This is the repository for the fifth assignment in the course Language Analytics from the bachelors elective course Cultural Data Science at Aarhus University
The code was written independently by me. Help Sources freeCodeCamp.org: "Natural Language Processing with spaCy & Python - Course for Beginners" https://www.youtube.com/watch?v=dIUTsFT2MeQ
As your everyday linguist, I wanted to test out what the extension and limits of current models are, when trying having to a Danish conversation manuscript.
The transcript provided as a textfile 'anne_og_beate.txt' in the folder in. T
I have used the Danish spaCy model nlp = spacy.load("da_core_news_sm") to perform POS-tagging and dependency Parsing.
Part-Of-Speech (POS) tagging is a grammatical tagging of lexemes with the following categories:
ADJ: adjective
ADP: adposition
ADV: adverb
AUX: auxiliary
CCONJ: coordinating conjunction
DET: determiner
INTJ: interjection
NOUN: noun
NUM: numeral
PART: particle
PRON: pronoun
PROPN: proper noun
PUNCT: punctuation
SCONJ: subordinating conjunction
SYM: symbol
VERB: verb
X: other
source Universal POS tags
Dependency Parsing is a Natural Language Processing (NLP) texhnique that analyse the grammatical structures of the syntax, and how the words in a sentence depend on each other and form a structure. The dependy parsing both POS-tag, gives a dependy label and a head token.
ROOT: Root of a sentence or a clause acl: Clausal modifier of noun acomp: Adjectival complement advcl: Adverbial clause modifier advmod: Adverbial modifier agent: Agent amod: Adjectival modifier appos: Appositional modifier attr: Attribute aux: Auxiliary auxpass: Auxiliary (passive) case: Case marker cc: Coordinating conjunction ccomp: Clausal complement compound: Compound conj: Conjunct csubj: Clausal subject csubjpass: Clausal subject (passive) dative: Dative dep: Unspecified dependency det: Determiner dobj: Direct object expl: Expletive intj: Interjection mark: Marker meta: Meta modifier neg: Negation modifier nmod: Nominal modifier npmod: Noun phrase as adverbial modifier nsubj: Nominal subject nsubjpass: Nominal subject (passive) nummod: Numeric modifier oprd: Object predicate parataxis: Parataxis pcomp: Complement of preposition pobj: Object of preposition poss: Possession modifier preconj: Preconjunct predet: Predeterminer prep: Prepositional modifier prt: Particle punct: Punctuation quantmod: Modifier of quantifier relcl: Relative clause modifier xcomp: Open clausal complement source spaCy Label Scheme
Furthermore each lexeme, except for the root, in a sentence is assigned a 'head', meaniing the word that governs or controls the syntactic relationship with another word (the parent word).
For the sentiment analysis I have implented Sentida, a Danish sentiment analysis tool, which provide a positive score (<0.0 or a negative score >0.0)
To convert the .txt-file to a dataframe the pandas command pd.DataFrame() was used.
The data was retrieved from Samtalebank in the folder called "Sam2" which is a collection of transcriptions of Danish conversations. The transcripts are equipped with a set of symbols and special characters, that showcases notational symbols for pronunciational and temporal features of speech. Since no models are able to interpret these symbols (for now), the conventions for which symbols to use varies, and I am not a programmer these symbols have been removed by simply searching and replacing.
The transcripts where, when downloaded, in a '.cha' format, which is the format used for the transcription software CLAN. The files can be converted either by running the command CHAT2TEXT inside the software, or simply be changing the file extension and opening the file in a text-editor.
The scripts where made in python 3.10.7. Should any errors occur, contrary to expectation, when running the script, please try to install this python version.
From the command line, make sure your arel ocated in the main directory. Run the command bash setup.sh to install the required packages.
In the folder src, three pythonscripts are located: pos_tag.py, dendency_gone_wrong.py and sentiment.py. The pos_tag.py script outputs a .csv-table called Anne_og_Beate_POS.csv, with a column on the right called "POS" where it
The POS-tagging is working ok and Sentiment analysis with the Danis sentiment analysis also ok.
I have had problems in how I could make the output a CSV-file without overcomplicating things. Especially with the sentiment analysis it proved hard, since the sentida will not work with a pandas dataframe. This is also why the output only have two columns "Text" and "Sentiment". It's too bad, since thsi means the sentiment analysis was also run on the speaker. Luckily the speakers doesn't form any meaningful word, and should not skew the results too much.
