CS4051 Natural Language Processing

Question 1

What does it mean for an LLM to exhibit priming effects?

Answer 1

After seeing a sentence of a particular structure, the LLM is less surprised by an upcoming sentence of a similar structure

Question 2

Define the following terms? Morphology, Pragmatics, Semantics.

Answer 2

Morphology captures information about the different senses of a word in context.
Pragmatics is the study of meaning in context.
Semantics is the study of meaning references/truth.

Question 3

What does Zipf’s Law state and what does it imply in model development?

Answer 3

Zipf’s law states that the frequency of a word within a corpus is inversely proportional to its rank, implying that it is important for a model to be trained on a representative corpus, so that rarer words within the test set do no surprise the model.

Question 4

Define the following terms? Lexicon, Homonym, Word senses.

Answer 4

A lexicon is a collection of lexical entries (words).
Homonyms are words which have the same spelling/pronunciation but different meanings, for example “saw” (the cutting tool) and “saw” (the past tense of see).
Word senses are all the different meanings of a word, for example the senses of “saw” would include the meanings mentioned above.

Question 5

What are stopwords (give and example) and how can they be identified in a corpus?

Answer 5

Stopwords are commonly occurring words which are usually irrelevant to most NLP tasks. A few examples are “of, the, a”. These can be usually identified as the terms with highest frequency in a corpus or by consulting a pre-computed list of stopwords.

Question 6

Identify and explain the text pre-processing technique concerned with breaking down text into larger chunks of words?

Answer 6

Sentence splitting consists of breaking the corpus down into sentences. A good heuristic used is to look at punctuation symbols. Abbreviation dictionaries can also be used to determine whether a punctuation symbol marks the end of a sentence.

Question 7

Identify and explain the text pre-processing technique concerned with breaking down text into words? Also define the vocabulary size?

Answer 7

Tokenisation is the task of converting a sentence/corpus into a sequence of tokens/features. Common methods include using spaces as token boundaries and handling edge cases using heuristics - for example “Mary’s” could be split into “Mary” and “s” or regarded as a single token.

The vocabulary size is simply the number of tokens in a sentence. This can vary based on edge case handling as shown in the given example.

Question 8

Identify and explain the text pre-processing technique concerned with identifying common names, places, etc. ?

Answer 8

Named Entity Recognition is concerned with identifying token which represent country, people, titles etc. The should be considered as one token, regardless of the number of words they are made up of. For example “New Mexico” should not be “New” and “Mexico”.

Question 9

What is a common set-up for model development? Please explain the advantages and disadvantages of this setup in regard to model performance. Hint : 3 datasets are usually involved.

Answer 9

A common setup is the training, dev-test, and test datasets split. The training set is used for model training/fitting, while the dev-test set is used to compute the prediction error during model selection, and the test-set is used to compute the generalization error before model deployment.

An advantage of using the additional dev-test set is that the test set is kept hidden from the model for as long as possible, meaning that there is a lower risk of over-fitting.

A requirement of this approach is that the dev-test must be disjoint from the test-set to avoid over-fitting, but still be representative to avoid bias.

Question 10

What is a useful technique for model evaluation when a good training, dev-test, test split is not achievable? Hint : give the answer in terms of “k-fold”.

Answer 10

Cross-validation allows to test a model when there is not enough quality data for a 3-part split. The general approach is to conduct a k-fold cross-validation on the training dataset, where the training set is split into k equal chunks. k-1 are used for model training and 1 chunk used for calculating the prediction error. The process is repeated k times and the prediction errors averaged together.

Question 11

What are 4 common intrinsic evaluation metrics when evaluating a machine learning model? Provide formulae for all of them and explain them.

Answer 11

Accuracy measures the percentage of correctly classified inputs in the test set. The “baseline” accuracy is the number of occurrences of the majority class in the test set divided by the size of the test set. For unbalanced datasets, if the model’s accuracy > baseline accuracy, it suggests that the model is likely not guessing the most probable class each time but actually using the training features for the classification task.

Precision is the ratio of items predicted to belong to a certain class which actually belong to that class. It is calculated by P = TP / (TP + FP).

Recall is the number of correctly classified items mentioned in the provided gold-standards. It is calculated by R = TP / (TP + FN)

F1 gives a metric to combine both precision and recall. It is computed as F1 = 2PR/(P + R)

Question 12

How would you interpret a Confusion Matrix to extract the accuracy of the model, and precision and recall for a particular class?

Answer 12

The accuracy is given by sum(numbers along top-left bottom-right diagonal)/(total number of instances).

The precision for a class is the number appearing on the diagonal / sum of the column indexed by that class

The recall for a class is the number appearing on the diagonal / sum of row indexed by that class.

