the syntax of English due to their inability to represent long-distance dependencies and parenthetical constructions, the approach thrived in practical applications7. Statistics on n‑grams of adjacent letters and phomenes proved useful for optical character recognition and speech recognition as early as the 1970s (Robertson & Willett, 1998). When textual data became abundant in the 1990s, using n‑grams of words turned into a well-established technique in applied language modeling tasks, such as part of speech tagging (Brants, 2000), or for more linguistically oriented tasks like collocation extraction (Manning & Schütze, 1999).

35The computational and conceptual simplicity of the “default” bi‑ and tri‑grams of adjacent words favored their usage wherever the performance of the resulting model was acceptable. Yet it is clear that related words are not necessarily positioned next to each other. As computational power and storage capacity grew, the idea of using word skipgrams gained traction as a way to capture long-distance dependencies (Guthrie et al., 2006). Concgrams, suggested in (Cheng, Greaves & Warren, 2006), relaxed constraints not only on the distance between collocates, but also on their relative order, thereby going even further to alleviate the problem of variability in surface structures. These generalizations of n‑grams clearly widen the scope of syntactic phenomena that n‑grams are able to represent, but simultaneously introduce much more noise in the frequency data.

36The “noise” here means unrelated or indirectly related words appearing together in the n‑gram, while n‑grams carrying information on meaningful regularities would constitute a useful “signal”. In the example (1) above ‘You should know that a greedy woman won’t cook a good sauce’, the skipgram mùso–tóbi (‘woman’ – ‘to cook’) reflects a very relevant subject–verb relation, while the structurally identical skipgram mùso–dɔ́n (‘woman’ – ‘to know’) conveys only a much less direct link between the verb of the main clause and the subject of the embedded clause. The problem is that the two skipgrams and the co-occurrence events they represent get the same weight (the latter skipgram will even get more weight if we take distance between words in text into account). Hence the problem of noise is a direct consequence of the simplistic way of treating context relationships that reduces any syntactic structure to plain word sequence.

37Two opposite ways to maintain a decent signal-to-noise ratio in n‑gram data can be attested in the recent literature. One way is to collect ever more data to let the noisy co-occurrences be dwarfed by the relevant ones. This became a trend after the advent of neural networks in language modeling that followed the success of word2vec (Mikolov et al., 2013). The other approach is to reduce noise sources in terms of the surface structure by adding more linguistic structure to the input data. The idea of building syntactic n‑grams based on relations in the syntactic tree rather than the word sequence is characteristic of this latter position (Sidorov et al., 2014). The downside of the first method is that it requires large amounts of textual data to be available for training. The obvious drawback of the second is that a reliable syntactic parser is required for it to work. Both of these are serious, if not blocking, limitations in the context of low-resourced languages.

38The positional skipgrams method suggested in this paper sits somewhere in between the above approaches in terms of balancing signal and noise in the n‑gram data. Contrary to the word2vec approach, positional skipgrams do require linguistically annotated data for the input, but the annotation can be rather shallow, like part of speech tags in the examples above. By using tags and their relative positions, the skipgrams are able to capture the signal on syntactic regularities from the tags. Part-of-speech tags and other morphological annotation, in their turn, accumulate information from the dictionary that relates wordforms to lemmas and grammatical categories for many infrequent words for which textual examples in a corpus would be insufficient to infer categories with statistical models. Another source of information contained in the tags is language knowledge added by human annotators in case if annotation was checked manually. That is exactly the kind of signal for which sparse lexical data would be insufficient in the absence of the huge training corpora. At the same time, given the current state of the art in part of speech tagging, it seems reasonable to assume that such annotation can be obtained for significant amounts of text even for lower-resourced languages. Collecting more shallow-annotated data of this kind can compensate for the noisy way of capturing morphosyntactic structures offered by n‑grams. Positional skipgrams are not an exception in this respect. But given the higher frequency and lower diversity of the part of speech tags, the signal can be expected to overcome noise much sooner compared to training on raw words.

39To recapitulate, the positional skipgrams were introduced here for the case of a low-resourced language in order to alleviate the problem of data sparsity. Word-level skipgrams deliberately throw out information encoded in the exact positioning of words in the data, regarding it as noise. This may indeed work when there is a large amount of data. But when the data is relatively scarce, sampling errors with sparse lexical items is a very serious concern. Traditionally, co-occurrence analysis treats lexicon as signal and syntax as noise, whereas in this work I suggest to switch sides. By tracking positional co-occurrences with higher-level grammatical categories with positional skipgrams, patterns of the local surface structures emerge that stem from the information obtained using dictionaries and human language competence. This can provide a statistical signal that is more reliable and broad than plain lexical co-occurrence.