A problem though, is the fact that the transcript is not always written not as real words, but in some contexts rely on the way they sound such as å being og(and). This mostly occurs with conjunctions, and more complicated words, if pronounced 'normally' follow the danish dictionary. If I wanted to improve things, I probably should've made a stopwordlist. Though I tried to remove trailing white-space, it seems that some trailing white-space still persist.
The Pos tagging seems to work rather well with only minor mistakes, both due to the pronouncial spellings and interjections such as 'øh' or 'øhm'(uhm). When these interjections occur, it seems to give it a random POS-tag.These words should've been sorted out in a non-existent data augmentation process. Example:
| 1 | Speaker | Text | Sentiment |
|---|---|---|---|
| 5 | BE | å så tænker jeg hva fanden jeg skal lave den her øh feature om | ADP ADV VERB PRON VERB NOUN PRON AUX VERB DET ADV NOUN NOUN ADP |
The sentence 'å så tænker jeg hva fanden jeg skal lave den her øh feature om' can roughly be translated to 'and then I think what the fuck should I make this particular uhm feature about' The words are parsed mostly correct. The 'å' which in correct ortographic spelling would be an 'og' (and) is said to be an adposition though it is and adverb. This is understandable, since the model is trained on ortographic spelling. 'hva' is furthermore an auditory spelling, which in ortographic manners would have been an 'hvad'(what) is a pronoun but is labeles as a verb. Here it can also be seen how the 'øh' is parsed as an adverb.
The results are, given the fact that this is not ortographic data, pretty okay, and the mdoel is able to grasp the big picture.
The dependency parsing csv-file is rather messy. The dependency parsing do POS tagging like seen in the previous section and also some dependency labeling and assigning each word a 'head'. The 'Head' column is not very succesful- Example
| 1 | Speaker | Text | Token | POS | Dependency Label | Head |
|---|---|---|---|---|---|---|
| 2 | *AN | havde du egentlig øh nået a lave dine feature | " | AUX PRON ADV VERB VERB X VERB DET NOUN | aux nsubj advmod ROOT xcomp nmod:poss punct | det obl øh øh øh øh øh feature a feature nået |
The sentence "havde du egentlig øh nået a lave dine feature" can roughly be translated to: "did you actually uhm get to make yours feature" If we take a look at the dependency labeling, it's surprisingly okay. 'havde' is an auxiliary verb (modal verb) that modifies 'nået'. 'du' is the nominal subject to 'havde' 'egentlig' is and adverbial that modifies 'nået' 'øh' is skipped, which is good. 'nået' is the root verb. 'a lave' is correctly labelled as a open clausal complement that introduces the subordinate clause, though misspelled and should've been 'at lave' 'dine' is a noun modifier, and due to the misspelling that should've been 'your' (your) it is labelled as possesive. 'feature' is labelled as punctuation, though it is a direct object. all in all the dependency labelling was pretty good although with minor mistakes, and a bit hard to analyse given the way I outputted the whole sentence followed by the entire dependency labelling. The 'Head' column on the other hand makes no sense.
An example from the Anne_og_Beate_sentiment.csv
| 1 | Text | Sentiment |
|---|---|---|
| 377 | *AN :ah men det var så klamt også bare da jeg så billedet var jeg | -0.4833333333333333 |
| 379 | *AN : ved at brække mig igen | -0.533333333333334 |
| 385 | *AN :hold kæft jeg synes det var godt de andre altså | 0.26111111111111146 |
| 386 | :sådan Anders Fogh gjorde det var helt vildt godt | 0.3500000000000003 |
Since the sentence i line 377 and 379 contains the lexemes 'klamt' and 'brække', it would make senese that these are rated negatively, while the sentences in line 385 and 386 contains the lexemes 'godt' and 'vildt godt' these rated as positive. Furthermore I am impressed that the sentiment analysis rate line 386 higher than 385, where the adverb 'vildt' modifies that adjective 'godt', and (probably) by that rates the sentence as more postive than without the adverb. The results are far from perfect, but it works to an extend.
For any questions, please write Nikita J. Myrting on the email: 201906799@post.au.dk.