Question 13

Decision Trees are a popular machine learning classifier. Explain how they work, its advantages and disadvantages?

Answer 13

Decision trees are flowcharts composed of decision nodes which check feature values and leaf nodes which assign labels to instances. It uses features which best split the data.

Advantages include high interpretability and suitability for classifying hierarchical data.

Disadvantages are that the amount of training data decreases at lower levels, making it a model prone to over-fitting, and that it forces features to be checked in a particular order event if these are independent of each other.

Question 14

Naive Bayes is a popular machine learning classifier. Explain how it works, including the underlying Naive Bayes assumption, and how smoothing works and what issue it solves?

Answer 14

The prior probability for each label is computed as the label frequency “P(label)”, then the contribution of each feature is combined to the previous computation to obtain a likelihood estimate for each label “P(feature | label) for all features”. Finally, the label with the highest likelihood is chosen as the classification result.

The Naive Bayes assumption states that given a label, all features are statistically independent of each other, meaning that the classifier treats the corpus as a “bag of words”.

Smoothing solves the issue of zero-counts, which is when a feature never occurs in the corpus. By adding a constant value to the counts, the probability is never 0.

Question 15

Explain what causal language modelling is.

Answer 15

Causal language modelling predicts the next token in a sequence of tokens incrementally, usually basing the prediction off of a limited history of tokens

Question 16

Explain how n-gram language models work? Also mention the underlying assumption used to limit the required generated token history.

Answer 16

N-gram models select the most likely token based-off of the previous n-1 tokens in the sequence, using the Markov assumption which implies that only a limited history of tokens is relevant for selecting the next one.

Question 17

Explain the problem of Sparsity in Language, relating your answer to n-gram models?

Answer 17

The capability of the model to generalize depends on how representative the training data is. According to Zipf’s law, many n-grams can appear in the test set but not in the training set. A possible solution for this issue could be to use smoothing.

Question 18

Explain what the Maximum Likelihood Estimation (MLE) is, relating it back to the Markov assumption in language models?

Answer 18

MLE refers to the process by which a model estimates the probability of a sequence of tokens occurring, using the Markov assumption to reduce the number of parameters required in computation.

Question 19

Perplexity is a common evaluation metric for Language Models. Please explain what it is and how it can be interpreted?

Answer 19

Perplexity is the inverse probability of the test corpus according to the model, normalized by the corpus size. Lower perplexity means that the model can more easily predict the next word in the test corpus.

Question 20

Word embedding is a common prerequisite for auto-completion tasks. Name and explain two auto-completion tasks. Do these preserve order information of the generated output? Finally, name a popular pre-trained word embedding model.

Answer 20

Two word embedding tasks are characterised by Skip-gram (given a word, predict its surrounding context) and CBOW (given a context, predict the missing word). Both of these do not preserve the order information, making it a limitation. Popular pre-trained embeddings are GloVe and Word2Vec.

Question 21

Sentence embedding is used for clustering and retrieval tasks. Explain the two naive approaches used, as well as the 3 families of neural language models used for sentence embeddings.

Answer 21

The first naive approach is to treat each sentence in a corpus as word, and run a CBOW model on it to encode the context. This is challenging due to extreme data sparsity.

The second naive approach is to embed each sentence as the mean of all embeddings of words within that sentence. Limitations include weighting-down stopwords and the fact that word ordering is ignored by the embedding.

The 3 families of models are Long-Short Term Memory (LSTM), Transformers, and LLMs

Question 22

Name and explain the main usage of two popular LLMs used for sentence embedding?

Answer 22

GPT is an autoregressive model used for text generation, while BERT is used for masked word and next-sentence prediction.

Question 23

Define morphemes and lexemes?

Answer 23

Morphemes are the smallest possible unit of language, even if it cannot stand on its own. For example “un” in “unusual” is a morpheme.
Lexemes are units of lexical meaning from which a set of words can be reached from. For example the lexeme of “running”, “ran”, “runner” is “run”

Question 24

Explain what a Stemmer and Lemmatiser do?

Answer 24

Stemmer removes affixes from words, leaving the stem behind. For example stemming “unpleasantly” yields “pleasant”.
Lemmatiser map a word to its lemma, a.k.a. its dictionary entry. For example lemmatising “leaving” yields “leave”.

Question 25

Explain what POS tagging consists of? Explain the main challenge of POS tagging?

Answer 25

Part-of-speech tagging involves assigning a lexical category to each word in a corpus. POS tagging can then be used to define rules about the syntax of a sentence and use them to generate other sentences of the same structure. Difficulty of POS arises when considering context.