Conclusion

40This article presented a new dataset built using a morphologically-annotated and manually disambiguated subcorpus of the Bambara Reference Corpus, and demonstrated several ways in which this dataset may help in formulating linguistic hypotheses about various contrasts present in textual data. This quantitative data (with metadata) enables testing hypotheses and building models based on lexico-grammatical distributions in various parts of the corpus. The goal of publication of this dataset is to provide a wider audience with access to the data on linguistic regularities observed in the Bambara corpus for linguistic research and development of NLP applications.

41The data is organized in the form of frequency lists of positional skipgrams, which is a framework for building n‑grams suggested in this article. These n‑grams capture information about co-occurrence of lexical items with grammatical categories at various relative positions. Researchers are free to download the data and to use it both as an aid for linguistic queries for the corpus, and as a basis for building applications for the natural language processing tasks.

42The approach to create an n‑gram corpus suggested in this article is suited well to less-resourced settings where overall textual data is not easily available but some annotated texts are present. For linguists, positional skipgrams may serve as an exploratory tool which, much like the concordance, reorganizes the textual data in a non-linear fashion in order to reveal regularities. This is intended not to replace, but to supplement other views of the corpus, including online search and concordancing.

References

43Brants Thorsten, 2000, “TnT: a statistical part-of-speech tagger” in Proceedings of the sixth conference on Applied natural language processing, Association for Computational Linguistics, pp. 224‑231.

44Brants Thorsten & Franz Alex, 2006, The Google Web 1T 5‑gram corpus version 1.1. LDC2006T13.

45Brown Peter F, Desouza Peter V, Mercer Robert L, Della Pietra Vincent J & Lai Jenifer C, 1992, “Class-based n‑gram models of natural language” in Computational linguistics 18(4), pp. 467‑479.

46Cavnar William B, Trenkle John M et al., 1994, “N‑gram-based text categorization” in Proceedings of SDAIR-94, 3rd annual symposium on document analysis and information retrieval, pp. 161‑175.

47Cheng Winnie, Greaves Chris & Warren Martin, 2006, “From n‑gram to skipgram to concgram” in International journal of corpus linguistics 11(4), pp. 411‑433.

48Chomsky Noam, 1956, “Three models for the description of language” in IRE Transactions on information theory 2(3), pp. 113‑124.

49Church Kenneth, 2011, “A pendulum swung too far” in Linguistic Issues in Language Technology 6(5), pp. 1‑27.

50Davydov Artem, 2010, “Towards the Manding corpus: Texts selection principles and metatext markup” in Proceedings of the Second Workshop on African Language Technology AfLaT, pp. 59‑62.

51Gries Stefan Th & Divjak Dagmar, 2009, “Behavioral profiles: a corpus-based approach to cognitive semantic analysis” in New directions in cognitive linguistics, pp. 57‑75.

52Guthrie David, Allison Ben, Liu Wei, Guthrie Louise & Wilks Yorick, 2006, A Closer Look at Skip-gram Modelling. Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06), European Language Resources Association (ELRA), Genoa, URL: http://www.lrec‑conf.org/proceedings/lrec2006/pdf/357_pdf.pdf.

53Hanks Patrick, 1996, “Contextual dependency and lexical sets” in International Journal of Corpus Linguistics 1(1), pp. 75‑98.

54Jaggi Martin, Uzdilli Fatih & Cieliebak Mark, 2014, “Swiss-chocolate: Sentiment detection using sparse SVMs and part-of-speech n‑grams” in Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pp. 601‑604.

55Kopotev Mikhail, Pivovarova Lidia, Kochetkova Natalia & Yangarber Roman, 2013, “Automatic Detection of Stable Grammatical Features in N‑Grams” in NAACL HLT 2013: Proceedings of the 9th Workshop on Multiword Expressions, Association for Computational Linguistics, Atlanta, pp. 73‑81, URL: https://www.aclweb.org/​anthology/​W13-1011.

56Koppel Moshe & Schler Jonathan, 2003, “Exploiting stylistic idiosyncrasies for authorship attribution” in Proceedings of IJCAI’03 Workshop on Computational Approaches to Style Analysis and Synthesis, vol. 69, pp. 72‑80.

57Levy Omer, Goldberg Yoav & Dagan Ido, 2015, “Improving Distributional Similarity with Lessons Learned from Word Embeddings” in Transactions of the Association for Computational Linguistics 3, pp. 211‑225, DOI:10.1162/tacl_a_00134.