Question 26

Explain what syntax trees are?

Answer 26

Syntax trees represent how grammar rules combine to form sentences

Question 27

Explain what depending parsing consists of?

Answer 27

The task of obtaining semantic structure of a sentence by analyzing the relationship between words. Verbs are often the root of the tree, since they are central to their clause (sentence).

Question 28

What does it mean for a grammar to be context free?

Answer 28

Context-free grammars allow for expansions of productions/rules regardless of the context.

Question 29

Explain the most basic version of a language model? Include “cross-entropy” in your answer.

Answer 29

The idea behind language models is to assign probabilities to symbols of a language, based on the product of preceding tokens. Cross-Entropy measures how close the distribution of tokens learned by the model is to the true token distribution for the language learnt.

Question 30

What is one major issue with word embedding models such as Word2Vec?

Answer 30

Word2Vec embeds all senses of a word to the same vector, so the past tense of “see” and the cutting tool “saw” are both embedded the same.

Question 31

What are two desirable properties of word embeddings?

Answer 31

Words with similar meanings should be close in the vector space (for all languages), and syntax/morphology should sometimes be preserved as well.

Question 32

What is a metric used to measure similarity between embeddings? Provide the formula

Answer 32

Cosine similarity between two embeddings, measured as the dot product of the two vectors divided by both their magnitude.

Question 33

What is contextualized word embedding and what model is capable of producing these?

Answer 33

Contextualized embedders such as BERT can embed words without disregarding their context.

Question 34

Name and explain a heuristic used for document embedding?

Answer 34

TF*IDF (term-frequency, inverse-document-frequency) is a heuristic which given a term/query and a collection of documents, attributes higher weights to documents containing a high term frequency and for which the term is common in the document, but uncommon in all other documents in the corpus.

Question 35

What is a technique in bilingual embeddings which can help improve monolingual embeddings?

Answer 35

Direct Transfer consists of training an embedding model on a resource-rich language and apply it to a resource poor language. A common limitation with this approach are out-of-vocab words

Question 36

Explain the naive approach for training bilingual embeddings and its main bottleneck?

Answer 36

The common approach is to learn a monolingual embedding in a language and map them to the target language using a dictionary. A bottleneck arises if there is not entry in the dictionary for a particular token, meaning a mapping cannot be acheived.

Question 37

Name and explain how two common bilingual embedding models work?

Answer 37

BiSkip uses word-aligned and sentence-aligned texts, then runs a SKIP-GRAM model using words from both languages as context.

BiVCD merges and shuffles aligned documents, then runs a monolingual embedding on this output.

Question 38

What is the task of natural language generation?

Answer 38

Turning structured data into human-readable text.

Question 39

What are the two approaches of natural language generation? Discuss advantages and disadvantages.

Answer 39

Rule-based generation requires manual effort to define rules. It does not generalize well however it produces high quality outputs for the target domain.
Data-driven generation requires no rule-definition and generalises well, however the output quality is not a guarantee and there is a risk of picking up bias from the training data.

Question 40

What is the task of decoding in LLM generation? Name the most common decoding strategy

Answer 40

Decoding is the task of choosing a word to generate based on the model’s probabilities. Sampling is a common decoding approach.

Question 41

Name and briefly describe the 6 sampling methods used in decoding during LLM generation?

Answer 41

Random sampling : randomly choose a word from the model’s probability distribution
Greedy sampling : always choose the most probable word
Beam Search sampling : expand X sequences and pick the one with the highest cumulative probability.
Top-k sampling : choose a word at random from the top-k most probable ones
Nucleus sampling : choose a word at random from the top-p percent of the probability mass
Temperature sampling : normalize logits by a temperature before applying a soft-max function. As temperature decreases, higher probability words are pushed towards 1 and lower probability ones towards 0.

Question 42

Explain the In-context learning and few-shot prompting are?

Answer 42

In-context learning (prompting) is the task of improving the model’s performance without using any gradient-descent based updates on its parameters, usually by using prompts to generate a context for the LLM to work in.

Few-shot prompting consists including labelled example answers to guide the model.

Question 42

Explain broadly how fine-tuning works and what it is used for?

Answer 42

Fine tuning consists of adapting the parameters of a pre-trained model to complete tasks in a specific domain.

Question 43

Explain what reference metrics are and provide two example metrics?

Answer 43

Reference metrics assess the similarity between generated text for a task and the human-written output. Two common metrics are BLEU to evaluate cross-language machine translations and BERTscore which computes token-wise similarity, taking semantics into account.

Question 44

Explain the task of reference resolution. What are the two things required to accomplish this task?

Answer 44

Reference resolution refers to the task of determining the preceding topic being refereed to in a dialogue context. It relies on correct understanding of speaker common ground and speaker intent

Question 45

What are speech acts? Give some example

Answer 45

Speech acts convey speaker intent. Examples of these are warnings, requests, and invitations.

Question 46

Explain what we mean by Abductive Reasoning?

Answer 46

Inference to the most plausible explanation

Question 47

Explain what Information Retrieval is and what are the two approaches to achieve it?

Answer 47

Information Retrieval refers to the task of returning documents from a corpus which are relevant to answering a user query.

The first approach is to encode the documents as sparse vectors, weight them using TF-IDF and then computing the most relevant one by cosine similarity.

The second approach involves encoding the documents and query as dense vectors using BERT and then computing the similarity within the embedding space

Question 48

What is RAG? Explain what it is and its main usage.

Answer 48

Retrieval Augmented Generation uses information retrieval to gather relevant documents, then feed them to a LLM to generate an answer.

The main benefit of RAG is that answers are more likely to be grounded in truth.

Question-Answering systems use RAG in the Retriever-Reader architecture.

Question 49

What is alignment in dialogue?

Answer 49

The phenomenon by which speakers adapt to each other to aid mutual understanding

Question 50

What are dialogue acts?

Answer 50

These are generalized speech acts which also represent grounding and may require inference by one of the actors in the dialogue

Question 51

What is the most trivial approach to dialogue generation?

Answer 51

Annotate each utterance/turn with the speaker and concatenate these together to generate a dialogue context.

Question 52

What is the main challenge of dialogue?

Answer 52

The main challenge in dialogue is exploiting common ground to understand context in order to generate an utterance/response.

Question 53

What is Surprisal in dialogue, and what can be said about surprisal between speakers?

Answer 53

A measure of a speaker’s processing effort. Surprisal has been observed to converge between speakers as the dialogue progresses.

Question 54

What are constructions in dialogues, and what is facilitated by construction repetition in dialogue?

Answer 54

Constructions are utterances of 3+ tokens which occur 3+ times within the dialogue.

Constructions repetition facilitates processing in task-oriented dialogues, especially since repetition facilitates information delivery rate between speakers.

Question 55

What is one of the most significant indicators of task success in dialogue?

Answer 55

Dialogue length

Question 56

What is the task of machine translation? Explain it and name a few common challenges.

Answer 56

Machine translation is the process of using computational techniques to translate text or speech from one language to another. Common challenges include interpreting creativity and style (especially when translating creating pieces such as poetry) and translating from/to low resource languages

Question 57

Give an example of when biases may be learned when translating from a language like Hungarian to English?

Answer 57

A possible bias can occur when translating from a gender-neutral language such as Hungarian to a gendered language such as English. For example, “she is a nurse” and “he is a doctor”.

Question 58

Name and explain the 2 extrinsic machine translation evaluation metrics?

Answer 58

Two common metrics are Adequacy (how well the translation captures the meaning of the source sentence) and Fluency (how fluent the translation is in the target language)

Question 59

Explain the main idea behind intrinsic metrics for evaluation of machine translation. Name two common scoring metrics and broadly explain what they do?

Answer 59

These measure the similarity between the translation output and a human-generated gold standard by calculating the overlap between the two.

BLEU is a scoring metric which captures lexical relatedness and penalizes translations which are too short.

BERTscore captures semantic relatedness. It works by encoding both translations using BERT and computing the token-wise cosine similarity.

Question 60

What is the main objective of summarisation? Explain the three stages of a typical summarisation process.

Answer 60

Summarisation aims to produce a condensed version of a document which contains information relevant to answering a user query.

The three stages of a summarisation task are
(1) pre-processing - parse document and extract required contents
(2) feature design and sentence scoring - extract features such as length and use them to score sentences
(3) post-processing - sentence simplification, redundancy removal, and reordering.

Question 61

Explain the difference between extractive vs abstractive summarisation?

Answer 61

Extractive summarisation produces summaries consisting entirely of material copied from the input document

Abstractive summarisation consists of material not in the input document i.e., paraphrased

Question 62

Name and explain a popular metric for intrinsic evaluation of summarisation outputs?

Answer 62

ROUGE is regarded as a convenient metric to use when human evaluation is not possible. The core idea is to compute the percentage of bi-grams contained in the human-defined gold-standard which also appear in the generated summarisation.

Question 63

What is temporality?

Answer 63

Temporality is defined as knowledge change over time

Question 64

Name and explain the three categories of temporal hallucination?