58Ling Wang, Dyer Chris, Black Alan W & Trancoso Isabel, 2015, “Two/Too Simple Adaptations of Word2Vec for Syntax Problems” in Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, Denver, pp. 1299‑1304, DOI:10.3115/v1/N15-1142.

59Manning Christopher D & Schütze Hinrich, 1999, Foundations of statistical natural language processing, MIT press.

60Maslinsky Kirill, 2014, “Daba: a model and tools for Manding corpora” in TALN-RECITAL 2014 Workshop TALAf 2014: Traitement automatique des langues africaines (TALAf 2014: African Language Processing), pp. 114‑122.

61Mikolov Tomas, Sutskever Ilya, Chen Kai, Corrado Greg S & Dean Jeff, 2013, “Distributed representations of words and phrases and their compositionality” in Advances in neural information processing systems, pp. 3111‑3119.

62Monroe Burt L., Colaresi Michael P. & Quinn Kevin M., 2008, “Fightin’ Words: Lexical Feature Selection and Evaluation for Identifying the Content of Political Conflict” in Political Analysis 16(4), pp. 372‑403, DOI:10.1093/pan/mpn018.

63Plecháč Petr, 2019, Atribuce autorství básnických textů [Authorship attribution of poetic texts]. Univerzita Karlova phd, Praha, URL: https://dspace.cuni.cz/​handle/​20.500.11956/​115525.

64Robertson Alexander M & Willett Peter, 1998, “Applications of n-grams in textual information systems” in Journal of Documentation 54(1), pp. 48‑67.

65Rosenfeld Ronald, 2000, “Two decades of statistical language modeling: Where do we go from here?” in Proceedings of the IEEE 88(8), pp. 1270‑1278.

66Schnoebelen Tyler & Silge Julia, 2019, tidylo: Tidy Log Odds Ratio Weighted by Uninformative Prior, URL: http://github.com/​juliasilge/​tidylo.

67Sidorov Grigori, Velasquez Francisco, Stamatatos Efstathios, Gelbukh Alexander & Chanona‑Hernández Liliana, 2014, “Syntactic n-grams as machine learning features for natural language processing” in Expert Systems with Applications 41(3), pp. 853‑860.

68Stefanowitsch Anatol & Gries Stefan Th, 2003, “Collostructions: Investigating the interaction of words and constructions” in International journal of corpus linguistics 8(2), pp. 209‑243.

69Taylor Paul & Black Alan W, 1998, “Assigning phrase breaks from part-of-speech sequences” in Computer Speech & Language 12(2), pp. 99‑117.

70Vydrin Valentin, 2013, “Bamana Reference Corpus (BRC)” in Vargas-Sierra Chelo (ed.), Procedia – Social and Behavioral Sciences: Corpus Resources for Descriptive and Applied Studies. Current Challenges and Future Directions: Selected Papers from the 5th International Conference on Corpus Linguistics (CILC2013), vol. 95, Elsevier, Alicante, pp. 75‑80, URL: http://www.sciencedirect.com/​science/​journal/​18770428.

Haut de page

Notes

1 The corpus search interface as well as general info about the corpus are available online at: http://cormand.huma-num.fr/. The new dataset is available at http://cormand.huma-num.fr/ngrams.

2 The term n‑gram presupposes a variable number of co-occurring words, but in the data and in the examples discussed in this article the n is always limited to two.

3 Depending on the task at hand, it may be convenient to record reverse co-occurrence events (pm:1 — mùso, etc.) simultaneously to simplify further processing.

4 See the full list of the glosses for grammatical morphemes and auxiliaries for Bambara at: http://cormand.huma-num.fr/gloses.html.

5 See information on the dictionary at http://cormand.huma-num.fr/bamadaba.html.

6 The computation was performed using the tidylo R package (Schnoebelen & Silge, 2019).

7 See Church (2011) for a discussion of the n‑gram based language models in a wider context of rationalist/empiricist debate in computational linguistics.

Haut de page

Pour citer cet article

Référence électronique

Kirill Maslinsky, « Positional skipgrams for Bambara: a resource for corpus-based studies »Mandenkan [En ligne], 62 | 2019, mis en ligne le 07 mai 2020, consulté le 18 mai 2024. URL : http://journals.openedition.org/mandenkan/2119 ; DOI : https://doi.org/10.4000/mandenkan.2119

Haut de page

Auteur

Kirill Maslinsky

Institute of Russian Literature (Pushkinskij Dom) RAS, Saint Petersburg, Russia

Haut de page

Droits d’auteur

CC-BY-NC-SA-4.0

Le texte seul est utilisable sous licence CC BY-NC-SA 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.

Haut de page
Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search