Answer 64

Omission occurs when LLMs provide an incomplete or partly correct answer for a query with multiple correct answers.
Fabrication occurs when LLMs provide a made-up answer for queries with no answer.
Misattribution occurs when LLMs answer with the wrong entity for a query whose answer is a proper noun.

Question 65

What is the difference between 1-hop and 2-hop temporal complexity?

Answer 65

In 1-hop, the timestamp of the answer is explicitly included in the query, while in 2-hop this is implicit.

Question 66

Name and explain the three intrinsic evaluation metrics for temporal queries?

Answer 66

Exact match yields true if the ground truth is contained within the model answer.
F1 measures overlap between the LLM’s answer and the ground truth.
PEDANTS measure the correlation of model output with a human-labelled QA dataset

Question 67

Explain what these evaluation results might tell us about the answer of an LLM to a temporal query?

1. F1 score = PEDANTS score and both F1 and PEDANTS are larger than the EM score
2. PEDANTS > F1 > EM
3. F1 > PEDANTS > EM
4. F1 = EM and both F1 and EM are larger than PEDANTS

Answer 67

1. The query doesn’t support order swapping - ground truth expects “x … y” therefore “y … x” is not allowed.
2. Model answer is correct but verbose
3. Model is likely guilty of omission of information from the answer
4. Model is likely guilt of fabricating the answer

Question 68

What is a possible explanation for worse results about a temporal query concerned with the distant past, but better results for queries concerned with more recent times?

Answer 68

Possible data shortage about what happened a long time ago, there could be some gaps in the data or different formats of data may have been gathered. In general, increased data availability results in better answers.

Question 69

LLMs struggle to understand dates when answering temporal queries. Name a technique by which information can be obtained from elsewhere and used to improve the semantics of the output?

Answer 69

RAG - retrieval augmented generation

Question 70

What are some reasons for carrying out intrinsic evaluation in NLP? Name 3 ways in wh

Answer 70

Reasons for intrinsic evaluation are assessing performance, assessing generalization, identifying bias and fairness. Common methods for intrinsic evaluation can be using performance metrics (BLUE, perplexity, etc.), using dataset splits, and using human evaluation such as labeling

Question 71

What are some reasons for carrying out extrinsic evaluation in NLP? Name some common task success measures when evaluating with participants.

Answer 71

Reasons for extrinsic evaluation include ensuring ethical and safe deployment, assessing levels of trust and UX quality, and ensuring real-word applicability. Common success measures include user satisfaction score, task outcome, and participant behavior during the interaction with the model.

Question 72

Explain the difference between the three types of biases in NLP systems? Data bias, Model bias, Evaluation bias

Answer 72

Data bias is embedded in the training data, Model bias is developed or amplified during learning, and Evaluation bias is usually induced by human annotators or biased evaluation metrics.

Question 73

What is a procedure used to tackle data issues in NLP datasets? Hint : this is something that comes included with the training dataset

Answer 73

Attaching a Datasheet to a dataset outlines its motivation, composition, and collection process.

Question 74

Data labeling when done by human annotators can induce bias. Explain how this might occur and what is a metric used to quantify the level of agreement between annotators? What is the threshold value for this metric such that the annotation is agreed upon?

Answer 74

Humans may have different agreements on what is considered “harmful” for example, hence if a group of annotators is under sensitive, then the data will result contain this preference/bias. Cohen’s Kappa is a statistical measure to quantify the agreement level between two annotators. A k >= 0.44 is acceptable.

Question 75

What is the main concern in regard to sustainability when developing LLMs? What is an unusual technique used to combat growing computing power requirements during model training?

Answer 75

Developing LLMs require a lot of power and resources, hence it can be harmful to the environment and require a lot of budget. A popular technique is to use the student-teacher training architecture, where a smaller model (the student) distills knowledge from the larger model (the teacher) to train itself.

Question 76

What is interpretability and why is it important? Give examples of interpretable models.

Answer 76

interpretability is defined as the degree to which a human can consistently predict a model’s output/decision process. Interpretability leads to increased model transparency. KNNs and Decision Trees are known for their interpretability.

Question 77

Black-box models can be hard to interpret. Name a typical example of a black box model, and name a common post-hoc technique to improve black-box model interpretability.

Answer 77

Neural networks and ensemble models are typically hard to interpret. Local interpretation techniques such as local surrogate models (LIME) can help make these more interpretable by replacing some internal components with interpretable models. For example Random Forests with Decision Trees (known for their interpretability)

Question 78

What is an advantage of using attribution techniques when calculating the interpretability of a model?

Answer 78

Attribution techniques allow us to investigate what aspects of the context lead to a specific output, often achieved by weighting the influence of each token in the input on the output.